Bayesian Deep Learning for InSAR-Based Geohazard Assessment With Uncertainty Quantification

Geological hazards in mountainous basins commonly arise from overlapping landslide, subsidence, and uplift processes, yet susceptibility mapping remains limited by incomplete inventories and poorly quantified predictive uncertainty. This study proposes a Bayesian deep-learning framework for probabilistic multihazard susceptibility mapping by integrating small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) deformation measurements with geological, geomorphological, edaphic, climatic, land-use, and multispectral covariates. The training inventory is strengthened through physically guided label refinement based on slope and deformation thresholds derived from SBAS-InSAR, and its reliability is independently evaluated prior to model training. Bayesian convolutional and multilayer perceptron (MLP) models with Monte Carlo (MC) dropout are then used to estimate class probabilities together with predictive uncertainty. Independent validation of the integrated inventory yields an overall reliability of 79.48%. Under the original validation setting, the Bayesian convolutional neural network (CNN) achieves an accuracy of 0.94 and an AUCROC of 0.99, while spatial block cross-validation confirms a more conservative but methodologically stronger estimate of generalization under clustered deformation samples. The resulting susceptibility maps show spatial agreement of 76% and 72% with independent landslide and subsidence inventories, respectively. High-susceptibility zones indicate that subsidence affects 11.6% of the basin, landslides 2.2%, and uplift 1.3%, all with strong spatial clustering (Moran's I = 0.62-0.90). Compared with deterministic baselines, the Bayesian framework improves probability calibration and enables spatial separation of observation-driven and model-driven uncertainty. A decision-grade reliability mask derived from the validation coverage-risk frontier further allows unreliable predictions to be excluded explicitly. These results demonstrate that uncertainty-aware SBAS-InSAR susceptibility mapping can provide more reliable and operationally interpretable products for regional hazard assessment, planning, and early warning.