Identifying landslides in northern Chilean Patagonia using deep learning classification over Sentinel images

Landslide databases are essential for the proper development of predictive models. In the Patagonia of Chile, few works have advanced in this line, which affects the ability of civil authorities to prepare for landslides hazards that cause significant damage to communities and ecosystems, even more so considering the expected increase in landslide activity because of climate change. Mitigation techniques have been investigated in various studies to manage these natural hazards using different artificial intelligence techniques. However, the predictive capacity of the models depends to a great extent on the quality of the landslide databases necessary for training; also, they are generally not available. In this work, we want to address the efficiency of deep learning models to detect landslides in Patagonia. We compare various artificial intelligence-based models to detect landslides using optical images from the Sentinel 2 constellation. To do this, we generated a database of 10,000 landslides for northern Patagonia in Chile (42-45 ° S). Later we implemented different deep learning models combining topographic and optical data from the constellation Sentinel 2 for the segmentation of landslides. The original database generated does not include temporality; however, using this trained DL model, we will be able to address this goal in future works. Finally, this research seeks to be the basis for developing future early response models for landslides and the generation of automated maps of landslides after disasters in Chilean Patagonia.