A Machine Learning Framework for Predicting Anomalous Radionuclide Transfer in Phosphate-Rich Semi-Arid Soils

This study developed a machine learning framework for predicting anomalous 232Th transfer in phosphate-rich semi-arid soils using gamma-ray spectrometry data of 24 soil and 26 crop samples (beans, maize, pepper) collected from eight Local Government Areas (LGAs) of Sokoto and Kebbi States. The five algorithms evaluated are as follows: Linear Regression, Ridge, Lasso, Random Forest and Gradient Boosting of which the linear regression model exhibited an excellent performance of R2 = 0.997, RMSE = 0.018 and MAE = 0.015 and significantly outperformed tree-based ensemble methods. Feature importance analysis revealed that soil thorium concentration was the major predictor (31.8%), followed by crop thorium concentration (7.7%) and thorium-to-radium ratio (7.1%). Model robustness was confirmed by bootstrap confidence intervals (R2 95% CI: 0.992-0.999). Logistic regression model achieved 100% classification accuracy for anomalous samples, though this result requires cautious interpretation given the limited sample size. These computationally simple, low-cost models provide a scalable tool to screen agricultural systems at risk of high radionuclide transfer. However, the framework’s broader applicability is constrained by the absence of independent validation datasets, and further external validation is necessary before operational deployment. Regulatory authorities may use these models for preliminary screening, but site-specific measurements remain essential for confirmatory assessments in resource-limited environments.