Answers are grounded in this paper and cite their sources.
Accurate prediction of spatial yield variations within individual fields is crucial for precision agriculture, as it enables optimized resource allocation and targeted crop management. In this study, we propose a novel framework that leverages remote sensing data and Graph Attention Networks (GATv2) to predict fine-scale yield variations for winter wheat at a high resolution (10 m × 10 m). The objectives of our research are twofold: (i) to develop an integrated, multi-modal prediction model that embeds temporal information directly into a graph-based architecture to capture both global and local spatiotemporal dependencies, and (ii) to rigorously evaluate the model’s performance in post-harvest yield estimation and pre-harvest yield forecasting. Our approach fuses high-resolution Sentinel-2 imagery, spectral indices, soil characteristics, and weather dynamics within a unified graph structure, eliminating the need for separate temporal models while dynamically adjusting the influence of neighboring nodes via attention mechanisms. Experimental results demonstrate competitive performance, with normalized RMSE values of 11.5% for absolute yield and 9.6% for yield variation, alongside R 2 scores of 80.7% and 86.9%, respectively, in post-harvest yield estimation. Moreover, our framework successfully forecasts intra-field yield variability up to a year in advance (nRMSE of 11.4%), underscoring its robustness and stability across diverse data conditions. By identifying stable, field-specific factors governing spatial yield variability, the model highlights the separability of yield variation from overall yield levels. This capability provides actionable insights for both immediate interventions and strategic planning, enabling optimized resource allocation, reduced waste, and minimized environmental impacts from over-fertilization. These results further underscore the potential of graph-based machine learning to transform precision agriculture through scalable, and high-resolution yield prediction.
No takes yet. Share an insight, caveat, or question.
Åström et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: