An advanced geochemical assessment of the Djeffara shallow aquifer using stable isotopes, GIS based tools, and deep learning algorithms

Abstract

The Djeffara shallow aquifer, a vital water resource in southern Tunisia, is facing increasing pressure due to overexploitation and climate change. This study aims to investigate the geochemical processes within this aquifer system by combining hydrochemical, statistical and stable isotopes techniques, interpolation methods and deep learning approach. Supervised and unsupervised statistical techniques were employed to optimize the prediction model and assess its performance. t-Distributed Stochastic Neighbor Embedding was also used as an innovative dimensionality reduction. Stable isotope analysis was conducted to determine the origin and mixing processes of groundwater. Integration of stable isotope data, GIS based interpolation methods and deep learning modeling provided a comprehensive understanding of the geochemical processes in the Djeffara aquifer. Our results indicated that the t-SNE plot revealed a clear grouping of three distinct clusters, confirming the results of cluster analysis and principal component analysis. Deep neural network model was developed to predict groundwater salinity. The relationship between measured and calculated salinity showed a strong linear correlation with a high coefficient (0.987) indicating a strong linear association between predicted and actual values. This was supported by low RMSE (1776.189) and MAPE (19.411), suggesting that the model’s predictions are generally close to the observed values. The deep neutral network model used here with a structure of (11:5:4:1). Stables isotopes analyses indicated that groundwater in the Djeffara plain is governed by evaporation, dissolution and mixing with paleo-waters. This knowledge can inform sustainable water management strategies, including optimal water allocation, artificial recharge, and salinity prediction based on available data.