Predicting River Eutrophication by Integrating Interpretable Machine Learning and the PLUS Model in the Chaohu Lake Basin, China

Investigating the influence of landscape evolution on river eutrophication is critical for optimizing spatial patterns to improve water quality. Machine learning (ML) models can capture the complex relationship between landscape metrics and water quality, but their black-box property restricts the interpretability of the underlying mechanisms and makes it difficult to forecast future trends in water quality. To address this, we developed a novel framework that, for the first time, couples an interpretable ML model with the Patch-generating Land Use Simulation (PLUS) model for eutrophication index (EI) prediction. This approach elucidates the response of river eutrophication to landscape dynamics and forecasts future river EI trends. The random forest regression (RFR) model outperformed other algorithms in quantifying these relationships (R2 = 0.934 for training, 0.711 for testing). SHAP analysis revealed that landscape metrics contributed 81.78% to the river EI, far exceeding climate factors (18.22%). Consequently, landscape evolution emerged as the dominant explanatory factor. Scenario simulations indicated that while the ecological protection (EP) scenario effectively mitigates river eutrophication, the urban development (UD) scenario significantly exacerbates it. Specifically, under the UD scenario, the average EI in urban sub-watersheds is projected to reach 60.78 by 2040, approaching heavy eutrophic levels. Our findings inform spatial optimization strategies for river eutrophication management and facilitate the design of targeted, localized water ecological protection policies in subtropical monsoonal basins.