Accurate surface soil moisture retrieval is essential for precision agriculture, yet challenging in broad-leaved crops like sunflowers due to canopy attenuation. This study proposes an explainable stacked ensemble learning framework using Sentinel-1 SAR and Sentinel-2 optical data for surface SM retrieval in Zile, Türkiye. To address the common issue of spatial data leakage in agricultural machine learning, a strict parcel-based GroupKFold cross-validation strategy was implemented. An ensemble of Random Forest, LightGBM and XGBoost, optimized via Optuna, achieved the best performance (R2=0.711, RMSE =8.447%). Crucially, SHAP interpretability analysis revealed that temporal features acted as the dominant predictors over SAR backscatter, reflecting the strong continental drying cycle. These findings emphasize the importance of rigorous spatial validation and interpretability when applying machine learning to regional crop monitoring.