Explainable Deep Learning (DL) and GIS-Based Spatial Assessment of Urban River Water Quality

Rivers are essential freshwater resources; however, rapid urbanization, industrial discharge, and untreated municipal wastewater have significantly degraded river water quality in densely populated cities such as Dhaka, Bangladesh, threatening environmental sustainability and public health. This study aims to develop an interpretable Deep Learning (DL) and Geographic Information Systems (GIS) based framework for predicting the Water Quality Index (WQI) and identifying the dominant physicochemical and heavy metal parameters influencing water quality in the Turag, Balu, Shitalakshya, and Buriganga Rivers located in Dhaka, Bangladesh. A total of 7,616 datasets collected during the 2020–2024 monsoon season were analyzed, including eight physicochemical and eight heavy metal parameters. The Interquartile Range (IQR) method was applied for outlier detection, while normalization techniques improved data quality and predictive performance. Three DL models, namely Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (Bi-LSTM), were developed, and GridSearchCV was employed for hyperparameter optimization. Model performance was evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). The developed ANN, LSTM, and Bi-LSTM models achieved R² values of 0.96, 0.99, and 0.65, respectively, with LSTM providing the best predictive performance. SHapley Additive exPlanations (SHAP) analysis identified Dissolved Oxygen (DO), turbidity, transparency, and Chemical Oxygen Demand (COD) as the most influential variables governing WQI prediction, while GIS-based Inverse Distance Weighting (IDW) interpolation revealed significant spatial pollution hotspots across the river systems. Although the framework demonstrated strong predictive capability, the study was limited to monsoon-season observations. Future research should incorporate multi-seasonal datasets and real-time monitoring systems to improve long-term prediction and support sustainable river management.