Evaluating the reliability of data interpolation and machine learning methods for water quality management: a SWAT model comparison

Abstract

Due to data scarcity and the time-consuming nature of process-based modeling, SWAT often faces challenges in its application. This study evaluates the reliability of simple spatial interpolation using monitoring data, combined with advanced machine learning techniques, including Self-Organizing Maps (SOM), Generalized Additive Models (GAMs), Receiver Operating Characteristic (ROC) analysis, and K-means clustering, to identify critical source areas (CSAs), key stressors, and thresholds for water quality management. Similar to SWAT-based analyses, the study found that forest cover and human-modified land use significantly affect total nitrogen (TN) and total phosphorus (TP) levels, while also revealing population density as an additional influential factor. GAMs showed that human-disturbed land use drives TN pollution, and population density is key to TP enrichment. ROC analysis identified thresholds of 40.91% for forest cover (close to SWAT results) and 10.21% for human-disturbed areas, which is lower than SWAT-based estimates. A population threshold of 239 significantly impacted TP, a factor not identified by SWAT modeling. K-means clustering highlighted clusters 1, 4, and 5 as high-priority areas, and SWAT modeling indicated that managing these clusters—covering 47.39% of the watershed—could mitigate 42.66% of TN and 41.34% of TP. While this approach cannot fully replace SWAT modeling, but simple and time saving, it proves to be helpful for identifying CSAs and informing water quality management strategies.