近年来，深度学习在多时相遥感影像变化检测任务中表现出巨大的潜力。充分的训练样本是深度学习技术能够有效挖掘遥感影像变化特征的重要前提，然而当前有限的公开标注数据集还不能满足实际应用中各种变化类型检测的需求。在双时相变化检测中，由于地表覆盖变化通常只占据少部分区域，能够获取的变化样本常常数量很少，且与不变化样本相比存在严重的不平衡问题；而在时序变化检测中，变化时间的定位需要每一期观测的标签，这种包含变化时间的时序样本更加难以获取。因此，如何在小样本情况下有效训练与应用变化检测网络是急需突破的难题。相比变化检测样本，单时相地表覆盖分类样本的获取难度要低得多；充分训练的地表覆盖分类网络所挖掘的信息可为变化检测提供重要的先验特征。基于此，本文从地表覆盖分类后验概率的角度出发，将地表覆盖分类后验概率引入到双时相变化检测中，降低对变化检测样本的依赖。针对时序变化检测问题，本研究通过非监督变化时间提取方法，将其分类后验概率拓展到时间序列变化检测中，规避了时序样本标签的需求，大大增加了算法实用性。本文研究工作主要包括以下方面：

 (1) 针对双时相变化检测，本文提出了一种基于分类后验概率空间的孪生Nested-UNet变化检测网络（Siamese Nested-UNet for change detection in Posterior Probability Space，SNU-PS），通过结合两期地表覆盖分类后验概率信息，降低对变化检测样本的依赖。该方法首先基于地表覆盖分类样本训练高分辨率网络（High-Resolution Network，HRNet），得到双时相影像的地物分类后验概率；然后将后验概率图像输入到孪生Nested-UNet变化检测网络（Siamese Nested-UNet for change detection，SNUNet-CD）中以获取变化检测结果。在SpaceNet7 和HRSCD数据集上测试的结果表明，SNU-PS能够充分利用地表覆盖的语义信息，在不同变化检测训练样本数量水平下，保持稳定的变化检测精度；相比分类后比较（Post Classification Comparison，PCC）、分类后融合孪生网络（Deep Siamese Postclassification Fusion Network，PCFN）和SNUNet-CD三种方法，具备更高与更稳定的变化检测精度，特别在样本数量不足时，优势更为明显。

(2)针对时序变化检测问题，本文结合非监督的移动加和（Moving sum，MOSUM）时序断点检测算法，将SNU-PS拓展到时序数据上，提出了基于分类后验概率框架的半监督时序变化检测方法（A Semi-supervised time series change detection method based on Posterior Probability Framework，STSCD-PF）。该方法主要包括三步：a.基于单期地表覆盖分类样本训练HRNet，并应用到每期影像生成相应的地表覆盖分类后验概率时序；b.利用SNU-PS算法检测时序第一期和最后一期影像对比的变化区域；c.针对提取出的变化区域，通过MOSUM算法逐像元提取后验概率时序断点，作为变化时间；d.应用主值滤波减少像元级断点识别的椒盐噪声。在SpaceNet7和DynamicEarthNet数据集上，将本文提出的STSCD-PF算法与PCC和lxastro0算法进行了比较，实验结果表明，STSCD-PF在变化空间位置与变化时间的检测精度上都要优于PCC和lxastro0。较于传统方法，STSCD-PF可以更精确地提取变化位置、变化时间与变化类型，受多期影像分类不一致问题的影响较小，同时不需要时序样本标签进行训练，具备良好的实用性。

In recent years, deep learning has demonstrated remarkable potential for change detection in multi-temporal remote sensing images. However, annotated datasets, which are crucial for training change detection networks, are often limited in various change detection tasks within practical applications. Since land cover change typically occupies only a small portion of an image, the number of changed samples is frequently quite limited, resulting in a significant imbalance between changed and unchanged samples. Consequently, effectively training change detection networks with small and imbalanced change detection samples presents an urgent challenge. For time series change detection, identifying the change time necessitates a time series of land cover labels, which are even more challenging to acquire. Collecting change detection samples is more difficult than obtaining land cover classification samples at a single time. Leveraging adequate land cover classification samples, a well-trained land cover segmentation network can provide valuable prior features for change detection. This paper introduces the posterior probability of land cover classification into dual-phase change detection to reduce reliance on change detection samples. Combined with an unsupervised time node extraction method, this approach is extended to time series change detection, significantly enhancing the algorithm's practicality. The research presented in this paper focuses on the following aspects:

(1) We propose a method called Siamese Nested-UNet for Change Detection in Posterior Probability Space (SNU-PS), which aims to reduce dependence on change detection samples by utilizing the posterior probability information of the segmentation network. The method initially trains a high-resolution network (HRNet) based on land cover classification samples to obtain the bi-temporal image's posterior probability. Subsequently, the posterior probability images are input into a Siamese Nested-UNet for change detection (SNUNet-CD) to obtain change detection results. To simplify network complexity and reduce training difficulty, the training of the semantic segmentation network and change detection network is carried out step by step without interactions during their respective training stages. Since the posterior probability image already contains the land cover's semantic information, the change detection samples' requirement is reduced as the change detection network does not need to extract semantic features from multi-spectral images. Change detection experiments using the SpaceNet7 and HRSCD datasets demonstrate that SNU-PS can effectively utilize the semantic information provided by the land cover segmentation network and maintain stable change detection accuracy when trained with various change detection sample sizes. Compared to Post Classification Comparison (PCC), Post-Classification Fusion Network (PCFN), and SNUNet-CD, SNU-PS achieves higher accuracy and improved stability, particularly when the sample size is small.

(2) For time series change detection, we propose a semi-supervised time series change detection method based on a posterior probability framework (STSCD-PF), which extends SNU-PS to time series data by integrating the Moving Sum (MOSUM) algorithm, capable of unsupervisedly detecting breakpoints in a time series. The method primarily involves four steps: a. HRNet is trained with land cover classification samples from either period and then applied to each image in the time series to generate corresponding posterior probability time series of land cover classification; b. SNU-PS is utilized to detect the spatial location of changed objects based on the first and last images in the time series; c. the change time of detected changed pixels is further identified by detecting the breakpoint of the posterior probability time series using the MOSUM algorithm; d. major filtering is applied to reduce the "salt and pepper" noise induced by pixel-based breakpoint detection. The proposed STSCD-PF method is compared with two typical methods, PCC and lxastro0, based on change detection experiments using the SpaceNet7 and DynamicEarthNet datasets. The experimental results reveal that STSCD-PF outperforms PCC and lxastro0 in change location accuracy and change time accuracy. Compared to traditional methods, STSCD-PF can more accurately identify the change area, change time, and change type, with less sensitivity to inconsistencies in multi-temporal image classification. Furthermore, STSCD-PF, as a semi-supervised method, does not require time-series label samples for training, which greatly expands its application scope in land cover monitoring.