A parsimonious model for predicting the NO3−-N concentration in shallow groundwater in intensive agricultural areas using few easily accessible indicators and small datasets based on machine learning

Abstract

Groundwater nitrate poses a health risk to humans, and monitoring groundwater quality is time-consuming and expensive. To improve the efficiency and reduce the costs of monitoring groundwater nitrate, a parsimonious model was developed using easily accessible predictors and small datasets based on random forest (RF), neural network (NN), and support vector machine (SVM) algorithms. Groundwater NO3−-N concentrations of 1196 groundwater samples from intensive agricultural areas along with data on 13 predictors were obtained for the construction of prediction models. R2 and RMSE metrics were used to evaluate the models, with an emphasis on a minimal number of predictors and a low data size. The results showed that even with few readily available predictors (conductivity, EC; chemical nitrogen fertilization rate, NF; and precipitation, PPT) and a small dataset (260–700 observations), the three models achieved good precision, with R2 and RMSE values of 0.70–0.80 and 6.87–8.13, respectively. In these prediction models, EC was the primary predictor acting as an intrinsic factor, NF acted as a loading factor to supplement the blind area of EC for NO3−-N prediction, and PPT was introduced because of its accessibility to improve predictive performance. The final predictors applied in this study link the two currently dominant prediction methods by combining the high predictive accuracy of the intrinsic predictor with the accessibility of the extrinsic predictor, and the results demonstrate the practicability and applicability advantages of this model in groundwater NO3−-N prediction. This model provides technical support for the rapid and low-cost realization of the online monitoring of NO3−-N pollution in shallow groundwater in intensive agricultural areas.