The Role of Machine Learning to Predict Potentially Harmful Elements Remobilisation and Bioaccessibility Change During Flooding

Abstract Climate change and urbanisation have exacerbated flooding in many regions. Flooding alters soil chemical properties, impacting the remobilisation and bioaccessibility of potentially harmful elements (e.g. lead), which are toxic to humans and wildlife. The physicochemical processes driving these changes and how to model them remain unclear. This study uses machine learning to quantify flooding’s effect on the remobilisation and bioaccessibility of these elements. Soil samples from two distinct catchments, the Clyde and Humber, were analysed under varying pre-and-post-flooding simulation patterns or regimes. Random forests, XGBoost, and neural networks were trained using soil porewater physicochemical data from the Clyde to predict remobilisation and bioaccessibility changes in the Humber. Our findings identified four phases we term - first lag, critical, subcritical, and second lag - characterising shifts in porewater physicochemistry during flooding. The critical phase (days 2-4) showed the most significant porewater physicochemical changes, with manganese, pH, and iron acting as key drivers in influencing porewater solubility and bioaccessibility. Random forest and XGBoost achieved high accuracy in predicting remobilisation (R2 = > 0.80) and intestinal bioaccessibility (R2 = > 0.80) in the Clyde and generalised well to the Humber with no statistically significant deviations from the laboratory flooding results. Random forests and XGBoost models, trained using soil porewater physicochemical data, particularly during the critical phase, can offer novel ways of accurately estimating potentially harmful element dynamics during flooding.