Water-Matching CAM: A Novel Class Activation Map for Weakly-Supervised Semantic Segmentation of Water in SAR Images

Abstract

Recently, semantic segmentation of water in synthetic aperture radar (SAR) images has attracted the attention of more and more scholars. However, existing methods usually require many accurate manually labeled pixel-level water annotations of SAR images, which leads to the problem that they are often time-consuming and costly. To mitigate this problem, we apply the weakly-supervised semantic segmentation (WSSS) and class activation maps (CAMs) to the semantic segmentation of water in SAR images. To address the issues of incomplete activation and false positives associated with applying existing CAM methods to semantic segmentation of water in SAR images, we propose a novel water-matching CAM to generate accurate CAMs of water. Water-matching CAM includes a multilevel water-backscatter guided module (MWGM) and a nonwater targets consistency module (NTCM). MWGM introduces a priori information on water backscatter for multilevel CAM generation, which can generate complete water CAMs using features at four different depths. NTCM further improves the performance of water CAM by subjecting nonwater targets to feature consistency constraints, which can effectively alleviate the issue of false positives. Then, we utilize the CAMs to generate pseudolabels to train the semantic segmentation of water models. Experiments on three datasets of SAR images taken by the GF-3 and Sentinel-1 satellite verify the validity of water-matching CAM. Our method achieves state-of-the-art performance compared to other CAM-based WSSS methods

Keywords

• Class activation map (CAM)
• synthetic aperture radar (SAR)
• water semantic segmentation
• weakly-supervised learning