这是用户在 2024-10-30 16:14 为 https://app.immersivetranslate.com/pdf-pro/de9885e9-3653-450d-89eb-db282528cfd5 保存的双语快照页面,由 沉浸式翻译 提供双语支持。了解如何保存?

Taylor & Francis IAT E X IAT E X IAT_(E)X\mathrm{IAT}_{\mathbf{E}} \mathbf{X} template for authors (Interact layout + Chicago author-date reference style)
泰勒与弗朗西斯 IAT E X IAT E X IAT_(E)X\mathrm{IAT}_{\mathbf{E}} \mathbf{X} 作者模板(互动布局 + 芝加哥作者-日期参考风格)

Shuning Xie a a ^(a){ }^{\mathrm{a}}, Junjun Yin a a ^(a){ }^{\mathrm{a}}, Luobing Chen a a ^(a){ }^{\mathrm{a}} and Jian Yang b b ^(b){ }^{\mathrm{b}}
谢书宁 a a ^(a){ }^{\mathrm{a}} , 尹俊俊 a a ^(a){ }^{\mathrm{a}} , 陈罗兵 a a ^(a){ }^{\mathrm{a}} 和 杨健 b b ^(b){ }^{\mathrm{b}}
a a ^(a){ }^{a} School of Computer & Communication Engineering, University of Science and Technology Beijing, Beijing, China; b ^("b "){ }^{\text {b }} Department of Electronic Engineering, Tsinghua University, Beijing, China
a a ^(a){ }^{a} 北京科技大学计算机与通信工程学院,中国北京; b ^("b "){ }^{\text {b }} 清华大学电子工程系,中国北京

ARTICLE HISTORY 文章历史

Compiled October 29, 2024
编译于 2024 年 10 月 29 日

Abstract 摘要

Synthetic aperture radar (SAR) is a widely utilized technology in remote sensing, recognized for its high resolution and all-weather, day-and-night operational capabilities. Change detection, which aims to identify differences in the same area across different time intervals, serves as a critical application of SAR technology. This review systematically examines SAR-based change detection methods, providing a comprehensive classification and evaluation of their performance. The paper first outlines the common SAR systems and datasets used in change detection. It then delves into the foundational process of change detection, analyzing the analytical units and technical methodologies employed, with a focus on recent advancements. Furthermore, the paper identifies current challenges in SAR change detection, offering insights into future research directions. Finally, the review concludes with a discussion of emerging trends in the field and their potential implications for realworld applications.
合成孔径雷达(SAR)是一种广泛应用于遥感的技术,以其高分辨率和全天候、昼夜操作能力而闻名。变化检测旨在识别同一区域在不同时间间隔内的差异,是 SAR 技术的重要应用。本文系统地审查了基于 SAR 的变化检测方法,提供了全面的分类和性能评估。文章首先概述了变化检测中常用的 SAR 系统和数据集。接着深入探讨了变化检测的基础过程,分析了所采用的分析单元和技术方法,重点关注近期的进展。此外,文章识别了 SAR 变化检测中的当前挑战,并提供了未来研究方向的见解。最后,综述讨论了该领域的新兴趋势及其对实际应用的潜在影响。

KEYWORDS 关键词

Synthetic Aperture Radar(SAR); Change detection; Multi-temporal data; Change detection process; Change detection methods; Remote sensing
合成孔径雷达(SAR);变化检测;多时相数据;变化检测过程;变化检测方法;遥感

1. Introduction 1. 引言

Change detection technology holds substantial practical significance across various remote sensing application fields. Remote sensing technology encompasses diverse methods for observing the Earth. Among these, Synthetic Aperture Radar(SAR) satellites stand out as an active microwave remote sensing imaging system, offering distinct advantages such as all-weather, day-and-night operation, strong penetration capabilities, and extensive imaging coverage(Xue-song and Siwei (2020)). Compared to optical remote sensing technology, SAR demonstrates superior accuracy and reliability under challenging weather conditions, making it a vital tool for detecting surface changes. Since the 1990s, the development of radar remote sensing systems has accelerated, with the successive launch of numerous SAR satellites. Consequently, SAR-based change detection has emerged as a significant and expanding field of international research.In recent years, SAR change detection methods have advanced significantly. Polarisation is a fundamental characteristic of electromagnetic waves, providing crucial information beyond frequency, amplitude, and geometric orientation. When a SAR satellite
变化检测技术在各类遥感应用领域具有重要的实际意义。遥感技术涵盖了多种观察地球的方法。其中,合成孔径雷达(SAR)卫星作为一种主动微波遥感成像系统,具有全天候、昼夜运行、强穿透能力和广泛成像覆盖等显著优势(Xue-song 和 Siwei (2020))。与光学遥感技术相比,SAR 在恶劣天气条件下表现出更高的准确性和可靠性,使其成为检测地表变化的重要工具。自 1990 年代以来,雷达遥感系统的发展加速,许多 SAR 卫星相继发射。因此,基于 SAR 的变化检测已成为一个重要且不断扩展的国际研究领域。近年来,SAR 变化检测方法取得了显著进展。极化是电磁波的一个基本特征,提供了超越频率、幅度和几何方向的重要信息。当 SAR 卫星

transmits a signal to the Earth’s surface, different targets produce distinct echo sig-nals-polarisation information-due to electromagnetic field oscillations in various directions. Polarimetric SAR (PolSAR) systems transmit and receive signals using orthogonal polarisations, enabling them to capture not only the amplitude, phase, and frequency of the target’s scattered echoes but also its polarisation properties. This greatly enhances the radar’s capability for target interpretation and analysis.Owing to the distinct ways in which various land features interact with polarization, PolSAR can extract more comprehensive information by analyzing the polarization characteristics of electromagnetic waves. Consequently, PolSAR has garnered substantial attention in recent years, emphasizing its potential to improve the detection and classification of land features(Wang (2016)).
向地球表面发射信号,不同的目标产生不同的回波信号-极化信息-由于电磁场在不同方向上的振荡。极化合成孔径雷达(PolSAR)系统使用正交极化发射和接收信号,使其能够捕捉目标散射回波的幅度、相位和频率,以及其极化特性。这大大增强了雷达对目标解释和分析的能力。由于各种地表特征与极化的相互作用方式不同,PolSAR 可以通过分析电磁波的极化特性提取更全面的信息。因此,PolSAR 近年来受到广泛关注,强调其在改善地表特征的检测和分类方面的潜力(王,2016)。
The SAR system primarily consists of airborne and satellite platforms. In 1985, the Jet Propulsion Laboratory (JPL) successfully developed the first airborne polarimetric SAR system. With the ongoing advancements in SAR and PolSAR theories and technologies, various satellites such as ALOS-1/ALOS-2/PALSAR, Radarsat-2, TerraSAR/DEM-X, Cosmo-SkyMed, and Gaofen-3 have been launched globally.Tables 1 and 2 provide an overview of representative airborne and spaceborne PolSAR systems, outlining their key missions, frequency bands, and polarisation modes.
SAR 系统主要由机载和卫星平台组成。1985 年,喷气推进实验室(JPL)成功开发了第一个机载极化 SAR 系统。随着 SAR 和 PolSAR 理论与技术的不断进步,全球已发射了多颗卫星,如 ALOS-1/ALOS-2/PALSAR、Radarsat-2、TerraSAR/DEM-X、Cosmo-SkyMed 和高分-3。表 1 和表 2 概述了代表性的机载和空间 PolSAR 系统,列出了它们的主要任务、频段和极化模式。
Table 1. Typical Airborne Polarimetric SAR Systems
表 1. 典型的空中极化合成孔径雷达系统
Polarimetric SAR System 极化合成孔径雷达系统 Country 国家 Frequency Band 频段 Launch Year 发售年份 Polarisation Mode 极化模式
AIRSAR USA P/L/C 1987 Dual Polarisation 双极化
ADTS-SAR USA Ka  1987 Fully Polarimetric 全极化
E-SAR Germany 德国 P/L/C/X 1988 Dual Polarisation 双极化
Convair 580 康维尔 580 Canada 加拿大 C/X 1990 Dual Polarisation 双极化
C/X-SAR Canada 加拿大 C 1990 Multi-polarisation 多极化
Pi-SAR Japan 日本 L/X 1996 Fully Polarimetric 全极化
EMISAR Denmark 丹麦 L/C 信用证 1997 Fully Polarimetric 全极化
TUAVR USA Ku  2001 Multi-polarisation 多极化
F-SAR Germany 德国 P/L/S/C/X 2006 Fully Polarimetric 全极化
UAVSAR USA L 2007 Fully Polarimetric 全极化
High-Resolution Fully Polarimetric System
高分辨率全极化系统
China 中国 X 2008 Fully Polarimetric 全极化
Pi-SAR2 Japan 日本 X 2010 Fully Polarimetric 全极化
Polarimetric SAR System Country Frequency Band Launch Year Polarisation Mode AIRSAR USA P/L/C 1987 Dual Polarisation ADTS-SAR USA Ka 1987 Fully Polarimetric E-SAR Germany P/L/C/X 1988 Dual Polarisation Convair 580 Canada C/X 1990 Dual Polarisation C/X-SAR Canada C 1990 Multi-polarisation Pi-SAR Japan L/X 1996 Fully Polarimetric EMISAR Denmark L/C 1997 Fully Polarimetric TUAVR USA Ku 2001 Multi-polarisation F-SAR Germany P/L/S/C/X 2006 Fully Polarimetric UAVSAR USA L 2007 Fully Polarimetric High-Resolution Fully Polarimetric System China X 2008 Fully Polarimetric Pi-SAR2 Japan X 2010 Fully Polarimetric| Polarimetric SAR System | Country | Frequency Band | Launch Year | Polarisation Mode | | :---: | :---: | :---: | :---: | :---: | | AIRSAR | USA | P/L/C | 1987 | Dual Polarisation | | ADTS-SAR | USA | Ka | 1987 | Fully Polarimetric | | E-SAR | Germany | P/L/C/X | 1988 | Dual Polarisation | | Convair 580 | Canada | C/X | 1990 | Dual Polarisation | | C/X-SAR | Canada | C | 1990 | Multi-polarisation | | Pi-SAR | Japan | L/X | 1996 | Fully Polarimetric | | EMISAR | Denmark | L/C | 1997 | Fully Polarimetric | | TUAVR | USA | Ku | 2001 | Multi-polarisation | | F-SAR | Germany | P/L/S/C/X | 2006 | Fully Polarimetric | | UAVSAR | USA | L | 2007 | Fully Polarimetric | | High-Resolution Fully Polarimetric System | China | X | 2008 | Fully Polarimetric | | Pi-SAR2 | Japan | X | 2010 | Fully Polarimetric |
Table 2. Typical Spaceborne Polarimetric SAR Systems
表 2. 典型的空间极化合成孔径雷达系统
Polarimetric SAR System 极化合成孔径雷达系统 Country 国家 Frequency Band 频段 Launch Year 发售年份 Polarisation Mode 极化模式
Radarsat-2 雷达卫星-2 Canada 加拿大 C 2007 Fully Polarimetric 全极化
TerraSAR-X Germany 德国 X 2007 Multi-polarisation 多极化
COSMO-SkyMed Italy 意大利 X 2007 Multi-polarisation 多极化
RISAT-1 India 印度 C 2012 Fully Polarimetric 全极化
Kompsat-5 South Korea 韩国 X 2013 Multi-polarisation 多极化
ALOS-PALSAR-1/2 Japan 日本 L 2006 / 2014 2006 / 2014 2006//20142006 / 2014 Fully Polarimetric 全极化
Sentinel-1 哨兵-1 ESA C 2014 Dual Polarisation 双极化
Gaofen-3 (GF-3) 高分三号 (GF-3) China 中国 C 2016 Fully Polarimetric 全极化
ICEYE Finland 芬兰 X 2018 Fully Polarimetric 全极化
NovaSAR-1 UK S 2018 Fully Polarimetric 全极化
PAZ Spain 西班牙 X 2018 Multi-polarisation 多极化
SAOCOM Argentina 阿根廷 L 2018 Fully Polarimetric 全极化
Polarimetric SAR System Country Frequency Band Launch Year Polarisation Mode Radarsat-2 Canada C 2007 Fully Polarimetric TerraSAR-X Germany X 2007 Multi-polarisation COSMO-SkyMed Italy X 2007 Multi-polarisation RISAT-1 India C 2012 Fully Polarimetric Kompsat-5 South Korea X 2013 Multi-polarisation ALOS-PALSAR-1/2 Japan L 2006//2014 Fully Polarimetric Sentinel-1 ESA C 2014 Dual Polarisation Gaofen-3 (GF-3) China C 2016 Fully Polarimetric ICEYE Finland X 2018 Fully Polarimetric NovaSAR-1 UK S 2018 Fully Polarimetric PAZ Spain X 2018 Multi-polarisation SAOCOM Argentina L 2018 Fully Polarimetric| Polarimetric SAR System | Country | Frequency Band | Launch Year | Polarisation Mode | | :---: | :---: | :---: | :---: | :---: | | Radarsat-2 | Canada | C | 2007 | Fully Polarimetric | | TerraSAR-X | Germany | X | 2007 | Multi-polarisation | | COSMO-SkyMed | Italy | X | 2007 | Multi-polarisation | | RISAT-1 | India | C | 2012 | Fully Polarimetric | | Kompsat-5 | South Korea | X | 2013 | Multi-polarisation | | ALOS-PALSAR-1/2 | Japan | L | $2006 / 2014$ | Fully Polarimetric | | Sentinel-1 | ESA | C | 2014 | Dual Polarisation | | Gaofen-3 (GF-3) | China | C | 2016 | Fully Polarimetric | | ICEYE | Finland | X | 2018 | Fully Polarimetric | | NovaSAR-1 | UK | S | 2018 | Fully Polarimetric | | PAZ | Spain | X | 2018 | Multi-polarisation | | SAOCOM | Argentina | L | 2018 | Fully Polarimetric |
Change detection involves identifying and determining state changes in objects or phenomena by analyzing observational data from the same area at different times. This technology is extensively applied across various fields, including urban development monitoring and planning, disaster and environmental surveillance, farmland and for-
变化检测涉及通过分析来自同一区域在不同时间的观测数据,识别和确定对象或现象的状态变化。这项技术广泛应用于多个领域,包括城市发展监测与规划、灾害与环境监测、农田和森林管理等。

est management, agricultural surveys, updating of fundamental geographic databases, land degradation and desertification assessment, as well as monitoring of marine, inland water, and coastal environments. Furthermore, change detection plays a crucial role in wetland management, natural disaster response, military reconnaissance, and combat effectiveness evaluation.
最佳管理、农业调查、基础地理数据库的更新、土地退化和沙漠化评估,以及海洋、内陆水域和沿海环境的监测。此外,变化检测在湿地管理、自然灾害响应、军事侦察和作战效果评估中发挥着至关重要的作用。
Although there is a substantial body of literature reviewing change detection methods(Tewkesbury et al. (2015)), most of these reviews primarily focus on optical imagery. Given that SAR images are influenced by multiplicative speckle noise and geometric distortions inherent to SAR imaging, many optical change detection techniques are not directly applicable to SAR data. Consequently, a systematic and in-depth exploration of SAR-specific change detection methods remains lacking(Zuxun et al. (2022)). This paper aims to address this gap by providing a comprehensive summary and analysis of SAR-based change detection techniques. In light of the existing focus on optical imagery, this review centers on SAR change detection methods, offering a classification and performance evaluation of these techniques. Additionally, the paper highlights the challenges unique to SAR change detection and explores potential future research directions.
尽管有大量文献回顾了变化检测方法(Tewkesbury 等,2015),但这些回顾主要集中在光学影像上。由于 SAR 图像受到乘法斑点噪声和 SAR 成像固有的几何失真的影响,许多光学变化检测技术并不直接适用于 SAR 数据。因此,针对 SAR 特定变化检测方法的系统深入探索仍然缺乏(Zuxun 等,2022)。本文旨在填补这一空白,通过提供 SAR 基础变化检测技术的全面总结和分析。鉴于现有文献对光学影像的关注,本文重点讨论 SAR 变化检测方法,提供这些技术的分类和性能评估。此外,本文还强调了 SAR 变化检测独特的挑战,并探讨了未来的研究方向。
The structure of this paper is as follows: Section I introduces the fundamental process of change detection. Section II summarizes and discusses the methods and performance of SAR image change detection techniques developed in recent years. Section III analyzes the development trends in SAR image change detection and explores future research directions. Finally, Section IV provides a comprehensive summary.
本文的结构如下:第一部分介绍了变化检测的基本过程。第二部分总结并讨论了近年来开发的 SAR 图像变化检测技术的方法和性能。第三部分分析了 SAR 图像变化检测的发展趋势,并探讨了未来的研究方向。最后,第四部分提供了全面的总结。

2. Change detection process and pre-processing
2. 变化检测过程和预处理

2.1. Change detection process
2.1. 变化检测过程

Remote sensing image change detection typically relies on analyzing multiple images of the same area captured at different times to evaluate the changes and extent of surface features over time. The main challenges faced in current change detection include limited data availability, significant discrepancies between various data sources, and the complexity and diversity of land cover types. Consequently, performing change detection on diverse land features across varying scenarios necessitates the use of tailored techniques and processing methods to effectively address these challenges and enhance detection accuracy.A typical change detection process involves the following steps: (a) establish the spatial alignment between images from two different time phases, (b) identify the target features for change detection, © indicate the changes observed(Jabari and Zhang (2016)). (a)Perform data pre-processing, (b)model training, ©result prediction(Lian, ShiQi, and LeQing (2020)); (a)Conduct pre-processing, (b)feature extraction, ©feature evaluation(Runbo (2023)). In summary, the change detection process follows a general workflow, which is optimized and refined based on specific applications, as illustrated in Figure 1. The typical workflow includes: (a)data selection and pre-processing, (b)selection of change detection techniques,© result prediction and accuracy assessment.
遥感图像变化检测通常依赖于分析在不同时间拍摄的同一区域的多幅图像,以评估表面特征随时间的变化及其程度。目前变化检测面临的主要挑战包括数据可用性有限、不同数据源之间存在显著差异,以及土地覆盖类型的复杂性和多样性。因此,在不同场景下对多样化土地特征进行变化检测需要使用量身定制的技术和处理方法,以有效应对这些挑战并提高检测准确性。典型的变化检测过程包括以下步骤:(a)建立来自两个不同时间阶段的图像之间的空间对齐,(b)识别变化检测的目标特征,(c)指示观察到的变化(Jabari 和 Zhang(2016))。 (a)进行数据预处理,(b)模型训练,(c)结果预测(Lian, ShiQi 和 LeQing(2020));(a)进行预处理,(b)特征提取,(c)特征评估(Runbo(2023))。 总之,变化检测过程遵循一个通用工作流程,该流程根据具体应用进行了优化和改进,如图 1 所示。典型的工作流程包括:(a) 数据选择和预处理,(b) 变化检测技术的选择,(c) 结果预测和准确性评估。

2.2. Typical change detection data
2.2. 典型变化检测数据

As SAR technology advances, the availability of SAR and PolSAR data has increased, with continuous improvements in resolution. Based on the quantity and temporal
随着合成孔径雷达(SAR)技术的进步,SAR 和极化合成孔径雷达(PolSAR)数据的可用性增加,分辨率不断提高。基于数量和时间的

Figure 1. Change Detection Workflow
图 1. 变化检测工作流程

characteristics of the datasets, change detection can be categorized into two-time-phase change detection and time series change detection (instantaneous sequence trajectory analysis). The two-time-phase change detection method utilizes two or more remote sensing images, typically acquired over intervals of months or years. This approach aims to minimize the influence of environmental factors, such as seasonal variations and differences in solar altitude, on the detection results. However, a practical challenge is selecting the appropriate image acquisition time points to ensure the accuracy and stability of the change detection process. In contrast, time series change detection involves using a sequence of remote sensing images captured at consistent time intervals to identify and monitor land surface changes through time series analysis and model development(Liu et al. (2023)). This method focuses on identifying trends and dynamic processes, making it suitable for long-term monitoring. Its advantage lies in its ability to evaluate the rate, direction, and magnitude of changes over time, thereby minimizing the influence of environmental variations on detection results(Dongmei, Hui, and Yao (2017)), and providing more systematic and comprehensive information about the changes. However, it requires a substantial amount of multi-temporal data, which places high demands on the spatio-temporal consistency and continuity of the data. Table 3 summarizes these two methods, including a brief description, as well as their respective advantages and disadvantages.
数据集的特征,变化检测可以分为两阶段变化检测和时间序列变化检测(瞬时序列轨迹分析)。两阶段变化检测方法利用两幅或更多的遥感图像,通常是在几个月或几年的时间间隔内获取。该方法旨在最小化环境因素(如季节变化和太阳高度差异)对检测结果的影响。然而,实际挑战在于选择合适的图像获取时间点,以确保变化检测过程的准确性和稳定性。相比之下,时间序列变化检测涉及使用在一致时间间隔内捕获的一系列遥感图像,通过时间序列分析和模型开发来识别和监测地表变化(刘等,2023)。该方法侧重于识别趋势和动态过程,适合于长期监测。 其优势在于能够评估随时间变化的速率、方向和幅度,从而最小化环境变化对检测结果的影响(董梅、辉和姚(2017)),并提供关于变化的更系统和全面的信息。然而,它需要大量的多时相数据,这对数据的时空一致性和连续性提出了较高的要求。表 3 总结了这两种方法,包括简要描述以及各自的优缺点。
Table 3. Summary of Data Selection for Change Detection
表 3. 变化检测的数据选择总结
Type 类型 Description 描述 Advantages 优势 Disadvantages 缺点

双时相变化检测
Bi-temporal
Change Detection
Bi-temporal Change Detection| Bi-temporal | | :---: | | Change Detection |

在特定时间间隔内获取的两幅或更多图像
Two or more images
acquired at specific
time intervals
Two or more images acquired at specific time intervals| Two or more images | | :---: | | acquired at specific | | time intervals |

减少季节变化和太阳高度差异的影响
Reduces impact of
seasonal variations
and solar altitude
differences
Reduces impact of seasonal variations and solar altitude differences| Reduces impact of | | :---: | | seasonal variations | | and solar altitude | | differences |

具有挑战性的图像选择;依赖于特定时间点的数据;难以准确捕捉动态变化
Challenging image
selection; depends on
data from specific time
points; difficult to
capture dynamic
changes accurately
Challenging image selection; depends on data from specific time points; difficult to capture dynamic changes accurately| Challenging image | | :---: | | selection; depends on | | data from specific time | | points; difficult to | | capture dynamic | | changes accurately |

时间序列变化检测
Time-series
Change Detection
Time-series Change Detection| Time-series | | :---: | | Change Detection |

时间序列遥感图像
Time-series remote
sensing images
Time-series remote sensing images| Time-series remote | | :---: | | sensing images |

适合长期监测;最小化环境干扰
Suitable for
long-term
monitoring;
minimizes
environmental
interference
Suitable for long-term monitoring; minimizes environmental interference| Suitable for | | :---: | | long-term | | monitoring; | | minimizes | | environmental | | interference |

需要大量数据;需要高时空一致性;数据质量和连续性影响有效性
Requires large data
volumes; high
spatio-temporal
consistency needed;
data quality and
continuity impact
effectiveness
Requires large data volumes; high spatio-temporal consistency needed; data quality and continuity impact effectiveness| Requires large data | | :---: | | volumes; high | | spatio-temporal | | consistency needed; | | data quality and | | continuity impact | | effectiveness |
Type Description Advantages Disadvantages "Bi-temporal Change Detection" "Two or more images acquired at specific time intervals" "Reduces impact of seasonal variations and solar altitude differences" "Challenging image selection; depends on data from specific time points; difficult to capture dynamic changes accurately" "Time-series Change Detection" "Time-series remote sensing images" "Suitable for long-term monitoring; minimizes environmental interference" "Requires large data volumes; high spatio-temporal consistency needed; data quality and continuity impact effectiveness"| Type | Description | Advantages | Disadvantages | | :---: | :---: | :---: | :---: | | Bi-temporal <br> Change Detection | Two or more images <br> acquired at specific <br> time intervals | Reduces impact of <br> seasonal variations <br> and solar altitude <br> differences | Challenging image <br> selection; depends on <br> data from specific time <br> points; difficult to <br> capture dynamic <br> changes accurately | | Time-series <br> Change Detection | Time-series remote <br> sensing images | Suitable for <br> long-term <br> monitoring; <br> minimizes <br> environmental <br> interference | Requires large data <br> volumes; high <br> spatio-temporal <br> consistency needed; <br> data quality and <br> continuity impact <br> effectiveness |
Multi-platform, multi-band, and multi-polarimetric SAR images offer a rich data source for change detection. Various organizations and institutions have released public datasets to encourage the broader application of SAR data across different domains. However, an analysis of the current state of SAR change detection indicates that publicly available SAR datasets for practical change detection applications remain relatively scarce, with much of the research still dependent on self-collected data. Table 4 provides a summary of commonly used bistatic datasets in the SAR change
多平台、多频段和多极化的合成孔径雷达(SAR)图像为变化检测提供了丰富的数据源。各种组织和机构已发布公共数据集,以鼓励 SAR 数据在不同领域的更广泛应用。然而,对当前 SAR 变化检测状态的分析表明,适用于实际变化检测应用的公开 SAR 数据集仍然相对稀缺,许多研究仍依赖于自收集的数据。表 4 提供了 SAR 变化检测中常用的双静态数据集的总结。

detection field(Gao et al. (2016)).
检测领域(Gao et al. (2016))。
Table 4. Common SAR Image Change Detection Dataset
表 4. 常见 SAR 图像变化检测数据集
Data 数据 Sensor 传感器 Size (pixels) 大小(像素) Date 日期 Pre Image, Post Image & Ground Truth Image
前图像、后图像和真实图像
Ottawa 渥太华 RADARSAT SAR  RADARSAT   SAR  {:[" RADARSAT "],[" SAR "]:}\begin{aligned} & \text { RADARSAT } \\ & \text { SAR } \end{aligned} 350 × 290 350 × 290 350 xx290350 \times 290

