<p>Precise agricultural yield prediction is crucial for food security and economic planning. However, existing models face challenges due to the complexity and high dimensionality of agricultural data. In this study, we developed a model capable of analyzing feature importance. We then refined the dataset by removing negatively contributing features, which, in many cases, enhanced prediction accuracy. This effect is particularly evident in our limited dataset, though it may be less significant with larger datasets. To achieve feature reduction, we incorporate Shapley Additive exPlanations (SHAP), which enhances model interpretability by identifying and eliminating features that negatively impact predictions. Our study, based on a multivariate dataset from rice fields in Vietnam, integrates optical data from Landsat and radar data from Sentinel-1. Analyzing key radar polarization bands (VV and VH) reveals their strong influence on yield prediction. By removing 15–20 non-contributory features, our approach improves prediction accuracy by 2–3% while significantly reducing computational costs and training time. This efficiency gain ensures scalability and practical applicability, particularly in resource-limited environments. Our findings underscore the potential of XAI-driven feature selection in agricultural modeling, offering a robust, interpretable, and computationally efficient solution for yield forecasting.</p>