Landslide mapping from post-event single-temporal polarimetric SAR image by a deep learning method exploiting a morphological model

Accurate and timely mapping of landslides after the event (e.g., earthquake) is crucial for effective rescue operations and comprehensive disaster assessment. While optical images are often obstructed by clouds and fog, synthetic aperture radar (SAR) can identify landslides independently of weather conditions. In this study, we propose a deep learning method which exploits a morphological model (DLM) to achieve accurate landslide identification using only a post-event single-temporal polarimetric SAR image. The SAR scattering mechanisms and polarimetric characteristics of various ground objects are thoroughly analyzed to select optimal polarimetric parameters for deep learning. To accurately map landslide shapes and extract boundaries, we introduce a Majority Voting mechanism and a morphological optimization model. We have used one quad-pol ALOS-2 image for landslide mapping and achieved an overall accuracy of 95.24 % with the proposed method. Additionally, considering the limited availability of quad-pol SAR data, we have employed dual-pol ALOS-2 and Sentinel-1 data to assess the method's usability with dual-pol data. The dual-pol ALOS-2 image achieved an overall accuracy of 89.78 %, while Sentinel-1 image effectively captured the general landslide shape with an overall accuracy of 76.32 %. This demonstrates the high applicability of the proposed method for landslide mapping using a single post-event polarimetric SAR image, enhancing the timeliness of SAR-based landslide mapping and improving emergency response and post-disaster rescue capabilities. • Proposed a landslide mapping method using one post-event PolSAR image. • Designed a deep learning method exploiting a morphological model (DLM). • Explored landslide characteristics through polarimetric parameter analysis. • Provided solutions for post-disaster assessment using different polarimetric modes.