1997 年 5 月,1997 年 8 月
May 1997,
August 1997
May 1997, August 1997| May 1997, | | :--- | | August 1997 |
Yellow1 黄色 1 Radarsat-2 雷达卫星-2 289 × 257 289 × 257 289 xx257289 \times 257

2008 年 6 月,2009 年 6 月
June 2008,
June 2009
June 2008, June 2009| June 2008, | | :--- | | June 2009 |
Yellow2 黄色 2 Radarsat-2 雷达卫星-2 289 × 257 289 × 257 289 xx257289 \times 257

2008 年 6 月,2009 年 6 月
June 2008,
June 2009
June 2008, June 2009| June 2008, | | :--- | | June 2009 |
Yellow3 黄 3 Radarsat-2 雷达卫星-2 291 × 444 291 × 444 291 xx444291 \times 444

2008 年 6 月,2009 年 6 月
June 2008,
June 2009
June 2008, June 2009| June 2008, | | :--- | | June 2009 |
Yellow4 黄 4 Radarsat-2 雷达卫星-2 450 × 280 450 × 280 450 xx280450 \times 280

2008 年 6 月,2009 年 6 月
June 2008,
June 2009
June 2008, June 2009| June 2008, | | :--- | | June 2009 |
Bern 伯尔尼 ERS-2 301 × 301 301 × 301 301 xx301301 \times 301

1999 年 4 月,1999 年 5 月
April 1999,
May 1999
April 1999, May 1999| April 1999, | | :--- | | May 1999 |
 卡提奥斯国家公园
Katios
National Park
Katios National Park| Katios | | :--- | | National Park |
Sentinel 1 A 哨兵 1A 879 × 1319 879 × 1319 879 xx1319879 \times 1319

2019 年 3 月,2019 年 4 月
March 2019,
April 2019
March 2019, April 2019| March 2019, | | :--- | | April 2019 |
Atlantico 大西洋 ALOS/PALSAR 729 × 1056 729 × 1056 729 xx1056729 \times 1056

2010 年 4 月,2011 年 3 月
April 2010,
March 2011
April 2010, March 2011| April 2010, | | :--- | | March 2011 |
San Francisco 旧金山 ERS-2 SAR 256 × 256 256 × 256 256 xx256256 \times 256 August 2003 , May 2004  August  2003 ,  May  2004 {:[" August "2003","],[" May "2004]:}\begin{aligned} & \text { August } 2003, \\ & \text { May } 2004 \end{aligned}
Wen Chuan 温川 ESA/ASAR 301 × 442 301 × 442 301 xx442301 \times 442 March 2008 , June 2008  March  2008 ,  June  2008 {:[" March "2008","],[" June "2008]:}\begin{gathered} \text { March } 2008, \\ \text { June } 2008 \end{gathered}
Data Sensor Size (pixels) Date Pre Image, Post Image & Ground Truth Image Ottawa " RADARSAT SAR " 350 xx290 "May 1997, August 1997" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=249&width=670&top_left_y=417&top_left_x=1054 Yellow1 Radarsat-2 289 xx257 "June 2008, June 2009" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=231&width=657&top_left_y=665&top_left_x=1067 Yellow2 Radarsat-2 289 xx257 "June 2008, June 2009" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=196&width=584&top_left_y=895&top_left_x=1140 Yellow3 Radarsat-2 291 xx444 "June 2008, June 2009" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=316&width=659&top_left_y=1094&top_left_x=1054 Yellow4 Radarsat-2 450 xx280 "June 2008, June 2009" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=143&width=657&top_left_y=1418&top_left_x=1054 Bern ERS-2 301 xx301 "April 1999, May 1999" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=214&width=624&top_left_y=1555&top_left_x=1054 "Katios National Park" Sentinel 1 A 879 xx1319 "March 2019, April 2019" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=85&width=573&top_left_y=1772&top_left_x=1151 Atlantico ALOS/PALSAR 729 xx1056 "April 2010, March 2011" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=165&width=402&top_left_y=1912&top_left_x=1322 San Francisco ERS-2 SAR 256 xx256 " August 2003, May 2004" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=156&width=637&top_left_y=2076&top_left_x=1065 Wen Chuan ESA/ASAR 301 xx442 " March 2008, June 2008" https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=159&width=670&top_left_y=2229&top_left_x=1054| Data | Sensor | Size (pixels) | Date | Pre Image, Post Image & Ground Truth Image | | :---: | :---: | :---: | :---: | :---: | | Ottawa | $\begin{aligned} & \text { RADARSAT } \\ & \text { SAR } \end{aligned}$ | $350 \times 290$ | May 1997, <br> August 1997 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=249&width=670&top_left_y=417&top_left_x=1054) | | Yellow1 | Radarsat-2 | $289 \times 257$ | June 2008, <br> June 2009 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=231&width=657&top_left_y=665&top_left_x=1067) | | Yellow2 | Radarsat-2 | $289 \times 257$ | June 2008, <br> June 2009 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=196&width=584&top_left_y=895&top_left_x=1140) | | Yellow3 | Radarsat-2 | $291 \times 444$ | June 2008, <br> June 2009 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=316&width=659&top_left_y=1094&top_left_x=1054) | | Yellow4 | Radarsat-2 | $450 \times 280$ | June 2008, <br> June 2009 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=143&width=657&top_left_y=1418&top_left_x=1054) | | Bern | ERS-2 | $301 \times 301$ | April 1999, <br> May 1999 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=214&width=624&top_left_y=1555&top_left_x=1054) | | Katios <br> National Park | Sentinel 1 A | $879 \times 1319$ | March 2019, <br> April 2019 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=85&width=573&top_left_y=1772&top_left_x=1151) | | Atlantico | ALOS/PALSAR | $729 \times 1056$ | April 2010, <br> March 2011 | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=165&width=402&top_left_y=1912&top_left_x=1322) | | San Francisco | ERS-2 SAR | $256 \times 256$ | $\begin{aligned} & \text { August } 2003, \\ & \text { May } 2004 \end{aligned}$ | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=156&width=637&top_left_y=2076&top_left_x=1065) | | Wen Chuan | ESA/ASAR | $301 \times 442$ | $\begin{gathered} \text { March } 2008, \\ \text { June } 2008 \end{gathered}$ | ![](https://cdn.mathpix.com/cropped/2024_10_29_817d829f2314a28a1176g-05.jpg?height=159&width=670&top_left_y=2229&top_left_x=1054) |

2.3. Pre-processing 2.3. 预处理

The core objective of change detection is to identify areas in the image that have experienced significant changes, while filtering out insignificant or low-confidence ‘pseudo’ changes. In practice, pre-processing the image is essential prior to change detection to minimize the impact of these ‘pseudo’ changes and enhance the accuracy of detection results. Given the variations between different remote sensing satellites, as well as geometric distortions in images captured by the same sensor, image data must undergo pre-processing before performing change detection(Ren Qiuru (2021)).Common pre-processing steps primarily involve image registration and correction. Registration aligns images geometrically across different time periods to ensure that pixels correspond to the same ground targets, thereby eliminating discrepancies caused by varying viewing angles or platform positions. Typical methods include feature point matching and geometric model correction.Geometric correction aims to eliminate distortions resulting from terrain variations and the side-looking imaging mode, ensuring that the image accurately represents the true shape of the ground. Radiometric correction, on the other hand, is used to adjust for brightness and contrast variations caused by differences in imaging conditions, thereby maintaining the comparability of images across time periods(Lee and Pottier (2009)).In addition, data preprocessing involves filtering to address speckle noise, which is a characteristic feature of SAR images. Filtering is essential in SAR image processing as it enhances the accuracy of change detection. Unlike the additive noise present in optical images, SAR image noise is multiplicative in nature(Rignot and van Zyl (1993)). Common filtering techniques include Lee filtering(Lee (1980)), Frost filtering(Frost et al. (1982)), Gamma MAP filtering(Lopes, Touzi, and Nezry (1990)), and the improved Sigma filtering method(Lee et al. (2008)).
变化检测的核心目标是识别图像中经历显著变化的区域,同时过滤掉不显著或低置信度的“伪”变化。在实际操作中,进行图像预处理是变化检测前的必要步骤,以最小化这些“伪”变化的影响并提高检测结果的准确性。考虑到不同遥感卫星之间的差异,以及同一传感器捕获的图像中的几何畸变,图像数据必须在进行变化检测之前进行预处理(任秋如,2021)。常见的预处理步骤主要包括图像配准和校正。配准在不同时间段之间几何对齐图像,以确保像素对应于相同的地面目标,从而消除由于视角或平台位置变化引起的差异。典型的方法包括特征点匹配和几何模型校正。几何校正旨在消除由于地形变化和侧视成像模式造成的畸变,确保图像准确地表示地面的真实形状。 辐射校正则用于调整由于成像条件差异引起的亮度和对比度变化,从而保持不同时间段图像的可比性(Lee 和 Pottier(2009))。此外,数据预处理涉及过滤,以解决斑点噪声,这是 SAR 图像的一个特征。过滤在 SAR 图像处理中的重要性在于它提高了变化检测的准确性。与光学图像中存在的加性噪声不同,SAR 图像噪声本质上是乘法性的(Rignot 和 van Zyl(1993))。常见的过滤技术包括 Lee 过滤(Lee(1980))、Frost 过滤(Frost 等(1982))、Gamma MAP 过滤(Lopes、Touzi 和 Nezry(1990))以及改进的 Sigma 过滤方法(Lee 等(2008))。

3. Change detection techniques
3. 变化检测技术

Change detection technology can be categorized based on the application purpose into the following types: binary change detection, change type identification, and change trend analysis.
变化检测技术可以根据应用目的分为以下几种类型:二元变化检测、变化类型识别和变化趋势分析。
Binary change detection is used to determine whether a change has occurred within an image, producing a binary output that differentiates between changed and unchanged areas. Change type identification goes a step further by categorizing the specific type of change within the detected areas, such as changes in land use or vegetation cover. Change trend analysis, also referred to as time series change detection, focuses on analyzing the evolution of land features over time. This approach consists of two components: time series processing and trend modeling.
二元变化检测用于确定图像中是否发生了变化,产生一个二元输出,以区分变化和未变化的区域。变化类型识别更进一步,通过对检测到的区域内的具体变化类型进行分类,例如土地利用或植被覆盖的变化。变化趋势分析,也称为时间序列变化检测,专注于分析土地特征随时间的演变。该方法由两个组成部分构成:时间序列处理和趋势建模。

3.1. Unit of analysis 3.1. 分析单位

The advancement of modern remote sensing and image processing technologies has enabled change detection to be performed across multiple frameworks. The choice of analysis units is fundamental in change detection and has a profound influence on the outcomes. Currently, pixels and objects are the two principal types of analysis units, from which various other categories have evolved. The selection of analysis units can be categorized into four levels: ‘pixel-region-object-scene’.As depicted in Figure 2 , the relationship between a pixel and its surroundings gradually extends through these levels, evolving from ‘isolated - surroundings - proximity - overall’. Accordingly,
现代遥感和图像处理技术的进步使得可以在多个框架中进行变化检测。分析单元的选择在变化检测中是基础,并对结果产生深远的影响。目前,像素和对象是两种主要的分析单元,从中演变出各种其他类别。分析单元的选择可以分为四个层次:‘像素-区域-对象-场景’。如图 2 所示,像素与其周围环境之间的关系通过这些层次逐渐扩展,演变为‘孤立-周围-接近-整体’。因此,

the degree of information utilization progresses from ‘pixel - feature - object - higher dimension’.
信息利用的程度从“像素 - 特征 - 物体 - 更高维度”逐步提升。

Figure 2. Development of Semantic Information in Change Detection
图 2. 变化检测中语义信息的发展

3.1.1. Pixel 3.1.1. 像素

A pixel is the smallest unit of an image(Fisher (1997)). Pixel-based change detection methods operate by analyzing individual pixels to quickly identify areas of change through pixel-by-pixel operations, using intensity differences to extract change information. This approach is straightforward and efficient, making it particularly suitable for scenarios where pixel intensity is highly correlated with changes in the area(Fulong, Hong, and Chao (2007)). However, there are significant limitations to using pixels as the sole analysis unit. For instance, Chen et al. (2012)demonstrated that pixels can only capture limited features and cannot effectively utilize contextual information; they are also highly susceptible to noise interference, which may result in inaccurate information(Hussain et al. (2013)). Furthermore, with the increasing prevalence of high-resolution and ultra-high-resolution images, the computational complexity of pixel-level methods has grown substantially. In summary, while pixel-level change detection methods are simple and fast, their applicability is diminishing due to limitations in feature extraction, noise sensitivity, and increased computational complexity in high-resolution imagery.
像素是图像的最小单位(Fisher (1997))。基于像素的变化检测方法通过分析单个像素,快速识别变化区域,采用逐像素操作,利用强度差异提取变化信息。这种方法简单高效,特别适合像素强度与区域变化高度相关的场景(Fulong, Hong, 和 Chao (2007))。然而,单独使用像素作为分析单位存在显著的局限性。例如,Chen 等(2012)证明像素只能捕捉有限的特征,无法有效利用上下文信息;它们也对噪声干扰高度敏感,这可能导致信息不准确(Hussain 等(2013))。此外,随着高分辨率和超高分辨率图像的日益普及,像素级方法的计算复杂性显著增加。 总之,尽管像素级变化检测方法简单且快速,但由于特征提取的局限性、对噪声的敏感性以及高分辨率图像中计算复杂性的增加,它们的适用性正在减弱。
To overcome the limitations of using single pixels, pixel-centric sliding window methods and kernel filter techniques have been proposed. The sliding window method enhances feature extraction by analyzing texture and contextual information within a local region, using a fixed-size window centered around each pixel. The kernel filter method, on the other hand, suppresses noise and accentuates edge features by applying a specific kernel function to the local area, thereby improving the robustness of feature detection.Despite these advancements, the information extracted using these methods remains limited and incomplete, constrained by the window size or kernel
为了克服使用单个像素的局限性,提出了以像素为中心的滑动窗口方法和核滤波技术。滑动窗口方法通过分析局部区域内的纹理和上下文信息,使用围绕每个像素的固定大小窗口来增强特征提取。另一方面,核滤波方法通过对局部区域应用特定的核函数来抑制噪声并强调边缘特征,从而提高特征检测的鲁棒性。尽管这些进展取得了一定成效,但使用这些方法提取的信息仍然有限和不完整,受限于窗口大小或核的影响。

scale, which may result in issues such as blurred boundaries and missing features. To address the challenge of determining the optimal window size(Warner (2011)), extended approaches such as adaptive windows(Kanade and Okutomi (1991)) and multi-scale windows(Liang and Anderson (2007)) have been introduced to enhance flexibility and detection accuracy. While sliding window and kernel filter methods improve change detection accuracy through local analysis, they are still restricted by the choice of window size and kernel function, leading to incomplete boundary and feature information.
尺度可能导致模糊的边界和缺失的特征等问题。为了解决确定最佳窗口大小的挑战(Warner (2011)),引入了自适应窗口(Kanade 和 Okutomi (1991))和多尺度窗口(Liang 和 Anderson (2007))等扩展方法,以增强灵活性和检测准确性。虽然滑动窗口和核滤波方法通过局部分析提高了变化检测的准确性,但仍然受到窗口大小和核函数选择的限制,导致边界和特征信息不完整。

3.1.2. Superpixel 3.1.2. 超像素

In 2003, Ren and Malik (2003)defined an image segment composed of adjacent pixels with similar texture, color, brightness, and other characteristics as a “superpixel”. The superpixel method employs semantic region segmentation to more effectively preserve boundaries and extract comprehensive features, thereby addressing the limitations of traditional pixel-based methods. Superpixel-based SAR change detection has achieved significant progress in recent years(Zhang, Yin, and Yang (2021)).Representative superpixel segmentation methods include normalized cuts (Ncut), graph-based methods, superpixel lattice methods, watershed methods, MeanShift methods, Turbopixels methods, and simple linear iterative clustering (SLIC). The first three are graph theory-based approaches, while the latter four rely on gradient descent clustering. Superpixels group pixels based on pixel correlation, enhancing the consistency of classified regions, effectively reducing redundant information and noise interference, and thereby minimizing the complexity of subsequent image processing(Chun-Yao, Jun-Zhou, and Wei (2014)).However, the superpixel method has limitations in preserving boundaries and local details, making precise and efficient segmentation an ongoing research focus. While the superpixel approach enhances classification consistency and feature extraction efficiency by aggregating similar pixels into semantic regions, further improvements are necessary to optimize boundary retention and detail preservation.
在 2003 年,Ren 和 Malik(2003)将由相邻像素组成、具有相似纹理、颜色、亮度和其他特征的图像片段定义为“超像素”。超像素方法采用语义区域分割,更有效地保留边界并提取综合特征,从而解决了传统基于像素的方法的局限性。基于超像素的 SAR 变化检测近年来取得了显著进展(Zhang、Yin 和 Yang(2021))。代表性的超像素分割方法包括归一化切割(Ncut)、基于图的方法、超像素格方法、分水岭方法、均值漂移方法、涡轮像素方法和简单线性迭代聚类(SLIC)。前三种是基于图论的方法,而后四种依赖于梯度下降聚类。超像素根据像素相关性对像素进行分组,增强了分类区域的一致性,有效减少冗余信息和噪声干扰,从而最小化后续图像处理的复杂性(Chun-Yao、Jun-Zhou 和 Wei(2014))。然而,超像素方法在保持边界和局部细节方面存在局限性,使得精确和高效的分割成为一个持续的研究重点。虽然超像素方法通过将相似像素聚合成语义区域来增强分类一致性和特征提取效率,但仍需进一步改进以优化边界保留和细节保留。

3.1.3. Object 3.1.3. 对象

In 2000, Baatz and Schpe (2000)introduced the fractal net evolution approach (FNEA) algorithm, which established the foundation for object-level change detection in SAR images. An “object” is defined as a cluster of pixels characterized by polarization properties, spatial homogeneity, and geometric shape consistency. This representation captures the structural and scattering features in SAR images, such as the double-bounce scattering effect from buildings, thereby enhancing the accuracy and robustness of change detection(Shao et al. (2019)).Object-level change detection methods can be categorized into direct object comparison methods and post-classification comparison methods(Warner (2011)). The direct object comparison method identifies change areas by analyzing the characteristic variations of the same object in images acquired at different time phases. In contrast, the post-classification comparison method classifies the images first and then detects changes by comparing the classification outcomes.With the advancement of deep neural network technology, the object-oriented approach has further evolved. By integrating the polarization information and geometric characteristics of PolSAR images and leveraging deep learning models to extract abstract features of the target at multiple levels, it is possible to accurately identify changes in complex scenes. Object-level change detection methods demonstrate improved robust-
在 2000 年,Baatz 和 Schpe(2000)提出了分形网演化方法(FNEA)算法,为 SAR 图像中的对象级变化检测奠定了基础。“对象”被定义为由极化特性、空间均匀性和几何形状一致性特征的像素簇。该表示法捕捉了 SAR 图像中的结构和散射特征,例如建筑物的双重反射散射效应,从而提高了变化检测的准确性和鲁棒性(Shao 等,2019)。对象级变化检测方法可以分为直接对象比较方法和后分类比较方法(Warner,2011)。直接对象比较方法通过分析在不同时间阶段获取的图像中同一对象的特征变化来识别变化区域。相反,后分类比较方法首先对图像进行分类,然后通过比较分类结果来检测变化。随着深度神经网络技术的发展,面向对象的方法进一步演变。 通过整合 PolSAR 图像的极化信息和几何特征,并利用深度学习模型在多个层次上提取目标的抽象特征,可以准确识别复杂场景中的变化。对象级变化检测方法显示出更强的鲁棒性

ness against noise and clutter interference in SAR images; however, challenges persist in refining boundaries and detecting small targets.
在合成孔径雷达图像中抵御噪声和杂乱干扰的有效性;然而,在细化边界和检测小目标方面仍然存在挑战。

3.2. Models and methods for change detection
3.2. 变化检测的模型和方法

Numerous models and methods exist for SAR image change detection(Singh (1989); Warner (2011)), each with its own strengths and limitations. For instance, the direct comparison method is effective for detecting changes in simple scenes, while the polarimetric feature method fully utilizes the polarimetric information in SAR images to enhance detection accuracy. Despite these advancements, a comprehensive and systematic review of SAR image change detection methods is still lacking. Therefore, this paper classifies SAR image change detection methods based on the characteristics of SAR data, combined with the classification criteria of general change detection methods. According to the core principles of change detection models, these methods are grouped into five categories: direct comparison method, image transformation method, polarization feature method, classification comparison method, and machine learning method. The focus is placed on the application of these methods in various scenarios and their potential for integration, as illustrated in Figure 3.
众多模型和方法存在于 SAR 图像变化检测中(Singh(1989);Warner(2011)),每种方法都有其自身的优缺点。例如,直接比较方法在简单场景中有效检测变化,而极化特征方法充分利用 SAR 图像中的极化信息以提高检测精度。尽管有这些进展,关于 SAR 图像变化检测方法的全面系统评审仍然缺乏。因此,本文根据 SAR 数据的特征以及一般变化检测方法的分类标准对 SAR 图像变化检测方法进行了分类。根据变化检测模型的核心原则,这些方法被分为五类:直接比较方法、图像变换方法、极化特征方法、分类比较方法和机器学习方法。重点放在这些方法在各种场景中的应用及其整合潜力,如图 3 所示。

Figure 3. Classification of SAR Image Change Detection Methods
图 3. SAR 图像变化检测方法的分类

The direct comparison method calculates the difference between pixels of two images using arithmetic operations(Singh (1989)). The image transformation method transforms the image into a feature space, where the transformed results are analyzed to extract change information. The polarimetric feature method includes hypothesis testing based on statistical modeling and methods that analyze physical scattering characteristics, as well as combinations of these approaches(Liu et al. (2021)). The classification comparison method involves classifying the images first, followed by comparison and analysis of the classification results. Machine learning methods are also extensively applied in change detection, with traditional models such as support vector machines and decision trees being typical examples. In recent years, change detection techniques based on deep neural networks have gradually become a research focus. Ta-
直接比较方法通过算术运算计算两幅图像像素之间的差异(Singh (1989))。图像变换方法将图像转换为特征空间,在该空间中分析变换结果以提取变化信息。极化特征方法包括基于统计建模的假设检验和分析物理散射特性的各种方法,以及这些方法的组合(Liu et al. (2021))。分类比较方法首先对图像进行分类,然后比较和分析分类结果。机器学习方法在变化检测中也得到了广泛应用,传统模型如支持向量机和决策树是典型例子。近年来,基于深度神经网络的变化检测技术逐渐成为研究重点。

ble 5 summarizes the five main change detection methods, detailing their advantages, disadvantages, and representative studies.
表 5 总结了五种主要的变化检测方法,详细说明了它们的优点、缺点和代表性研究。
Table 5. Summary of Change Detection Methods Based on SAR Imagery
表 5. 基于 SAR 影像的变化检测方法总结
Method 方法 Subclass 子类 Subclass Description 子类描述 Advantages 优势 Limitations 限制 Application 应用程序
 直接比较
Direct
Comparison
Direct Comparison| Direct | | :--- | | Comparison |
 差异方法
Difference
Method
Difference Method| Difference | | :--- | | Method |
Measures the intensity difference between images.
测量图像之间的强度差异。
Simple and quick to calculate
简单快速计算
Limited to binary change detection; sensitive to noise
仅限于二元变化检测;对噪声敏感
Coppin Bauer (1996)  Coppin   Bauer (1996)  {:[" Coppin "],[" Bauer (1996) "]:}\begin{aligned} & \text { Coppin } \\ & \text { Bauer (1996) } \end{aligned}
Ratio Method 比率法 Calculates the ratio of intensities for each pixel.
计算每个像素的强度比。
Effective at reducing radiometric differences
有效减少辐射差异
Sensitive to noise in low-intensity areas
对低强度区域的噪音敏感
Conradsen et al. (2003) Conradsen 等人 (2003)
Regression Analysis 回归分析 Establishes regression models to account for gradual temporal changes.
建立回归模型以考虑逐渐的时间变化。
Suitable for gradual change detection
适合渐变变化检测

