YieldSAT: A multimodal benchmark dataset for high-resolution crop yield prediction

• Paper url: CVPR website

This work introduces YieldSAT, a large-scale multimodal benchmark dataset for high-resolution crop yield prediction. Covering 2,173 expert-curated fields across Argentina, Brazil, Uruguay, and Germany, the dataset provides over 12.2 million yield samples at 10 m spatial resolution, paired with multispectral satellite imagery and auxiliary environmental data. The study formulates crop yield prediction as a pixel-level regression task and benchmarks deep learning models and multi-modal fusion architectures across diverse crops and climate zones. It also highlights severe distribution shifts in real-world yield data and explores a domain-informed Deep Ensemble approach that substantially improves predictive performance, providing a foundation for scalable, high-resolution yield prediction research.

Recommended citation: Miranda, Miro, et al. "YieldSAT: A multimodal benchmark dataset for high-resolution crop yield prediction." arXiv preprint arXiv:2604.00940 (2026).