Machine learning prediction of absorbed gamma dose rate and radiological hazard assessment at a Nigerian steel dumpsite.

Industrial dumpsites are potential hotspots of environmental radiation exposure, yet predictive assessments of dose distribution remain limited in Nigeria contexts. This study evaluates the use of supervised machine learning techniques to predict absorbed gamma dose rate at the NAK Steel dumpsite in Katsina State, Nigeria, using a dataset of 36 soil sampling locations (n = 36). The predicted absorbed dose rates were subsequently used to estimate annual gonadal dose (AGD) as an organ-specific dose indicator and excess lifetime cancer risk (ELCR) as a radiological risk metric. Four learning algorithms: support vector machine (SVM), linear regression (LR), random forest (RF), and decision tree (DT), were trained and tested under identical preprocessing and validation conditions. Across all performance metrics, RF consistently outperformed the other techniques. It achieved the lowest prediction errors (RMSE = 0.0131; MAE = 0.0101), the highest coefficient of determination (R2 = 0.902), and the strongest hotspot classification performance (accuracy = 0.89; kappa = 0.77; AUC = 0.93). In contrast, LR and SVM tended to smooth spatial variability and underestimate localized anomalies, while DT produced abrupt spatial patterns suggestive of overfitting. Predicted AGD values ranged from 0.03 to 0.27 mSv y⁻1, while the corresponding ELCR estimates reflected the spatial variability in annual effective dose with the highest values occurring near the principal hotspot (S11). These findings demonstrate that random forest provides a reliable framework for predicting absorbed gamma dose rate in heterogeneous industrial environments and for supporting subsequent radiological dose and risk assessment using internationally accepted conversion models. Although the relatively small dataset (n = 36) limits model generalization, the proposed framework provides a practical tool for environmental radiation monitoring and decision-making in data-sparse regions.