需要一个大而一致的数据集以确保准确性
Requires a
large and consistent dataset for accuracy
Requires a large and consistent dataset for accuracy| Requires a | | :--- | | large and consistent dataset for accuracy |
Zhuang, Ddeng, and Fan (2016); Zhao et al. (2021)
壮、邓和范(2016);赵等(2021)
 图像转换
Image
Transformation
Image Transformation| Image | | :--- | | Transformation |
CVA Uses Change Vector Analysis for multidimensional feature space analysis.
使用变化向量分析进行多维特征空间分析。
Effective at detecting multidimensional changes
有效检测多维变化
Complexity in result interpretation; requires normalization
结果解释的复杂性;需要标准化
Bayarjargal et al. (2006); Bovolo, Marchesi, and Bruzzone (2011)
Bayarjargal 等人 (2006); Bovolo, Marchesi 和 Bruzzone (2011)
PCA Reduces data dimensionality to focus on significant change information.
减少数据维度,以关注重要的变化信息。
Reduces redundancy and enhances significant features
减少冗余,增强重要特征
May obscure minor but relevant changes
可能掩盖一些微小但相关的变化

李、龚和张(2019);施瓦茨等(2020)
Li, Gong, and
Zhang (2019);
Schwartz et al.
(2020)
Li, Gong, and Zhang (2019); Schwartz et al. (2020)| Li, Gong, and | | :--- | | Zhang (2019); | | Schwartz et al. | | (2020) |
K-T Kauth-Thomas transformation, particularly for vegetation index detection.
Kauth-Thomas 变换,特别用于植被指数检测。
Highly quad\quad ef- fective for vegetation monitoring
高度 quad\quad 有效的植被监测
Limited use in non-vegetative change detection
在非植被变化检测中的有限使用
Rogan, Franklin, and Roberts ( 2002 )  Rogan,   Franklin,   and   Roberts  ( 2002 ) {:[" Rogan, "," Franklin, "],[" and "," Roberts "],[(2002),]:}\begin{array}{lr} \text { Rogan, } & \text { Franklin, } \\ \text { and } & \text { Roberts } \\ (2002) & \end{array}
Texture Analysis 纹理分析 Examines spatial distribution of pixel intensities for texture changes.
检查纹理变化的像素强度的空间分布。
Suitable for textural and structural changes
适合纹理和结构变化
Window size and texture scale choice impact accuracy
窗口大小和纹理缩放选择影响准确性
Li, Gong, and Zhang (2019); Ansari, Buddhiraju, and Malhotra (2020)
李、龚和张(2019);安萨里、布迪拉朱和马尔霍特拉(2020)

极化特征方法
Polarization
Feature
Method
Polarization Feature Method| Polarization | | :--- | | Feature | | Method |

统计分析基础的方法
Statistical
Analysis-
Based Methods
Statistical Analysis- Based Methods| Statistical | | :--- | | Analysis- | | Based Methods |
Analyzes polarimetric statistical distributions for change detection.
分析极化统计分布以进行变化检测。
Provides detailed polarization insights
提供详细的极化洞察
Requires highquality, multipolarization data
需要高质量的多极化数据
Inglada and Mercier ( 2007 ) ; Zhao et al. ( 2020 )  Inglada   and   Mercier  ( 2007 ) ;  Zhao et al.  ( 2020 ) [" Inglada "," and "],[" Mercier ",(2007);],[" Zhao et al. ",(2020)]\begin{array}{lr} \hline \text { Inglada } & \text { and } \\ \text { Mercier } & (2007) ; \\ \text { Zhao et al. } & (2020) \end{array}

目标散射特征基础方法
Target
Scattering
Characteristic-
Based Meth-
ods
Target Scattering Characteristic- Based Meth- ods| Target | | :--- | | Scattering | | Characteristic- | | Based Meth- | | ods |
Focuses on physical scattering properties to detect surface changes.
专注于物理散射特性以检测表面变化。
Sensitive to structural characteristics
对结构特征敏感
Computationall intensive and datadependent
计算密集型和数据依赖型
Han, Cong, and Zhang (2013); Sun et al. (2021)
韩、从和张(2013);孙等(2021)
Fusion Methods 融合方法 Combines multiple data types or features to improve detection accuracy.
结合多种数据类型或特征以提高检测准确性。
Enhances ro- bustness and accuracy  Enhances   ro-   bustness   and   accuracy  [" Enhances "," ro- "],[" bustness "," and "],[" accuracy "]\begin{array}{lr} \hline \text { Enhances } & \text { ro- } \\ \text { bustness } & \text { and } \\ \text { accuracy } \end{array} Increases complexity and processing time
增加复杂性和处理时间
Ratha et al. (2017); Ma et al. (2019)
拉塔等人(2017);马等人(2019)
Classification Comparison
分类比较
Direct Target Comparison 直接目标比较 Compares classified target areas directly between time phases.
直接比较不同时间阶段的分类目标区域。
Simple interpretation and visual analysis
简单解释和视觉分析
Prone to misclassification errors
容易发生误分类错误

吴等人(2017);赵等人(2017)
Wu et al. (2017);
Zhao et al. (2017)
Wu et al. (2017); Zhao et al. (2017)| Wu et al. (2017); | | :--- | | Zhao et al. (2017) |
Target Classification Comparison
目标分类比较
Compares classifications across time phases for change identification.
比较不同时间阶段的分类以识别变化。
Effective for categorical change detection
有效的类别变化检测
Dependent on classifier quality
依赖于分类器质量

布里斯科, r r rr 乌拉比和多布森(1983);阿克巴里,杜尔杰里斯和埃尔托夫(2016)
Brisco, r r rr Ulaby,
and Dobson
(1983); Akbari,
Doulgeris, and
Eltoft (2016)
Brisco, r Ulaby, and Dobson (1983); Akbari, Doulgeris, and Eltoft (2016) | Brisco, $r$ | Ulaby, | | :--- | ---: | | and | Dobson | | (1983); | Akbari, | | Doulgeris, | and | | Eltoft (2016) | |
 机器学习
Machine
Learning
Machine Learning| Machine | | :--- | | Learning |
Deep Learning 深度学习 Utilizes neural networks for automated feature extraction and classification.
利用神经网络进行自动特征提取和分类。
Suitable for complex, highdimensional data
适合复杂的高维数据
Data-intensive and requires high computational resources
数据密集且需要高计算资源
Zhang et al. (2021a); Weng et al. (2024)  Zhang et al.   (2021a); Weng   et al. (2024)  {:[" Zhang et al. "],[" (2021a); Weng "],[" et al. (2024) "]:}\begin{aligned} & \text { Zhang et al. } \\ & \text { (2021a); Weng } \\ & \text { et al. (2024) } \end{aligned}
SVM Support r r rr Vector Machine used for boundary-based classi- fication.
支持 r r rr 向量机用于基于边界的分类。
Effective for binary or linear separable problems
有效于二元或线性可分问题

在处理多类和非线性问题时有限
Limited in
handling
multi-class
and non-linear
problems
Limited in handling multi-class and non-linear problems| Limited in | | :--- | :--- | | handling | | multi-class | | and non-linear | | problems |
Wang, Yang, and Jiao (2016); Wieland, Liu, and Yamazaki (2016)
王、杨和焦(2016);维兰德、刘和山崎(2016)

决策树/随机森林
Decision
Tree/Random Forest
Decision Tree/Random Forest| Decision | | :--- | | Tree/Random Forest |
Utilizes hierarchical decision structures fqr classification.
利用分层决策结构进行分类。
Interpretability and adaptability for various tasks
可解释性和适应性用于各种任务
Prone to overfitting in highdimensional spaces
在高维空间中容易过拟合
Simard, Saatchi, and De Grandi (2000); Mastro et al. (2022)
西马尔、萨奇和德格兰迪(2000);马斯特罗等(2022)
Method Subclass Subclass Description Advantages Limitations Application "Direct Comparison" "Difference Method" Measures the intensity difference between images. Simple and quick to calculate Limited to binary change detection; sensitive to noise " Coppin Bauer (1996) " Ratio Method Calculates the ratio of intensities for each pixel. Effective at reducing radiometric differences Sensitive to noise in low-intensity areas Conradsen et al. (2003) Regression Analysis Establishes regression models to account for gradual temporal changes. Suitable for gradual change detection "Requires a large and consistent dataset for accuracy" Zhuang, Ddeng, and Fan (2016); Zhao et al. (2021) "Image Transformation" CVA Uses Change Vector Analysis for multidimensional feature space analysis. Effective at detecting multidimensional changes Complexity in result interpretation; requires normalization Bayarjargal et al. (2006); Bovolo, Marchesi, and Bruzzone (2011) PCA Reduces data dimensionality to focus on significant change information. Reduces redundancy and enhances significant features May obscure minor but relevant changes "Li, Gong, and Zhang (2019); Schwartz et al. (2020)" K-T Kauth-Thomas transformation, particularly for vegetation index detection. Highly quad ef- fective for vegetation monitoring Limited use in non-vegetative change detection " Rogan, Franklin, and Roberts (2002) " Texture Analysis Examines spatial distribution of pixel intensities for texture changes. Suitable for textural and structural changes Window size and texture scale choice impact accuracy Li, Gong, and Zhang (2019); Ansari, Buddhiraju, and Malhotra (2020) "Polarization Feature Method" "Statistical Analysis- Based Methods" Analyzes polarimetric statistical distributions for change detection. Provides detailed polarization insights Requires highquality, multipolarization data " Inglada and Mercier (2007); Zhao et al. (2020)" "Target Scattering Characteristic- Based Meth- ods" Focuses on physical scattering properties to detect surface changes. Sensitive to structural characteristics Computationall intensive and datadependent Han, Cong, and Zhang (2013); Sun et al. (2021) Fusion Methods Combines multiple data types or features to improve detection accuracy. " Enhances ro- bustness and accuracy " Increases complexity and processing time Ratha et al. (2017); Ma et al. (2019) Classification Comparison Direct Target Comparison Compares classified target areas directly between time phases. Simple interpretation and visual analysis Prone to misclassification errors "Wu et al. (2017); Zhao et al. (2017)" Target Classification Comparison Compares classifications across time phases for change identification. Effective for categorical change detection Dependent on classifier quality "Brisco, r Ulaby, and Dobson (1983); Akbari, Doulgeris, and Eltoft (2016) " "Machine Learning" Deep Learning Utilizes neural networks for automated feature extraction and classification. Suitable for complex, highdimensional data Data-intensive and requires high computational resources " Zhang et al. (2021a); Weng et al. (2024) " SVM Support r Vector Machine used for boundary-based classi- fication. Effective for binary or linear separable problems "Limited in handling multi-class and non-linear problems" Wang, Yang, and Jiao (2016); Wieland, Liu, and Yamazaki (2016) "Decision Tree/Random Forest" Utilizes hierarchical decision structures fqr classification. Interpretability and adaptability for various tasks Prone to overfitting in highdimensional spaces Simard, Saatchi, and De Grandi (2000); Mastro et al. (2022)| Method | Subclass | Subclass Description | Advantages | Limitations | Application | | :---: | :---: | :---: | :---: | :---: | :---: | | Direct <br> Comparison | Difference <br> Method | Measures the intensity difference between images. | Simple and quick to calculate | Limited to binary change detection; sensitive to noise | $\begin{aligned} & \text { Coppin } \\ & \text { Bauer (1996) } \end{aligned}$ | | | Ratio Method | Calculates the ratio of intensities for each pixel. | Effective at reducing radiometric differences | Sensitive to noise in low-intensity areas | Conradsen et al. (2003) | | | Regression Analysis | Establishes regression models to account for gradual temporal changes. | Suitable for gradual change detection | Requires a <br> large and consistent dataset for accuracy | Zhuang, Ddeng, and Fan (2016); Zhao et al. (2021) | | Image <br> Transformation | CVA | Uses Change Vector Analysis for multidimensional feature space analysis. | Effective at detecting multidimensional changes | Complexity in result interpretation; requires normalization | Bayarjargal et al. (2006); Bovolo, Marchesi, and Bruzzone (2011) | | | PCA | Reduces data dimensionality to focus on significant change information. | Reduces redundancy and enhances significant features | May obscure minor but relevant changes | Li, Gong, and <br> Zhang (2019); <br> Schwartz et al. <br> (2020) | | | K-T | Kauth-Thomas transformation, particularly for vegetation index detection. | Highly $\quad$ ef- fective for vegetation monitoring | Limited use in non-vegetative change detection | $\begin{array}{lr} \text { Rogan, } & \text { Franklin, } \\ \text { and } & \text { Roberts } \\ (2002) & \end{array}$ | | | Texture Analysis | Examines spatial distribution of pixel intensities for texture changes. | Suitable for textural and structural changes | Window size and texture scale choice impact accuracy | Li, Gong, and Zhang (2019); Ansari, Buddhiraju, and Malhotra (2020) | | Polarization <br> Feature <br> Method | Statistical <br> Analysis- <br> Based Methods | Analyzes polarimetric statistical distributions for change detection. | Provides detailed polarization insights | Requires highquality, multipolarization data | $\begin{array}{lr} \hline \text { Inglada } & \text { and } \\ \text { Mercier } & (2007) ; \\ \text { Zhao et al. } & (2020) \end{array}$ | | | Target <br> Scattering <br> Characteristic- <br> Based Meth- <br> ods | Focuses on physical scattering properties to detect surface changes. | Sensitive to structural characteristics | Computationall intensive and datadependent | Han, Cong, and Zhang (2013); Sun et al. (2021) | | | Fusion Methods | Combines multiple data types or features to improve detection accuracy. | $\begin{array}{lr} \hline \text { Enhances } & \text { ro- } \\ \text { bustness } & \text { and } \\ \text { accuracy } \end{array}$ | Increases complexity and processing time | Ratha et al. (2017); Ma et al. (2019) | | Classification Comparison | Direct Target Comparison | Compares classified target areas directly between time phases. | Simple interpretation and visual analysis | Prone to misclassification errors | Wu et al. (2017); <br> Zhao et al. (2017) | | | Target Classification Comparison | Compares classifications across time phases for change identification. | Effective for categorical change detection | Dependent on classifier quality | Brisco, $r$ Ulaby, <br> and Dobson <br> (1983); Akbari, <br> Doulgeris, and <br> Eltoft (2016) | | Machine <br> Learning | Deep Learning | Utilizes neural networks for automated feature extraction and classification. | Suitable for complex, highdimensional data | Data-intensive and requires high computational resources | $\begin{aligned} & \text { Zhang et al. } \\ & \text { (2021a); Weng } \\ & \text { et al. (2024) } \end{aligned}$ | | | SVM | Support $r$ Vector Machine used for boundary-based classi- fication. | Effective for binary or linear separable problems | Limited in <br> handling <br> multi-class <br> and non-linear <br> problems | Wang, Yang, and Jiao (2016); Wieland, Liu, and Yamazaki (2016) | | | Decision <br> Tree/Random Forest | Utilizes hierarchical decision structures fqr classification. | Interpretability and adaptability for various tasks | Prone to overfitting in highdimensional spaces | Simard, Saatchi, and De Grandi (2000); Mastro et al. (2022) |

3.2.1. Direct comparison method
3.2.1. 直接比较法

Direct comparison is a SAR image change detection method based on arithmetic operations. This approach swiftly extracts changes in the backscattering coefficient of targets at the same location by performing pixel-by-pixel subtraction or division operations on dual-time-phase SAR images. The primary direct comparison techniques include the difference method, ratio method, and regression analysis method(Singh (1989)), each optimized to meet specific change detection requirements.
直接比较是一种基于算术运算的 SAR 图像变化检测方法。该方法通过对双时相 SAR 图像进行逐像素的减法或除法运算,迅速提取同一位置目标的后向散射系数变化。主要的直接比较技术包括差分法、比率法和回归分析法(Singh (1989)),每种方法都经过优化以满足特定的变化检测需求。
The difference method is the most fundamental direct comparison technique, obtaining change information through pixel-by-pixel subtraction of dual-time-phase images. While this method is widely applied in optical imagery, it is less effective for SAR images due to the presence of multiplicative noise(Rignot and van Zyl (1993)). As a result, the ratio method has become the primary approach for change detection in SAR images. Various ratio-based detection operators have been further developed, such as the logarithmic ratio detector(Ma, Gong, and Zhou (2012)), the normalized average ratio(Hou et al. (2014)) and the Gaussian Log-Ratio Detector(Conradsen et al. (2003)).
差分方法是最基本的直接比较技术,通过对双时间相位图像逐像素相减来获取变化信息。虽然该方法在光学图像中广泛应用,但由于存在乘法噪声,它在合成孔径雷达(SAR)图像中的效果较差(Rignot 和 van Zyl (1993))。因此,比例方法已成为 SAR 图像变化检测的主要方法。进一步开发了各种基于比例的检测算子,如对数比率检测器(Ma, Gong, 和 Zhou (2012))、归一化平均比率(Hou 等 (2014))和高斯对数比率检测器(Conradsen 等 (2003))。
Image regression analysis employs a regression model to describe the relationships between pixels in SAR images across different time phases, addressing variations in mean and variance between images. This method uses sample points to estimate the parameters of a linear regression model through least squares, with the time phase 1 image serving as a predictor for the time phase 2 image. The predicted result is then compared to the actual time phase 2 image, with pixels showing significant differences in feature information flagged as change indicators. Image regression analysis is particularly suited for cases where mean and variance differ across multiple time-lapse images, as it helps mitigate the influence of factors such as illumination, viewing angle, and climatic conditions on detection outcomes. This approach is effective for identifying large-scale changes but has limitations in detecting finer, small-scale changes.
图像回归分析采用回归模型来描述不同时间阶段 SAR 图像中像素之间的关系,解决图像之间均值和方差的变化。该方法使用样本点通过最小二乘法估计线性回归模型的参数,以时间阶段 1 的图像作为时间阶段 2 图像的预测器。然后将预测结果与实际的时间阶段 2 图像进行比较,像素中显示出显著特征信息差异的部分被标记为变化指示器。图像回归分析特别适用于多个时间间隔图像中均值和方差不同的情况,因为它有助于减轻照明、视角和气候条件等因素对检测结果的影响。这种方法有效识别大规模变化,但在检测更细微的小规模变化方面存在局限性。
Direct comparison methods can quickly detect areas of change by performing direct numerical calculations on images from different time periods. The primary advantage of these methods lies in their simplicity, making them well-suited for large-scale change detection tasks. Notably, the ratio method demonstrates strong noise immunity in SAR images. However, direct comparison methods are highly dependent on image quality and viewing angle, and they face limitations when addressing complex surface changes. Additionally, while regression analysis can effectively capture details in highnoise environments, it presents challenges in terms of applicability. As a result, the direct comparison method is ideal for straightforward change detection scenarios but requires integration with other approaches to enhance performance in more complex settings.
直接比较方法可以通过对不同时间段的图像进行直接数值计算,快速检测变化区域。这些方法的主要优点在于其简单性,使其非常适合大规模变化检测任务。值得注意的是,比例法在合成孔径雷达(SAR)图像中表现出强大的抗噪声能力。然而,直接比较方法高度依赖于图像质量和观察角度,并且在处理复杂表面变化时面临限制。此外,尽管回归分析可以有效捕捉高噪声环境中的细节,但在适用性方面存在挑战。因此,直接比较方法非常适合简单的变化检测场景,但在更复杂的环境中需要与其他方法结合以提高性能。

3.2.2. Image transformation methods
3.2.2. 图像变换方法

Image transformation methods work by converting images into a feature space and analyzing the transformed results to extract pixels that indicate change. Common transformation techniques include Change Vector Analysis (CVA), Principal Component Analysis (PCA), Kauth-Thomas Transformation (KT Transformation), texture analysis, Multivariate Alteration Detection (MAD), and Slow Feature Analysis (SFA). These methods effectively extract or enhance change information, demonstrating high detection accuracy in multidimensional feature spaces. However, during the feature extraction process, important change information may be obscured by numerous transformed features, potentially affecting detection outcomes. Therefore, when applying
图像变换方法通过将图像转换为特征空间并分析变换结果来提取指示变化的像素。常见的变换技术包括变化矢量分析(CVA)、主成分分析(PCA)、Kauth-Thomas 变换(KT 变换)、纹理分析、多变量变化检测(MAD)和慢特征分析(SFA)。这些方法有效提取或增强变化信息,在多维特征空间中表现出高检测准确性。然而,在特征提取过程中,重要的变化信息可能会被众多变换特征所掩盖,从而可能影响检测结果。因此,在应用时

image transformation methods, it is crucial to carefully select the feature space and transformation parameters to ensure that key change information is accurately extracted and retained.
图像变换方法中,仔细选择特征空间和变换参数至关重要,以确保准确提取和保留关键信息变化。
Change Vector Analysis (CVA) is a multidimensional spatial analysis method that identifies areas of change by calculating the vector difference of each pixel across images from different time phases. CVA is particularly effective in detecting the direction and magnitude of change within a multidimensional feature space, making it well-suited for complex scene analysis. For instance, Shen, Guo, and Liao (2007) employed CVA to detect inundation changes, calculating change vectors from multi-temporal and multipolarization SAR images to create a cosine image that highlights change features. Decision trees were subsequently applied to extract regions and types of change. CVA can effectively extract target change characteristics, allowing for accurate identification. For example, Qi and Yeh (2013) combined CVA with post-classification analysis to detect land use and land cover (LULC) changes, using the coherence matrix in CVA to locate change areas and post-classification analysis to identify change types. However, CVA generally uses the data from the initial time period as a baseline reference vector (Cohen (1998)), which can result in different types of change being represented by similar magnitudes and directions, affecting detection accuracy. Additionally, CVA is sensitive to noise and requires high-precision pre-processing and image correction.
变化矢量分析(CVA)是一种多维空间分析方法,通过计算不同时间阶段图像中每个像素的矢量差异来识别变化区域。CVA 在检测多维特征空间内变化的方向和幅度方面特别有效,使其非常适合复杂场景分析。例如,沈、郭和廖(2007)采用 CVA 检测淹没变化,从多时相和多极化 SAR 图像中计算变化矢量,创建一个突出变化特征的余弦图像。随后应用决策树提取变化的区域和类型。CVA 能够有效提取目标变化特征,从而实现准确识别。例如,齐和叶(2013)将 CVA 与后分类分析相结合,以检测土地利用和土地覆盖(LULC)变化,利用 CVA 中的相干矩阵定位变化区域,并通过后分类分析识别变化类型。 然而,CVA 通常使用初始时间段的数据作为基线参考向量(Cohen (1998)),这可能导致不同类型的变化以相似的大小和方向表示,从而影响检测准确性。此外,CVA 对噪声敏感,需要高精度的预处理和图像校正。
Principal Component Analysis (PCA) reduces the dimensionality of original image data through a linear transformation, extracting principal components that best capture the variance within the data. PCA can enhance prominent change features, minimize data redundancy, and increase the efficiency of change detection. Traditional PCA methods project two images into a PCA feature space to extract differences, while differential PCA first computes the difference between images before applying PCA transformation. For example, Cheng et al. (2013) applied PCA to differential image blocks generated by the absolute logarithmic ratio operator to extract feature vectors, then clustered these vectors with the k-means algorithm to produce a change detection map. To capture polarization feature information unique to PolSAR images, Imani (2023) refined three popular transformation-based feature extraction methods. They incorporated polarization scattering characteristics, such as randomness and scattering mechanisms, to define the scattering coefficient and embedded scattering information into the PCA, Linear Discriminant Analysis (LDA), and Locality Preserving Projections (LPP) calculations, enhancing the utilization of scattering data. The Kauth-Thomas (KT) transform, similar to PCA, is based on the statistical characteristics of image data, reducing redundancy due to band correlation and providing feature information with physical significance. While the KT transform has proven effective for detecting changes in vegetation and land surface features, it is less suited to changes that are not primarily spectral, making it less commonly used in SAR image change detection.
主成分分析(PCA)通过线性变换减少原始图像数据的维度,提取最佳捕捉数据方差的主成分。PCA 可以增强显著变化特征,最小化数据冗余,并提高变化检测的效率。传统的 PCA 方法将两幅图像投影到 PCA 特征空间中以提取差异,而差分 PCA 则首先计算图像之间的差异,然后再应用 PCA 变换。例如,Cheng 等人(2013)将 PCA 应用于由绝对对数比率算子生成的差分图像块,以提取特征向量,然后使用 k 均值算法对这些向量进行聚类,生成变化检测图。为了捕捉 PolSAR 图像独特的极化特征信息,Imani(2023)改进了三种流行的基于变换的特征提取方法。 他们结合了极化散射特性,如随机性和散射机制,以定义散射系数,并将散射信息嵌入到主成分分析(PCA)、线性判别分析(LDA)和局部保持投影(LPP)计算中,从而增强了散射数据的利用。Kauth-Thomas(KT)变换类似于 PCA,基于图像数据的统计特性,减少了由于波段相关性造成的冗余,并提供了具有物理意义的特征信息。虽然 KT 变换在检测植被和地表特征变化方面已被证明有效,但对于那些主要不是光谱变化的情况,它的适用性较差,因此在 SAR 图像变化检测中使用较少。
Texture analysis leverages image characteristics such as color, texture, shape, and spatial relationships to comprehensively assess global or local (pixel-level and neighborhood) contextual information, effectively mitigating coherent speckle noise interference in SAR images. Color features, as global descriptors, capture surface characteristics of features in an image or region; however, due to the inherent properties of SAR, they are typically applied in fusion studies with optical or very-high-resolution (VHR) images. Texture features are also global descriptors, representing characteristics of features within an image or image region. Common texture descriptors include the Local Binary Pattern (LBP) and its derivatives (Kim, Park, and Lee (2023)), and the Gray-Level Co-Occurrence Matrix (GLCM) (Huifu, Kazhong, and Hongdong (2016)).
纹理分析利用图像特征,如颜色、纹理、形状和空间关系,全面评估全局或局部(像素级和邻域)上下文信息,有效减轻合成孔径雷达(SAR)图像中的相干斑点噪声干扰。颜色特征作为全局描述符,捕捉图像或区域中特征的表面特征;然而,由于 SAR 的固有特性,它们通常与光学或超高分辨率(VHR)图像的融合研究一起应用。纹理特征也是全局描述符,代表图像或图像区域内特征的特征。常见的纹理描述符包括局部二值模式(LBP)及其衍生物(Kim, Park, and Lee (2023)),以及灰度共生矩阵(GLCM)(Huifu, Kazhong, and Hongdong (2016))。
Shape features are divided into contour and region features: contour features (edge features) represent the outline of a region, while region features encompass edge information and internal shape attributes. Texture analysis generally considers multiple image features; for example, Gupta, Singh, and Kumar (2023) utilized a random forest to select an optimal feature set, incorporating polarization, texture, color, and wavelet features, which, combined with an OTSU threshold, generated classification rules. This adaptive land cover classification approach based on SAR data enhanced classification accuracy and adaptability. Balling, Herold, and Reiche (2023) introduced a method that combines SAR backscatter with GLCM texture features to detect forest disturbances. However, texture analysis is sensitive to noise and depends on the selected texture scale.
形状特征分为轮廓特征和区域特征:轮廓特征(边缘特征)表示区域的轮廓,而区域特征则包含边缘信息和内部形状属性。纹理分析通常考虑多个图像特征;例如,Gupta、Singh 和 Kumar(2023)利用随机森林选择最佳特征集,结合极化、纹理、颜色和小波特征,并与 OTSU 阈值结合,生成分类规则。这种基于 SAR 数据的自适应土地覆盖分类方法提高了分类准确性和适应性。Balling、Herold 和 Reiche(2023)提出了一种将 SAR 后向散射与 GLCM 纹理特征相结合的方法,以检测森林干扰。然而,纹理分析对噪声敏感,并且依赖于所选的纹理尺度。
Image transformation methods excel at enhancing change information by transforming feature spaces, demonstrating significant advantages in multidimensional feature extraction. However, methods like Change Vector Analysis (CVA) and Principal Component Analysis (PCA) face limitations in handling noise and extracting nonlinear features. Although the Kauth-Thomas (KT) transformation and texture analysis methods perform well in specific scenarios, their applicability remains limited. Therefore, applying image transformation methods requires careful consideration of the specific characteristics of the target scene, along with appropriate pre-processing and parameter selection, to achieve optimal detection outcomes.
图像变换方法在通过变换特征空间来增强变化信息方面表现出色,在多维特征提取中具有显著优势。然而,变化矢量分析(CVA)和主成分分析(PCA)等方法在处理噪声和提取非线性特征方面存在局限性。尽管 Kauth-Thomas(KT)变换和纹理分析方法在特定场景中表现良好,但它们的适用性仍然有限。因此,应用图像变换方法时,需要仔细考虑目标场景的具体特征,以及适当的预处理和参数选择,以实现最佳检测结果。

3.2.3. Polarization Feature Method
3.2.3. 极化特征方法

Research on change detection methods for PolSAR images primarily focuses on two main approaches: statistical analysis-based methods, which involve hypothesis testing, and target scattering characteristic-based methods, which leverage polarization features.
对 PolSAR 图像变化检测方法的研究主要集中在两种主要方法:基于统计分析的方法,涉及假设检验,以及基于目标散射特征的方法,利用极化特征。
Statistical analysis-based methods for change detection in PolSAR images assess the similarity of corresponding regions in two images by utilizing a statistical model of PolSAR data and deriving a test statistic for change detection. Lee and Grunes (1992) established the complex Wishart distribution, which models the covariance and coherence matrices of PolSAR data, and experimentally validated this distribution using statistical modeling of actual PolSAR measurements. However, this distribution is applicable primarily to homogeneous regions. Various statistical similarity-based detection operators have been developed, such as the likelihood ratio test based on the complex Wishart distribution (Yin and Yang (2016)), the change detection operator focused on optimal polarimetric contrast enhancement (Akbari et al. (2017)), and the Hotelling-Lawley Trace (HLT) detection operator based on the Fisher-Snedecor distribution (Blaschke (2005)). These methods exploit the statistical characteristics of SAR data to enhance detection accuracy. For example, Bouhlel et al. (2022) modeled wavelet coefficients with multivariate probability distributions, further leveraging statistical data information. Additionally, Chen (2023) introduced an HLT-WTMF algorithm for PolSAR change detection based on HLT difference maps and wavelet domain TMF, which significantly improves detection sensitivity and accuracy. Hypothesis testing methods have also been adapted for image segmentation to enable region- or object-based change detection. Recent advancements in region-based PolSAR change detection include approaches such as a unipolar method using multi-scale segmentation (Gong et al. (2016)), a method combining superpixel segmentation with a deep confidence network (Ban and Jacob (2013)), and a PolSAR change detection approach based on enhanced watershed segmentation and Markov random fields
基于统计分析的 PolSAR 图像变化检测方法通过利用 PolSAR 数据的统计模型评估两幅图像中对应区域的相似性,并推导出用于变化检测的检验统计量。Lee 和 Grunes(1992)建立了复杂的 Wishart 分布,该分布对 PolSAR 数据的协方差和相干矩阵进行建模,并通过对实际 PolSAR 测量的统计建模进行了实验验证。然而,该分布主要适用于均匀区域。已经开发了各种基于统计相似性的检测算子,例如基于复杂 Wishart 分布的似然比检验(Yin 和 Yang(2016))、专注于最佳极化对比度增强的变化检测算子(Akbari 等(2017))以及基于 Fisher-Snedecor 分布的 Hotelling-Lawley Trace(HLT)检测算子(Blaschke(2005))。这些方法利用 SAR 数据的统计特性来提高检测准确性。例如,Bouhlel 等人。 (2022)使用多元概率分布对小波系数进行了建模,进一步利用统计数据。此外,陈(2023)提出了一种基于 HLT 差异图和小波域 TMF 的 PolSAR 变化检测算法 HLT-WTMF,显著提高了检测灵敏度和准确性。假设检验方法也已被调整用于图像分割,以实现基于区域或对象的变化检测。最近在基于区域的 PolSAR 变化检测方面的进展包括使用多尺度分割的单极方法(龚等)。(2016)),一种将超像素分割与深度置信网络相结合的方法(Ban 和 Jacob(2013)),以及一种基于增强的分水岭分割和马尔可夫随机场的 PolSAR 变化检测方法

(Omati and Sahebi (2018)). Furthermore, Zhang et al. (2022a) refined segmentation techniques by applying an improved SLIC algorithm for segmenting PolSAR images, using the average coherence matrix as the detection unit. They compared the effectiveness of three classic test statistics at both regional and pixel levels, demonstrating the robustness of these statistics across different scales .
(Omati 和 Sahebi(2018))。此外,Zhang 等(2022a)通过应用改进的 SLIC 算法对 PolSAR 图像进行分割,使用平均相干矩阵作为检测单元,进一步完善了分割技术。他们在区域和像素级别比较了三种经典检验统计量的有效性,展示了这些统计量在不同尺度上的稳健性。
When the type of a local object changes, the target echo’s scattering characteristics also alter. Consequently, polarization feature-based analysis methods are frequently employed for change detection 1(Novak, Burl, and Irving (1993)). Polarization decomposition is a key technique for extracting polarization characteristics. By decomposing the target into multiple scattering components that reveal the physical scattering mechanisms through various decomposition models, polarization decomposition provides a comprehensive representation of the target’s scattering characteristics. Polarization decomposition methods can be categorized by purpose - either scattering model-based or scattering vector-based-and by data type, with coherent and incoherent decompositions being the primary categories. Coherent decomposition techniques include Pauli decomposition 1(Cloude and Pottier (1996)), Krogager decomposition (Cameron and Leung (1990)), and Cameron decomposition 1; incoherent decomposition methods consist of Huynen decomposition, Cloude-Pottier decomposition (Cloude and Pottier (1997)), Freeman three-component decomposition (Freeman and Durden (1998)), and Yamaguchi’s four-component decomposition (Yamaguchi et al. (2005)), among others. Early studies often utilized simple polarimetric features as operators for change detection, but as research advanced, various target decomposition methods were applied to enhance detection accuracy. Additionally, other attributes such as texture and statistical features have been integrated with polarization features to improve detection performance through feature fusion. For example, Ma et al. (2019) employed an incoherent polarization decomposition combined with a gray-level cooccurrence matrix to extract 19 polarimetric and 8 textural features, achieving a fusion of polarization and texture features for building extraction from PolSAR images. Liu et al. (2020) applied SAR image scattering feature decomposition for RGB color synthesis to visually differentiate intact and damaged buildings, facilitating building damage assessment. Yu and Wei (2023) introduced a PolSAR change detection approach that combines polarimetric cross-correlation and spatial context information. This method leverages the HLT statistic to analyze scattering correlation and amplitude ratios across different phases, generating a difference map. Using a Markov Random Field (MRF) and Simulated Annealing (SA) algorithm, the difference map undergoes binary segmentation to delineate changed areas effectively.
当局部物体的类型发生变化时,目标回波的散射特性也会改变。因此,基于极化特征的分析方法常被用于变化检测 1(Novak, Burl 和 Irving (1993))。极化分解是提取极化特性的关键技术。通过将目标分解为多个散射分量,这些分量通过各种分解模型揭示物理散射机制,极化分解提供了目标散射特性的全面表示。极化分解方法可以按目的分类——散射模型基础或散射向量基础——以及按数据类型分类,主要分为相干和非相干分解。 相干分解技术包括保利分解(Cloude 和 Pottier(1996))、Krogager 分解(Cameron 和 Leung(1990))和 Cameron 分解;非相干分解方法包括 Huynen 分解、Cloude-Pottier 分解(Cloude 和 Pottier(1997))、Freeman 三分量分解(Freeman 和 Durden(1998))以及 Yamaguchi 的四分量分解(Yamaguchi 等(2005))等。早期研究通常利用简单的极化特征作为变化检测的算子,但随着研究的进展,各种目标分解方法被应用于提高检测精度。此外,纹理和统计特征等其他属性也与极化特征结合,通过特征融合来改善检测性能。例如,Ma 等(2019)采用了一种非相干极化分解,结合灰度共生矩阵提取了 19 个极化特征和 8 个纹理特征,实现了从 PolSAR 图像中提取建筑物的极化和纹理特征的融合。刘等。 (2020)应用 SAR 图像散射特征分解进行 RGB 颜色合成,以直观区分完整和受损建筑,促进建筑损坏评估。余和魏(2023)提出了一种结合极化交叉相关和空间上下文信息的 PolSAR 变化检测方法。该方法利用 HLT 统计量分析不同相位的散射相关性和幅度比,生成差异图。通过使用马尔可夫随机场(MRF)和模拟退火(SA)算法,差异图经过二值分割,有效划定变化区域。
Change detection methods for PolSAR images primarily fall into two categories: statistical analysis-based methods and target scattering characteristic-based methods. Hypothesis-testing methods apply statistical tests to assess image differences, providing theoretical rigor and making them suitable for high-noise images, though they depend on specific distributional assumptions. Polarization feature-based methods utilize polarization information for change detection, making them effective in complex scenes, but they are sensitive to noise and computationally intensive. Combined methods leverage the statistical robustness of hypothesis testing alongside the detailed scattering information from polarization features, enhancing detection accuracy and robustness. However, these combined approaches require optimized computation to manage complexity effectively.
PolSAR 图像的变化检测方法主要分为两类:基于统计分析的方法和基于目标散射特征的方法。假设检验方法应用统计检验来评估图像差异,提供理论严谨性,使其适用于高噪声图像,尽管它们依赖于特定的分布假设。基于极化特征的方法利用极化信息进行变化检测,使其在复杂场景中有效,但对噪声敏感且计算量大。组合方法结合了假设检验的统计稳健性和极化特征的详细散射信息,提高了检测的准确性和稳健性。然而,这些组合方法需要优化计算以有效管理复杂性。
The classification comparison method is widely used for change detection in remote sensing images and is typically divided into direct classification comparison, postclassification comparison (PCC), and joint classification comparison (JCC). Each of
分类比较方法广泛用于遥感图像的变化检测,通常分为直接分类比较、后分类比较(PCC)和联合分类比较(JCC)。每个

these approaches offers unique advantages and limitations, making it essential to select the appropriate method based on the specific characteristics of the data and the requirements of the application.
这些方法各有独特的优点和局限性,因此根据数据的具体特征和应用的要求选择合适的方法至关重要。
The direct object comparison method is widely used for generating land use and land cover change maps. By analyzing multi-temporal datasets, this approach inputs images as classified features into a classifier, utilizing either supervised or unsupervised techniques to yield consistent land change results (Lunetta et al. (2002)). Currently, this method is the primary choice for change detection, especially in object-based change detection applications. Some methods that perform well in pixel-based change detection-such as independent component analysis, multivariate change detection, and change vector analysis - can also deliver optimal results when adapted for objectoriented change detection (Zhao and Zhao (2018)). For example, Qi and Yeh (2013) applied object-oriented CVA change detection to PolSAR imagery, achieving high identification accuracy.Typically, the direct target comparison method involves three main steps: (1) image pre-processing, (2) difference map generation, and (3) difference map analysis. The primary limitation of this approach is the dependence on the quality of the difference map, which heavily influences the final detection outcome. In SAR image change detection, the main factors affecting difference map quality are speckle noise suppression and the accuracy of difference map analysis.
直接对象比较方法广泛用于生成土地利用和土地覆盖变化地图。通过分析多时相数据集,该方法将图像作为分类特征输入分类器,利用监督或无监督技术产生一致的土地变化结果(Lunetta 等,2002)。目前,该方法是变化检测的主要选择,特别是在基于对象的变化检测应用中。一些在基于像素的变化检测中表现良好的方法,如独立成分分析、多变量变化检测和变化矢量分析,在适应于面向对象的变化检测时也能提供最佳结果(Zhao 和 Zhao,2018)。例如,Qi 和 Yeh(2013)将面向对象的 CVA 变化检测应用于 PolSAR 影像,取得了高识别准确率。通常,直接目标比较方法包括三个主要步骤:(1)图像预处理,(2)差异图生成,以及(3)差异图分析。该方法的主要限制在于对差异图质量的依赖,这对最终检测结果有很大影响。 在合成孔径雷达图像变化检测中,影响差异图质量的主要因素是斑点噪声抑制和差异图分析的准确性。
The target classification comparison method, commonly known as the postclassification comparison (PCC) method, represents a typical supervised change detection approach. This method has distinct advantages: it is less sensitive to registration accuracy, the classification algorithms are well-established and straightforward to implement, and it effectively tracks changes in land feature categories, making it widely applicable in change detection (Wu et al. (2017)). Given SAR imagery’s susceptibility to coherent speckle noise and the complexities in feature interpretation, the Wishart classification method, based on maximum likelihood estimation Brisco, Ulaby, and Dobson (1983); Lee, Grunes, and Kwok (1994), has been developed and extensively applied. For instance, Gomez et al. (2015) introduced a classification method leveraging random distances within random matrix space, governed by the complex Wishart distribution, and embedded this into a partial differential equation framework to achieve improved speckle noise suppression.Additionally, the target classification comparison method frequently incorporates machine learning to extract features from change samples, framing the change detection task as a classification problem. Qi and Yeh (2013), for example, employed an object-oriented classification method combining polarimetric decomposition, decision trees, and support vector machines for PCC change detection, enabling efficient real-time monitoring of land use and land cover changes. Akbari, Doulgeris, and Eltoft (2016) utilized an object-oriented, non-Gaussian clustering algorithm for monitoring Arctic glaciers, which demonstrated enhanced performance in complex environments and yielded higher monitoring accuracy. However, a notable limitation of the target classification comparison method is the potential accumulation of classification errors from each individual image, which can influence the overall results. Thus, this method demands high classification accuracy across both images.
目标分类比较方法,通常称为后分类比较(PCC)方法,代表了一种典型的监督变化检测方法。该方法具有明显的优势:对配准精度的敏感性较低,分类算法成熟且易于实施,并且能够有效跟踪土地特征类别的变化,使其在变化检测中广泛适用(Wu et al. (2017))。考虑到 SAR 影像对相干斑点噪声的敏感性以及特征解释的复杂性,基于最大似然估计的 Wishart 分类方法(Brisco, Ulaby, 和 Dobson (1983); Lee, Grunes, 和 Kwok (1994))已被开发并广泛应用。例如,Gomez et al. (2015) 引入了一种利用随机矩阵空间内随机距离的分类方法,该方法受复杂 Wishart 分布的支配,并将其嵌入到偏微分方程框架中,以实现更好的斑点噪声抑制。此外,目标分类比较方法通常结合机器学习从变化样本中提取特征,将变化检测任务框定为分类问题。例如,Qi 和 Yeh(2013)采用了一种结合极化分解、决策树和支持向量机的面向对象分类方法进行 PCC 变化检测,实现了对土地利用和土地覆盖变化的高效实时监测。Akbari、Doulgeris 和 Eltoft(2016)利用了一种面向对象的非高斯聚类算法监测北极冰川,在复杂环境中表现出更好的性能,并提高了监测精度。然而,目标分类比较方法的一个显著局限性是每个单独图像的分类错误可能会累积,从而影响整体结果。因此,该方法要求在两幅图像中都具有高分类精度。
Joint Classification Comparison (JCC) represents an adaptive classification method that enhances the conventional PCC by integrating temporal information to assess similarity between adjacent datasets, thereby mitigating the impact of cumulative errors. Although JCC has shown promise, its application in change detection remains underexplored Han and Zhou (2015). Li et al. (2009) employed a K-means-based JCC method for change detection in single-polarization SAR imagery, achieving favorable outcomes. Similarly, Zhao et al. (2017) developed a change detection framework for
联合分类比较(JCC)是一种自适应分类方法,通过整合时间信息来评估相邻数据集之间的相似性,从而增强传统的 PCC,减轻累积误差的影响。尽管 JCC 显示出良好的前景,但其在变化检测中的应用仍然未被充分探索(Han 和 Zhou,2015)。Li 等(2009)采用基于 K 均值的 JCC 方法进行单极化 SAR 影像的变化检测,取得了良好的结果。同样,Zhao 等(2017)开发了一个变化检测框架。
PolSAR images, leveraging joint classification with hypothesis testing and K&I similarity constraints. Nonetheless, due to the distinct characteristics of SAR data, the adoption of JCC methods in SAR imagery change detection remains limited.
PolSAR 图像,利用假设检验和 K&I 相似性约束的联合分类。然而,由于 SAR 数据的独特特性,JCC 方法在 SAR 影像变化检测中的应用仍然有限。
Classification comparison methods are extensively applied in change detection, each offering unique advantages. Direct classification comparison methods are straightforward and efficient for generating change information but face challenges with noise and non-linear feature variations. The PCC method minimizes the impact of registration errors and can effectively monitor categorical changes in land features, though classification errors may accumulate and influence the final results. The joint classification method incorporates temporal information with adaptive classification, alleviating error accumulation; however, its effectiveness in SAR imagery remains limited. Furthermore, the success of classification comparison methods relies heavily on classification accuracy and data quality. Therefore, when implementing these methods, selecting an appropriate approach based on data characteristics and conducting necessary preprocessing and parameter adjustments are essential.
分类比较方法广泛应用于变化检测,各自具有独特的优势。直接分类比较方法简单高效,能够生成变化信息,但在噪声和非线性特征变化方面面临挑战。PCC 方法最小化了配准误差的影响,能够有效监测土地特征的类别变化,尽管分类错误可能会累积并影响最终结果。联合分类方法结合了时间信息和自适应分类,减轻了错误累积;然而,其在 SAR 影像中的有效性仍然有限。此外,分类比较方法的成功在很大程度上依赖于分类准确性和数据质量。因此,在实施这些方法时,根据数据特征选择合适的方法,并进行必要的预处理和参数调整是至关重要的。

3.2.4. Machine learning 3.2.4. 机器学习

The previously discussed data analysis methods and detection models primarily fall within the scope of machine learning but focus more on feature extraction and change detection strategies rather than classification methods based on direct training. This section emphasizes the application of classification techniques in change detection. Machine learning methods, as versatile tools, are widely adopted across numerous applications and are integral to change detection processes. Common machine learning approaches in change detection include traditional algorithms and deep learning models. Typical traditional methods encompass support vector machines (SVM)(Cortes and Vapnik (1995)), decision trees (DT), and random forests (RF)(Surhone et al. (2010)) .
之前讨论的数据分析方法和检测模型主要属于机器学习的范畴,但更侧重于特征提取和变化检测策略,而不是基于直接训练的分类方法。本节强调分类技术在变化检测中的应用。机器学习方法作为多功能工具,广泛应用于众多领域,并且是变化检测过程中的重要组成部分。变化检测中常见的机器学习方法包括传统算法和深度学习模型。典型的传统方法包括支持向量机(SVM)(Cortes 和 Vapnik(1995))、决策树(DT)和随机森林(RF)(Surhone 等(2010))。
Support Vector Machine (SVM) is a supervised, non-parametric statistical learning algorithm well-suited for nonlinear classification tasks(Vapnik and Vapnik (1998)). Through the use of kernel functions, SVM maps sample data into a high-dimensional feature space, showing strong advantages for nonlinear mapping, particularly in cases with limited samples(Joachims (1999)). The strengths of SVM include low training costs, high computational efficiency, and robust performance in high-dimensional problem spaces; however, its theoretical complexity and sensitivity to missing data and parameter tuning can pose challenges. Fukuda, Katagiri, and Hirosawa (2002) pioneered the application of SVM in constructing high-dimensional spaces to differentiate various land features based on scattering characteristics by defining appropriate hyperplanes. Consequently, SVM has gained widespread use in change detection and image classification tasks. In SAR image change detection, SVM-based classifiers commonly leverage multi-temporal polarization features, texture attributes, and scattered intensity as input data, with kernel functions facilitating classification by mapping features into high-dimensional spaces to distinguish between changed and unchanged regions. For instance, Zhang et al. (2009) introduced an SVM classification method incorporating the Multi-Component Scattering Model (MCSM) and Gray-Level CoOccurrence Matrix (GLCM) texture features. To enhance detection accuracy, SVM is frequently integrated with dimensionality reduction techniques like PCA, optimizing feature selection and increasing robustness in complex SAR datasets. Niu and Ban (2013) utilized SVM to classify multi-temporal PolSAR polarization features, integrating shape and contextual information to refine classification results, achieving effective urban land cover mapping.
支持向量机(SVM)是一种监督式、非参数的统计学习算法,非常适合非线性分类任务(Vapnik 和 Vapnik(1998))。通过使用核函数,SVM 将样本数据映射到高维特征空间,显示出在非线性映射方面的强大优势,特别是在样本有限的情况下(Joachims(1999))。SVM 的优点包括低训练成本、高计算效率以及在高维问题空间中的稳健性能;然而,其理论复杂性以及对缺失数据和参数调优的敏感性可能带来挑战。Fukuda、Katagiri 和 Hirosawa(2002)开创性地将 SVM 应用于构建高维空间,以通过定义适当的超平面区分各种土地特征,基于散射特性。因此,SVM 在变化检测和图像分类任务中得到了广泛应用。 在 SAR 图像变化检测中,基于 SVM 的分类器通常利用多时相极化特征、纹理属性和散射强度作为输入数据,核函数通过将特征映射到高维空间来促进分类,以区分变化和未变化区域。例如,Zhang 等(2009)提出了一种结合多组分散射模型(MCSM)和灰度共生矩阵(GLCM)纹理特征的 SVM 分类方法。为了提高检测精度,SVM 常常与主成分分析(PCA)等降维技术结合,优化特征选择并增强复杂 SAR 数据集中的鲁棒性。Niu 和 Ban(2013)利用 SVM 对多时相 PolSAR 极化特征进行分类,结合形状和上下文信息以细化分类结果,实现了有效的城市土地覆盖制图。
A decision tree is a supervised learning algorithm that generates a hierarchical structure based on a sequence of decision rules, mapping object attributes to object values. In change detection, each node in the decision tree represents a test of a specific attribute, each branch represents a test outcome, and the leaf nodes represent classification labels or distributions (Larose and Larose (2014)). For instance, Simioni et al. (2020) employed a decision tree classification method to delineate swamp areas in multi-frequency SAR images. Decision trees offer advantages such as interpretability and simplicity in computation; however, they are prone to overfitting and may converge to local optima. Random forests, on the other hand, are data-driven, non-parametric classifiers that construct an ensemble of decision trees, improving robustness to noisy data and allowing for feature importance estimation. Random forests exhibit fast learning speeds and high computational efficiency (Lei et al. (2019)). For example, Wang et al. (2016) applied random forests to classify an enhanced feature set, achieving high accuracy. Song et al. (2022) further improved the High-Order Hybrid Discriminative Random Field (HO-HDRF) by integrating the Two-Layer Random Forest (TL-RF), creating the RF-HoDRF model, which effectively leverages discriminative texture and high-order structural features, enhancing accuracy in change detection.
决策树是一种监督学习算法,它基于一系列决策规则生成层次结构,将对象属性映射到对象值。在变化检测中,决策树中的每个节点代表对特定属性的测试,每个分支代表测试结果,叶节点代表分类标签或分布(Larose 和 Larose (2014))。例如,Simioni 等人 (2020) 使用决策树分类方法在多频率 SAR 图像中划定沼泽区域。决策树具有可解释性和计算简单等优点;然而,它们容易过拟合,并可能收敛到局部最优解。随机森林则是一种数据驱动的非参数分类器,它构建了一组决策树,提高了对噪声数据的鲁棒性,并允许特征重要性估计。随机森林表现出快速的学习速度和高计算效率(Lei 等人 (2019))。例如,Wang 等人 (2016) 应用随机森林对增强特征集进行分类,取得了高准确率。宋等人。 (2022)通过整合双层随机森林(TL-RF)进一步改进了高阶混合判别随机场(HO-HDRF),创建了 RF-HoDRF 模型,该模型有效利用了判别纹理和高阶结构特征,提高了变化检测的准确性。
In recent years, deep learning networks have demonstrated substantial advancements in the domain of change detection. These deep learning-based methods can autonomously extract change features directly from bi-temporal or time-series remote sensing imagery and use these features to segment images and generate change maps. The features learned by deep learning approaches exhibit robust adaptability, efficiently handling complex scenes and diverse change patterns. A variety of deep learning-based change detection models have been developed, such as Convolutional Neural Networks (CNNs) (Vinholi et al. (2020)), Deep Belief Networks (DBNs) (Samadi, Akbarizadeh, and Kaabi (2019)), Regions with Convolutional Neural Networks features (R-CNNs) (Sivapriya and Mohamed (2022)), and Convolutional Wavelet Neural Networks (CWNNs) (Zhang et al. (2021b)).Deep learning can be effectively integrated with other change detection techniques to enhance both detection accuracy and efficiency while retaining the inherent advantages of deep learning. For instance, Zhang et al. (2021c) proposed a deep spatio-temporal co-occurrence-aware CNN, introducing a 3D co-occurrence matrix as an auxiliary feature to effectively capture bi-temporal information, thereby improving change detection accuracy. Wang et al. (2022) developed a Graph-based Knowledge Supplement Network (GKS Net), which enhances contextual information by integrating discriminative knowledge from labeled datasets through a graph transmission module, effectively reducing noise interference and bridging feature correlations between datasets. Furthermore, Zhang et al. (2023) introduced ACAHNet, a hybrid model combining CNN and Transformer architectures with feature fusion mid-way through change detection. The Asymmetric Multi-Head Cross-Attention (AMCA) module in this model reduces the computational complexity of the Transformer, optimizing efficiency without compromising effectiveness.
近年来,深度学习网络在变化检测领域取得了显著进展。这些基于深度学习的方法可以自主从双时相或时间序列遥感影像中提取变化特征,并利用这些特征对图像进行分割和生成变化图。深度学习方法所学习的特征表现出强大的适应性,能够高效处理复杂场景和多样的变化模式。已经开发出多种基于深度学习的变化检测模型,如卷积神经网络(CNNs)(Vinholi 等,2020),深度置信网络(DBNs)(Samadi、Akbarizadeh 和 Kaabi,2019),具有卷积神经网络特征的区域(R-CNNs)(Sivapriya 和 Mohamed,2022),以及卷积小波神经网络(CWNNs)(Zhang 等,2021b)。深度学习可以有效地与其他变化检测技术结合,以提高检测的准确性和效率,同时保留深度学习的固有优势。例如,Zhang 等。 (2021c)提出了一种深度时空共现感知卷积神经网络,引入了一个 3D 共现矩阵作为辅助特征,以有效捕捉双时间信息,从而提高变化检测的准确性。王等人(2022)开发了一种基于图的知识补充网络(GKS Net),通过图传输模块整合来自标记数据集的区分知识,增强上下文信息,有效减少噪声干扰,并弥合数据集之间的特征关联。此外,张等人(2023)推出了 ACAHNet,这是一种结合卷积神经网络和变换器架构的混合模型,在变化检测过程中进行特征融合。该模型中的非对称多头交叉注意力(AMCA)模块降低了变换器的计算复杂性,优化了效率而不影响效果。
Despite notable successes, deep learning-based models face limitations in feature extraction, especially under restricted labeled data. Consequently, unsupervised or self-supervised methods have gained traction. For example, Li, Cao, and Meng (2024) applied a base model for change detection, proposing a Contrastive Language-Image Pretraining (CLIP)-based model. This approach extracts generalizable features via a frozen base model and uses a bridging module to align and inject features into a dualtwin network, enabling flexible feature extraction and change detection. Additionally, deep learning combined with clustering methods has shown promise in deriving cluster-
尽管取得了显著成功,基于深度学习的模型在特征提取方面仍面临限制,尤其是在标记数据受限的情况下。因此,无监督或自监督的方法逐渐受到关注。例如,Li、Cao 和 Meng(2024)应用了一个基础模型进行变化检测,提出了一种基于对比语言-图像预训练(CLIP)的模型。这种方法通过冻结的基础模型提取可泛化的特征,并使用桥接模块对齐并注入特征到双重网络中,从而实现灵活的特征提取和变化检测。此外,深度学习与聚类方法相结合在推导聚类方面显示出前景。

friendly representations(Zhang et al. (2018)). For instance, Li et al. (2019) introduced an unsupervised spatial fuzzy clustering technique that generates pseudo-labels, filtering samples to train a CNN directly on SAR images, thus bypassing the need for difference maps and improving detection robustness. However, while the CNN process is nominally unsupervised, its training phase remains supervised. Dong et al. (2021) proposed a multi-scale self-attention (SA) deep clustering method, using octave convolution with a K-means++ algorithm for end-to-end learning. By incorporating octave convolution and SA mechanisms, this method captures critical spatial structure information, while a multi-scale fusion module enhances contextual and semantic feature integration, reducing noise impact and refining feature differentiation.
友好的表示(Zhang et al. (2018))。例如,Li et al. (2019) 引入了一种无监督的空间模糊聚类技术,该技术生成伪标签,过滤样本以直接在 SAR 图像上训练 CNN,从而绕过差异图的需求并提高检测的鲁棒性。然而,尽管 CNN 过程名义上是无监督的,但其训练阶段仍然是有监督的。Dong et al. (2021) 提出了一个多尺度自注意力(SA)深度聚类方法,使用八度卷积与 K-means++算法进行端到端学习。通过结合八度卷积和 SA 机制,该方法捕捉关键的空间结构信息,而多尺度融合模块增强了上下文和语义特征的整合,减少了噪声影响并精炼了特征区分。
Machine learning methods in change detection have gained substantial traction due to their capability to handle complex data patterns and automate feature extraction. Traditional algorithms, such as SVM and decision trees, have shown robustness in binary and multi-class classifications. Random forests, in particular, offer high accuracy and can handle multi-dimensional data while reducing overfitting. In recent years, deep learning models, especially CNNs, have become a focus in the field due to their proficiency in learning high-level representations and detecting subtle changes. However, these models demand large amounts of labeled data and computational resources, which may pose challenges, particularly in SAR image processing where data availability and quality are limited. Consequently, the integration of traditional machine learning techniques with deep learning approaches, along with transfer learning and self-supervised learning strategies, holds promise for advancing SAR image change detection applications.
机器学习方法在变化检测中获得了显著关注,因为它们能够处理复杂的数据模式并自动提取特征。传统算法,如支持向量机(SVM)和决策树,在二分类和多分类中表现出强大的鲁棒性。随机森林尤其提供了高准确性,能够处理多维数据,同时减少过拟合。近年来,深度学习模型,特别是卷积神经网络(CNN),因其在学习高级表示和检测细微变化方面的能力而成为该领域的焦点。然而,这些模型需要大量标记数据和计算资源,这可能在合成孔径雷达(SAR)图像处理中特别具有挑战性,因为数据的可用性和质量有限。因此,传统机器学习技术与深度学习方法的结合,以及迁移学习和自监督学习策略的应用,展现了推动 SAR 图像变化检测应用的前景。

3.2.5. Change type identification
3.2.5. 变更类型识别

Binary change detection primarily determines whether a change has occurred in an image, categorizing areas as either “changed” or “unchanged.” Change type identification, however, goes beyond this binary approach to determine the specific change types and processes within altered regions. Building upon binary detection, change type recognition initially classifies multi-temporal SAR images to label ground object categories, which are then compared to analyze shifts in classification labels within areas of change. For instance, Niu and Ban (2013) employed SVM for urban land cover mapping, while classification comparison methods and deep learning networks have shown strong capabilities in identifying change types. Qi and Yeh (2013), for example, applied post-classification comparison (PCC) to identify types of land change.Additionally, change type recognition methods share similarities with image feature classification techniques, which are effective for detecting and identifying specific change patterns. Consequently, feature classification techniques based on SAR images can be integrated into change type recognition methods to enhance detection accuracy and detail identification.
二元变化检测主要确定图像中是否发生了变化,将区域分类为“已变化”或“未变化”。然而,变化类型识别超越了这种二元方法,以确定改变区域内的具体变化类型和过程。在二元检测的基础上,变化类型识别最初对多时相 SAR 图像进行分类,以标记地面物体类别,然后进行比较以分析变化区域内分类标签的变化。例如,Niu 和 Ban(2013)采用支持向量机(SVM)进行城市土地覆盖制图,而分类比较方法和深度学习网络在识别变化类型方面表现出强大的能力。例如,Qi 和 Yeh(2013)应用后分类比较(PCC)来识别土地变化类型。此外,变化类型识别方法与图像特征分类技术有相似之处,这些技术在检测和识别特定变化模式方面有效。因此,基于 SAR 图像的特征分类技术可以集成到变化类型识别方法中,以提高检测准确性和细节识别。

3.3. Difference map analysis
3.3. 差异图分析

In dual-temporal SAR image change detection, generating a difference map is a critical step, and difference map-based change detection remains the primary approach for dual-temporal analysis (Cohen (1998)). This method highlights areas of change by calculating the difference between two temporal images, providing a foundational basis for subsequent change identification and classification (Wang et al. (2023)). The difference map serves as a similarity measure between the images, employing features
在双时相 SAR 图像变化检测中,生成差异图是一个关键步骤,基于差异图的变化检测仍然是双时相分析的主要方法(Cohen (1998))。该方法通过计算两幅时相图像之间的差异来突出变化区域,为后续的变化识别和分类提供了基础(Wang et al. (2023))。差异图作为图像之间的相似性度量,采用特征。

such as test statistics derived from hypothesis testing or detection sub-targets from direct comparison methods. Change detection results are then produced by analyzing and classifying the difference map.Since SAR images are influenced by speckle noise, the difference map must exhibit high effectiveness and robustness to ensure reliable change detection. The accuracy of classification is significantly affected by the quality of the difference map and the selected analysis method. Common approaches to analyzing difference maps include thresholding and clustering techniques.
例如,从假设检验中得出的测试统计量或通过直接比较方法检测的子目标。变化检测结果是通过分析和分类差异图生成的。由于 SAR 图像受到斑点噪声的影响,差异图必须表现出高效性和鲁棒性,以确保可靠的变化检测。分类的准确性受到差异图质量和所选分析方法的显著影响。分析差异图的常见方法包括阈值处理和聚类技术。
The thresholding method segments a difference map into two pixel classes by selecting an optimal threshold. This method is advantageous due to its simplicity and fast computation speed; however, accurately determining the optimal threshold is challenging, often leading to limited accuracy and potential loss of change information. Common unsupervised optimal threshold selection techniques include the ExpectationMaximization (EM) algorithm (Su et al. (2014)) and Otsu’s Method (OTSU) (Balling, Herold, and Reiche (2023)), among others.
阈值法通过选择最佳阈值将差异图分割为两类像素。该方法因其简单性和快速计算速度而具有优势;然而,准确确定最佳阈值具有挑战性,常常导致精度有限和潜在的变化信息丢失。常见的无监督最佳阈值选择技术包括期望最大化(EM)算法(Su 等,2014)和大津法(OTSU)(Balling、Herold 和 Reiche,2023)等。
The clustering method applies a clustering algorithm to the difference map to obtain two cluster centers, corresponding to the unchanged and changed classes. Subsequently, the difference map is segmented into these two regions using a nearest-neighbor approach. Clustering methods can be categorized as either hard or fuzzy clustering. A standard hard clustering method is K-means clustering (KM) (Celik (2012)), which iteratively identifies suitable cluster centers by maximizing inter-cluster distances while minimizing intra-cluster distances. However, misclassification can occur. For fuzzy clustering, a typical method is fuzzy C-means clustering (FCM) (Ghosh, Mishra, and Ghosh (2011)), which incorporates fuzzy set theory into the K-means algorithm to create an affiliation matrix, enhancing classification accuracy. Yet, FCM remains vulnerable to noise and other interference factors, impacting its stability. To address this, advanced methods have been developed, including Fast Generalized Fuzzy C-means (FGFCM) (Cai, Chen, and Zhang (2007)), Fuzzy Local Information C-means (FLICM) (Krinidis and Chatzis (2010)), and Markov random field-enhanced FCM (Chen, Zhang, and Yin (2012)).
聚类方法将聚类算法应用于差异图,以获得两个聚类中心,分别对应于未变化和变化的类别。随后,使用最近邻方法将差异图分割为这两个区域。聚类方法可以分为硬聚类和模糊聚类。标准的硬聚类方法是 K 均值聚类(KM)(Celik(2012)),它通过最大化聚类间距离并最小化聚类内距离来迭代识别合适的聚类中心。然而,可能会发生误分类。对于模糊聚类,典型的方法是模糊 C 均值聚类(FCM)(Ghosh、Mishra 和 Ghosh(2011)),它将模糊集理论纳入 K 均值算法中,以创建隶属矩阵,从而提高分类准确性。然而,FCM 仍然容易受到噪声和其他干扰因素的影响,影响其稳定性。为了解决这个问题,已经开发了先进的方法,包括快速广义模糊 C 均值(FGFCM)(Cai、Chen 和 Zhang(2007))、模糊局部信息 C 均值(FLICM)(Krinidis 和 Chatzis(2010))以及马尔可夫随机场增强的 FCM(Chen、Zhang 和 Yin(2012))。

3.4. Temporal change detection
3.4. 时间变化检测

With advancements in time series analysis and detection technology, SAR image analysis for temporal change trends has emerged as a cutting-edge research field. SAR timeseries change detection primarily identifies alterations in ground objects by examining the change patterns and trends in multi-temporal SAR images. This approach often utilizes time-series data to track variations in surface scattering characteristics, thereby distinguishing long-term and short-term changes within a specific region. A fundamental technique in SAR time-series change detection is the difference map generation. Shen et al. (2023) employed the Log-Ratio operator for logarithmic transformation of time-series images, effectively suppressing coherent speckle noise and creating a difference map. They further introduced an improved Constant False Alarm Rate (CFAR) algorithm to enhance detection accuracy. Similarly, Quin et al. (2013) utilized statistical detection methods, such as the generalized likelihood ratio test (GLRT), to identify change regions by comparing scattering properties across multiple time phases, while enabling automatic threshold adjustment for improved accuracy and efficiency.To refine the identification of change areas, superpixel-based segmentation methods for time-series PolSAR images have also been explored. For instance, Ye et al. (2024) applied the SLIC model for adaptive superpixel segmentation, integrating polariza-
随着时间序列分析和检测技术的进步,SAR 图像分析在时间变化趋势方面已成为一个前沿研究领域。SAR 时间序列变化检测主要通过检查多时相 SAR 图像中的变化模式和趋势来识别地面物体的变化。这种方法通常利用时间序列数据来跟踪表面散射特征的变化,从而区分特定区域内的长期和短期变化。SAR 时间序列变化检测中的一个基本技术是差异图生成。Shen 等人(2023)采用对数比率算子对时间序列图像进行对数变换,有效抑制了相干斑点噪声并创建了差异图。他们进一步引入了一种改进的恒虚警率(CFAR)算法,以提高检测精度。同样,Quin 等人(2013)利用统计检测方法,如广义似然比检验(GLRT),通过比较多个时间阶段的散射特性来识别变化区域,同时实现自动阈值调整以提高准确性和效率。为了精确识别变化区域,基于超像素的时间序列 PolSAR 图像分割方法也得到了探索。例如,叶等人。(2024)应用了 SLIC 模型进行自适应超像素分割,整合了极化-

tion covariance matrices from multiple time phases to improve change detection accuracy and achieve temporal consistency by balancing polarization and spatial distance. In complex scenes, time-series change detection can leverage time-series modeling to capture long-term change patterns. Li et al. (2023) proposed a model that converts change detection into time-series classification, suitable for large-scale SAR datasets, effectively identifying various change patterns and reducing noise interference. Additionally, multi-feature fusion algorithms have been developed to enhance automation and detection precision. Colin Koeniguer and Nicolas (2020), for example, applied a GLRT-based approach that incorporates diverse statistical computations along with frequency and intensity information from the time series, enabling dynamic analysis of change regions and optimizing detection efficiency with automated thresholding and dynamic assessment.
从多个时间阶段提取协方差矩阵,以提高变化检测的准确性,并通过平衡极化和空间距离实现时间一致性。在复杂场景中,时间序列变化检测可以利用时间序列建模来捕捉长期变化模式。Li 等人(2023)提出了一种将变化检测转化为时间序列分类的模型,适用于大规模 SAR 数据集,有效识别各种变化模式并减少噪声干扰。此外,还开发了多特征融合算法,以增强自动化和检测精度。例如,Colin Koeniguer 和 Nicolas(2020)应用了一种基于 GLRT 的方法,该方法结合了多种统计计算以及来自时间序列的频率和强度信息,使变化区域的动态分析成为可能,并通过自动阈值和动态评估优化检测效率。
SAR temporal change detection holds critical applications in environmental change monitoring and disaster assessment. Existing methods demonstrate high accuracy and efficiency in handling large-scale data. However, SAR image data is often limited in availability, inherently complex, and prone to registration errors and noise interference. To address the challenges posed by data scarcity and intricate features, further advancements in data utilization and feature extraction techniques are essential.
SAR 时序变化检测在环境变化监测和灾害评估中具有重要应用。现有方法在处理大规模数据时表现出高准确性和效率。然而,SAR 图像数据的可用性通常有限,内在复杂,并且容易受到配准错误和噪声干扰。为了解决数据稀缺和复杂特征带来的挑战,进一步改进数据利用和特征提取技术是必不可少的。

4. Results of change detection and accuracy evaluation
4. 变化检测和准确性评估的结果

Accuracy assessment in change detection is generally divided into qualitative and quantitative evaluations. Current evaluation metrics for SAR image change detection focus primarily on pixel-level accuracy, with a noticeable gap in criteria specifically designed for object-level change analysis, rendering the overall set of evaluation metrics relatively limited.
变化检测中的准确性评估通常分为定性和定量评估。目前,SAR 图像变化检测的评估指标主要集中在像素级准确性上,而专门针对对象级变化分析的标准明显不足,使得整体评估指标相对有限。
Qualitative evaluation assesses the effectiveness of change detection methods by comparing the derived difference map with a ground truth map, though obtaining a reliable ground truth map is challenging. Typically, three approaches are employed: conducting field surveys, which yield high accuracy but are resource-intensive; extracting change information from high-resolution images taken on similar dates, a method that is both convenient and precise yet necessitates geometric correction; and visual interpretation, which is straightforward but demands a high level of expertise from interpreters. For SAR images, generating an accurate ground truth map remains particularly difficult, as the precision of this map heavily influences the final assessment results. Consequently, optical remote sensing images are frequently integrated with Google Earth data to produce a binary black-and-white reference map, enhancing both the reliability and accuracy of the evaluation.
定性评估通过将生成的差异图与真实情况图进行比较来评估变化检测方法的有效性,尽管获得可靠的真实情况图具有挑战性。通常采用三种方法:进行实地调查,虽然准确性高但资源密集;从在相似日期拍摄的高分辨率图像中提取变化信息,这种方法既方便又精确,但需要进行几何校正;以及视觉解读,这种方法简单但要求解读者具备高水平的专业知识。对于合成孔径雷达(SAR)图像,生成准确的真实情况图尤其困难,因为该图的精度对最终评估结果有很大影响。因此,光学遥感图像通常与谷歌地球数据结合,以生成二元黑白参考图,从而提高评估的可靠性和准确性。
Confusion matrices are frequently employed to assess the accuracy of binary classification problems, as illustrated in Table 6. In this context, TP (True Positive) represents the count of positive samples accurately classified as positive, while FN (False Negative) indicates the number of positive samples incorrectly classified as negative. Within change detection applications, a ‘positive sample’ denotes a ‘pixel that has changed’, whereas a ‘negative sample’ refers to a ‘pixel that has remained unchanged’.
混淆矩阵常用于评估二分类问题的准确性,如表 6 所示。在这个上下文中,TP(真正例)表示被准确分类为正例的正样本数量,而 FN(假负例)表示被错误分类为负例的正样本数量。在变化检测应用中,“正样本”指的是“已变化的像素”,而“负样本”指的是“未变化的像素”。
Table 6. Confusion Matrix
表 6. 混淆矩阵

