Ambient dose rate variation in the Fukushima region visualized using explainable AI techniques

Yoshida, Ryu*; Kurikami, Hiroshi  ; Nagao, Fumiya; Takahashi, Shigeo*; Sanada, Yukihisa   

Following the Fukushima Daiichi Nuclear Power Station accident in 2011, ambient dose rates in the surrounding region have gradually declined due to radioactive decay and decontamination efforts. However, spatial variations in dose rate reduction remain insufficiently understood, particularly in forested areas where contamination persists. This study investigates long-term trends in ambient dose rate changes using explainable AI techniques. A 12-year integrated dose rate map, combining fixed-point, walk, carborne, and airborne survey data, was used to analyze temporal and spatial patterns. We developed a predictive model using Light Gradient Boosting Machine (LightGBM) to estimate dose rate reduction ratios based on geographic and environmental features. SHapley Additive Explanations (SHAP) were applied to quantify the contribution of each variable and enhance model interpretability. Our findings revealed that land use significantly influences dose rate reduction, with urban and agricultural areas showing faster declines due to infrastructure and human activity including decontamination works, while forests exhibit slower reductions. Notably, topographical features such as elevation and slope affect dose rate trends in undisturbed forests, with valleys and depressions showing stagnation. This study provided the first visual validation of area-wide decontamination effects and demonstrates the utility of explainable AI in environmental radiation analysis. The proposed approach offers a robust framework for geospatial interpretation and supports future policymaking for regional recovery and forest utilization.