This study proposes a novel hybrid modeling framework that integrates satellite-derived data with Group Method of Data Handling (GMDH) models to enhance groundwater level prediction in arid and data-scarce regions, specifically the Kahoorestan Plain in southern Iran. The objectives were: (1) to develop new models suitable for prediction in sparse data regions, (2) to analyze different optimization methods (Genetic Algorithm and Harmony Search) for improving GMDH performance, and (3) to provide trend analysis of groundwater depletion over time and space. Input data included GRACE-JPL gravity data, potential evapotranspiration, TRMM precipitation, temperature, and soil moisture from satellites. Three model variants were applied: standard GMDH, GMDH-GA, and GMDH-HS. The GMDH-HS model outperformed others with 27–39% lower mean error than standard GMDH, achieving R² values between 0.82 and 0.94 across 15 observation wells. Decadal forecasts (10-year projections) indicated critical groundwater declines in the eastern region, with an anticipated drop below surface level to -1.851 meters, signaling severe aquifer stress. These findings are essential for developing tailored intervention plans, including zoned extraction policies and early warning systems, for conservation efforts in at-risk areas.