Table 6. Confusion Matrix
表 6. 混淆矩阵
Prediction 预测 Total 总计
Positive 积极 Negative 负面
Actual 实际 Positive 积极 True Positive (TP) 真正阳性 (TP) False Negative (FN) 假阴性 (FN) T
Negative 负面 False Positive (FP) 假阳性 (FP) True Negative (TN) 真阴性 (TN) F
Total 总计 P N T + F T + F T+F\mathrm{T}+\mathrm{F} or P +N
T + F T + F T+F\mathrm{T}+\mathrm{F} 或 P + N
Prediction Total Positive Negative Actual Positive True Positive (TP) False Negative (FN) T Negative False Positive (FP) True Negative (TN) F Total P N T+F or P +N| | | Prediction | | Total | | :---: | :---: | :---: | :---: | :---: | | | | Positive | Negative | | | Actual | Positive | True Positive (TP) | False Negative (FN) | T | | | Negative | False Positive (FP) | True Negative (TN) | F | | | Total | P | N | $\mathrm{T}+\mathrm{F}$ or P +N |
Quantitative evaluation is typically conducted in conjunction with a confusion matrix. The accuracy of change detection is calculated by thoroughly analyzing the ground truth map and detection result map, utilizing metrics such as accuracy, detection rate, and other quantitative indicators. Common evaluation metrics include accuracy (ACC), precision, true positive rate (TPR), false alarm ratio (FA), missed alarm ratio (MA), Kappa coefficient (KC), and area under the curve (AUC).
定量评估通常与混淆矩阵一起进行。变化检测的准确性通过全面分析真实情况图和检测结果图来计算,利用准确率、检测率和其他定量指标等指标。常见的评估指标包括准确率(ACC)、精确度、真正率(TPR)、误报率(FA)、漏报率(MA)、卡帕系数(KC)和曲线下面积(AUC)。
The accuracy metric indicates the proportion of correctly classified samples among all samples. Generally, higher accuracy reflects better model performance; however, it may be an inadequate measure for assessing performance in cases of imbalanced classification.
准确率指标表示所有样本中正确分类样本的比例。通常,较高的准确率反映了更好的模型性能;然而,在不平衡分类的情况下,它可能是评估性能的不足衡量标准。
ACC = TP + TN P + N = TP + TN TP + FP + FN + TN ACC = TP + TN P + N = TP + TN TP + FP + FN + TN ACC=(TP+TN)/(P+N)=(TP+TN)/(TP+FP+FN+TN)\mathrm{ACC}=\frac{\mathrm{TP}+\mathrm{TN}}{P+N}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}+\mathrm{TN}}
Precision represents the proportion of true positives among the samples classified as positive. Generally, a higher precision indicates better model performance.
精确度表示在被分类为正样本的样本中真实正样本的比例。通常,较高的精确度表示模型性能更好。
Precision = TP P = TP TP + FP  Precision  = TP P = TP TP + FP " Precision "=(TP)/(P)=(TP)/(TP+FP)\text { Precision }=\frac{\mathrm{TP}}{P}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}
The True Positive Rate (TPR), also known as the detection rate, recall rate, or simply recall, represents the proportion of actual positive samples that are correctly classified as positive. Generally, a higher TPR indicates that the model successfully identifies more positive samples, reflecting stronger performance.
真正阳性率(TPR),也称为检测率、召回率或简单地称为召回,表示被正确分类为阳性的实际阳性样本的比例。通常,较高的 TPR 表明模型成功识别了更多的阳性样本,反映出更强的性能。
TPR = TP T = TP TP + FN TPR = TP T = TP TP + FN TPR=(TP)/(T)=(TP)/(TP+FN)\mathrm{TPR}=\frac{\mathrm{TP}}{T}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}
False Alarm Rate (FAR) represents the probability of negative samples being incorrectly classified as positive. A lower FAR indicates fewer misclassifications among negative samples, contributing to a model’s robustness.
误报率(FAR)表示负样本被错误分类为正样本的概率。较低的 FAR 表示负样本中的错误分类较少,有助于提高模型的鲁棒性。
FA = FP F = FP FP + TN FA = FP F = FP FP + TN FA=(FP)/(F)=(FP)/(FP+TN)\mathrm{FA}=\frac{\mathrm{FP}}{F}=\frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TN}}
False Negative Rate (FNR), also known as the miss rate, represents the proportion of positive samples that are incorrectly classified as negative. A lower FNR indicates fewer missed detections, improving the model’s sensitivity to actual changes.
假阴性率(FNR),也称为漏检率,表示被错误分类为负样本的正样本比例。较低的 FNR 表示漏检较少,提高了模型对实际变化的敏感性。
MA = FN T = FN TP + FN MA = FN T = FN TP + FN MA=(FN)/(T)=(FN)/(TP+FN)\mathrm{MA}=\frac{\mathrm{FN}}{T}=\frac{\mathrm{FN}}{\mathrm{TP}+\mathrm{FN}}
Overall Error (OE) represents the sum of false positives (FP) and false negatives (FN). A lower OE indicates a more accurate result from the difference map analysis, with fewer misclassifications.
总体误差(OE)表示假阳性(FP)和假阴性(FN)的总和。较低的 OE 表示差异图分析的结果更准确,错误分类更少。
OE = FP + FN OE = FP + FN OE=FP+FN\mathrm{OE}=\mathrm{FP}+\mathrm{FN}
The overall error rate, also referred to as Positive Errors (PE), serves as a metric for assessing overall performance.
整体错误率,也称为正错误(PE),作为评估整体表现的指标。
PE = OE N = FP + FN N PE = OE N = FP + FN N PE=(OE)/(N)=(FP+FN)/(N)\mathrm{PE}=\frac{\mathrm{OE}}{N}=\frac{\mathrm{FP}+\mathrm{FN}}{N}
The F1 score is a statistical metric used to assess the accuracy of binary classification models. This measure considers both the precision and recall of a model, serving as a harmonic mean of these two metrics. The F1 score ranges of [ 0 , 1 ] [ 0 , 1 ] [0,1][0,1], with 1 indicating the highest accuracy and 0 the lowest.
F1 分数是用于评估二元分类模型准确性的统计指标。该指标同时考虑了模型的精确度和召回率,作为这两个指标的调和平均值。F1 分数的范围是 [ 0 , 1 ] [ 0 , 1 ] [0,1][0,1] ,其中 1 表示最高准确性,0 表示最低准确性。
F 1 = 2 precision recall precision + recall F 1 = 2  precision   recall   precision  +  recall  F_(1)=2*(" precision "*" recall ")/(" precision "+" recall ")F_{1}=2 \cdot \frac{\text { precision } \cdot \text { recall }}{\text { precision }+ \text { recall }}
More generally, the F β F β F_(beta)F_{\beta} score is defined as a metric that allows weighting recall more heavily than precision (or vice versa) by adjusting the parameter β β beta\beta. The F β F β F_(beta)F_{\beta} score generalizes the F 1 F 1 F_(1)F_{1} score, where β β beta\beta determines the relative importance of recall to precision. Specifically, a larger β β beta\beta gives more importance to recall, while a smaller β β beta\beta emphasizes precision.
更一般地说, F β F β F_(beta)F_{\beta} 分数被定义为一种度量,允许通过调整参数 β β beta\beta 来更重视召回率而不是精确率(或反之亦然)。 F β F β F_(beta)F_{\beta} 分数是对 F 1 F 1 F_(1)F_{1} 分数的推广,其中 β β beta\beta 决定了召回率与精确率的相对重要性。具体来说,较大的 β β beta\beta 更重视召回率,而较小的 β β beta\beta 则强调精确率。
F β = ( 1 + β 2 ) precision recall ( β 2 precision ) + recall F β = 1 + β 2  precision   recall  β 2  precision  +  recall  F_(beta)=(1+beta^(2))*(" precision "*" recall ")/((beta^(2)*" precision ")+" recall ")F_{\beta}=\left(1+\beta^{2}\right) \cdot \frac{\text { precision } \cdot \text { recall }}{\left(\beta^{2} \cdot \text { precision }\right)+\text { recall }}
The Kappa coefficient (Shen et al. (2023)) is a comprehensive metric that takes into account both true positive and false positive rates, providing an overall assessment of change detection accuracy. It is calculated within the range of [ 1 , 1 ] [ 1 , 1 ] [-1,1][-1,1] but typically falls between [ 0 , 1 ] [ 0 , 1 ] [0,1][0,1] in practical applications. A higher Kappa coefficient signifies greater accuracy in change detection performance.
Kappa 系数(Shen 等,2023)是一种综合指标,考虑了真正率和假正率,提供了变化检测准确性的整体评估。它的计算范围在 [ 1 , 1 ] [ 1 , 1 ] [-1,1][-1,1] 之间,但在实际应用中通常落在 [ 0 , 1 ] [ 0 , 1 ] [0,1][0,1] 之间。更高的 Kappa 系数意味着变化检测性能的准确性更高。
Kappa = PAR PRE 1 PRE PRA = TP + TN P + N  Kappa  =  PAR   PRE  1 PRE  PRA  = TP + TN P + N {:[" Kappa "=(" PAR "-" PRE ")/(1-PRE)],[" PRA "=(TP+TN)/(P+N)]:}\begin{aligned} \text { Kappa } & =\frac{\text { PAR }- \text { PRE }}{1-\mathrm{PRE}} \\ \text { PRA } & =\frac{\mathrm{TP}+\mathrm{TN}}{P+N} \end{aligned}
PRE = P × T + N × F ( P + N ) 2 = ( TP + FP ) × ( TP + FN ) + ( FN + TN ) × ( FP + TN ) ( P + N ) 2 PRE = P × T + N × F ( P + N ) 2 = ( TP + FP ) × ( TP + FN ) + ( FN + TN ) × ( FP + TN ) ( P + N ) 2 PRE=(P xx T+N xx F)/((P+N)^(2))=((TP+FP)xx(TP+FN)+(FN+TN)xx(FP+TN))/((P+N)^(2))\mathrm{PRE}=\frac{P \times T+N \times F}{(P+N)^{2}}=\frac{(\mathrm{TP}+\mathrm{FP}) \times(\mathrm{TP}+\mathrm{FN})+(\mathrm{FN}+\mathrm{TN}) \times(\mathrm{FP}+\mathrm{TN})}{(P+N)^{2}}
The Receiver Operating Characteristic (ROC) curve (Krogager (1990)) is a comprehensive tool for evaluating classifier performance. This curve plots the false alarm rate on the X -axis and the detection rate on the Y -axis, spanning from the origin to ( 1 , 1 ) ( 1 , 1 ) (1,1)(1,1). The Area Under the Curve (AUC), which represents the area between the curve and the baseline ( y = 0 y = 0 y=0\mathrm{y}=0 to x = 1 x = 1 x=1\mathrm{x}=1 ), ranges from 0 to 1 . Generally, a higher AUC indicates superior classifier performance.
接收者操作特征(ROC)曲线(Krogager(1990))是评估分类器性能的综合工具。该曲线在 X 轴上绘制假警报率,在 Y 轴上绘制检测率,从原点延伸到 ( 1 , 1 ) ( 1 , 1 ) (1,1)(1,1) 。曲线下的面积(AUC)表示曲线与基线( y = 0 y = 0 y=0\mathrm{y}=0 x = 1 x = 1 x=1\mathrm{x}=1 )之间的面积,范围从 0 到 1。通常,较高的 AUC 表示更优越的分类器性能。
Sequential change detection poses greater challenges in evaluation compared to twophase change detection. Schroeder et al. (2011) suggested using ground truth data or historical archive data as reference data for sequential change detection outcomes. However, there is often a shortage of historical data and change truth maps that align with the study period’s detection results. Consequently, high-spatial-resolution optical remote sensing images, such as those from Google Earth, are frequently employed as supplementary tools. These images facilitate obtaining reference data through visual interpretation of original time series data, using methods like the TimeSync algorithm and sampling techniques.
序列变化检测在评估方面比两阶段变化检测面临更大的挑战。Schroeder 等(2011)建议使用地面真实数据或历史档案数据作为序列变化检测结果的参考数据。然而,通常缺乏与研究期间检测结果相一致的历史数据和变化真实图。因此,高空间分辨率的光学遥感图像,如来自谷歌地球的图像,常被用作补充工具。这些图像通过对原始时间序列数据的视觉解读,利用时间同步算法和抽样技术等方法,便于获取参考数据。
Accuracy evaluation metrics for temporal change detection results are relatively limited. Since the perturbation map in temporal change detection encompasses both spatial and temporal information, evaluation metrics are categorized into spatial domain and temporal domain accuracy evaluations. Generally, spatial domain accuracy is considered more critical than temporal domain accuracy, and in most instances, spatial domain evaluation suffices to assess temporal change detection results. Key spatial domain metrics include user accuracy (UA) and producer accuracy (PA). User accuracy, analogous to recall, reflects the likelihood that a detected disturbed pixel is correctly classified, while producer accuracy, also referred to as mapping accuracy and similar to precision, indicates the probability that the classification aligns with reference data or actual conditions. Temporal domain metrics encompass temporal accuracy, which leverages time-sensitive data to evaluate incremental changes preceding the disturbance.
时间变化检测结果的准确性评估指标相对有限。由于时间变化检测中的扰动图既包含空间信息又包含时间信息,评估指标被分为空间域和时间域的准确性评估。一般来说,空间域的准确性被认为比时间域的准确性更为重要,在大多数情况下,空间域评估足以评估时间变化检测结果。关键的空间域指标包括用户准确率(UA)和生产者准确率(PA)。用户准确率类似于召回率,反映了检测到的扰动像素被正确分类的可能性,而生产者准确率,也称为映射准确率,类似于精确度,表示分类与参考数据或实际情况一致的概率。时间域指标包括时间准确率,它利用时间敏感数据来评估扰动前的增量变化。

5. Conclusions 5. 结论

SAR image change detection holds significant importance within remote sensing image processing. This paper provides a comprehensive overview of frequently used SAR datasets and the standard procedures for SAR change detection, including in-depth analysis of essential components: analysis units, detection methodologies, and target change types. Commonly employed evaluation metrics are also discussed. By separating analysis units from change detection methods, this paper aims to clarify methodological approaches, minimize redundancies, and clearly outline the change detection process, thereby providing a structured pathway for advancing research in SAR change detection.
SAR 图像变化检测在遥感图像处理中的重要性不言而喻。本文提供了常用 SAR 数据集和 SAR 变化检测标准程序的全面概述,包括对基本组成部分的深入分析:分析单元、检测方法和目标变化类型。还讨论了常用的评估指标。通过将分析单元与变化检测方法分开,本文旨在澄清方法论,减少冗余,并清晰地概述变化检测过程,从而为推动 SAR 变化检测研究提供结构化的路径。
The summary of analysis units suggests that selecting an analysis unit should align closely with the experiment’s objectives, with a clear understanding of the strengths and limitations of each type. Additionally, the analysis unit best suited to the image characteristics should be prioritized. However, achieving an optimal analysis unit remains a technical challenge. Pixels, as fundamental units, are applicable across diverse scenarios, while superpixels and image objects, although theoretically ideal, may yield
分析单元的总结建议,选择分析单元应与实验目标紧密对齐,并清楚了解每种类型的优缺点。此外,应优先考虑最适合图像特征的分析单元。然而,实现最佳分析单元仍然是一个技术挑战。像素作为基本单元,适用于各种场景,而超像素和图像对象虽然在理论上理想,但可能会产生。

results dependent on segmentation quality, potentially impacting subsequent analyses. Therefore, further research is necessary to refine analysis unit separation techniques, and ongoing exploration is required to better understand the applicability, advantages, and limitations of different analysis units.
结果依赖于分割质量,可能影响后续分析。因此,进一步研究是必要的,以完善分析单元分离技术,并且需要持续探索,以更好地理解不同分析单元的适用性、优点和局限性。
Following the determination of criteria for selecting analysis units, this article delves into SAR-based change detection methods. These methods are structured around distinct analysis units, with a detailed summary of their applications, strengths, and limitations in the context of SAR image characteristics. This classification provides insight into selecting optimal detection methods for varying analysis units, aiming to enhance both the accuracy and reliability of change detection.
在确定分析单元选择标准后,本文深入探讨了基于合成孔径雷达(SAR)的变化检测方法。这些方法围绕不同的分析单元构建,并详细总结了它们在 SAR 图像特征背景下的应用、优点和局限性。这一分类为选择适合不同分析单元的最佳检测方法提供了见解,旨在提高变化检测的准确性和可靠性。
This paper categorizes change detection methods into five distinct types. Machine learning, currently a central focus in research, is designated as an independent category to allow for a more comprehensive analysis of its methodologies. Simultaneously, considering the unique characteristics of SAR and PolSAR images, polarimetric feature decomposition is highlighted as a separate category to clarify the application of statistical modeling and polarimetric decomposition techniques in SAR change detection. Among change detection techniques, post-classification change detection remains widely favored, especially with the integration of CNNs, which have demonstrated improved accuracy and performance. However, the majority of machine learning approaches are supervised, and due to the limited availability of SAR datasets, models are often impacted by a scarcity of labeled samples. Furthermore, the inherent unpredictability of change location and direction contributes to deep learning models’ challenges with generalizability. Addressing these limitations presents a critical avenue for future research.
本文将变化检测方法分为五种不同类型。机器学习目前是研究的中心焦点,被指定为一个独立类别,以便对其方法进行更全面的分析。同时,考虑到 SAR 和 PolSAR 图像的独特特性,极化特征分解被强调为一个单独的类别,以澄清统计建模和极化分解技术在 SAR 变化检测中的应用。在变化检测技术中,后分类变化检测仍然广受欢迎,特别是与 CNN 的结合,已显示出提高的准确性和性能。然而,大多数机器学习方法是监督的,由于 SAR 数据集的有限可用性,模型通常受到标记样本稀缺的影响。此外,变化位置和方向的固有不可预测性使得深度学习模型在泛化能力上面临挑战。解决这些局限性为未来的研究提供了一个重要的方向。
Although numerous methods have been developed for SAR image change detection, no single approach is universally applicable across all scenarios. The challenges posed by limited SAR datasets and the generally small data volume restrict the effective use of many supervised methods. Thus, dataset expansion and small-sample learning remain areas of high research value. Furthermore, the unique target scattering characteristics of SAR images limit the efficacy of directly applying change detection techniques from other types of remote sensing imagery, as method migration requires substantial theoretical backing. SAR images contain a wealth of extractable features for change detection; however, reliably capturing stable features amidst speckle noise remains a primary focus for future work. Additionally, as the analysis unit progresses from the pixel to the object level, change detection methods have also evolved. Nonetheless, accuracy evaluation largely relies on pixel-level metrics, underscoring the need for an improved, comprehensive evaluation framework.
尽管已经开发了多种 SAR 图像变化检测方法,但没有一种方法可以普遍适用于所有场景。有限的 SAR 数据集和通常较小的数据量对许多监督方法的有效使用构成了限制。因此,数据集扩展和小样本学习仍然是具有高研究价值的领域。此外,SAR 图像独特的目标散射特性限制了直接应用其他类型遥感影像变化检测技术的有效性,因为方法迁移需要 substantial 理论支持。SAR 图像包含丰富的可提取特征用于变化检测;然而,在斑点噪声中可靠捕捉稳定特征仍然是未来工作的主要重点。此外,随着分析单位从像素级别向对象级别的进展,变化检测方法也在不断演变。然而,准确性评估在很大程度上依赖于像素级别的指标,这突显了改进和全面评估框架的必要性。

6. Acknowledgments 6. 致谢

This study was funded by NSFC Grant no.62171023,NSFC under Grant no. 62222102 ,and the Fundamental Research Funds for the Central Universities under Grant no.FRF-TP- 22-005C1.
本研究得到了国家自然科学基金(NSFC)资助,资助号为 62171023,NSFC 资助号为 62222102,以及中央高校基本科研业务费资助,资助号为 FRF-TP-22-005C1。

References 参考文献

Akbari, V., S. Anfinsen, A. Doulgeris, T. Eltoft, G. Moser, and S. Serpico. 2017. “Polarimetric SAR change detection with the complex Hotelling-Lawley trace statistic.” IEEE Transac-
Akbari, V., S. Anfinsen, A. Doulgeris, T. Eltoft, G. Moser, 和 S. Serpico. 2017. “使用复杂的霍特林-劳利迹统计量进行极化合成孔径雷达变化检测。” IEEE 交易-

tions on Geoscience and Remote Sensing To be included in a future issue.
地球科学与遥感的讨论将被纳入未来一期。

Akbari, V., A. P. Doulgeris, and T. Eltoft. 2016. “Post-classification Change Detection in Arctic Glaciers by Multi-polarization SAR.” Multitemporal Remote Sensing: Methods and Applications 125-144.
Akbari, V., A. P. Doulgeris, 和 T. Eltoft. 2016. “通过多极化 SAR 对北极冰川进行后分类变化检测。” 多时相遥感:方法与应用 125-144.

Ansari, R. A., K. M. Buddhiraju, and R. Malhotra. 2020. “Urban change detection analysis utilizing multiresolution texture features from polarimetric SAR images.” Remote Sensing Applications: Society and Environment 20: 100418.
Ansari, R. A., K. M. Buddhiraju, 和 R. Malhotra. 2020. “利用极化合成孔径雷达图像的多分辨率纹理特征进行城市变化检测分析。” 遥感应用:社会与环境 20: 100418.

Baatz, M., and A. Schpe. 2000. “An optimization approach for high quality multi-scale image segmentation.” In Beiträge zum AGIT-Symposium, .
Baatz, M. 和 A. Schpe. 2000. “一种高质量多尺度图像分割的优化方法。” 在 AGIT 研讨会论文集。

Balling, J., M. Herold, and J. Reiche. 2023. “How textural features can improve SAR-based tropical forest disturbance mapping.” International Journal of Applied Earth Observation and Geoinformation 124: 103492.
Balling, J., M. Herold, 和 J. Reiche. 2023. “纹理特征如何改善基于 SAR 的热带森林干扰制图。” 应用地球观测与地理信息国际期刊 124: 103492.

Ban, Y., and A. Jacob. 2013. “Object-based fusion of multitemporal multiangle Envisat ASAR and HJ-1B multispectral data for urban land-cover mapping.” IEEE Transactions on Geoscience and Remote Sensing 51 (4): 1998-2006.
班, Y., 和 A. 雅各布. 2013. “基于对象的多时相多角度 Envisat ASAR 和 HJ-1B 多光谱数据融合用于城市土地覆盖制图。”《IEEE 地球科学与遥感学报》51 (4): 1998-2006.

Bayarjargal, Y., A. Karnieli, M. Bayasgalan, et al. 2006. “A comparative study of NOAA-AVHRR derived drought indices using change vector analysis.” Remote Sensing of Environment 105 (1): 9-22.
Bayarjargal, Y., A. Karnieli, M. Bayasgalan, 等. 2006. “使用变化矢量分析的 NOAA-AVHRR 衍生干旱指数的比较研究。”环境遥感 105 (1): 9-22.

Blaschke, T. 2005. “Towards a framework for change detection based on image objects.” Göttinger Geographische Abhandlungen 113: 1-9.
布拉施克,T. 2005。“基于图像对象的变化检测框架。”哥廷根地理论文 113: 1-9。

Bouhlel, N., V. Akbari, S. Méric, et al. 2022. “Multivariate statistical modeling for multitemporal SAR change detection using wavelet transforms and integrating subband dependencies.” IEEE Transactions on Geoscience and Remote Sensing 60: 1-18.
Bouhlel, N., V. Akbari, S. Méric, 等. 2022. “使用小波变换和集成子带依赖的多时相 SAR 变化检测的多变量统计建模。”《IEEE 地球科学与遥感学报》60: 1-18.

Bovolo, F., S. Marchesi, and L. Bruzzone. 2011. “A framework for automatic and unsupervised detection of multiple changes in multitemporal images.” IEEE Transactions on Geoscience and Remote Sensing 50 (6): 2196-2212.
Bovolo, F., S. Marchesi, 和 L. Bruzzone. 2011. “多时相图像中多重变化的自动和无监督检测框架。”《IEEE 地球科学与遥感学报》50 (6): 2196-2212.

Brisco, B., F. T. Ulaby, and M. C. Dobson. 1983. “Spaceborne SAR data for land-cover classification and change detection.” In 1983 International Geoscience and Remote Sensing Symposium (IGARSS’83), .
布里斯科,B.,F. T. 乌拉比,和 M. C. 多布森。1983。“用于土地覆盖分类和变化检测的空间 borne SAR 数据。”在 1983 年国际地球科学与遥感研讨会(IGARSS’83)上。

Burges, C. J. C. 1998. “A tutorial on support vector machines for pattern recognition.” Data Mining and Knowledge Discovery 2 (2): 121-167.
Burges, C. J. C. 1998. “支持向量机模式识别教程。” 数据挖掘与知识发现 2 (2): 121-167.

Cai, Weiling, Songcan Chen, and Daoqiang Zhang. 2007. “Fast and robust fuzzy C-means clustering algorithms incorporating local information for image segmentation.” Pattern Recognition 40 (3): 825-838.
蔡伟玲,陈松灿,张道强。2007。“结合局部信息的快速稳健模糊 C 均值聚类算法用于图像分割。”模式识别 40 (3): 825-838。

Cameron, W. L., and L. K. Leung. 1990. “Feature motivated polarization scattering matrix decomposition.” In IEEE International Radar Conference, 549-557. IEEE.
卡梅伦,W. L.,和 L. K. 梁。1990。“特征驱动的极化散射矩阵分解。”在 IEEE 国际雷达会议,549-557。IEEE。

Celik, T. 2012. “Unsupervised change detection of satellite images using local gradual descent.” IEEE Transactions on Geoscience and Remote Sensing 50 (5): 1919-1929.
Celik, T. 2012. “使用局部渐进下降的卫星图像无监督变化检测。”《IEEE 地球科学与遥感学报》50 (5): 1919-1929.

