Classification of Salt–Alkali‐Tolerant Rice Cultivars Using Hyperspectral Reflectance and Machine Learning

Aiming at the problems of low efficiency, heavy workload, long cycle, and destructiveness in traditional methods for distinguishing salt–alkali tolerance among different rice cultivars, this research proposes a classification method combining hyperspectral technology with machine learning algorithms. The Analytical Spectral Devices (ASD) FieldSpec 4 hyperspectral instrument was employed to collect leaf reflectance spectral data of different rice cultivars. After removing abnormal data, smoothing, and noise reduction, the combined preprocessing method of standard normal variable (SNV) transformation and first‐order derivative (FD) was applied to the original spectral data to eliminate noise and enhance subtle spectral features. Furthermore, Competitive Adaptive Reweighted Sampling (CARS) and Random Forest (RF) were separately used for feature extraction to achieve data dimensionality reduction and select key spectral features. On this basis, the performance of Back Propagation Neural Network (BPNN), Support Vector Machine (SVM), and K‐Nearest Neighbor (KNN) algorithms was systematically compared. These algorithms were combined with the two feature extraction methods to construct classification models for rice cultivars’ salt–alkali tolerance. Experimental results showed that the CARS feature extraction method could more effectively extract spectral features related to salt–alkali tolerance traits. The model combining CARS with KNN achieved the best performance, with a classification accuracy of 97.14% and a Kappa coefficient of 96.88%. The hyperspectral CARS‐KNN classification model for salt–alkali‐tolerant rice developed in this research provides technical support for rice breeding against salt–alkali stress and establishes a methodological basis for the precise discrimination of salt–alkali‐tolerant cultivars in future research.