Improving groundwater level prediction based on optimized deep learning model with Coati algorithm

Accurate groundwater level (GWL) estimation is of great significance to attain integrated water resources management and sustainable development goals. Recently, in the field of artificial intelligence (AI), nature-inspired optimization has been implemented in hydrology for training (calibration) of both statistical and process-based models. The presented study investigates the predictive performance of a hybrid deep learning technique, namely Convolutional Neural Network (CNN), with Coati Optimization Algorithm (COA) (CNN-COA hybridized model) in GWL modeling. Outcomes of CNN-COA are compared with improved CNN-HHO (Harris Hawks Optimization), Support Vector Regression (SVR) hybridized with COA and HHO (SVR-COA, SVR-HHO), and standalone CNN and SVR for GWL estimation. Evaluation statistics such as Nash-Sutcliffe efficiency (NSE), determination coefficient (R²), Root Mean Squared Error (RMSE), and mean absolute error (MAE) were used to assess the model's accuracy. The results show that the CNN-COA achieves greater accuracy than the other methods due to its effective exploration of the search space, avoiding local minima. In particular, among the three algorithms, the CNN-COA offered the best accuracy for all observation wells, with best results in Patyale Chak (NSE: 0.9932, MAE: 0.0468, R²: 0.9953, and RMSE: 0.0978) and worst in Badsoo (NSE: 0.9923, MAE: 1.0051, R²: 0.9945, and RMSE: 1.0517). Outcomes demonstrate substantial enhancement in prediction accuracy, highlighting the efficiency of the proposed methodology for GWL prediction.