Groundwater Potential Mapping in the Chirchik Lowland (Tashkent Region, Uzbekistan) Using Metaheuristic-Tuned Machine-Learning and Deep-Learning Models

Groundwater screening is important in water-stressed Central Asian lowlands, where field investigations are costly, and model rankings can depend on validation choices.This study evaluated relative groundwater potential in the Chirchik lowland groundwater-monitoring district of northeastern Uzbekistan using 129 productive wells, 129 background points, and 13 geological, land-surface, climatic, and terrain factors.Five model-optimizer configurations were compared: convolutional neural network-long shortterm memory (CNN-LSTM) with Tabu Search (TS), Super Learner with the Equilibrium Optimizer (EO), kernel logistic regression (KLR) with Simulated Annealing (SA), balanced random forest (BRF) with the Whale Optimization Algorithm (WOA), and frequency ratio (FR) with the Firefly Algorithm (FA).Because each reported pairing was selected using the same 78-point holdout set used for reporting, the estimates are exploratory.Holdout area under the curve (AUC) was 0.87 for CNN-LSTM, 0.82 for KLR, Super Learner, and BRF, and 0.80 for FR.Spatial-block AUC was 0.73, 0.76, 0.79, 0.77, and 0.75, respectively, reversing the nominal ranking.Fifteen predictor reorderings reduced CNN-LSTM AUC to 0.67 ± 0.07.Cross-model integrated-gradient attribution assigned the largest mean shares to geology and land cover, although modelspecific profiles differed substantially.Percentile-ranked maps placed the upper classes in the centralsouthwestern alluvial lowland and northeastward river corridor, whereas eastern and northeastern mountains were predominantly low or very low.The results provide a regional screening basis for groundwater development planning and management in the Tashkent region, subject to field verification.External wells, nested validation, and probability calibration remain necessary before operational use.

Hali tarjima qilinmagan