A spatially and temporally transferrable bi-temporal conditional generative adversarial network for landslide extraction

• Propose a bi-temporal conditional GAN framework for landslide extraction. • Propose non-local and adaptive convolution modules for spatial refinement. • Improve boundary accuracy and overall IoU by over 11% compared to SOTA. • Enable stable cross-year and cross-region generalization without fine-tuning. • Alleviate class imbalance via contour-based potential landslide selection. Accurate landslide extraction from remote sensing imagery is critical for hazard assessment, yet existing methods often struggle to generalize across different times, regions, and sensor modalities. This limitation largely stems from complex terrain, spectral confusion with background surfaces, and severe class imbalance in training data. To address these challenges, we propose BTLE-cGAN, a bi-temporal landslide extraction framework using conditional generative adversarial networks. Unlike prior approaches, BTLE-cGAN directly models the mapping task in an adversarial setting to jointly optimize spatial accuracy and structural realism. The generator integrates a Multi-scale Spectral–Spatial Convolution (MSSConv) module for morphological variation, a Non-Local Context Enhancement (NLCE) module to model long-range dependencies, and a Landslide-Adaptive Convolution (LAC) module that adjusts filters based on predicted shapes to refine boundaries. The discriminator leverages bi-temporal difference features to enforce semantic consistency and topological alignment. To alleviate sample imbalance, we introduce a Potential Landslide Extraction (PLE) strategy that generates contour-guided, land-cover-filtered candidate regions, thereby improving the representativeness and efficiency of both training and inference. Extensive evaluations across regions, time periods, and sensors show BTLE-cGAN consistently outperforms 11 state-of-the-art frameworks, achieving up to 11% average improvement in IoU. It maintains robust performance under temporal shifts, sensor variation, and spatial transfers, offering a scalable, generalizable solution for high-precision landslide monitoring. Our method advances bi-temporal change detection for geohazard mapping and supports future applications in dynamic environments.