RMAU-NET: A Residual-Multihead-Attention U-Net Architecture forLandslide Segmentation and Detection from Remote Sensing Images

Abstract In recent years, landslide disasters have reportedfrequently due to the extreme weather events of droughts,floods , storms, or the consequence of human activities such asdeforestation, excessive exploitation of natural resources. How-ever, automatically observing landslide is challenging due tothe extremely large observing area and the rugged topographysuch as mountain or highland. This motivates us to proposean end-to-end deep-learning-based model which explores theremote sensing images for automatically observing landslideevents. By considering remote sensing images as the inputdata, we can obtain free resource, observe large and roughterrains by time. To explore the remote sensing images, weproposed a novel neural network architecture which is for twotasks of landslide detection and landslide segmentation. Weevaluated our proposed model on three different benchmarkdatasets of LandSlide4Sense, Bijie, and Nepal. By conductingextensive experiments, we achieve F1 scores of 98.23, 93.83 forthe landslide detection task on LandSlide4Sense, Bijie datasets;mIoU scores of 63.74, 76.88 on the segmentation tasks regardingLandSlide4Sense, Nepal datasets. These experimental resultsprove potential to integrate our proposed model into real-lifelandslide observation systems.