Patagonian Andes Landslides Inventory: The Deep Learning’s Way to Their Automatic Detection

Abstract

Landslide inventories are crucial to studying the dynamics, associated risks, and effects of these geomorphological processes on the evolution of mountainous landscapes. The production of landslide maps is mainly based on manual visual interpretation methods of aerial and satellite images combined with field surveys. In recent times, advances in machine learning methods have made it possible to explore new semi-automated landslide detection methodologies using remotely detected images. In this sense, developing new artificial intelligence models based on Deep Learning (DL) opens up an excellent opportunity to automate this arduous process. Although the Andes mountain range is one of the most geomorphologically active areas on the planet, the few investigations that use DL mainly focus on mountain ranges in Europe and Asia. One of the main reasons is the low density of landslide data available in the Andean areas, making it difficult to experiment with DL models requiring large data volumes. In this work, we seek to narrow the existing gap in the availability of landslide inventories in the area of the Patagonian Andes. In addition, the feasibility and efficiency of DL techniques are studied to develop landslide detection models in the Andes from the generated datasets. To achieve this goal, we generated in a manual process a datasets of 10,000 landslides for northern Chilean Patagonia (42–45°S), being the largest freely accessible landslide datasets in this region. We implement a machine learning model, through DL, to detect landslides in optical images of the Sentinel-2 constellation using a model based on the DeepLabv3+ architecture, a state-of-the-art deep learning network for semantic segmentation. Our results indicate that the algorithm detects landslides with an accuracy of 0.75 at the object level. For its part, the segmentation reaches a precision of 0.86, a recall of 0.74, and an F1-score of 0.79. The correlation of the segmentation measured through the Matthews correlation coefficient shows a value of 0.59, and the geometric similarity of the correctly detected landslides measured through the Jaccard score reaches 0.70. Although the model shows a good response in the testing area, errors are generated that can be explained by geometric and spectral relationships, which should be solved through new training approaches and data sets.