A Weakly Supervised Landslide Extraction Method from Remote Sensing Images Using Deep Attention and Multi-feature Fusion

Abstract

The semantic segmentation model significantly improves the accuracy of landslide extraction from very high resolution (VHR) remote sensing images, but it requires manual sketching of many pixel-level annotations. Pixel-level annotation needs can be solved by weakly supervised learning based on image-level annotations. We propose a weakly supervised strategy combining deep attention and multi-feature fusion for landslide extraction. By obtaining high quality class activation maps (CAMs), an accurate landslide extraction model can be trained. Many experiments on the VHR remote sensing images after the Jiuzhaigou earthquake show that the proposed strategy can obtains more complete and accurate CAMs, and the landslide extraction accuracy is better than mainstream weakly supervised methods and achieved results comparable to strong supervised method.