Data Science-Driven Agricultural Yield Prediction Using a Transformer-Based Ensemble Model

Abstract

Accurate prediction of crop yields is vital for addressing global food security challenges amidst climate variability and increasing agricultural demands. This study introduces a novel data sciencedriven approach for maize yield prediction in the US Corn Belt, leveraging a Transformer-based ensemble model that integrates a Vision Transformer (ViT) for spatial data, a Temporal Transformer for timeseries data, and a Light Gradient Boosting Machine (LightGBM) for tabular feature interactions. Utilizing a comprehensive dataset from 2005 to 2024, including USDA yield records, NOAA weather data, Sentinel-2 satellite imagery, and IoT soil sensor measurements, the proposed model achieves a relative root-meansquare error (RRMSE) of 6.12% on test data, surpassing conventional machine learning and deep learning methods. The implementation employs advanced feature engineering, including phenological and interaction terms, to capture complex agro-environmental dynamics. This paper details the methodology, algorithm, framework, architecture, workflow, and experimental results, offering a scalable solution for precision agriculture.