Explainable machine learning to predict root biomass of field crops using UAV multispectral data.

Area of Science:

• Agricultural Science
• Plant Biology
• Remote Sensing

Background:

• Below-ground biomass is crucial for crop performance, especially in water-limited environments.
• Current field-scale root biomass monitoring relies on destructive, labor-intensive methods.
• Non-destructive techniques are needed for efficient root phenotyping and crop improvement.

Purpose of the Study:

• To develop and validate explainable machine learning models for estimating root biomass.
• To utilize Unmanned Aerial Vehicle (UAV) multispectral imagery and canopy phenotypic traits for this estimation.
• To assess the potential for large-scale, non-destructive root biomass assessment in semi-arid agriculture.

Main Methods:

• Collected 405 samples of maize, millet, and sorghum during the 2024 growing season.
• Evaluated eight machine learning algorithms, focusing on Random Forest and XGBoost.
• Employed SHAP analysis to interpret model predictions and identify key phenotypic drivers.

Main Results:

• Random Forest and XGBoost models achieved high predictive accuracy for root biomass (R² = 0.763 for millet, 0.688 for maize, 0.659 for sorghum).
• Leaf area was the most influential predictor across all species.
• Leaf water content and chlorophyll traits showed species-specific drought adaptation effects.

Conclusions:

• UAV-based multispectral phenotyping with interpretable machine learning enables non-destructive field-scale root biomass estimation.
• The 'cost of indirect inference' (15-25% R² reduction) highlights challenges in predicting below-ground traits from canopy data.
• This approach shows promise for precision agriculture and crop improvement in semi-arid regions, pending broader validation.