Evaluating Conventional and Machine Learning Models for Soil Moisture Effects on Radon Exhalation and Indoor Radon Levels

This study compares the performance of machine-learning linear regression and random forest models with the conventional correlation analysis in the prediction of the influence of soil moisture on radon exhalation and indoor radon levels.

Radon exhalation rates from the soil were experimentally determined using the sealed-can method. Soil moisture content was estimated using their wet and dry masses. Conventional correlation analysis was conducted to assess the relationships between moisture content and radon parameters. Linear regression and random forest machine learning models were applied to evaluate their predictive performance.

Conventional correlation revealed a strong negative association between soil moisture and radon exhalation (R=-0.82), and a weaker association with indoor radon (R=-0.30). The linear regression analysis showed limited predictive capacity for moisture and radon exhalation rate, with a training correlation of 0.42, and a negative testing coefficient of -16.0. The random forest showed higher values of 0.65 and -5.52 for the training correlation and testing coefficient, respectively, indicating poor overfitting potential. Between moisture content and indoor radon, the linear regression yielded a training correlation of 0.42 with a -2.17 testing coefficient, while the random forest returned 0.65 and -1.22, respectively.

The results confirm that soil moisture influences radon exhalation. However, both models exhibited weak predictive performanc and, poor generalization,highlighting the complexity of radon-moisture interactions.

This work re-emphasizes the need for improvement in predictive models, such as the use of non-linear algorithms, consideration of additional environmental factors, and enhanced validation strategies to improve accuracy in predictive correlation studies on radon.