Early Yield Prediction for Sugar Beet Fields using Satellite Data - Learnings from Specialized Vision Transformers

This study investigates the effectiveness of Vision Transformers (ViTs) for predicting sugar beet harvest yields using only optical Sentinel-2 satellite imagery. Unlike many existing models that integrate multi-source data like climate, soil, or UAV imagery, this research focuses on a purely optical, single-label-per-field regression task. The authors systematically evaluate key model design choices—specifically patch size, spectral band selection, and positional encodings—to determine how to best adapt standard computer vision architectures for agricultural remote sensing.

The researchers found that standard design choices for natural image processing are often suboptimal for agricultural yield prediction. Specifically, using smaller patch sizes (e.g., ViT/2) allows the model to capture critical, fine-grained spatial variations within fields that larger patches overlook. Furthermore, the study demonstrates that training on all available Sentinel-2 spectral bands, rather than relying on hand-crafted vegetation indices (like NDVI or EVI), yields superior performance. This suggests that end-to-end deep learning can discover more nuanced, task-specific spectral patterns than traditional index-based approaches. Additionally, the authors found that relative positional encodings provided no performance benefit, indicating that for this specific task, the model does not require explicit spatial awareness of patch locations to estimate yield.

By isolating these design factors, the study provides a roadmap for building more efficient and accurate remote sensing models. The authors successfully demonstrated that their optimized model could identify a significant portion of low-yield fields early in the growth cycle. This capability offers a scalable, cost-effective tool for agricultural monitoring, enabling stakeholders to detect underperforming fields well before harvest without needing expensive auxiliary data sources.