Predicting groundwater depth is crucial for sustainable water management in agricultural regions facing intensive irrigation and rapid urbanization. Traditional data-driven models lack physical interpretability, while physical models like SWAT+ often struggle with non-stationary inputs, limiting their accuracy. This study proposes a novel Physics-Dominated Signal Decomposition Optimization (PD-SDO) framework that integrates SWAT+ with Extreme-point Symmetric Mode Decomposition (ESMD) to integrate physical mechanisms with data optimization. The framework is specifically designed for intensively managed mixed land-use regions, exemplified by the study area—a national strategic zone for agricultural development. Key innovations include: 1) ESMD-optimized inputs that extract high-frequency components (the first intrinsic mode function (Intrinsic Mode Function 1, IMF1), representing the highest-frequency component of the decomposed precipitation series) from precipitation to capture extreme rainfall pulses (>50 mm/d), thereby enhancing SWAT+'s adaptability to non-stationary climatic forcings prevalent in monsoon-influenced areas; 2) sensitivity-guided analysis identifying ALPHA_BF as the dominant parameter controlling groundwater recharge, which, combined with spatial heterogeneity analysis, reveals dual-mode drainage dynamics in loess-dominated unsaturated zones; 3) A closed-loop optimization where residual analysis dynamically adjusts ESMD thresholds, enabling model self-optimization; 4) Quantified land-use impacts showing contrasting hydrological effects under afforestation and urbanization scenarios, providing direct insights for spatial planning. The framework significantly improves simulation accuracy for groundwater depth, achieving a Nash-Sutcliffe Efficiency (NSE) of 0.94 and a Mean Absolute Error (MAE) of 8.02 m, which represents a 23% gain over the standalone SWAT+. It also enhances computational efficiency by 40% with minimal accuracy loss and provides 2.3× greater physical interpretability than pure black-box models. PD-SDO offers a replicable solution for high-fidelity groundwater modeling in rapidly changing mixed land-use zones, effectively balancing mechanistic fidelity with predictive performance. The framework and its findings provide a scientific basis for optimizing water allocation and land use management within strategic agricultural regions.