A dynamic approach for corn yield prediction to ensure agricultural resilience in the U.S. Midwest

Accurate prediction of crop yield is essential for food security, with regional weather playing a key role in agricultural productivity. Machine learning ( ML ) and deep learning ( DL ) models have proven effective in capturing complex patterns and delivering high-precision yield estimates. To improve their generalizability, these data-driven models often require retraining with site-specific information. This study builds on our previous ML -based methodologies by integrating spatial effects through a Geo-Spatial Module ( GSM ), enhancing model applicability across the U.S. Midwest corn belt. The GSM enables the model to leverage spatial relationships among counties, allowing it to make informed predictions even in regions with data-scarce area. We developed a robust and adaptive neural network model designed to forecast crop yields through dynamic data integration. Using 14 years of daily weather and yield data from over 1 , 000 counties, the model was trained to capture both temporal and spatial variability. To optimize performance, we employed Bayesian Optimization to fine-tune the model architecture, improving its ability to learn intricate patterns and deliver accurate predictions. This integrated framework—combining neural networks, spatial modeling via GSM , and efficient Bayesian Optimization —reflects our commitment to developing scalable, precise, and context-aware tools for modern agriculture. • Spatially distributed corn yield prediction. • Adaptive neural network-based model. • Geo-spatial module. • US Midwest corn belt.