Change-driven Hysteresis and Inference-ensemble Pseudo-labeling for Landslide Segmentation

Accurate landslide segmentation from multitemporal remote sensing remains challenging due to strong class imbalance, heterogeneous background disturbances, and ambiguous boundaries induced by shadows, vegetation dynamics, and sensor noise. We propose Change-driven Hysteresis and Inference-ensemble Pseudo-labeling for Landslide Segmentation (CHIPS), a semi-supervised framework that leverages change cues to scale high-quality supervision while controlling error propagation in pseudo-label learning. CHIPS uses LandTrendr-derived spectral change descriptors as the primary representation and couples them with a hysteresis-based selection strategy that assigns pseudo-labels via dual thresholds, enabling confident positives and confident negatives while deferring uncertain pixels to an ignored set. This design explicitly balances precision–recall trade-offs and mitigates confirmation bias under severe foreground sparsity. To further stabilize learning, we integrate an inference-ensemble mechanism that aggregates multiple stochastic predictions (e.g., perturbation- or dropout-based) to estimate pixel-wise confidence and improve pseudo-label reliability. A teacher–student training scheme with exponential moving average supervision combines supervised segmentation loss with pseudo-label and consistency objectives under a scheduled ramp-up. Experiments on a large-scale landslide dataset constructed from change-map patches demonstrate that CHIPS consistently improves intersection-over-union and boundary delineation over fully supervised baselines and common semi-supervised alternatives, particularly in challenging terrain and low-label regimes. The proposed framework offers a practical and scalable solution for regional landslide mapping using change-driven priors and robust pseudo-labeling.