Physics-Aware Deep Learning Reconstructs Ground Contamination from Sparse UAV Radiation Measurements over the Fukushima Ukedo Basin Without Field Training
Abstract
Aerial radiation surveys produce sparse trajectories that must be reconstructed into contamination maps. Conventional aerial interpolators—inverse distance weighting (IDW) and ordinary kriging—treat observations as local ground samples, ignoring that each measurement integrates radiation over an extended footprint A(x, y) = (C * K)(x, y). The resulting double-blurring imposes a second smoothing on already-convolved values, causing systematic underprediction regardless of measurement density. We cast reconstruction as inverse deconvolution. A physics-aware encoder–decoder receives five channels (sparse measurements, IDW baseline, land–water scalar prior, measurement mask, water mask) and learns to invert K under a forward-consistency loss. The network is pretrained on synthetic data and deployed without fine-tuning. At a 50% random within-system holdout over the 2213-point Ukedo benchmark trajectory, 25 runs achieve a mean root-mean-square error (RMSE) of 705.4 ± 102.8 counts per second (CPS) versus 916.8 ± 34.2 (IDW) and 832.4 ± 31.3 (Kriging), with directional improvement over IDW in 25/25 runs. In a three-model ensemble diagnostic, among held-out points exceeding T = 6000 CPS (n = 64 at split seed 10, near the IDW ceiling), the U-Net recovers approximately 80% while IDW and kriging both fall to approximately 0%. The operational value lies in high-intensity hotspot recovery. These gains apply to dense-trajectory, within-coverage reconstruction; under large-gap extrapolation beyond the observed trajectory, the advantage over conventional interpolation is drastically reduced, and spatially independent validation remains an open challenge.