Predicting the Spatial Distribution of Phosphorus Concentration in Quaternary Sedimentary Aquifers Using Simple Field Parameters

Geogenic phosphorus (P) in groundwater has been found in different regions, posing a risk for surface water eutrophication. However, the prediction of groundwater P distribution is less studied. In this study, three machine learning-based regression models including random forest regression (RFR), support vector regression (SVR), and multiple linear regression (MLR) were applied to predict the spatial distribution of geogenic P in alluvial-lacustrine sedimentary aquifers of the central Yangtze River basin, with readily accessible field parameters, such as pH, Eh, EC, depth, NH4+-N and Fe2+. The results indicate that the RFR model achieves the best fit by three times 10-fold cross-validation with the highest R-2 (0.49) and explanatory variance (0.52), the lowest root mean square error (0.48) and mean absolute error (0.30), producing a groundwater P distribution that is highly consistent with the observed results. According to the prediction results, the areas with high P (> 0.4 mg/L) groundwater and abnormally high P (> 1 mg/L) groundwater account for 55% and 11% of the whole study area in JH-DT-P, respectively. Meanwhile, NH4+-N and Fe2+ are the two most prominent indicating factors of P enrichment in groundwater, and NH4+-N has higher relative importance than Fe2+. The wider validity of the model was suggested by its successful application to two regions in Bangladesh with similar hydrogeological conditions.