Rapid mapping of landslides using satellite SAR imagery: A progressive learning approach

Abstract

Rapid detection of landslides after an exceptional event is critical for planning effective disaster management. Previous works have typically used machine learning-based methods, including the recently popular deep-learning approaches, to identify characteristics surface features from satellite remote sensing data, especially from optical images. However, data acquisition from optical images is not possible in cloudy conditions, leading to unpredictable delays in any mapping task from future events. These methods also rely on large manually labelled inventories for training, which is often not available before the event. In this work, we propose an active training strategy to generate a landslide map after an event using the first available synthetic-aperture radar (SAR) image and improve it once subsequent cloud-free optical images are acquired. The proposed active learning workflow can start with a small ($\sim 100m^2$) and incomplete inventory,- and can grow the extent and completeness in iterative steps with manual updates after each step. This significantly reduces the slow manual mapping typically required for generating a large training inventory. We designed our experiments to map the landslides triggered by the $M_w$ 6.6 Hokkaido Eastern Iburi earthquake of 2018 in Japan using sequentially ALOS-2 (SAR) and PlanetScope (Optical) scenes in the order they are acquired. The choice of active learning prioritizes speed over accuracy. However, we note only a modest reduction in performance ($\sim 10\text{\%}$ drop in F1 and MCC scores), with our method allowing a preliminary landslide inventory to be completed within a single day. This is of major importance in disaster response, improving performance and reducing the potential subjectivity associated with manual mapping.