Optimizing water quality monitoring in an arid wetland using interpretable machine learning: From identifying dominant parameters to simplifying the water quality index

• Sustained water quality recovery has been observed, with marked spatial heterogeneity. • The Gradient Boosting Tree model outperformed five other machine learning models in WQI prediction. • SHAP analysis reveals nonlinear dynamics and parameter interactions in water quality systems. • The novel six-parameter index reduces water quality monitoring costs without compromising diagnostic. Arid inland wetlands are critical ecosystems vital for maintaining regional ecological balance. Understanding the spatiotemporal evolution and driving mechanisms of their water quality is essential for conserving these fragile environments. This study focused on the Heihe Wetland in China. Based on long-term monthly water quality monitoring data (2010–2023) from four monitoring sites, we (1) analyzed spatiotemporal variations and long-term trends in water quality; (2) developed ensemble machine learning models to accurately predict the Water Quality Index (WQI); (3) identified dominant parameters and explored their nonlinear driving mechanisms; (4) constructed a simplified, cost-effective Minimum Water Quality Index (WQI min ). The results indicated significant spatial heterogeneity (Kruskal-Wallis test, p < 0.05) and seasonal variation in water quality, except for water temperature (WT, p > 0.05). The Locally Estimated Scatterplot Smoothing (LOESS) regression curve visualized a sustained improving trend in WQI over the study period, which was statistically confirmed as significant by the Mann–Kendall test (p < 0.05). The Gradient Boosting Tree (GBT) model achieved the highest predictive accuracy (R 2 = 0.972, MAE = 0.528, MSE = 0.827). Via SHAP (Shapley Additive Explanations) − based interpretation, we identified NH 3 -N, TP, COD Mn , Cr 6+ , DO, and BOD 5 as the dominant parameters driving water quality variations. Notable nonlinear threshold effects (e.g., NH 3 -N at ∼ 0.25 mg/L) and synergistic interactions (e.g., NH 3 -N and TP, r = 0.59) were detected. Strong agreement with the full WQI (R 2 = 0.78) was shown by the simplified WQI min model incorporating these dominant parameters, significantly reducing monitoring requirements while retaining accuracy. This study demonstrates the effective integration of interpretable machine learning into water quality assessment, providing a practical, scalable framework for monitoring and managing wetlands in arid regions.