A robust SAR change detection pipeline for landslide mapping using tile-wise HFEM and graph-cut Refinement

Landslide detection from SAR images remains challenging due to speckle noise, varied terrain conditions, and limited annotated data. In this work, a hybrid model with a Histogram Feature Extraction Module (HFEM) and Graph-Cut Refinement is used to improve change detection. The dataset consists of pre- and post-event SAR images of the 2021 Haiti earthquake with ground-truth annotations. The HFEM is used to acquire discriminative spatial features and suppress noise, and Graph-Cut Refinement is used to regularize segmentation and make it more consistent. The experimental results verify that the proposed approach obtains high recall and comparative precision, balanced F1-score, and fewer error regions than baseline approaches. The qualitative findings also demonstrate the steadiness of the forecast change maps. Overall, the system is very suitable for geospatial mapping of big landslides and supports disaster relief and risk management uses.