Estimation of environmental radioactivity levels using machine learning and the creation of a high-resolution radiation map of Turkey: development of an online computational tool

Accurate mapping of the spatial distribution of natural radiation is a fundamental prerequisite for public health and environmental risk management. However, the existing environmental radiation maps of Türkiye have limited spatial resolution and are inadequate for representing local-scale variations. The primary objective of this study is to predict the concentrations of terrestrial primordial radionuclides (226Ra, 232Th, 40K) across Türkiye at a high spatial resolution of 300x300 m² using modern data science techniques, to generate a national radiation map based on these predictions, and to present the results on an interactive online platform. A multi-step hybrid methodology was implemented to achieve this goal. Firstly, the Ambient Gamma Dose Rate (AGDR) and its associated spatial uncertainty were modeled using data from the Radiation Monitoring and Warning System Network (RADISA) with Sequential Gaussian Simulation (SGS). Secondly, the Cosmic Gamma Dose Rate (CGDR) was predicted for the entire country via a machine learning surrogate model that emulates the computationally expensive EXPACS (EXcel-based Program for calculating Atmospheric Cosmic-ray Spectrum) physical model. Finally, these spatial layers were combined with geographic variables (latitude, longitude, altitude) and data digitized from the Radioactivity Atlas of Türkiye to develop the final radionuclide prediction models. During the model development process, the Extreme Gradient Boosting (XGBoost) algorithm was identified as the top-performing model. The study resulted in the production of Türkiye's first high-resolution maps for Ra-226, Th-232, K-40, and the derived Terrestrial Gamma Dose Rate (TGDR). It was observed that the spatial patterns in the maps show strong agreement with the country's known uranium/thorium deposits and major geological formations. The model's validity was also tested against an external, point-based measurement dataset not used during the training process, where it demonstrated significant accuracy at both district and point scales.