Modeling Long-term Groundwater Levels By Exploring Deep Bidirectional Long Short-Term Memory using Hydro-climatic Data

Inevitable issues concerning the sustainability of groundwater resources are crucial under the present climatic situation. Therefore, the prevision of groundwater environments may able to reinforce the management system. In this respect present study considered a new method to predict long-term groundwater level framework as an alternative option of expensive physical models. The proposed Bidirectional Long Short-Term Memory (BLSTM) model can efficiently capture Spatio-temporal features from historical data. A highway LSTM network is also introduced within the architecture of the model to optimize the analysis. The relative performance of the proposed BLSTM with the highway LSTM (BHLSTM) network compared with simple BLSTM. Stack size increment of the BHLSTM and BLSTM layers can enhance the learning ability and improve by incorporating straight LSTM at the top of the architecture. The proposed model was applied to predict the groundwater level exemplary of the Varuna River basin for twenty years. The model incorporates the historical annual average of total precipitation, temperature, relative humidity, actual evapotranspiration, and groundwater level data to develop and validate the models. The result shows that the signals are captured reasonably well by a stack of four BHLSTM and straight LSTM models in forecasting groundwater levels. The predicted water level range (0—20 mbgl) has four categories low, medium, high, and very high which eventually, illustrates the water-threatened situation in upcoming years in the study area. It is also recommended to exploring this proposed method for further improvements and extensions towards interpreting spatial features.