YieldSAT: Multimodal crop yield prediction benchmark dataset accepted at CVPR 2026

Our paper “YieldSAT: A Multimodal Benchmark Dataset for High-Resolution Crop Yield Prediction” has been accepted at CVPR 2026, the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

Crop yield prediction requires substantial data to train scalable models, but existing datasets are scarce, low in quality, or limited to regional levels or single crop types. YieldSAT addresses this gap as the first multimodal dataset for crop yield prediction at both field and subfield (pixel) levels, combining combine harvester yield data, Sentinel-2 time series, weather, soil, and topography information.

Key facts

• 2,173 expert-curated fields across Argentina, Brazil, Uruguay, and Germany
• Over 12.2 million yield samples at 10 m spatial resolution
• 113,555 labeled Sentinel-2 images spanning entire growing seasons
• 4 crop types (corn, rapeseed, soybeans, wheat) across 9 years (2016–2024)
• 72 features per sample from satellite, weather, soil, and topography data

We demonstrate the potential of large-scale, high-resolution yield prediction as a pixel regression task by benchmarking various deep learning models and data fusion architectures, and highlight open challenges arising from severe distribution shifts under real-world conditions — mitigated by a domain-informed Deep Ensemble approach.

The work is also featured on our publications page.