DL-AISLE: A Deep Learning framework using Active Learning on Satellite imagery for Landslide identification

A landslide database is of utmost importance for hazard management as well as early warning systems. Historically landslides were manually identified by ground surveys or remote sensing data, but with development in satellite technology open-source satellite imagery has emerged as a preferred data source for landslide identification due to its cost effectiveness. On the other hand, an increase in computing power made computer vision methods especially deep learning popular for satellite image segmentation. Deep learning models require a large amount of data to reach operational performances, however there is very little labelled landslide data present. Labelling satellite imagery is costly and time consuming. Active learning remedies this by optimally selecting the data to label thereby maximizing the performance of the model given the limited data. In this study we present an active learning-based framework to train a segmentation model to identify landslides. The pre- and post-landslide images from sentinel 2 are merged with terrain features to create input data bands. The model is tested on a test database using metrics like IOU. The methodology has been developed with an application in India but can be applied globally.