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

Understanding below-ground biomass dynamics is essential for improving crop performance in water-limited regions. Yet field-scale root monitoring remains constrained by destructive and labor-intensive sampling. This study presents explainable machine learning models to estimate root biomass of maize, millet, and sorghum using UAV multispectral imagery and key canopy phenotypic traits. Across 405 samples collected during the 2024 growing season, eight algorithms were evaluated, among which Random Forest and XGBoost achieved the highest predictive accuracy (R² = 0.763 for millet, 0.688 for maize, and 0.659 for sorghum). SHAP analysis revealed that leaf area was the dominant predictor across all crops, with 2-3 times greater influence than other traits, while leaf water content and chlorophyll-related parameters exhibited species-specific effects associated with drought adaptation. Under the conditions tested, these results suggest that UAV-based multispectral phenotyping, combined with interpretable machine learning, can enable non-destructive estimation of root biomass at the field scale. Within the limits of this single-site, single-season study, the approach demonstrates potential for large-scale root phenotyping and for supporting crop improvement in semi-arid regions. We quantify a 15-25% reduction in R² relative to above-ground trait prediction, which we term the 'cost of indirect inference'-highlighting the inherent challenge of estimating below-ground biomass from canopy-level data. These findings offer insights for precision agriculture, subject to broader validation.