Multi-hazard road-network criticality from earth observation data using spatial graph attention networks with conformal uncertainty

Coastal deltaic road networks in India experience compounding hazards of monsoon flooding, prolonged soil-moisture saturation, and heat stress. However, many graph-learning approaches to network vulnerability assessment validate with random train–test splits that do not account for spatial dependence between adjacent road segments leading to insufficient quantification of failure probability without any spatially calibrated uncertainty. This study attempts to address these gaps by developing a spatially aware Graph Attention Network (GAT) that estimates road-segment failure probability from multi-hazard exposure and network-topological covariates. Due to the lack of geocoded road-closure records, multi-hazard threshold rules were used to generate training labels, with model predictions interpreted as disruption susceptibility rather than observed failure probabilities. Model performance was evaluated using random and spatial-block cross-validation, while split-conformal prediction was used to quantify predictive uncertainty. The framework was applied to the East Godavari–Krishna deltaic corridor of coastal Andhra Pradesh using multi-source climate, satellite, terrain, road-network and population data (i.e., IMD, NASA SMAP, ERA5-Land, CHIRPS, Copernicus DEM, Sentinel-2, OpenStreetMap, and WorldPop) from 2016 to 2025. The GAT achieved an area under the receiver operating characteristic curve (AUC) of 0.923 ± 0.001 under random cross-validation, and 0.682 ± 0.126 under spatial-block cross-validation. Under spatial-block cross-validation, Random Forest outperformed the GAT in all five folds, achieving an AUC of 0.746 ± 0.134 using the same edge-level features without network topology. Message-passing ablation indicated a marginal improvement from real network topology, with an AUC of 0.639 without message passing compared with 0.624 using randomly shuffled neighbours. A population-weighted Monte Carlo simulation showed that failure-probability-weighted removal produced greater network efficiency loss than random removal across all tested removal fractions (1%–20%), with the largest difference observed at 10% removal, where the loss ratio reached 1.65 times. A consequence-sensitive risk ranking combining failure probability and demand-weighted network flow identified a top 5% of segments whose protection avoided 56% of simulated accessibility loss, compared with 15% for failure-probability ranking, 18% for hazard-exposure ranking, and 43% for edge betweenness.