A HYBRID DEEP LEARNING APPROACH FOR GROUNDWATER LEVEL FORECASTING USING MULTIVARIATE GEOSPATIAL AND METEOROLOGICAL DATA

Abstract

Groundwater level prediction is a major challenge due to the complex and nonlinear behavior of hydrological systems, particularly in regions such as Chhattisgarh, where groundwater is a primary source of irrigation and domestic supply. In this study, a hybrid Particle Swarm Optimization-Long Short-Term Memory (PSO-LSTM) model was developed for groundwater level forecasting and compared with traditional Artificial Neural Network (ANN) and standard LSTM models. The multivariate dataset included geographical coordinates, elevation, soil type, meteorological variables (precipitation, temperature, diurnal temperature range, and vapour pressure), and historical groundwater level observations collected from five major divisions of Chhattisgarh. Model performance was evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), coefficient of determination (R²), and Cosine Similarity (CS). The hybrid PSO-LSTM model achieved superior performance with testing values of MAE = 0.676, MSE = 0.678, RMSE = 0.824, R² = 0.876, and CS = 0.989, outperforming both ANN and standard LSTM models. The proposed model reduced RMSE by approximately 31% compared to ANN and 16% compared to standard LSTM, demonstrating its effectiveness in capturing nonlinear relationships and long-term temporal dependencies in groundwater systems. The empirical results of the proposed framework can be extended to other regions using relevant multivariate datasets, making it a reliable tool for groundwater management and planning. Accurate groundwater level prediction can support water harvesting, agriculture, industrialization, and sustainable water resource management. Therefore, these findings can significantly contribute to predicting the groundwater level effectively.