Chen, Gang, Geoffrey, J., Hay, Luis, M., T., Carvalho, and Michael and. 2012. “Object-based change detection.” International Journal of Remote Sensing 33 (14): 4434-4457.
陈刚,杰弗里·J·海,路易斯·M·T·卡瓦略,迈克尔. 2012. “基于对象的变化检测。”《国际遥感杂志》33 (14): 4434-4457.

Chen, Junxin. 2023. “Polarization SAR image change detection based on HLT difference map and triple Markov field.” PhD diss., Xidian University. Accessed on 2023-12-28.
陈俊欣. 2023. “基于 HLT 差异图和三重马尔可夫场的极化 SAR 图像变化检测.” 博士论文, 西安电子科技大学. 访问日期: 2023-12-28.

Chen, Y., R. Zhang, and D. Yin. 2012. “Multi-polarimetric SAR image compression based on sparse representation.” Science China Information Sciences 55: 1888-1897.
陈勇,张锐,尹东。2012。“基于稀疏表示的多极化 SAR 图像压缩。”中国科学信息科学 55: 1888-1897。

Cheng, Y. Q., H. C. Li, T. Celik, et al. 2013. “FRFT-based improved algorithm of unsupervised change detection in SAR images via PCA and K-means clustering.” In 2013 IEEE International Geoscience and Remote Sensing Symposium - IGARSS, 1952-1955. IEEE.
程, Y. Q., H. C. 李, T. Celik, 等. 2013. “基于 FRFT 的 SAR 图像无监督变化检测的改进算法,通过 PCA 和 K 均值聚类。” 在 2013 年 IEEE 国际地球科学与遥感会议 - IGARSS, 1952-1955. IEEE.

Chun-Yao, W., C. Jun-Zhou, and L. I. Wei. 2014. “Review on superpixel segmentation algorithms.” Application Research of Computers 31 (01): 6-12.
春耀,吴,君周,李伟。2014。“超像素分割算法综述。”计算机应用研究 31 (01): 6-12。

Cloude, S. R., and E. Pottier. 1996. “A review of target decomposition theorems in radar polarimetry.” IEEE Transactions on Geoscience and Remote Sensing 34 (2): 498-518.
Cloude, S. R. 和 E. Pottier. 1996. “雷达极化测量中的目标分解定理综述。”《IEEE 地球科学与遥感学报》34 (2): 498-518.

Cloude, S. R., and E. Pottier. 1997. “An entropy based classification scheme for land applications of polarimetric SAR.” IEEE Transactions on Geoscience and Remote Sensing 35 (1): 68-78.
Cloude, S. R. 和 E. Pottier. 1997. “一种基于熵的极化合成孔径雷达土地应用分类方案。”《IEEE 地球科学与遥感学报》35 (1): 68-78.
Cohen, J. 1960. “A coefficient of agreement for nominal scales.” Educational and Psychological Measurement 20 (1): 37-46.
科恩,J. 1960. “名义尺度的一致性系数。” 教育与心理测量 20 (1): 37-46.

Cohen, W. B. 1998. “Comparison of methods for detecting conifer forest change with Thematic Mapper imagery.”
科恩,W. B. 1998. “使用主题制图仪影像检测针叶林变化的方法比较。”

Colin Koeniguer, E., and J. M. Nicolas. 2020. “Change detection based on the coefficient of variation in SAR time-series of urban areas.” Remote Sensing 12 (13): 2089.
科林·科宁基尔,E.,和 J. M. 尼古拉斯。2020。“基于城市地区 SAR 时间序列变异系数的变化检测。”遥感 12 (13): 2089。

Conradsen, K., A. A. Nielsen, J. Schou, et al. 2003. “A test statistic in the complex Wishart distribution and its application to change detection in polarimetric SAR data.” IEEE Transactions on Geoscience and Remote Sensing 41 (1): 4-19.
Conradsen, K., A. A. Nielsen, J. Schou, 等. 2003. “复 Wishart 分布中的检验统计量及其在极化 SAR 数据变化检测中的应用。”《IEEE 地球科学与遥感学报》41 (1): 4-19.

Coppin, P. R., and M. E. Bauer. 1996. “Digital change detection in forest ecosystems with remote sensing imagery.” Remote Sensing Reviews 13 (3-4): 207-234.
Coppin, P. R. 和 M. E. Bauer. 1996. “利用遥感影像进行森林生态系统的数字变化检测。” 遥感评论 13 (3-4): 207-234.

Cortes, C., and V. Vapnik. 1995. “Support-vector networks.” Machine Learning 20: 273-297.
Cortes, C. 和 V. Vapnik. 1995. “支持向量网络。”机器学习 20: 273-297.

Dong, H., W. Ma, L. Jiao, et al. 2021. “A multiscale self-attention deep clustering for change detection in SAR images.” IEEE Transactions on Geoscience and Remote Sensing 60: 1-16.
董华,马伟,焦立,等。2021。“一种多尺度自注意力深度聚类用于 SAR 图像变化检测。”《IEEE 地球科学与遥感学报》60: 1-16。

Dongmei, Tang, Fan Hui, and Zhang Yao. 2017. “Review on Landsat Time Series Change Detection Methods.” Journal of Geo-Information Science .
董梅, 唐, 范辉, 张耀. 2017. “Landsat 时间序列变化检测方法综述.” 地理信息科学杂志.

Fisher, Peter F. 1997. “The pixel: A snare and a delusion.” International Journal of Remote Sensing 18: 679-685. https://api.semanticscholar.org/CorpusID:59468758.
费舍尔,彼得·F. 1997。“像素:一个陷阱和一个幻觉。”《国际遥感杂志》18:679-685。https://api.semanticscholar.org/CorpusID:59468758

Freeman, A., and S. L. Durden. 1998. “A three-component scattering model for polarimetric SAR data.” IEEE Transactions on Geoscience and Remote Sensing 36 (3): 963-973.
弗里曼,A.,和 S. L. 达登。1998。“一种用于极化合成孔径雷达数据的三分量散射模型。”《IEEE 地球科学与遥感学报》36 (3): 963-973。

Frost, Victor S., Josephine Abbott Stiles, K. Sam Shanmugan, and Julian C. Holtzman. 1982. “A Model for Radar Images and Its Application to Adaptive Digital Filtering of Multiplicative Noise.” IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-4: 157-166. https://api.semanticscholar.org/CorpusID:7447045.
Frost, Victor S., Josephine Abbott Stiles, K. Sam Shanmugan, 和 Julian C. Holtzman. 1982. “雷达图像模型及其在乘法噪声自适应数字滤波中的应用。” IEEE 模式分析与机器智能学报 PAMI-4: 157-166. https://api.semanticscholar.org/CorpusID:7447045

Fukuda, S., R. Katagiri, and H. Hirosawa. 2002. “Unsupervised approach for polarimetric SAR image classification using support vector machines.” In IEEE International Geoscience and Remote Sensing Symposium, Vol. 5, 2599-2601 vol.5.
福田, S., 片桐 R., 和 广泽 H. 2002. “使用支持向量机的无监督极化合成孔径雷达图像分类方法。” 载于《IEEE 国际地球科学与遥感会议》,第 5 卷, 2599-2601。

Fulong, Chen, Zhang Hong, and Wang Chao. 2007. “SAR change detection technology development review.” Journal of remote sensing technology and applications (001): 022.
傅龙,陈,张洪,和王超。2007。“SAR 变化检测技术发展回顾。”遥感技术与应用杂志(001):022。

Gao, Feng, Junyu Dong, Bo Li, Qizhi Xu, and Cui Xie. 2016. “Change detection from synthetic aperture radar images based on neighborhood-based ratio and extreme learning machine.” Journal of Applied Remote Sensing 10 (4): 046019.
高峰,董俊宇,李博,徐启志,谢翠。2016。“基于邻域比率和极限学习机的合成孔径雷达图像变化检测。”《应用遥感杂志》10 (4): 046019。

Ghosh, A., N. S. Mishra, and S. Ghosh. 2011. “Fuzzy clustering algorithms for unsupervised change detection in remote sensing images.” Information Science 181 (4): 699-715.
Ghosh, A., N. S. Mishra, 和 S. Ghosh. 2011. “用于遥感图像无监督变化检测的模糊聚类算法。” 信息科学 181 (4): 699-715.

Gomez, L., L. Alvarez, L. Mazorra, et al. 2015. “Classification of complex Wishart matrices with a diffusion-reaction system guided by stochastic distances.” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 373 (2056): 20150118.
戈麦斯, L., L. 阿尔瓦雷斯, L. 马佐拉, 等. 2015. “通过随机距离引导的扩散反应系统对复杂威沙特矩阵的分类。” 皇家学会 A 辑哲学交易: 数学、物理与工程科学 373 (2056): 20150118.
Gong, M., J. Zhao, J. Liu, et al. 2016. “Change detection in synthetic aperture radar images based on deep neural networks.” IEEE Transactions on Neural Networks and Learning Systems 27 (1): 125-138.
龚明,赵俊,刘杰,等。2016。“基于深度神经网络的合成孔径雷达图像变化检测。”《IEEE 神经网络与学习系统汇刊》27(1):125-138。

Gupta, S., D. Singh, and S. Kumar. 2023. “Development of an Ontology-Based Technique for Labeling Land Cover Classes with Minimum Utilization of SAR Features.” SN Computer Science 4 (6): 731.
Gupta, S., D. Singh, 和 S. Kumar. 2023. “基于本体的土地覆盖类别标记技术的开发,最小化 SAR 特征的利用。” SN 计算机科学 4 (6): 731.

Han, M., and Y. Zhou. 2015. “Joint-classification change detection based on improved fuzzy ARTMAP.” In 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 346-349. IEEE.
韩明和周勇. 2015. “基于改进模糊 ARTMAP 的联合分类变化检测.” 载于 2015 年 IEEE 国际地球科学与遥感会议(IGARSS),346-349. IEEE.

Han, Ping, Runmin Cong, and Weiran Zhang. 2013. “Based on weighted polarization of the difference degree of SAR image change detection algorithm.” Journal of Modern Radar 35 (12): 4. https://doi.org/10.3969/j.issn.1004-7859.2013.12.008.
韩平、丛润敏和张伟然。2013。“基于加权极化的 SAR 图像变化检测算法差异度。”现代雷达杂志 35 (12): 4. https://doi.org/10.3969/j.issn.1004-7859.2013.12.008
Hou, B., Q. Wei, Y. Zheng, et al. 2014. “Unsupervised change detection in SAR image based on gauss-log ratio image fusion and compressed projection.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7 (8): 3297-3317.
侯博, 魏青, 郑阳, 等. 2014. “基于高斯-对数比图像融合和压缩投影的 SAR 图像无监督变化检测.” IEEE 应用地球观测与遥感选刊 7 (8): 3297-3317.

Huifu, Z., D. Kazhong, and F. Hongdong. 2016. “SAR images unsupervised change detection
惠夫,Z.,D. 卡中,和 F. 洪东。2016。“SAR 图像无监督变化检测

based on combination of texture feature vector with maximum entropy principle.” Acta Geodaetica Et Cartographica Sinica 45 (3): 339.
基于纹理特征向量与最大熵原理的结合。”《测绘学报》45 (3): 339。

Hussain, Masroor, Dongmei Chen, Angela Cheng, Huimin Wei, and David Stanley. 2013. “Change detection from remotely sensed images: From pixel-based to objectbased approaches.” Isprs Journal of Photogrammetry and Remote Sensing 80: 91-106. https://api.semanticscholar.org/CorpusID:14283205.
侯赛因,马斯鲁尔,陈冬梅,郑安琪,魏惠敏,和大卫·斯坦利。2013 年。“从遥感图像中进行变化检测:从基于像素的方法到基于对象的方法。”《国际摄影测量与遥感杂志》80: 91-106。https://api.semanticscholar.org/CorpusID:14283205

Imani, M. 2023. “Modified PCA, LDA and LPP feature extraction methods for PolSAR image classification.” Multimedia Tools and Applications 1-22.
Imani, M. 2023. “用于 PolSAR 图像分类的改进 PCA、LDA 和 LPP 特征提取方法。” 多媒体工具与应用 1-22。

Inglada, J., and G. Mercier. 2007. “A new statistical similarity measure for change detection in multitemporal SAR images and its extension to multiscale change analysis.” IEEE Transactions on Geoscience and Remote Sensing 45 (5): 1432-1445.
Inglada, J. 和 G. Mercier. 2007. “一种用于多时相 SAR 图像变化检测的新统计相似性度量及其在多尺度变化分析中的扩展。”《IEEE 地球科学与遥感学报》45 (5): 1432-1445.

Jabari, Shabnam, and Yun Zhang. 2016. “RPC-based coregistration of VHR imagery for urban change detection.” Photogrammetric Engineering and Remote Sensing 82 (7): 521-534.
贾巴里、沙布南和张云。2016。“基于 RPC 的高分辨率影像核心注册用于城市变化检测。”摄影测量与遥感工程 82 (7): 521-534。

Joachims, T. 1999. “11 Making large-scale support vector machine learning practical.” In Advances in kernel methods: support vector learning, 169. MIT Press.
乔希姆斯, T. 1999. “11 使大规模支持向量机学习变得实用。” 收录于《核方法的进展:支持向量学习》,169. 麻省理工学院出版社。

Kanade, Takeo, and M. Okutomi. 1991. “A stereo matching algorithm with an adaptive window: theory and experiment.” Proceedings. 1991 IEEE International Conference on Robotics and Automation 1088-1095 vol.2. https://api.semanticscholar.org/CorpusID:22901323.
香田、武夫和大冢 M. 1991。“一种具有自适应窗口的立体匹配算法:理论与实验。”会议论文。1991 年 IEEE 国际机器人与自动化会议 1088-1095 卷 2。https://api.semanticscholar.org/CorpusID:22901323

Kim, M., S. E. Park, and S. J. Lee. 2023. “Detection of damaged buildings using temporal SAR data with different observation modes.” Remote Sensing 15 (2): 308.
金明洙, 朴世恩, 和李世俊. 2023. “利用不同观测模式的时序 SAR 数据检测受损建筑物。” 遥感 15 (2): 308.

Krinidis, S., and V. Chatzis. 2010. “A robust fuzzy local information c-means clustering algorithm.” IEEE Transactions on Image Processing 19 (5): 1328-1337.
Krinidis, S. 和 V. Chatzis. 2010. “一种稳健的模糊局部信息 c 均值聚类算法。”《IEEE 图像处理汇刊》19 (5): 1328-1337.

Krogager, E. 1990. “New decomposition of the radar target scattering matrix.” Electronics Letters 18 (26): 1525-1527.
克罗加杰,E. 1990. “雷达目标散射矩阵的新分解。” 电子快报 18 (26): 1525-1527.

Larose, D. T., and C. D. Larose. 2014. Discovering knowledge in data: an introduction to data mining. John Wiley and Sons.
拉罗斯,D. T.,和 C. D. 拉罗斯。2014。《数据中的知识发现:数据挖掘导论》。约翰·威利父子公司。

Lee, J. S., and M. R. Grunes. 1992. “Classification of multi-look polarimetric SAR data based on complex Wishart distribution.” In NTC-92: National Telesystems Conference, IEEE.
李, J. S., 和 M. R. Grunes. 1992. “基于复杂 Wishart 分布的多视极化 SAR 数据分类。” 在 NTC-92:国家远程系统会议,IEEE。

Lee, J. S., M. R. Grunes, and R. Kwok. 1994. “Classification of multi-look polarimetric SAR imagery based on complex Wishart distribution.” International Journal of Remote Sensing 15 (11): 2299-2311.
李, J. S., M. R. Grunes, 和 R. Kwok. 1994. “基于复杂 Wishart 分布的多视极化 SAR 图像分类。” 国际遥感杂志 15 (11): 2299-2311.

Lee, Jong-Sen. 1980. “Digital Image Enhancement and Noise Filtering by Use of Local Statistics.” IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-2: 165-168. https://api.semanticscholar.org/CorpusID:9060120.
李钟善. 1980. “通过使用局部统计进行数字图像增强和噪声过滤。”《IEEE 模式分析与机器智能学报》PAMI-2: 165-168. https://api.semanticscholar.org/CorpusID:9060120.

Lee, Jong-Sen, and Eric Pottier. 2009. “Polarimetric Radar Imaging: From Basics to Applications.” https://api.semanticscholar.org/CorpusID:128279998.
李钟善和埃里克·波蒂耶。2009 年。“极化雷达成像:从基础到应用。” https://api.semanticscholar.org/CorpusID:128279998

Lee, Jong-Sen, Jen-Hung Wen, Thomas L Ainsworth, Kun-Shan Chen, and Abel J Chen. 2008. “Improved sigma filter for speckle filtering of SAR imagery.” IEEE Transactions on Geoscience and Remote Sensing 47 (1): 202-213.
李钟森,温仁宏,托马斯·L·安斯沃斯,陈坤山,和陈阿贝尔·J。2008。“改进的西格玛滤波器用于合成孔径雷达图像的斑点滤波。”《IEEE 地球科学与遥感学报》47 (1): 202-213。

Lei, Ming, Weixin Tian, Dong Ren, and Ting Dong. 2019. “Forest change detection method for remote sensing images based on object classification.” Journal of Forest and Environment 39 (6): 641-646.
雷明、田维新、任东和董婷。2019。“基于对象分类的遥感影像森林变化检测方法。”森林与环境杂志 39 (6): 641-646。

Li, Jinji, Licheng Jiao, Xiangrong Zhang, et al. 2009. “SAR image change detection based on two-phase image joint classification.” Journal of Infrared and Millimeter Wave 28 (6): 466-471.
李金基,焦立成,张相荣等。2009。“基于两相图像联合分类的 SAR 图像变化检测。”红外与毫米波学报 28 (6): 466-471。

Li, K., X. Cao, and D. Meng. 2024. “A New Learning Paradigm for Foundation Model-Based Remote-Sensing Change Detection.” IEEE Transactions on Geoscience and Remote Sensing 62: 1-12.
李凯, 曹晓, 孟达. 2024. “基于基础模型的遥感变化检测的新学习范式。”《IEEE 地球科学与遥感学报》62: 1-12.

Li, Q., L. Gong, and J. Zhang. 2019. “A correlation change detection method integrating PCA and multi-texture features of SAR image for building damage detection.” European Journal of Remote Sensing 52 (1): 435-447.
李强,龚丽,张杰。2019。“一种结合主成分分析和 SAR 图像多纹理特征的相关性变化检测方法用于建筑损伤检测。”《欧洲遥感杂志》52 (1): 435-447。

Li, W., P. Ma, H. Wang, et al. 2023. “SAR-TSCC: A novel approach for long time series SAR image change detection and pattern analysis.” IEEE Transactions on Geoscience and Remote Sensing 61: 1-16.
李伟,马鹏,王辉等。2023。“SAR-TSCC:一种用于长时间序列 SAR 图像变化检测和模式分析的新方法。”《IEEE 地球科学与遥感学报》61:1-16。
Li, Y., C. Peng, Y. Chen, et al. 2019. “A deep learning method for change detection in synthetic aperture radar images.” IEEE Transactions on Geoscience and Remote Sensing 57 (8): 5751-5763.
李勇, 彭超, 陈阳, 等. 2019. “一种用于合成孔径雷达图像变化检测的深度学习方法.” IEEE 地球科学与遥感学报 57 (8): 5751-5763.

Lian, Gu, Xu ShiQi, and Zhu LeQing. 2020. “Detection of Building Changes in Remote Sensing Images via FlowS-Unet.” Acta Automatica Sinica 46 (6): 1291-1300.
连, 顾, 许世奇, 和朱乐青. 2020. “通过 FlowS-Unet 检测遥感图像中的建筑变化.” 自动化学报 46 (6): 1291-1300.

Liang, X. S., and D. G. M. Anderson. 2007. “Multiscale Window Transform.” SIAM Journal on Multiscale Modeling and Simulation 6 (2): 437-467.
梁, X. S., 和 D. G. M. 安德森. 2007. “多尺度窗口变换.” SIAM 多尺度建模与仿真期刊 6 (2): 437-467.

Liu, Shan, Fengli Zhang, Shiying Wei, Na Liu, and Yun Shao. 2020. “SAR image visual optimization and building damage assessment based on polarization decomposition combination.” Journal of Graduate School of Chinese Academy of Sciences 37 (6): 750-759.
刘珊,张风丽,魏世英,刘娜,邵云。2020。“基于极化分解组合的 SAR 图像视觉优化与建筑损伤评估。”中国科学院研究生院学报 37 (6): 750-759。

Liu, Sicong, Kecheng Du, Yongjie Zheng, Jin Chen, Peijun Du, and Xiaohua Tong. 2023. “Remote sensing change detection technology in the Era of artificial intelligence:Inheritance, development and challenges.” National Remote Sensing Bulletin https://api.semanticscholar.org/CorpusID:268470121.
刘思聪,杜克城,郑永杰,陈进,杜佩君,和童小华。2023。“人工智能时代的遥感变化检测技术:传承、发展与挑战。”《国家遥感快报》https://api.semanticscholar.org/CorpusID:268470121

Liu, Tao, Ziyuan Yang, Yanni Jiang, et al. 2021. “Review of ship detection in polarimetric synthetic aperture imagery.” Journal of Radars 10 (1): 1-19. https://doi.org/10.12000/JR20155.
刘涛,杨子元,姜妍妮等。2021。“极化合成孔径成像中的船舶检测综述。”《雷达学报》10 (1):1-19。https://doi.org/10.12000/JR20155

Lopes, Armand, Ridha Touzi, and Edmond Nezry. 1990. “Adaptive speckle filters and scene heterogeneity.” IEEE Transactions on Geoscience and Remote Sensing 28: 992-1000. https://api.semanticscholar.org/CorpusID:123450931.
洛佩斯,阿尔芒,里达·图齐和埃德蒙·内兹里。1990 年。“自适应散斑滤波器和场景异质性。”《IEEE 地球科学与遥感学报》28: 992-1000。https://api.semanticscholar.org/CorpusID:123450931

Lunetta, R. S., J. Ediriwickrema, D. M. Johnson, et al. 2002. “Impacts of vegetation dynamics on the identification of land-cover change in a biologically complex community in North Carolina, USA.” Remote Sensing of Environment 82 (2-3): 258-270.
Lunetta, R. S., J. Ediriwickrema, D. M. Johnson, 等. 2002. “植被动态对美国北卡罗来纳州生物复杂社区土地覆盖变化识别的影响。” 环境遥感 82 (2-3): 258-270.

Ma, J., M. Gong, and Z. Zhou. 2012. “Wavelet fusion on ratio images for change detection in SAR images.” IEEE Geoscience and Remote Sensing Letters 9 (6): 1122-1126.
马, J., 龚, M., 和周, Z. 2012. “基于比率图像的小波融合在 SAR 图像变化检测中的应用。” IEEE 地球科学与遥感快报 9 (6): 1122-1126.

Ma, Xiaoxiao, Bo Cheng, Yueming Liu, Shiai Cui, and Chenbin Liang. 2019. “PolSAR remote sensing image method for building extraction based on polarization and texture characteristics.” Journal of Remote Sensing 36 (5): 682-693.
马晓晓,博成,刘月明,崔世爱,梁晨彬。2019。“基于极化和纹理特征的 PolSAR 遥感图像建筑提取方法。”《遥感学报》36 (5): 682-693。

Mastro, P., G. Masiello, C. Serio, et al. 2022. “Change detection techniques with synthetic aperture radar images: Experiments with random forests and Sentinel-1 observations.” Remote Sensing 14 (14): 3323.
Mastro, P., G. Masiello, C. Serio, 等. 2022. “合成孔径雷达图像的变化检测技术:随机森林和哨兵-1 观测的实验。” 遥感 14 (14): 3323.

Niu, X., and Y. Ban. 2013. “Multi-temporal RADARSAT-2 polarimetric SAR data for urban land-cover classification using an object-based support vector machine and a rule-based approach.” International Journal of Remote Sensing 34 (1): 1-26.
牛晓, 班勇. 2013. “基于对象的支持向量机和规则基础方法的城市土地覆盖分类的多时相 RADARSAT-2 极化 SAR 数据.” 国际遥感杂志 34 (1): 1-26.

