Phenotyping terminal heat stress tolerance in wheat using UAV multispectral imagery to monitor leaf stay-greenness

Terminal high temperature is a significant constraint on compromising wheat yield and quality. Rapidly screening heat-tolerant varieties for breeding or cultivation is essential to secure wheat productivity under climate change. Our study introduces a novel method to distinguish different heat-tolerant wheat types based on spectral features extracted from multispectral images captured by UAV. Initially, we analyzed differences in SPAD values and SPAD dropping rate (SDR) between different heat-tolerant wheat types before and after high-temperature treatments. Subsequently, we analyzed the correlations between SPAD values and spectral indices, as well as between SDR and spectral indices variation rate (SIVR), both before and after high-temperature treatment. Ultimately, we constructed a classification model based on machine learning classifiers and examined the effects of time series data and various classification algorithms on the classification accuracy of the model. To evaluate the classification effectiveness of our newly proposed SDR, we conducted a comparative analysis of the performance of different classification models using SPAD values and SDR as screening traits for heat resistance. Subsequently, to determine the optimal input features for the classification model, we employed the recursive feature elimination (RFE) algorithm to select the input spectral indices and compared the results with those of a model that inputs all features. Our findings indicate that while both SPAD values and SDR are significantly difference between different heat tolerance types after high-temperature treatment. However, when SDR was used as a classification trait, the average overall accuracy (OA) was 71.88 %, and the average Kappa coefficient (kappa) was 0.55. In contrast, when SPAD was used as a classification trait, the average OA was 66.46 %, and the average kappa was 0.46. The model's overall performance was superior when using SDR as a classification trait. After applying the recursive feature elimination (RFE) algorithm for feature selection, the classification performance of the model was further enhanced, with the average OA and average kappa reaching 75.00 % and 0.60, respectively. The classification model exhibited optimal performance at 25 days after flowering (DAA). Among the four algorithms tested—decision tree (DT), random forest (RF), support vector machine (SVM), and logistic regression (LR)—the RF-based model achieved the highest accuracy, with an OA of 90.00 % and a kappa of 0.84. These results demonstrate that the SDR serves as a robust phenotypic trait for screening heat-tolerant wheat varieties. Furthermore, the proposed spectral index-based classification model provides a reliable decision-making tool for selecting and breeding heat-tolerant wheat germplasm.