KNN identification model of salt-alkali tolerant rice variety based on hyperspectral feature selection

To achieve rapid and non-destructive identification of salt-alkali tolerant rice varieties, this study proposes a novel method, to our knowledge, that integrates hyperspectral feature selection with machine learning. Twelve rice varieties cultivated in the saline-alkali soils of Ningxia were selected as research subjects. Hyperspectral data from leaves were collected during the grain-filling stage, from which 33 spectral feature parameters were extracted. The maximum relevance and minimum redundancy (MRMR) algorithm was employed for feature selection, identifying a core subset of 10 key features, including green peak position ( λ gp ), red edge position ( λ rb ), and maximum absorption depth (Dh). These features effectively highlight the discriminative spectral differences among varieties. Based on this feature set, the discriminative performance of four classification models—back propagation neural network (BPNN), support vector machine (SVM), K-nearest neighbor (KNN), and random forest (RF)—was systematically compared. The variety identification task was formulated as a 12-class supervised classification problem. The results demonstrate that the KNN model utilizing 10 optimized features attains optimal performance, with a test-set classification accuracy of 89.40% and a Kappa coefficient of 88.43%. It significantly outperforms other models and comparison results across different feature scales (5, 10, 15, and 33 parameters). This study confirms that integrating hyperspectral feature optimization with the KNN algorithm can establish an efficient and accurate identification model for saline-alkali tolerant Oryza sativa varieties, offering reliable technical support for variety resource identification and precision breeding.