CS-WSCDNet: Class Activation Mapping and Segment Anything Model-Based Framework for Weakly Supervised Change Detection
Lukang Wang, Min Zhang, Wenzhong Shi
Abstract
Change detection (CD) using deep learning techniques is a trending topic in the field of remote sensing. However, most existing networks require pixel-level labels for supervised learning, which is difficult and time-consuming to label all changed pixels from multi-temporal images. To address this challenge, we propose a novel framework for weakly supervised change detection (WSCD), namely CS-WSCDNet, which can achieve pixel-level results by training on samples with image-level labels. Specifically, the framework is built upon the localization capability of class activation mapping (CAM) and the powerful zero-shot segmentation ability of the foundation model, i.e., segment anything model (SAM). After training an image-level classifier to identify whether changes have occurred in the image pair, CAM is utilized to roughly localize the regions of change in the images pair. Subsequently, SAM is employed to optimize these rough regions and generate pixel-level pseudo-labels for changed objects. These pseudo-labels are then used to train a CD model at the pixel-level. To evaluate the effectiveness of CS-WSCDNet, experiments are conducted on two high-resolution remote sensing datasets. It shows that the proposed framework not only achieves state-of-the-art (SOTA) performance in WSCD tasks but also demonstrates the potential of weakly supervised learning in the field of CD. The demo codes are available at https://github.com/WangLukang/CS-WSCDNet.