<p>Food security is an increasingly pressing challenge that can be tackled through advanced technologies designed to monitor agroecosystems and facilitate timely, data-driven decisions. Integrating innovative technologies is crucial for enhancing agricultural sustainability and resilience, thereby contributing to global food security. This study investigates the importance of the feature combinations, incorporating (M1) vegetation indices combined with weather variables and (M2) M1, along with surface soil moisture (SSM) through implementing machine (support vector machine (SVM) and random forest (RF)) and deep (deep neural network (DNN)) learning models to predict in-season wheat yield. Results indicate that the optimal prediction window spans from the onset of stem elongation to the initiation of flowering (BBCH 30–61), during which the average R², RMSE, MAE, and IOA were 0.76, 484.04 kg ha⁻¹, 281.90 kg ha⁻¹, and 0.92, respectively. At this stage, the DNN achieved the highest performance with M1 (R² = 0.73, RMSE = 502.59 kg ha⁻¹, MAE = 338.51 kg ha⁻¹, IOA = 0.92) and improved further with M2 (R² = 0.85, RMSE = 379.93 kg ha⁻¹, MAE = 244.22 kg ha⁻¹, IOA = 0.96). These metrics improvements statistically enhanced model's performance when trained with M2 (<i>P</i> ≤ 0.05). The feature importance and Shapley Additive Explanations (SHAP) analysis revealed that both the choice of model and dataset influenced feature prioritization. Notably, Near-Infrared Reflectance of vegetation (NIRv), SSM and evapotranspiration (Evap) consistently appeared among the most influential features. These findings underscore the critical role of surface soil moisture in enhancing model accuracy across growth stages and emphasize the importance of selecting a well-balanced feature set rather than relying solely on complex models. A carefully curated set of features boosts model performance, ensuring more reliable yield predictions for various agricultural applications.</p>