Integrating remote sensing data assimilation, deep learning and large language model to interactive yield prediction for wheat breeding

Abstract

Improving yield is one of the core goals of crop breeding. By predicting the potential yield of different breeding materials, breeders can screen these materials at different growth stages to select the best-performing breeding materials. However, existing yield prediction methods struggle to balance accuracy, interpretability, and robustness, often facing trade-offs between model complexity, data requirements, and generalizability. To address the above challenges, this study proposed a new hybrid method integrating remote sensing data assimilation and deep learning. The leaf area index was assimilated into the calibrated and validated WOFOST crop model using a newly designed data assimilation algorithm. The dataset, including partial outputs from the WOFOST model, development day, and vegetation indexes (VIs), was used to train the Temporal Fusion Transformer model for wheat yield prediction. The results showed that the new hybrid method achieved the highest performance in wheat yield prediction for different breeding materials in different study areas (R² of 0.831 and RMSE of 372.8 kg/ha in Yuhang experiment; R² of 0.704 and RMSE of 605.3 kg/ha in Zijingang experiment), which was better than other process-based model-driven methods and data-driven methods. This showed that the hybrid method had superior applicability in accurate yield prediction. Other results showed that increasing the number of data collections during the growth stage could significantly improve the performance of yield prediction and reduce error. Data from the middle and late growth stages contributed more to prediction performance than data from the early stages. Physiological variables and some VIs were the most important factors for yield prediction, while morphological characteristics contributed less. The importance and impact of each feature varied at different growth stages, highlighting the complex nonlinear relationship between characteristics and yield. In addition, since existing methods are difficult to fully utilize multi-source heterogeneous data related to yield in the breeding process, and the usability and user-friendliness of yield prediction software are also insufficient, an interactive yield prediction website has been developed based on a new hybrid method, a large language model (Llama), and related technologies to assist breeding decisions. This study aims to improve the efficiency of breeding material screening, provide an accurate, user-friendly, and well-interpretable yield prediction tool for wheat breeding, and facilitate smart breeding and decision making.

Plant phenotyping relevance

コムギの収量という植物形質を、リモートセンシング・データ同化・深層学習で予測する手法を開発・比較検証し、育種向けソフトウェアも開発しているため、フェノタイピング手法が中心的である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.