Integrating geostatistical methods and neural networks for enhanced transmissivity mapping in the heterogeneous Annaba aquifer system (SE Algeria)

The heterogeneity of the Annaba plain aquifer, located in northeastern Algeria and characterized by a temperate Mediterranean climate with hot and dry summers, coupled with the concentrated distribution of experimental hydraulic parameters, poses a challenge in developing a reliable cartography of transmissivities, which is crucial for a robust numerical simulation model. For this purpose, two geostatistical methods, namely regression kriging (RK) and cokriging (CK), along with multilayer perceptron neural networks (MLP-NN), are employed to assess transmissivity across the aquifer system, utilizing readily available parameters such as transverse resistance and specific capacity. The best prediction method is determined based on the highest determination coefficient (R-2) and the lowest mean absolute errors (MAE) during the testing phase. The results of the generated map are used as inputs for the numerical model (Modflow). The dataset used for modeling, which was randomly split into training (48%), validation (27%), and testing (25%), comprises 47 transmissivity values determined from the interpretation of pump test data, 127 transverse resistance values obtained from a geophysical survey campaign, and 70 specific discharge values resulting from the interpretation of step-drawdown tests. The results indicates that the MLP-NN is the best prediction technique of transmissivities of the aquifer system with highest R2 (0.84) and lowest MAE (0.16). The integration of MLP-NN error prediction results into the numerical model leads to better reproduction of aquifer system heterogeneities, avoiding redundant simulations during model calibration. This research demonstrates that the best results are obtained by combining the deterministic approach of the numerical model with the probabilistic approach (MAE = 0.13 and R2 = 0.86), particularly the MLP-NN, which directly incorporates the spatial variability of all variables. This combined approach streamlines calibration and minimizes redundant simulations.