Groundwater level prediction using deep learning algorithms in a drought-prone area in Bangladesh.

Groundwater is a crucial resource for meeting daily needs such as drinking water and agricultural irrigation. The northwestern region of Bangladesh experiences significant water scarcity due to excessive use and the absence of accurate forecasting methods in this drought-prone area. Understanding groundwater level (GWL) forecasting is essential for efficient utilization and sustainable management. This study aimed to perform long-term time series forecasting of GWL using three innovative deep learning (DL) algorithms: Convolutional Neural Networks (CNNs), Gated Recurrent Unit (GRU), and Bidirectional Long-Short-Term Memory (BiLSTM) in the drought-prone northwestern region of Bangladesh. The primary objective was to improve the GWL prediction accuracy and support water resource management. Environmental data including maximum and minimum temperatures, total rainfall, and humidity, spanning from 1981 to 2017, were used. Monthly data from two climate stations-Dinajpur and Bogura-were collected, and the dataset was divided into a 70% training phase (1981-2006) and a 30% testing phase (2007-2017). Performance evaluation of these DL algorithms was conducted using six metrics such as root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), index of agreement (d), mean absolute percentage error (MAPE), and R2 score. The GRU model showed the best performance in Dinajpur during training, while BiLSTM performed best in Bogura. Among all tested models, BiLSTM demonstrated the highest accuracy during testing. In Dinajpur, it achieved RMSE = 0.415 m, MSE = 0.172 m2, MAE = 0.301 m, and R² = 0.89; in Bogura, it achieved RMSE = 0.311 m, MSE = 0.096 m2, MAE = 0.168 m, and R² = 0.93. These findings can assist policymakers and government authorities in making informed decisions for groundwater management in drought-prone areas.