Landslide Detection with U-net Based SAR2OPT Framework : A Case Study of 2024 Hualien Earthquake Wen-Hong Chen. 1, Shou-Hao Chiang. 2
1Department of Civil Engineering, National Central University , Taiwan
2Center for Space and Remote Sensing Research, National Central University , Taiwan
Abstract
The 2024 Hualien earthquake caused multiple landslide, highlighting the need for dependable and rapid disaster detection. Synthetic Aperture Radar (SAR) offers an all-weather means of observation, making it a staple in remote-sensing studies. Yet, because SAR data lack the rich color and fine texture of optical imagery, extracting landslide information from them alone remains difficult. In this study, we embedded the modified U-net framework that combined four bands Sentinel-1 imagery, including VV and VH polarization data from ascending and descending observations. Fusing both orbits mitigates geometric layover and shadowing in rugged terrain, thereby preserving surface backscatter integrity. A sliding-window scheme with overlapping patches is employed during the training phase to preserve spatial context and minimize edge discontinuities between patches- specifically, 128*128 patches with a 64-pixel overlap and 256*256 patches with a 128-pixel overlap were tested. After end-to-end training, the model can rapidly produce optical-like NDVI maps directly from dual-orbit Sentinel-1 SAR imagery, offering an alternative for post-earthquake surface monitoring when optical data are not timely available. This study adopts the April 2024 Hualien earthquake in northeastern Taiwan as its testbed. The goal is to verify whether the SAR-derived optical products maintain enough spatial detail and spectral fidelity to support landslide mapping, paving the way for their fusion with machine-learning models in fully automated landslide detection.
