SCGC-Net: Spatial Context-Guided Calibration Network for Multisource RSI Landslides Detection
Yukun Fan, Peifeng Ma, Q. Hu, Guiwei Liu, Zihuan Guo, Yixian Tang, Fan Wu, Hong Zhang
Abstract
Landslide is a common geological disaster, and rapid landslide extraction using high-resolution remote sensing imagery (RSI) is of great significance for emergency rescue and damage assessment. In RSI, landslides often have irregular shapes, large-scale variations, and are easily affected by environmental factors. Existing deep learning methods have limited ability in extracting multiscale features, integrating these features effectively, and adapting to complex environments, resulting in models that are not optimized for robustness. To overcome these challenges, this study proposes a spatial context-guided calibration network (SCGC-Net) for multisource remote sensing data. SCGC-Net introduces a novel combination of hybrid multiscale feature extraction, context-aware modulation of landslide characteristics, and a progressive feature calibration fusion strategy, enabling efficient feature extraction, accurate feature integration, and enhanced cross-domain generalization when working with multisource remote sensing data. SCGC-Net was tested on several datasets representing diverse geographical regions and imaging platforms, including the CAS Landslide Dataset (CLD), HR-GLDD, Bijie, and global very-high-resolution landslide mapping (GVLM). Experimental results indicate that SCGC-Net outperforms existing methods across all evaluation metrics and exhibits superior generalization performance in domain adaptation experiments.