Novak, L. M., M. C. Burl, and W. W. Irving. 1993. “Optimal polarimetric processing for enhanced target detection.” IEEE Transactions on Aerospace and Electronic Systems 29 (1): 234-244.
诺瓦克,L. M.,M. C. 伯尔,和 W. W. 欧文。1993。“增强目标检测的最佳偏振处理。”《IEEE 航空航天与电子系统学报》29 (1):234-244。
Omati, M., and M. R. Sahebi. 2018. “Change Detection of Polarimetric SAR Images Based on the Integration of Improved Watershed and MRF Segmentation Approaches.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 11 (11): 4170-4179. https://doi.org/10.1109/JSTARS.2018.2874517.
Omati, M.,和 M. R. Sahebi. 2018. “基于改进的分水岭和马尔可夫随机场分割方法的极化合成孔径雷达图像变化检测。”《IEEE 应用地球观测与遥感选定主题期刊》11 (11): 4170-4179. https://doi.org/10.1109/JSTARS.2018.2874517

Qi, Z., and A. G. O. Yeh. 2013. “Integrating change vector analysis, post-classification comparison, and object-oriented image analysis for land use and land cover change detection using RADARSAT-2 polarimetric SAR images.” Advances in Spatial Data Handling: Geospatial Dynamics, Geosimulation and Exploratory Visualization 107-123.
齐, Z., 和 A. G. O. 叶. 2013. “结合变化矢量分析、后分类比较和面向对象的图像分析,利用 RADARSAT-2 极化 SAR 图像进行土地利用和土地覆盖变化检测。” 空间数据处理进展:地理空间动态、地理模拟与探索性可视化 107-123.

Qi, Z., A. G. O. Yeh, X. Li, et al. 2015. “A three-component method for timely detection of land cover changes using polarimetric SAR images.” ISPRS Journal of Photogrammetry and Remote Sensing 107: 3-21.
齐, Z., 叶 A. G. O., 李 X., 等. 2015. “一种基于极化合成孔径雷达图像的及时检测土地覆盖变化的三组分方法。” ISPRS 摄影测量与遥感杂志 107: 3-21.

Quin, G., B. Pinel-Puyssegur, J. M. Nicolas, et al. 2013. “MIMOSA: An automatic change detection method for SAR time series.” IEEE Transactions on Geoscience and Remote Sensing 52 (9): 5349-5363.
Quin, G., B. Pinel-Puyssegur, J. M. Nicolas, 等. 2013. “MIMOSA:一种用于 SAR 时间序列的自动变化检测方法。”《IEEE 地球科学与遥感学报》52 (9): 5349-5363.

Ratha, D., S. De, T. Celik, et al. 2017. “Change detection in polarimetric SAR images using a geodesic distance between scattering mechanisms.” IEEE Geoscience and Remote Sensing Letters 14 (7): 1066-1070.
Ratha, D., S. De, T. Celik, 等. 2017. “利用散射机制之间的测地距离进行极化合成孔径雷达图像的变化检测。” IEEE 地球科学与遥感快报 14 (7): 1066-1070.
Ren, Xiaofeng, and Jitendra Malik. 2003. “Learning a classification model for segmentation.” Proceedings Ninth IEEE International Conference on Computer Vision 10-17 vol.1. https://api.semanticscholar.org/CorpusID:13571735.
任晓峰和吉滕德拉·马利克。2003。“学习用于分割的分类模型。”第九届 IEEE 国际计算机视觉会议论文集,第 1 卷,第 10-17 页。https://api.semanticscholar.org/CorpusID:13571735

Ren Qiuru, Wang Chuanjian, Yang Wenzhong. 2021. “Remote sensing image change detection review.” Journal of computer applications 41 (8): 12.
任秋如,王传健,杨文忠。2021。“遥感图像变化检测综述。”计算机应用杂志 41 (8): 12。

Rignot, E., and Jakob J. van Zyl. 1993. “Change detection techniques for ERS-1 SAR data.” IEEE Trans. Geosci. Remote. Sens. 31: 896-906. https://api.semanticscholar.org/CorpusID:2683179.
Rignot, E., 和 Jakob J. van Zyl. 1993. “ERS-1 SAR 数据的变化检测技术。” IEEE Trans. Geosci. Remote. Sens. 31: 896-906. https://api.semanticscholar.org/CorpusID:2683179

Rogan, J., J. Franklin, and D. A. Roberts. 2002. “A comparison of methods for monitoring multitemporal vegetation change using Thematic Mapper imagery.” Remote Sensing of Environment 80 (1): 143-156.
罗根,J.,J. 富兰克林,和 D. A. 罗伯茨。2002。“使用主题图像监测多时相植被变化的方法比较。”环境遥感 80 (1): 143-156。

Runbo, Xie. 2023. “Review of traditional methods of SAR image change detection.” Computer and Information Technology 31 (2): 13-14.
Runbo, Xie. 2023. “SAR 图像变化检测传统方法的回顾。”计算机与信息技术 31 (2): 13-14.

Samadi, F., G. Akbarizadeh, and H. Kaabi. 2019. “Change detection in SAR images using deep belief network: a new training approach based on morphological images.” IET Image Processing 13 (12): 2255-2264.
Samadi, F., G. Akbarizadeh, 和 H. Kaabi. 2019. “使用深度信念网络进行 SAR 图像变化检测:一种基于形态图像的新训练方法。” IET 图像处理 13 (12): 2255-2264.

Schroeder, T. A., M. A. Wulder, S. P. Healey, et al. 2011. “Mapping wildfire and clearcut harvest disturbances in boreal forests with Landsat time series data.” Remote Sensing of Environment 115 (6): 1421-1433.
施罗德, T. A., M. A. Wulder, S. P. Healey, 等. 2011. “利用 Landsat 时间序列数据绘制北方森林中的野火和清除采伐干扰。” 环境遥感 115 (6): 1421-1433.

Schwartz, C., L. P. Ramos, L. T. Duarte, et al. 2020. “Change detection in UWB SAR images based on robust principal component analysis.” Remote Sensing 12 (12): 1916.
施瓦茨, C., L. P. 拉莫斯, L. T. 杜阿尔特, 等. 2020. “基于稳健主成分分析的超宽带合成孔径雷达图像变化检测。” 遥感 12 (12): 1916.

Shao, Ningyuan, Huanxin Zou, Cheng Chen, Meilin Li, and Xianxiang Qin. 2019. “Change detection oriented superpixel cosegmentation algorithm for SAR images.” Systems Engineering and Electronics 41 (7): 1496-1503.
邵宁远、邹焕新、陈成、李美琳、秦先祥。2019。“面向变化检测的 SAR 图像超像素共同分割算法。”系统工程与电子学 41 (7): 1496-1503。

Shen, G., H. Guo, and J. Liao. 2007. “Change vector analysis method for inundation change detection using multi-temporal multi-polarized SAR images.” In MIPPR 2007: Remote Sensing and GIS Data Processing and Applications; and Innovative Multispectral Technology and Applications, Vol. 6790, 70-77. SPIE.
沈国,郭华,廖杰。2007。“基于多时相多极化 SAR 图像的淹没变化检测的变化矢量分析方法。”收录于 MIPPR 2007:遥感与 GIS 数据处理与应用;以及创新多光谱技术与应用,第 6790 卷,70-77 页。SPIE。

Shen, Wenjie, Yunzhen Jia, Yang Li, Yun Lin, Wen Jiang, and Yanping Wang. 2023. “Timeseries image change detection method of target-level spaceborne SAR based on Log-Ratio operator.” Industrial Control Computers (11).
沈文杰,贾云珍,李扬,林云,姜文,王艳平。2023。“基于对数比算子的目标级空间 SAR 时序图像变化检测方法。”工业控制计算机(11)。

Simard, M., S. S. Saatchi, and G. De Grandi. 2000. “The use of decision tree and multiscale texture for classification of JERS-1 SAR data over tropical forest.” IEEE Transactions on Geoscience and Remote Sensing 38 (5): 2310-2321.
Simard, M., S. S. Saatchi, 和 G. De Grandi. 2000. “使用决策树和多尺度纹理对热带森林的 JERS-1 SAR 数据进行分类。”《IEEE 地球科学与遥感学报》38 (5): 2310-2321.

Simioni, J. P. D., L. A. Guasselli, V. F. Nascimento, et al. 2020. “Integration of multi-sensor analysis and decision tree for evaluation of dual and quad-Pol SAR in L-and C-bands applied for marsh delineation.” Environment, Development and Sustainability 22: 5603-5620.
Simioni, J. P. D., L. A. Guasselli, V. F. Nascimento, 等. 2020. “多传感器分析与决策树的结合用于评估 L 波段和 C 波段的双极和四极 SAR 在沼泽划界中的应用。” 环境、发展与可持续性 22: 5603-5620.

Singh, A. 1989. “Review Article Digital change detection techniques using remotely-sensed data: International Journal of Remote Sensing: Vol 10, No 6.” International Journal of Remote Sensing https://doi.org/10.1080/01431168908903939.
Singh, A. 1989. “综述文章:使用遥感数据的数字变化检测技术:国际遥感杂志:第 10 卷,第 6 期。” 国际遥感杂志 https://doi.org/10.1080/01431168908903939

Sivapriya, M. S., and Fathimal P. Mohamed. 2022. “Ice Berg Detection in SAR Images Using Mask R-CNN.” In Emergent Converging Technologies and Biomedical Systems: Select Proceedings of ETBS 2021, 625-633. Springer Singapore.
Sivapriya, M. S. 和 Fathimal P. Mohamed. 2022. “使用 Mask R-CNN 在 SAR 图像中检测冰山。” 收录于《新兴融合技术与生物医学系统:ETBS 2021 会议论文集》,625-633 页。新加坡:施普林格。

Song, W., Y. Du, Y. Wu, et al. 2022. “RF-HoDRF: High-order hybrid discriminative random field improved by two-layer random forest for SAR image change detection.” IEEE Geoscience and Remote Sensing Letters 19: 1-5.
宋伟, 杜勇, 吴艳, 等. 2022. “RF-HoDRF:通过双层随机森林改进的高阶混合判别随机场用于 SAR 图像变化检测。” IEEE 地球科学与遥感快报 19: 1-5.

Su, Linzhi, Maoguo Gong, Bo Sun, et al. 2014. “Unsupervised change detection in SAR images based on locally fitting model and semi-EM algorithm.” International Journal of Remote Sensing 35 (2): 621-650.
苏,林志,毛国功,孙博等。2014。“基于局部拟合模型和半 EM 算法的 SAR 图像无监督变化检测。”《国际遥感杂志》35(2):621-650。

Sun, G., Y. Zhang, P. Song, et al. 2021. “Extraction of earthquake damaged buildings based on the change of coherence characteristics of SAR images.” In Proceedings of the 2021 5th International Conference on Electronic Information Technology and Computer Engineering, 396-401.
孙国辉, 张勇, 宋鹏, 等. 2021. “基于 SAR 图像相干特征变化的地震损坏建筑物提取.” 载于 2021 年第五届电子信息技术与计算机工程国际会议论文集, 396-401.

Surhone, L. M., M. T. Tennoe, S. F. Henssonow, et al. 2010. “Random Forest.” Machine
Surhone, L. M., M. T. Tennoe, S. F. Henssonow, 等. 2010. “随机森林。” 机器
Learning 45 (1): 5-32. 学习 45 (1): 5-32。
Tewkesbury, A. P., A. J. Comber, N. J. Tate, et al. 2015. “A critical synthesis of remotely sensed optical image change detection techniques.” Remote Sensing of Environment 160: 1-14.
Tewkesbury, A. P., A. J. Comber, N. J. Tate, 等. 2015. “遥感光学图像变化检测技术的批判性综合。”环境遥感 160: 1-14.

Vapnik, V. N., and V. Vapnik. 1998. Statistical learning theory.
瓦普尼克, V. N., 和 V. 瓦普尼克. 1998. 统计学习理论.

Vinholi, J. G., D. Silva, R. Machado, et al. 2020. “CNN-based change detection algorithm for wavelength-resolution SAR images.” IEEE Geoscience and Remote Sensing Letters 19: 1-5.
Vinholi, J. G., D. Silva, R. Machado, 等. 2020. “基于 CNN 的波长分辨率 SAR 图像变化检测算法。” IEEE 地球科学与遥感快报 19: 1-5.

Wang, J., F. Gao, J. Dong, et al. 2022. “Change detection from synthetic aperture radar images via graph-based knowledge supplement network.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 15: 1823-1836.
王, J., 高, F., 董, J., 等. 2022. “通过基于图的知识补充网络从合成孔径雷达图像中进行变化检测。” IEEE 应用地球观测与遥感精选主题期刊 15: 1823-1836.

Wang, S., S. Yang, and L. Jiao. 2016. “Saliency-guided change detection for SAR imagery using a semi-supervised Laplacian SVM.” Remote Sensing Letters 7 (11): 1043-1052.
王, S., 杨, S. 和 焦, L. 2016. “基于显著性引导的 SAR 图像变化检测方法:一种半监督拉普拉斯支持向量机。” 遥感快报 7 (11): 1043-1052.

Wang, Shaona, Yang Liu, Linlin Li, et al. 2023. “Are the SAR image change detection.” Laser Journal 44 (1): 8-15. https://doi.org/10.14016/j.carol carroll nki JGZZ.2023.01.008.
王少娜,杨柳,林琳,等。2023。“SAR 图像变化检测。”激光杂志 44 (1): 8-15。https://doi.org/10.14016/j.carol carroll nki JGZZ.2023.01.008。

Wang, W., X. Yang, X. Li, et al. 2016. “A fully polarimetric SAR imagery classification scheme for mud and sand flats in intertidal zones.” IEEE Transactions on Geoscience and Remote Sensing 55 (3): 1734-1742.
王伟, 杨晓, 李晓, 等. 2016. “一种用于潮间带泥沙平原的全极化 SAR 影像分类方案。”《IEEE 地球科学与遥感学报》55 (3): 1734-1742.

Wang, Xuesong. 2016. “Status and Prospects of Radar Polarimetry Techniques.” Journal of Radars 5 (2): 119-131. https://doi.org/10.12000/JR16039.
王雪松。2016。“雷达极化技术的现状与展望。”《雷达学报》5 (2): 119-131。https://doi.org/10.12000/JR16039

Warner, Timothy A. 2011. “Kernel-Based Texture in Remote Sensing Image Classification.” Geography Compass 5: 781-798. https://api.semanticscholar.org/CorpusID:129082022.
华纳,蒂莫西·A. 2011. “基于核的纹理在遥感图像分类中的应用。” 地理指南 5: 781-798. https://api.semanticscholar.org/CorpusID:129082022.

Weng, L., W. Yang, B. Hu, et al. 2024. “MDINet: Multi-Domain Incremental Network for Change Detection.” IEEE Transactions on Geoscience and Remote Sensing .
翁立,杨伟,胡斌等。2024。“MDINet:用于变化检测的多领域增量网络。”《IEEE 地球科学与遥感学报》。

Wieland, M., W. Liu, and F. Yamazaki. 2016. “Learning change from synthetic aperture radar images: Performance evaluation of a support vector machine to detect earthquake and tsunami-induced changes.” Remote Sensing 8 (10): 792.
Wieland, M., W. Liu, 和 F. Yamazaki. 2016. “从合成孔径雷达图像中学习变化:支持向量机检测地震和海啸引起变化的性能评估。” 遥感 8 (10): 792.

Wu, C., B. Du, X. Cui, et al. 2017. “A post-classification change detection method based on iterative slow feature analysis and Bayesian soft fusion.” Remote Sensing of Environment 199: 241-255.
吴, C., 杜, B., 崔, X., 等. 2017. “一种基于迭代慢特征分析和贝叶斯软融合的后分类变化检测方法。” 环境遥感 199: 241-255.

Xiaoxiao, M. A., C. Bo, L. I. U. Yueming, et al. 2019. “PolSAR remote sensing image method for building extraction based on polarization and texture characteristics.” Journal of University of Chinese Academy of Sciences 36 (5): 682.
小小, M. A., C. Bo, L. I. U. Yueming, 等. 2019. “基于极化和纹理特征的 PolSAR 遥感图像建筑提取方法。” 中国科学院大学学报 36 (5): 682.

Xue-song, Wang, and Chen Siwei. 2020. “Polarimetric Synthetic Aperture Radar Interpretation and Recognition: Advances and Perspectives.” https://api.semanticscholar.org/CorpusID:231801220.
薛松、王和陈思维。2020。“极化合成孔径雷达的解释与识别:进展与展望。” https://api.semanticscholar.org/CorpusID:231801220

Yamaguchi, Y., T. Moriyama, M. Ishido, et al. 2005. “Four-component scattering model for polarimetric SAR image decomposition.” Technical Report of Ieice Sane 104 (8): 1699-1706.
山口,Y.,森山,T.,石堂,M. 等。2005。“极化合成孔径雷达图像分解的四分量散射模型。”电子信息通信学会技术报告 104 (8):1699-1706。

Ye, J. W., C. C. Wang, H. Gao, et al. 2024. “Adaptive superpixel generation for time-series PolSAR images considering time-varying characteristics.” National Remote Sensing Bulletin 28 (4): 1066-1075. https://doi.org/10.11834/jrs.20221498.
叶, J. W., 王 C. C., 高 H., 等. 2024. “考虑时变特征的时间序列 PolSAR 图像自适应超像素生成。” 国家遥感快报 28 (4): 1066-1075. https://doi.org/10.11834/jrs.20221498.

Yin, J., and J. Yang. 2016. “A new change detector in polSAR imagery.” In Geoscience and Remote Sensing Symposium (IGARSS), 2016 IEEE International, 5662-5665. IEEE.
尹, J., 和 J. 杨. 2016. “一种新的极化合成孔径雷达图像变化检测器。” 在 2016 年 IEEE 国际地球科学与遥感研讨会(IGARSS),5662-5665. IEEE.

Yu, Xiaoping, and Jujie Wei. 2023. “A Multi-temporal Polarimetric SAR change detection Method combining polarimetric cross-correlation and spatial context information.” In The 6th Annual High Resolution Earth Observation Conference, Accessed on 2023-12-28.
余小平,魏菊杰。2023。“一种结合极化交叉相关和空间上下文信息的多时相极化 SAR 变化检测方法。”在第六届高分辨率地球观测会议上,访问日期:2023 年 12 月 28 日。

Zhang, H., M. Lin, G. Yang, et al. 2021a. “ESCNet: An end-to-end superpixel-enhanced change detection network for very-high-resolution remote sensing images.” IEEE Transactions on Neural Networks and Learning Systems 34 (1): 28-42.
张华, 林敏, 杨刚, 等. 2021a. “ESCNet:一种用于超高分辨率遥感图像的端到端超像素增强变化检测网络。”《IEEE 神经网络与学习系统汇刊》34 (1): 28-42.

Zhang, L., B. Zou, J. Zhang, et al. 2009. “Classification of polarimetric SAR image based on support vector machine using multiple-component scattering model and texture features.” EURASIP Journal on Advances in Signal Processing 2010: 1-9.
张, L., Zou, B., Zhang, J., 等. 2009. “基于支持向量机的极化合成孔径雷达图像分类,采用多组分散射模型和纹理特征。” EURASIP 信号处理进展杂志 2010: 1-9.

Zhang, P., M. Gong, H. Zhang, et al. 2018. “Unsupervised difference representation learning for detecting multiple types of changes in multitemporal remote sensing images.” IEEE Transactions on Geoscience and Remote Sensing 57 (4): 2277-2289.
张鹏,龚敏,张华等。2018。“无监督差异表示学习用于检测多时相遥感图像中的多种变化。”《IEEE 地球科学与遥感学报》57(4):2277-2289。
Zhang, X., S. Cheng, L. Wang, et al. 2023. “Asymmetric cross-attention hierarchical network based on CNN and transformer for bitemporal remote sensing images change detection.” IEEE Transactions on Geoscience and Remote Sensing 61: 1-15.
张晓, 程思, 王丽, 等. 2023. “基于 CNN 和变换器的非对称交叉注意力层次网络用于双时相遥感图像变化检测.” IEEE 地球科学与遥感学报 61: 1-15.

Zhang, X., H. Su, C. Zhang, et al. 2021b. “Robust unsupervised small area change detection from SAR imagery using deep learning.” ISPRS Journal of Photogrammetry and Remote Sensing 173: 79-94.
张, X., 苏, H., 张, C., 等. 2021b. “基于深度学习的 SAR 影像鲁棒无监督小区域变化检测.” ISPRS 摄影测量与遥感杂志 173: 79-94.

Zhang, X., X. Su, Q. Yuan, et al. 2021c. “Spatial-temporal gray-level co-occurrence aware CNN for SAR image change detection.” IEEE Geoscience and Remote Sensing Letters 19: 1-5.
张, X., 苏, X., 袁, Q., 等. 2021c. “基于时空灰度共生的卷积神经网络用于 SAR 图像变化检测。” IEEE 地球科学与遥感快报 19: 1-5.

Zhang, X., J. Yin, and J. Yang. 2021. “Comparison of change detectors by using satellite polarimetric SAR data.” In 2021 International Conference on Space, Aeronautical and Navigational Electronics, Vol. 121, 88-93. IEICE.
张晓,尹杰,杨杰。2021。“利用卫星极化合成孔径雷达数据的变化检测器比较。”载于 2021 年国际空间、航空和导航电子会议,卷 121,88-93。IEICE。

Zhang, X., J. Yin, J. Yang, et al. 2022a. “Region-based change detection for polarimetric SAR images.” In 2022 IEEE International Geoscience and Remote Sensing Symposium, 675-678. IEEE.
张, X., 尹, J., 杨, J., 等. 2022a. “基于区域的极化合成孔径雷达图像变化检测。” 载于 2022 年 IEEE 国际地球科学与遥感会议, 675-678. IEEE.

Zhang, Zuxun, Huiwei Jiang, Shiyan Pang, et al. 2022b. “Review and prospect in change detection of multi-temporal remote sensing images.” Acta Geodaetica et Cartographica Sinica 51 (7): 1091-1107. https://doi.org/10.11947/j.AGCS.2022.20220070.
张祖勋,姜辉伟,庞世彦等。2022b。“多时相遥感影像变化检测的回顾与展望。”《测绘学报》51 (7): 1091-1107。https://doi.org/10.11947/j.AGCS.2022.20220070

Zhao, J., Y. Chang, J. Yang, et al. 2020. “A novel change detection method based on statistical distribution characteristics using multi-temporal PolSAR data.” Sensors 20 (5): 1508.
赵, J., Y. 常, J. 杨, 等. 2020. “一种基于统计分布特征的多时相 PolSAR 数据的新型变化检测方法。”传感器 20 (5): 1508.

Zhao, J., J. Yang, Z. Lu, et al. 2017. “A novel method of change detection in bi-temporal PolSAR data using a joint-classification classifier based on a similarity measure.” Remote Sensing 9 (8): 846.
赵, J., 杨, J., 陆, Z., 等. 2017. “一种基于相似性度量的联合分类分类器在双时相 PolSAR 数据中变化检测的新方法。” 遥感 9 (8): 846.

Zhao, Min, and Yindi Zhao. 2018. “Object-oriented and multi-feature hierarchical change detection based on CVA for high-resolution remote sensing imagery.” Journal of Remote Sensing 22 (1): 119-131.
赵敏,赵银迪。2018。“基于 CVA 的面向对象和多特征分层变化检测用于高分辨率遥感影像。”遥感学报 22 (1): 119-131。

Zhao, R., G. H. Peng, W. Yan, et al. 2021. “Change detection in SAR images based on superpixel segmentation and image regression.” Earth Science Informatics 14: 69-79.
赵, R., 彭国华, 闫伟, 等. 2021. “基于超像素分割和图像回归的 SAR 图像变化检测。” 地球科学信息 14: 69-79.

Zhuang, Huifu, Kazhong Ddeng, and Hongdong Fan. 2016. “SAR Images Unsupervised Change Detection Based on Combination of Texture Feature Vector with Maximum Entropy Principle.” Acta Geodaeticaet Cartographica Sinica 45 (3): 339-346. https://doi.org/10.11947/j.AGCS.2016.20150022.
庄辉福、卡中邓和洪东范。2016。“基于纹理特征向量与最大熵原理组合的 SAR 图像无监督变化检测。”《测绘学报》45 (3): 339-346。https://doi.org/10.11947/j.AGCS.2016.20150022

Zuxun, ZHANG, JIANG Huiwei, PANG Shiyan, and HU Xiangyun. 2022. “Review and prospect in change detection of multi-temporal remote sensing images.” Acta Geodaetica et Cartographica Sinica 51 (7): 1091.
张祖勋,姜辉伟,庞世燕,胡向云。2022。“多时相遥感影像变化检测的回顾与展望。”《测绘学报》51(7): 1091。