|Paper title||Rapid Mapping of Event Landslides Using an Active Learning Workflow|
|Form of presentation||Poster|
Earthquakes and extreme weather events are responsible for triggering a population of catastrophic landslides in mountainous regions which can damage infrastructure and cause fatalities. In the last decade, an exceptionally high distribution of fatal landslides was observed after the cloudburst event in North India (2013), the Nepal earthquake (2015), the Hokkaido Iburi-Tobu earthquake (2018), Storm Alex in French-Italian Alps (2020), among many others that forced the civil defense authorities to quickly map event landslides over large regions for planning an effective disaster response. These mapping efforts were aided by the increased availability of Earth observation (EO) images from many satellites orbiting on agile platforms or in large constellations, combined with the coordinated efforts of The International Charter Space and Major Disasters members. Now it is possible to obtain data from the affected region in a couple of hours. Synthetic aperture radar (SAR) sensors can even provide data sensed through the clouds during bad weather conditions. However, the landslide mapping process is still predominantly dependent on visual interpretation or semi-automated methods, which can cause a delay of a few days to many months till a near-complete inventory is available. Hence, there is an increased need for a data-agnostic method for rapid landslide mapping. In recent years, deep-learning based methods have shown unprecedented success in image classification and segmentation tasks. They have been adopted for mapping landslides in several scientific studies. However, most of these studies rely on an already existing large inventory for training the deep-learning models, making such methods unsuitable for a rapid mapping scenario.
This work presents an active learning workflow to generate a landslide map from the first available post-event EO data. The proposed method is a multi-step process where we start with an incomplete inventory covering a small region. In subsequent steps, we increase the coverage and accuracy of the landslide map with feedback from an expert operator. We apply our method to map landslides triggered by the Hokkaido Iburi-Tobu earthquake (Japan), which occurred on 5th September 2018. In the next days, the affected region was covered with clouds which prohibited the acquisition of useful data from optical satellites. Hence, we used ALOS-2 SAR data which was available one day after the event. Our results indicate that an active learning workflow has a small reduction in performance compared to a traditionally trained model but eliminates the need for a large inventory for training which is a bottleneck during rapid mapping scenarios.