Classification of wheat rust severity using UAV-based multispectral and thermal imaging

Timely identification of wheat genotypes resistant to stripe rust is crucial for effective breeding initiatives, timely intervention, and efficient disease control. This study leverages UAV-based thermal and multispectral sensors data, acquired over six growth intervals, to classify tolerant and susceptible genotypes. Two fields were used: one non-infected control and another infected with rust, for comparison of spectral indices. The study presents an extensive data processing pipeline that includes UAV data calibration, feature generation, multimodal feature selection (correlation filtering, Lasso Regression, Recursive Feature Elimination, and Mutual Information), and machine learning model training using RF, XGBoost, SVM, and MLP for classification across successive time frames. Spectral indices, including NDVI, CWSI, and TVI, were consistently selected as key features for model training across growth stages. The model’s performance improved gradually over time, with SVM achieving an optimal classification accuracy of 0.84 during 126 to 182 days after sowing (DAS). The SVM attained an accuracy of 0.81, while the MLP reached 0.82 during 126–153 DAS, demonstrating that early rust detection is feasible. The research findings suggest that integrating thermal and multispectral indices improves the accuracy of genotype classification for rust stress tolerance, with the highest accuracy observed between 126 and 182 days after sowing. The thermal index achieved an accuracy of 0.72, the multispectral indices 0.80, and the integrated approach reached 0.84, outperforming the use of individual sensors. This was also confirmed by the distribution of spectral indices of rust across different growth stages, where 144 to 182 DAS were key stages of spectral change in both rust and control crops, and CWSI, as a thermal index, detected stress symptoms before other vegetation indices. The study’s findings highlight the effectiveness of integrating thermal and multispectral sensors for non-invasive phenotyping, facilitating data-driven decision-making for the early detection of stripe rust symptoms.