Deep learning-based landslide mapping using multi-sensor satellite imagery

Abstract

Intensification and increased frequency of extreme weather events due to the changing climate coupled with population urbanization is believed to increase the landslide hazard worldwide. Landslides often occur unpredicted and may result into loss of human life and property. Timely delivered information on the landslide location and extent as well as on the type and grade of damage is crucial to enable fast crisis response, i.e., to support rescue and humanitarian relief operations. This study aims to examine the applicability of a convolutional neural network (CNN) based on the U-Net architecture for mapping landslides using freely available optical and synthetic aperture radar (SAR) data from the Sentinel-2/1 satellites. Following research questions are investigated: (1) How accurately can we map landslides using 10m spatial resolution remote sensing data? (2) Does the addition of more pre- or/and post-event SAR scenes help to increase classification accuracies? (3) Does the combination of optical and SAR features result in better accuracies compared to single sensor features? The investigation is done within the framework of Multisat4slows project (Multi-Satellite imaging for Space-based Landslide Occurrence and Warning Service), financed by the Helmholtz Imaging 2020 call.