Comparative study of hydrochemical and machine learning methods for Fe anomaly identification and NBLs estimation in shallow groundwater of representative hydrogeological units in the North China plain

Iron (Fe) in groundwater results from both natural sedimentation and human activities, impacting biogeochemical cycles and the migration of various components. Increased human activities have disrupted Fe concentrations, causing deviations from the natural state. Thus, identifying Fe anomalies and determining its natural background levels (NBLs) are crucial. In this study, based on subdivided NBLs units, four typical units were selected for anomaly identification using Iterative 2-Sigma method, modified hydrochemical method (MI-OPT), and Isolation Forest model. The results showed that the MI-OPT method showed stable performance and also identified anomalies related to hydrochemical indicators. The Isolation Forest model efficiently detected Fe anomalies through a machine learning-based partitioning approach. In contrast, the Iterative 2-Sigma method exhibited instability due to its dependence on data distribution. Based on the distribution characteristics of the remaining data, the anomaly identification results from the MI-OPT method were selected as the final reference for determining the NBLs of Fe in the four units, with the upper limits being 0.45 mg/L, 2.80 mg/L, 2.58 mg/L, and 1.59 mg/L, progressively transitioning from the recharge area to the runoff and discharge areas. Additionally, an integrated analysis incorporating information entropy, PPI values, and environmental pollution source data explained most of the detected anomalies, validating the reliability of the anomaly identification results. The methods and results presented in this study offer a new perspective on Fe anomaly identification in groundwater.