Nikhil Prakash , Andrea Manconi, Simon Loew

Landslides are a common secondary hazard triggered by major earthquakes or extreme rainfall events in steep terrains. Rapid mapping of event landslides is crucial to identify the areas affected by damages as well as for effective disaster response. Traditionally, such maps are generated with visual interpretation of remote sensing imagery (manned/unmanned airborne systems or spaceborne sensors) and/or using pixel-based and object-based methods exploiting data-intensive machine learning algorithms. Recent works have explored the use of convolutional neural networks (CNN), a deep learning algorithm, for mapping landslides from remote sensing data. These methods follow a standard supervised workflow involving a model trained using a landslide inventory over a relatively small region and then applied for prediction in the surrounding. Here, we propose a new strategy, i.e., a progressive CNN training relying on multiple event landslide inventories to build a generalized model that can be applied directly to a new, unexplored area. These inventories are spread across different geographic regions. We first proved the effectiveness of CNNs for mapping event landslides in four regions after earthquakes and/or extreme meteorological events. The best MCC score for each region ranged from 0.574 to 0.806. However, when mapping new unseen areas, we found that CNNs trained on a combination of multiple inventories have a better generalization performance with a bias towards higher precision. This combined training model could achieve the highest MCC score of 0.69 when mapping landslides in new unseen regions. Despite the expense of a slightly reduced accuracy, the main advantage of combined training is to overcome the requirement of a local training inventory for mapping event landslides triggered by future events. This can facilitate an automated pipeline providing fast response for the generation of landslide maps in the post-disaster phase.