Weakly Supervised Remote Sensing Image Semantic Segmentation With Pseudo-Label Noise Suppression
Abstract: Semantic segmentation of remote sensing images (RSIs) plays a crucial role in various applications, including urban planning and environmental monitoring. However, the high cost and complexity of obtaining detailed annotations for RSIs pose significant challenge. This issue necessitates the exploration of weakly supervised learning as an effective alternative, which utilizes more readily available, less granular forms of labeling. Yet, weakly supervised approaches face their own set of challenges, primarily due to scarcity of precise pixel-level labels which significantly hampers the model’s ability to learn accurate representations. In this article, we introduce a weakly supervised semantic segmentation (WSSS) approach for RSIs that leverages self-supervised learning (SSL) and pseudo-label noise mitigation to address these challenges. Our method leverages a self-supervised encoder for providing similarity information, which enhances feature representation in RSIs and enables the generation of more accurate pseudo-labels, thus reducing the noise in the pseudo-labels. Furthermore, we propose a refined loss function that incorporates gradient clipping and label smoothing to mitigate the impact of noisy labels, thereby improving the robustness and accuracy of the segmentation results. Extensive experiments on the ISPRS Potsdam, ISPRS Vaihingen, and iSAID datasets demonstrate that our approach achieves state-of-the-art (SOTA) performance, closely matching that of fully supervised methods. Our method not only reduces the dependency on expensive pixel-level annotations but also showcases the potential of SSL in enhancing WSSS tasks.
Loading