Lorenzo Nava , Alessandro Cesare Mondini, Kushanav Bhuyan, Chengyong Fang, Oriol Monserrat, Alessandro Novellino , Filippo Catani

Efficient response to large and widespread multiple landslide events (MLEs) demands rapid and effective landslide detection. Despite extensive efforts using optical remotely sensed imagery, limitations in global, day & night, and all-weather operational capabilities remain. To address these gaps, we introduce an approach that harnesses Deep Neural Networks (DNNs) and Synthetic Aperture Radar (SAR) backscatter data. This approach is designed through the analysis of 11 earthquake-induced MLEs, encompassing approximately 73 thousand landslides that occurred worldwide in a variety of different geo-settings. We test the reproducibility of the model results on unseen earthquake-induced landslides that occurred in Sumatra and Haiti. The top-performing model achieved a test F1-score of 82% in rapid assessment, indicating significant progress compared to previous attempts. The approach harnesses the cloud computing resources of Google Earth Engine for Sentinel-1 SAR image acquisition and processing, complemented by local computing resources to utilize advanced image classification DNNs capabilities. Through explainable artificial intelligence, our study underscores the efficacy of change detection bands in their superior discriminative capacity to delineate landslide features, surpassing the utilization of backscatter data alone. Moreover, we observe an improved ability to detect landslides within multi-temporal information stacks as opposed to single post-event SAR images. Finally, we introduce the SAR-LRA Tool in its Beta version, providing a valuable resource for rapid and comprehensive all-weather global landslide assessment. The systematic use of the Tool promises to facilitate the timely response to future MLEs. Our work establishes a robust foundation for future research endeavors, wherein SAR and DNNs can be harnessed to identify natural hazards and/or specific earth surface changes in mountainous regions. Given the frequent and increasing occurrence of MLEs, the development of a robust modeling approach is imperative to timely assess the spatial distribution of these phenomena. This research will pave the way for efficient rapid assessment of MLEs in the future.