A robust and explainable deep learning model based on an LSTM-CNN framework for reliable FDOM prediction in water quality monitoring: Incorporating SHAP analysis for enhanced interpretability

The quality and availability of water play a vital role in sustaining human life, driving economic growth, and maintaining public health and environmental balance. Moreover, the traditional methods of assessing water quality rely on costly and lengthy laboratory testing and statistical evaluations. Given the serious risks posed by poor water quality, there is an urgent need for faster and more economical assessment techniques. Therefore, precise forecasting of water quality parameters stands as a critical tool for enhancing both water resource management and efforts to combat contamination. To bridge these gaps, this study suggested multiple deep learning algorithms (LSTM, CNN, GRU, BiLSTM, BiGRU, and LSTM-CNN) to predict daily fluorescent dissolved organic matter (FDOM) concentrations. For a more accurate assessment, the results of the suggested models were compared with baseline models, including CART and MLR. The analysis incorporated using nine different scenarios from seven water quality parameters including discharge (Q), water temperature (Tw), specific conductivity (SC), dissolved oxygen (DO), pH, turbidity (TU), chlorophyll-a (Chl-a), and also, YY (year), MM (month), and DD (day), from two USGS monitoring stations (14211720 and 14203500) in Oregon, USA. In this study, the models' performance was evaluated using four metrics: root mean squared error (RMSE), Nash-Sutcliffe efficiency (NSE), correlation coefficient (R), and mean absolute error (MAE). Analyzing results from both stations, the hybrid LSTM-CNN model demonstrated superior FDOM prediction accuracy compared to standalone architectures (LSTM, CNN, GRU, BiLSTM, BiGRU). At USGS 14203500, LSTM-CNN achieved RMSE of 2.867 ppb QSE, MAE of 1.641 ppb QSE, and R of 0.965. For USGS 14211720, it yielded RMSE of 1.022 ppb QSE, MAE of 0.631 ppb QSE, and R of 0.989. Based on SHAP results, DO, pH, TU, and Chl-a were identified as the most important parameters for predicting FDOM in each model's performance. Finally, this study demonstrates that the hybrid LSTM-CNN approach effectively predicts FDOM concentrations, making it an efficient tool for water quality monitoring.