Integrating Multivariate Analysis and Explainable Machine Learning to Assess Eutrophication: Chlorophyll-a and Dissolved Oxygen Dynamics in Ziway and Langano Lakes, Ethiopia

Abstract Eutrophication, driven by rising chlorophyll-a concentrations and declining dissolved oxygen (DO), is an escalating threat to lake ecosystems worldwide, and multivariate statistical analysis combined with explainable machine learning offers a powerful framework for characterizing its dynamics and drivers to support evidence-based watershed management. The Ethiopian Rift Valley has experienced substantial anthropogenic pressure and climate variability, increasingly affecting Lake Ziway (450 km², mean depth 5 m) and Lake Langano (230–250 km², depth 46–55 m). This study assessed spatio-temporal dynamics of chlorophyll-a and dissolved oxygen (DO) at both lakes using Principal Component Analysis (PCA), Cluster Analysis, Mann-Kendall trend analysis, Random Forest (RF) machine learning, and SHAP-based Explainable Artificial Intelligence (XAI), with turbidity and total suspended solids (TSS) included as supporting covariates. Mean chlorophyll-a concentrations were 15.09–29.09 µg/L (Langano) and 19.63–37.83 µg/L (Ziway), indicating high phytoplankton productivity in both lakes, with pronounced increases during the wet season. Dissolved oxygen showed statistically significant declining trends (Tau = -0.526, Langano; -0.595, Ziway), while chlorophyll-a increased significantly (Sen’s slope: +0.963 and +1.249 µg/L, respectively). DO explained 30–68% of Random Forest model variability and was consistently among the top two predictors of chlorophyll-a at both lakes. RF models achieved high predictive accuracy (R² = 0.968–0.986; RMSE = 1.230–1.435). Both lakes were classified as mesotrophic during the dry season, shifting to eutrophic during the wet season, with Lake Ziway showing greater ecological pressure and variability than Lake Langano. Targeted watershed management interventions to curb nutrient loading and protect oxygen balance are recommended.