Ensuring national food security in the United States necessitates prompt and precise crop output forecasts to guide policy and resource distribution. This study introduces a predictive yield modeling system that amalgamates real-time agricultural intelligence derived from multispectral satellite imaging, meteorological data, soil properties, and historical yield data. The approach utilizes self-supervised foundation models and temporal deep learning architectures, such as long short-term memory (LSTM) networks, to capture spatial and temporal variability in crop yields. Preprocessing, feature fusion, and data harmonization facilitate robust modeling across diverse datasets. Assessment across principal U.S. agricultural regions indicates that LSTM models regularly surpass baseline linear and conventional machine learning models, attaining reduced RMSE and MAE while preserving elevated R². Spatial error mapping and decision-support outputs identify regions of increased uncertainty and facilitate actionable actions, such as early warning systems, targeted subsidies, and regional priorities for resource distribution. The methodology directly correlates predictive outputs with policy-relevant variables, offering a practical instrument for improving national food security and modernizing agricultural decision-making in the U.S.