Classification of drought resistance wheat varieties based on hyperspectral remote sensing combined with machine learning

The screening of new drought resistance wheat varieties is the most economical and effective method for combating drought. In this study, hyperspectral data of the wheat canopy at different growth stages and drought resistance indices during the maturity period were acquired, and spectral processing, feature selection and machine learning algorithms were integrated to construct a drought resistance wheat variety selection model. The results demonstrated that the drought resistance classification models for wheat varieties constructed based on different growth stages varied, with the flowering stage having the highest overall classification accuracy (OA=72.99%). Spectral processing and feature selection methods influence the accuracy of classification models, and the average classification accuracy across different processing methods follows continuous wavelet transform (CWT) > first derivative (FD) > multiplicative scatter correction (MSC) > original reflectance (OR). The sensitive bands from CWT were primarily concentrated in the 1850–1920 nm and 2490–2500 nm ranges. The average classification accuracy based on competitive adaptive reweighted sampling (CARS) was 3.91% and 5.53% higher than that of successive projection algorithm (SPA) and Relief-F, respectively. Among the three classification algorithms, the random forest (RF) algorithm demonstrated the best performance, whereas the convolutional neural network (CNN)) algorithm performed the worst. The CWT–CARS–RF classification model achieved the highest accuracy (OA=92.40%, kappa=0.895), which provides a theoretical basis and technical support for the rapid and accurate screening of germplasm resources.