关键词: 梨树病害, 高光谱成像, 特征波长, 判别模型, 机器学习

Abstract: Pear trees are affected by diseases during the growing season, and pesticide spraying is the main measure for disease control, while disease identification is a prerequisite to ensure accurate application. A method based on visible/NIR hyperspectral imaging combined with machine learning for the classification and detection of pear leaf diseases is proposed to achieve efficient identification of pear leaf diseases. Hyperspectral images of four types of samples, namely, healthy leaves, brown spots, black spots and sunburn disease, were collected under natural light conditions using a ground‑based hyperspectral imaging system. The average spectral data of the region of interest between 401 and 935 nm band were extracted to compare and analyse four types of Savitzky‑Golay smoothing (SG), standard normal transform (SNV), SG combined with first‑order differentiation and SG combined with second‑order differentiation. The results were compared, and finally the best classification and discrimination model for pear diseases was selected by using principal component analysis (PCA) and continuous projection algorithm (SPA). The results showed that the full‑spectrum data were best identified after the SNV pre‑processing, and the feature wavelengths extracted by the SPA algorithm performed better in both the SVM and BPNN models compared to the PCA algorithm. The best discriminative classification model for pear diseases was found to be SNV-SPA-SVM. The overall accuracy was 93.57% and the Kappa coefficient was 0.916 5. The use of visible/NIR hyperspectral technology can effectively classify and identify pear leaf diseases, providing a new method to achieve automatic diagnosis of pear leaf diseases in the field.

Key words: pear diseases, hyperspectral imaging, feature wavelengths, calibration models, machine learning