SGCT-Net: SAM-Guided Cross-Teaching Network for Weakly Supervised Semantic Segmentation for Generating High-Quality CAMs in High-Resolution Remote Sensing Imagery

Weakly supervised semantic segmentation (WSSS), particularly using image-level labels, has emerged as a pivotal solution to alleviate the burden of costly and labor-intensive pixel-level annotations typically required for remote sensing image interpretation. However, image-level label-based weakly supervised methods generally suffer from limitations such as poor class activation map (CAM) quality, pronounced confirmation bias, and unclear object boundaries. These drawbacks result primarily from factors including insufficient or excessive activation, confirmation bias during training, and boundary ambiguities in CAM, leading to unreliable pseudolabel generation and suboptimal segmentation performance. Addressing these challenges, this article proposes the SAM-guided cross-teaching network (SGCT-Net), a novel framework designed to generate high-quality CAMs for remote sensing images. SGCT-Net utilizes two separate student networks with nonshared parameters, facilitating mutual learning and effective suppression of confirmation bias and the dynamic threshold adjustment mechanism and loss-mixture noise estimation (LMNE)-based noise filtering are incorporated to refine pseudolabel accuracy. Furthermore, the segment anything model (SAM), a powerful general-purpose segmentation model, is introduced through the SAM-guided CAMs (SGC) module to enhance the CAMs by refining boundary delineations, thus further improving the accuracy of pseudolabels. Extensive experiments on benchmark remote sensing datasets, including ISPRS Potsdam, ISPRS Vaihingen, and iSAID, demonstrate the effectiveness of the proposed method. Extensive experiments on ISPRS Potsdam, ISPRS Vaihingen, and iSAID demonstrate that SGCT-Net achieves state-of-the-art segmentation performance with mIoU scores of 50.19%, 34.40%, and 36.71% respectively, significantly surpassing the baseline CTFA by margins of 6.38%, 13.56%, and 5.7%.