Recordning on ai.lu.se

Topic: Mapping the Future Harvest: Predicting Fine-Scale Field Yields with Satellites and Graph Attention Networks

When: 3 September at 12.00-13.00

Where: Online

Speakers: 

• Oskar Åstrom, PhD Student Computer Vision and Machine Learning, Lund University
• Alexandros Sopasakis, Mathematics, Lund University

Spoken language: English

Abstract

Understanding how crop yields vary within a single field can unlock major gains in agricultural efficiency, but predicting these patterns before harvest has remained a challenge. In this work, we combine high-resolution Sentinel satellite imagery, soil and weather data, and the power of Graph Attention Networks (GATv2) to forecast intra-field yield variations for winter wheat at a 10 × 10 m scale. Our approach embeds temporal information directly into a graph-based model, capturing both global and local spatiotemporal dependencies without the need for separate time-series modules. The result: post-harvest yield estimation with an R² of 86.9% for yield variation, and pre-harvest predictions (up to a year in advance) with an nRMSE of 11.4%. By isolating the stable, field-specific drivers of yield variation, the model distinguishes them from overall yield levels, enabling actionable, fine-grained interventions. This fusion of remote sensing and graph-based machine learning points toward a future where farmers can plan months ahead, allocate resources with surgical precision, and reduce environmental impacts, turning every pixel of the field into a data-driven decision.