A Multimodal Deep Learning Framework for Crop Yield Prediction from Satellite and Weather Data
Sofia R. Albrecht<sup>1,2,#</sup>, Daniel O. Katsaros<sup>1,#</sup>, Amara J. Okonkwo<sup>1,2</sup>, Henrik P. Solberg<sup>1</sup>, Leila M. Haddad<sup>1,2,*</sup>
<sup>1</sup>Department of Agricultural Data Science, Eastvale University of Applied Sciences, Korsholm, Finland; <sup>2</sup>Centre for Precision Agriculture Analytics, Eastvale University of Applied Sciences, Korsholm, Finland
Correspondence to: Dr. Leila M. Haddad, Department of Agricultural Data Science, Eastvale University of Applied Sciences, Korsholm, Finland; Email: l.haddad@eastvale.example.edu
ABSTRACT
Reliable crop yield prediction requires integrating static field characteristics with time-varying satellite and weather observations, yet most operational models treat these sources separately. This study proposes a multimodal deep learning architecture, combining a temporal convolutional network, a bidirectional gated recurrent unit (GRU) branch, and a cross-attention fusion module, for yield class prediction across heterogeneous inputs. The proposed model was evaluated against three established baselines (logistic regression, random forest, and CatBoost) on a combined field-level and seasonal time-series benchmark of 54,200 records. The proposed model achieved 93.2% accuracy and an F1-score of 0.925, outperforming the strongest baseline (CatBoost) by 5.2 percentage points, at a moderate increase in inference time (12.6 ms versus 8.1 ms). Ablation analysis showed that removing the bidirectional GRU branch reduced accuracy by 5.5 points and removing the cross-attention module reduced accuracy by 3.1 points, indicating that both temporal modeling and cross-modal fusion contribute meaningfully to overall performance. These results suggest that multimodal architectures offer a practical and accurate approach for yield prediction tasks involving mixed static and sequential data.
Keywords: Crop yield prediction · Precision agriculture · Multimodal learning · Cross-attention · Remote sensing · Deep learning
