EOMF-26. Identifying global landslides using satellite imagery and machine learning in GEE with SLID

Abstract
We develop an open-source, user-friendly tool named Satellite Landslide Identification and Detection (SLID) on Google Earth Engine (GEE) that locates landslides using Synthetic Aperture Radar (SAR), Enhanced Vegetation Index (EVI), and clustering (unsupervised machine learning) on the global level. Although there are many landslide detection methods, GEE facilitates cloud-based computing that allows for rapid processing without storage complications inside a user-friendly platform. Sentinel-1 A/B SAR and Sentinel-2 optical data are used for the analysis, with an amplitude change approach for the SAR dataset. In addition, we use a Digital Elevation Model (DEM) from the Shuttle Radar Topography Mission to calculate slope and mask flat areas to reduce false positives. We compare our findings to five global study case studies: 2015 Tan Fjord (Alaska), 2016 Matililja Valley (California), 2017 Santa Lucia (Chile), 2020 Bahrabise (Nepal), and 2020 Pettimudi (India). Various change ratios are calculated from each dataset and tested against each other to understand their significance and performance, including EVI versus Normalized Difference Vegetation Index (NDVI) and SAR ratio versus log ratio. In addition, different training point amounts and cluster numbers are analyzed to optimize identification performance for any global landslide. We learned that EVI identifies landslides pixels 1.3 times better than NDVI, SAR ratio identifies landslide pixels 2.74 times better than SAR log ratio. Additionally, 10,000 training points and 7 clusters outperformed other variations.