Bi-LSTM and partial mutual information selection-based forecasting groundwater salinization levels

Abstract

Fresh-saline groundwater is currently distributed in a highly heterogeneous way throughout the world. Groundwater salinization is a serious environmental issue that harms ecosystems and public health in coastal regions worldwide. Because of the complexities of groundwater salinization processes and the variables that influence them, it is still challenging to predict groundwater salinity concentrations precisely. This study compares cutting-edge machine learning (ML) algorithms for predicting groundwater salinity and identifying contributing factors. This study employs bi-directional long short-term memory (BiLSTM) to indicate the salinity of groundwater. The input variable selection problem has recently attracted attention in the time series modeling community because it has been shown that information-theoretic input variable selection algorithms provide a more accurate representation of the modeled process than linear alternatives. To generate a variety of sample combinations for training multiple BiLSTM models, the PMIS-selected predictors are used, and the predicted values from various BiLSTM models are also used to calculate the degree of prediction uncertainty for groundwater levels. The findings give policymakers insights for recommending groundwater salinity remediation and management strategies in the context of excessive groundwater exploitation in coastal lowland regions. To ensure sustainable groundwater management in coastal areas, it is essential to recognize the significant impact of human-caused factors on groundwater salinization.