Rice yield prediction using UAV-based multispectral imagery and AutoGluon across regions and field scales

Abstract

Introduction Accurate and transferable rice yield prediction is essential for precision agriculture and food security, yet existing remote sensing-based models often suffer from limited generalization across regions, cultivars, and field scales. Methods This study developed an interpretable automated machine learning framework for rice yield prediction using UAV-based multispectral imagery collected at the maturity stage. A total of 143 rice samples, including 79 experimental plots and 64 production fields across 15 counties in Sichuan Province, China, were investigated. 20 vegetation indices and 36 gray-level co-occurrence matrix texture features were extracted from multispectral orthomosaics, and three feature selection strategies: Pearson correlation coefficient (PCC), Random Forest feature importance (RF-I), and AutoGluon feature importance(AutoGluon-I), were systematically compared. Four regression approaches, including CatBoost, ExtraTrees, Random Forest, and an AutoGluon stacked ensemble, were evaluated using R 2 , RMSE, and MAE. Results The results showed that the AutoGluon ensemble consistently outperformed individual machine learning models, improving testset R 2 from 0.403-0.670 to 0.528-0.736. The best performance was achieved by combining Pearson correlation-based feature selection with AutoGluon, yielding a training R 2 of 0.821 and a test R 2 of 0.736, with RMSE and MAE values of 0.749 and 0.568 t ha -1 , respectively. Shapley Additive Explanations (SHAP) analysis further revealed that texture features, particularly red-band contrast and angular second moment features, contributed substantially to yield prediction, indicating the importance of canopy structural heterogeneity at maturity. Discussion Overall, the proposed PCC-AutoGluon-SHAP framework provides a lightweight, accurate, and interpretable approach for UAV-based rice yield estimation across heterogeneous field conditions, offering practical potential for scalable precision agriculture and regional yield monitoring.

Plant phenotyping relevance

UAVマルチスペクトル画像からイネの収量を推定する画像・特徴抽出・機械学習フレームワークを開発し、複数モデルと特徴選択法を比較検証しており、植物表現型取得・推定が研究の中心である。

Code and data availability

The paper describes UAV multispectral imagery, yield measurements, and a Python/AutoGluon workflow, but provides no public deposit, repository, or URL for any dataset, imagery, or code. The Data Availability Statement only promises raw data from authors on request, and the supplementary material link is not stated to包含

No evidence-backed public reproduction asset is currently recorded.