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

Area of Science:

• Agricultural Remote Sensing
• Machine Learning Applications
• Precision Agriculture

Background:

• Accurate rice yield prediction is crucial for global food security and precision agriculture.
• Existing remote sensing models often lack generalization across different regions, cultivars, and scales.
• UAV-based multispectral imagery offers a promising data source for crop monitoring.

Purpose of the Study:

• To develop an interpretable and automated machine learning framework for rice yield prediction.
• To enhance the accuracy and transferability of yield prediction models using UAV multispectral data.
• To compare different feature selection strategies and machine learning models for optimal performance.

Main Methods:

• Utilized UAV-based multispectral imagery from 143 rice samples across diverse fields in Sichuan Province, China.
• Extracted 20 vegetation indices and 36 texture features, comparing three feature selection methods (PCC, RF-I, AutoGluon-I).
• Evaluated four regression models (CatBoost, ExtraTrees, Random Forest, AutoGluon ensemble) using R², RMSE, and MAE metrics.

Main Results:

• The AutoGluon ensemble consistently outperformed individual models, achieving test set R² values between 0.528-0.736.
• The best performance (training R²=0.821, test R²=0.736) was obtained using Pearson correlation coefficient (PCC) feature selection with AutoGluon.
• Shapley Additive Explanations (SHAP) revealed significant contributions from texture features, highlighting canopy structural heterogeneity at maturity.

Conclusions:

• The proposed PCC-AutoGluon-SHAP framework offers a lightweight, accurate, and interpretable method for UAV-based rice yield estimation.
• This approach demonstrates improved generalization across heterogeneous field conditions.
• The framework has practical potential for scalable precision agriculture and regional yield monitoring.