Study region: Northern China's Agro-Pastoral Ecotone (APENC, 2000-2022) was selected as a climatically sensitive transition belt from semi-arid to semi-humid conditions. Study focus: We developed a kernel Temperature-Vegetation Dryness Index (kTVDI) that exploits the kernel Normalized Difference Vegetation Index to overcome spectral saturation, and coupled it with XGBoost-SHAP and Generalized Additive Models to decode nonlinear drought drivers. Validated against an independent soil moisture dataset, kTVDI demonstrated superior sensitivity (rho = -0.604 vs. -0.533 for TVDI) and stronger agreement with solar-induced chlorophyll fluorescence (SIF), supporting its use as an improved ecological drought indicator. Over 2000-2022 regional drought declined significantly (-1.328 x 10(-)(3) yr(-)(1), p < 0.01), but spatial heterogeneity persisted: south-eastern alleviation vs. north-western intensification. New hydrological insights for the region: Temperature universally dominates (32.92-52.24 % SHAP importance), yet land-use/land-cover restructures secondary controls: vegetated systems exhibit evapotranspiration-regulated water balance, whereas deserts display precipitation-limited dynamics amplified by human disturbance. Critical thresholds marking energy-water regime shifts are 6.6 degrees C for temperature, 387 mm yr(-)(1) for evapotranspiration and 54.2 % for relative humidity. Interaction analysis further reveals land-use-specific synergies (humidity-soil in croplands, soil-topography in forests, topography-precipitation-human activity in deserts), underscoring ecosystem-dependent vulnerability. This interpretable framework advances process-based drought understanding in ecotones and furnishes quantitative thresholds for adaptive management under global change.