Weakly Supervised Semantic Segmentation Using Visual Explainability

Abstract

In the realm of remote sensing, the task of semantically segmenting landslide images is traditionally reliant on supervised learning techniques. These methods necessitate extensive training datasets and meticulous pixel-level annotations—a process that demands considerable human labor and incurs high costs. To mitigate these challenges, we introduce an innovative approach that employs weakly supervised learning, integrating Class Activation Maps (CAMs) with a Cycle Generative Adversarial Network (CycleGAN). This novel methodology leverages image-level labels in lieu of pixel-level annotations. Initially, CAMs are utilized to locate the landslide’s rough area. Subsequently, CycleGAN generates a synthetic image devoid of landslides, which, when contrasted with the original, yields precise segmentation results. The efficacy of our approach is quantified using the mean Intersection-over-Union (mIOU) metric, demonstrating a superior performance with an mIOU of 0.228. Additionally, when juxtaposed with a U-Net network’s supervised learning technique, which scored an mIOU of 0.408, our results affirm the viability of weakly supervised learning for accurate landslide semantic segmentation in remote sensing imagery. This method significantly alleviates the burden of data annotation. Incorporating the advancements of Score-CAM [6], which surpasses Grad-CAM in object localization accuracy, we further refine our model. Score-CAM’s enhanced precision in identifying relevant features contributes to the improved segmentation of landslide areas, promising a new frontier in remote sensing image analysis.