Abstract

Freshwater is essential for sustaining ecosystems and economic activities, and water quality largely determines its potential uses. Among water quality parameters, turbidity is particularly important due to its direct influence on aquatic ecosystem health, water resource management, and human well-being. This study investigates the long-term water quality and quantity dynamics of Lake Chapo, southern Chile, with the aim of assessing its evolution over the past 20 years. Thirteen machine learning models were evaluated to estimate turbidity under three scenarios: (1) using in situ limnological variables, (2) using meteorological variables, and (3) using the combined dataset. Overall, all models exhibited strong predictive performance. The best results were obtained with the multilayer perceptron model (R-2 = 0.78, MAE = 0.69, RMSE = 0.95), the Gaussian process model (R-2 = 0.75, RMSE = 0.78, MAE = 0.97 across all cases), and the random forest model (R-2 = 0.75, RMSE = 0.81, MAE = 0.96 in Case 1). Correlation analyses revealed that precipitation and wind speed exert the strongest meteorological influence on turbidity variability. These findings demonstrate that combining machine learning techniques with environmental data enhances understanding of aquatic ecosystem dynamics and supports improved freshwater resource management.