Drought tolerance classification using unmanned aerial systems based on RGB and multispectral data

Drought poses a global threat to food security and demands intensified efforts from breeding programs. Yet the lack of efficient methods for selecting this trait increases the cost and time required to develop new cultivars. The goal of this work was to assess the feasibility of using spectral data from RGB or multispectral sensors for drought-tolerance classification across various machine-learning models under the most practical cross-validation scenarios typical in breeding programs. The genotypes were assessed during trials conducted under either optimal (irrigated) or drought-stress conditions across two years, and evaluated using up to 10 field traits to determine their drought-tolerance classification based on membership function values related to drought. RGB and multispectral vegetation indices collected during several flights throughout the crop cycle were used to train machine learning models. We found that drought trials offer the best training data. Specificity was the metric most affected by sensor type and the nature of the training data. The multispectral sensor outperformed the RGB sensor on most evaluation metrics in both years. AdaBoost and linear discriminant analysis models demonstrated the strongest consistency across all prediction scenarios. Together, they achieved an overall accuracy, specificity, and F1-Score of 0.71, 0.56, and 0.77, respectively. The most influential vegetation indices for model performance consistently included the NIR band. Spectral information, such as vegetation indices, is a useful tool for plant researchers to complement drought tolerance evaluations in the field. This data-driven approach facilitates automation, paving the way to speed genetic gains by including early assessments of drought tolerance in breeding pipeline, and improves resource utilization efficiency.