Satellite-based network-level pavement roughness estimation using dual-polarization Sentinel-1 SAR and deep learning

Timely pavement-condition information supports network-level maintenance planning, but repeated profiler surveys require dedicated field campaigns. This study develops a scripted workflow that converts openly available Sentinel-1 synthetic aperture radar (SAR) observations into segment-level International Roughness Index (IRI) estimates and pavement-management outputs. A seven-configuration experiment separates SAR inputs (S), pre-inspection engineering information (P), and contemporaneous condition indicators (C) according to their availability relative to the target inspection. The study comprises 8,188 flexible-asphalt pavement segments in Alaska, Utah, and West Virginia, including 1,228 corridor-held-out test observations. SAR alone achieved a mean absolute error (MAE) of 14.8 in/mi. Under retrospective, inspection-aligned evaluation, combining SAR with pre-inspection information reduced MAE to 13.4 in/mi, compared with 16.1 in/mi for pre-inspection information alone. Full fusion achieved an MAE of 12.9 in/mi, an RMSE of 17.2 in/mi, and a squared Pearson correlation of 0.69, indicating only a further 0.5 in/mi improvement over the same-campaign condition indicators. Matched SAR-only AnySat and SkySense adaptations achieved three-seed mean MAEs of 14.2 and 14.5 in/mi, respectively. For the full-fusion predictor, MC dropout and split-conformal intervals achieved 89.6% and 90.2% empirical coverage, respectively, at a nominal 90% level. Uncertainty-guided screening flagged 8.4% of test segments while capturing 35.2% of errors exceeding 25 in/mi. The results support using an availability-aware satellite screening layer to prioritize field measurements within the evaluated domain.