Designing advanced feature selection and uncertainty quantification-based deep learning approach to predict chlorophyll-a and water bloom risks in dam reservoir

Designing advanced feature selection and uncertainty quantification-based deep learning approach to predict chlorophyll-a and water bloom risks in dam reservoir

Predicting water quality indicators accurately is vital for the sustainable management of aquatic ecosystems, particularly in dam reservoirs that are highly vulnerable to environmental phenomena. Dissolved oxygen (DO) and chlorophyll-a (Chl-a) are essential indicators for evaluating ecosystem stability and water quality. In this study, an innovative and robust intelligent framework is designed using integrated uncertainty quantification and feature selection to predict DO, Chl-a, and bloom risk evaluation of dams. First, the individual machine learning and deep learning models, including Extreme Gradient Boosting (XGBoost), Convolutional Neural Networks (CNNs), Gated Recurrent Units (GRUs), Least Square Support Vector Regression (LSSVR), and Multi-Layer Perceptron (MLP) were assessed. Subsequently, the most effective models are then integrated to enhance predictive accuracy. The Boruta Feature Selection Approach (BFSA), Gamma Test, and Shapley Additive Explanations (SHAP) are used to select the most suitable and relevant features. Then the Monte-Carlo simulation is implemented for uncertainty analysis to evaluate the reliability of models' prediction by determining probability distribution functions. The hybrid XGBoost-CNNs achieved the highest performance in terms of R2 = 0.923, RMSE = 0.547 μg/l for Chl-a prediction, and CNNs obtained R2 = 0.995, RMSE = 0.143 ppm for DO prediction. The 95 % Prediction Uncertainty (95PPU) varied from 79.37 to 100, which shows strong predictive reliability. Also, d-factor values lower than 0.77 confirmed the model uncertainty is low. Furthermore, water bloom risk was assessed using the predicted Chl-a concentration. The analysis indicated no risk levels at reservoir depths of 0–5.5 m and 13.5–32 m, while low-risk levels were identified between 5.5 and 13.5 m. The maximum risk probability was 20.66 % when Chl-a concentrations were below 40 μg/l. The results highlight the effectiveness of hybrid artificial intelligence frameworks in enabling real-time water quality monitoring, early detection of harmful algal blooms, and promoting sustainable reservoir management.