Unsupervised SAR Change Detection Using Two-Stage Pseudo Labels Refining Framework
Sheng Fang, Chenxu Qi, Shuqi Yang, Zhe Li, Wenhao Wang, Yumeng Wang
Abstract
Unsupervised change detection (CD) in Synthetic Aperture Radar (SAR) imagery is pivotal for terrestrial observations, more so for disaster-related applications. However, most existing deep learning methods primarily emphasize the construction of diverse networks, often overlooking the critical aspect of refining pseudo labels. This letter proposes a two-stage pseudo labels refining (TSPLR) framework for SAR image unsupervised CD. During the first stage, Fuzzy-C-Means (FCM) clustering is employed on the bi-temporal SAR images to yield initial pseudo labels, categorizing high-confidence data as changed or unchanged and the rest as uncertain. A straightforward network is then trained first with confident data, with the training subsequently extended to the uncertain data. In the second stage, we start by comparing the predictions from the model trained during the first phase with the initial pseudo labels. Data with inconsistencies is added to the uncertain dataset, and some pixels filtered according to connectivity areas are selected as hard samples. Subsequently, the model is trained further using the updated dataset. This two-stage refinement process bolsters the credibility of pseudo labels and creates a more robust network training against speckle noise. The experimental results show that even with a simple network, the performance of the proposed TSPLR exceeds the current performance of SOTA. We will be making the source codes publicly accessible at https://github.com/sdust-mmlab.