AUTHOR=Elmotawakkil Abdessamad , Moumane Adil , Zahi Assia , Sadiki Abdelkhalik , Karkouri Jamal Al , Batchi Mouhcine , Bhagat Suraj Kumar , Tiyasha Tiyasha , Enneya Nourddine 

TITLE=Artificial intelligence for groundwater recharge prediction in an arid region: application of tabular deep learning models in the Feija Basin, Morocco

JOURNAL=Frontiers in Remote Sensing

VOLUME=Volume 6 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/remote-sensing/articles/10.3389/frsen.2025.1622360

DOI=10.3389/frsen.2025.1622360

ISSN=2673-6187

ABSTRACT=IntroductionGroundwater recharge mapping is crucial for sustainable water resource management in arid and semi-arid regions, particularly in hydro-climatically stressed areas such as the Feija Basin in southeastern Morocco. Characterized by shallow aquifers, irregular rainfall, and over-extraction for agriculture, this region faces increasing groundwater depletion. Recent extreme rainfall events during the 2024–2025 season have highlighted both the vulnerability and opportunity for recharge, emphasizing the need for data-driven, proactive strategies.MethodsThis study introduces a GeoAI-based framework combining remote sensing, geospatial analysis, and advanced artificial intelligence (AI) models to predict optimal groundwater recharge zones. Ten conditioning factors (e.g., elevation, slope, topographic wetness index, NDVI, rainfall, soil permeability, geomorphology) were used to construct the input dataset. Five AI models TabNet, TabTransformer, Multilayer Perceptron (MLP), CatBoost, and AdaBoost were trained and optimized using grid search and particle swarm optimization (PSO). Performance was evaluated using accuracy, AUC-ROC, Cohen’s Kappa, and feature importance. Spatial validation was conducted using in-situ borehole data.ResultsAmong the tested models, TabNet achieved the highest performance (accuracy = 97.8%, AUC = 0.99), followed closely by TabTransformer (accuracy = 97.6%). Both models demonstrated strong generalization and produced spatially coherent recharge maps. Predicted optimal zones corresponded with low-lying, vegetated, and permeable areas, aligning with known hydrogeological features.DiscussionThis study presents a novel application of tabular deep learning models in groundwater science, enhancing the precision and interpretability of recharge zone mapping. The results provide actionable insights for water resource planners, especially in light of recent anomalous hydrological events. The proposed framework supports the development of rainwater harvesting and artificial recharge systems to ensure long-term groundwater sustainability in climate-sensitive areas.