Hybrid deep learning and remote sensing for the delineation of artificial groundwater recharge zones

Abstract

The increase in water demand and the scarcity of fresh water in arid regions have contributed to the depletion of groundwater. Artificial Groundwater Recharge (AGR) is an advanced strategy that contributes to combating water shortage issues. Limited efforts have been exerted to evaluate and demarcate AGR potential zones in the United Arab Emirates (UAE). The current study aims to delineate AGR potential zone mapping using the traditional analytical hierarchy process (AHP) and a hybrid deep learning model namely, Convolutional Neural Network-Xtreme Gradient Boosting (CNN-XGB) was used for the optimal prediction-based suitability assessment. A total of nine hydrogeological factors were considered for AGR mapping. First, the influence of each parameter was determined based on expert opinion and literature reviews for the AHP approach (0.007 consistency ratio). Second, a hybrid CNN-XGB model (90.8 % accuracy) predicted the AGR and non-AGR classes as part of binary classification and generated an AGR potential zone map. Moreover, the contributing factors were analyzed deeply for the AGR site selection to understand the intercorrelation, importance, and prediction interaction. Using both approaches, a comparative assessment was conducted in the eastern, central, and western parts of Sharjah. The AGR zone based on the CNN-XGB model achieved a precision of (0.8168), recall (0.7873), and F1-score (0.8018). The critical contributing factors for AGR mapping were found to be geology (20%), geomorphology (15%), rainfall (10%), and groundwater level (10%). The AGR map is expected to help explore new sites with potentially higher favourability to retain water, deal with water scarcity, and improve water management in the UAE.