Integrating Multi-Source and Multi-Temporal UAV Observations to Improve Wheat Yield Prediction Using Machine...

Area of Science:

• Agricultural Science
• Remote Sensing Technology
• Data Analytics

Background:

• Accurate wheat yield estimation is crucial for effective precision agriculture and breeding programs.
• Traditional methods often lack the comprehensive data needed for precise yield prediction.
• Unmanned Aerial Vehicle (UAV) remote sensing offers high-resolution, multi-temporal, and multi-source data acquisition capabilities.

Purpose of the Study:

• To develop and evaluate UAV-based remote sensing models for non-destructive wheat yield estimation.
• To compare the effectiveness of single-temporal versus multi-temporal data fusion for yield prediction.
• To assess the performance of different feature types (vegetation indices, texture, color) and machine learning algorithms.

Main Methods:

• Collected UAV-based multispectral and RGB imagery across six key wheat growth stages.
• Extracted vegetation indices, texture, and color features from the imagery.
• Developed and validated yield prediction models using Random Forest (RF), XGBoost, and K-Nearest Neighbors (KNN) algorithms under single- and multi-temporal scenarios.

Main Results:

• Red-edge-based vegetation indices demonstrated high sensitivity to wheat yield, outperforming texture and color features.
• Multi-feature fusion improved prediction accuracy, especially during booting and flowering stages (R² = 0.53–0.67).
• Multi-temporal data fusion significantly enhanced yield estimation accuracy, reaching a maximum R² of 0.72 when integrating data from late-jointing, booting, and flowering stages.
• XGBoost and KNN algorithms showed superior accuracy and stability across most growth stages.

Conclusions:

• Integrating multi-source and multi-temporal UAV remote sensing data significantly improves the accuracy and robustness of wheat yield estimation.
• This approach provides valuable technical support for precision agriculture and phenotyping-assisted breeding.
• Red-edge indices and multi-temporal data fusion are key factors for accurate yield prediction.