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

Area of Science:

• Environmental Science
• Radiation Monitoring
• Artificial Intelligence

Background:

• The Fukushima Daiichi Nuclear Power Station accident in 2011 led to widespread radioactive material deposition.
• Ambient dose rates have decreased due to decay and decontamination, but spatial variations persist, especially in forests.

Purpose of the Study:

• To investigate long-term ambient dose rate changes using explainable AI (XAI).
• To analyze temporal and spatial patterns of dose rate reduction over 12 years.
• To identify factors influencing dose rate reduction, particularly in forested areas.

Main Methods:

• Integrated dose rate mapping using diverse survey data (fixed-point, walk, carborne, airborne).
• Development of a predictive model (Light Gradient Boosting Machine) for dose rate reduction ratios.
• Application of SHapley Additive exPlanations (SHAP) for model interpretability and variable contribution analysis.

Main Results:

• Land use significantly impacts dose rate reduction; urban and agricultural areas show faster declines than forests.
• Topographical features (elevation, slope) influence trends in forests, with valleys and depressions exhibiting stagnation.
• XAI effectively quantified variable contributions, revealing factors behind spatial variations in decontamination effectiveness.

Conclusions:

• Forests exhibit slower ambient dose rate reduction compared to urban and agricultural areas due to persistent contamination and topographical influences.
• Explainable AI is a valuable tool for analyzing environmental radiation data and validating decontamination effects.
• The proposed geospatial approach supports informed policymaking for regional recovery and forest management post-accident.