TSHFS : a two‐stage hyperspectral feature selection approach for congeneric rice planthopper species classification

BACKGROUND Rice planthoppers, particularly the congeneric species Nilaparvata lugens (Stål) , Nilaparvata bakeri (Muir) , and Nilaparvata muiri (China) , are significant pests in rice production. Their morphological similarity poses a considerable challenge for conventional classification methods. To address this issue, we propose a two‐stage hyperspectral feature selection (TSHFS) method. The TSHFS framework first employs competitive adaptive reweighted sampling (CARS) to identify informative spectral bands, followed by the adaptive gradient boosting decision tree for feature refinement and importance analysis. Additionally, first‐order derivative spectral preprocessing is applied to enhance spectral contrast and improve feature discrimination. RESULTS Extensive experiments on hyperspectral data from the three congeneric species demonstrate that TSHFS significantly enhances classification performance. Specifically, TSHFS achieves the highest accuracy of 86.42% with an F1 score of 86.41%, outperforming models trained on raw spectral data (accuracy 64.17%, F1 score 64.27%) and CARS‐selected features alone (accuracy 66.33%, F1 score 66.52%). Furthermore, integrating first‐order spectral preprocessing with CARS‐selected bands maintains a high accuracy of 85.63% while achieving effective feature selection. CONCLUSION TSHFS effectively selects relevant spectral features, removes redundancy, optimizes spectral data utilization, and reduces computational complexity while improving classification accuracy. The results are significant as they provide a reliable method for distinguishing morphologically similar rice planthopper species, which advances automated pest identification in agricultural applications and can improve pest management strategies in rice production.