Abstract:
Abstract: In order to achieve fast and efficient identification of tea diseases, the method of identifying tea diseases based on hyperspectral imaging technology was put forward. Four kinds of samples, including anthracnose, brown leaf spot disease, white star disease and healthy leaf, were collected in Pingshan tea plantation of Nanjing. Hyperspectral images of these samples, ranging from 358 to 1 021 nm, were collected by hyperspectral imaging system. Among them, there were 80 samples of anthracnose, 72 samples of brown leaf spot disease, 80 samples of white star disease and 60 samples of healthy leaves. The region of interest (ROI) was an area of 200 pixels × 200 pixels near the tip of the tea leaf. The average spectral reflectance curves of the effective band of ROI were extracted to analyze the spectral characteristics. For the purpose of decreasing the redundancy of hyperspectral data, and reducing the computational complexity, this study used principal component analysis (PCA) to process the original hyperspectral images, and obtained 4 kinds of principal component images for the samples with the maximum weight coefficients, and the wavelengths of 762, 700, 721, 719 nm corresponded were taken as the characteristic wavelengths. The test showed that direct use of 481 bands for the first PCA resulted in low calculation speed and low processing efficiency. Thus, the second principal components with the 4 characteristic wavelengths were employed, and the second principal component image was selected as the feature image through comparing the characteristics of lesion and non lesion regions. To get the accurate extraction of tea leaf spots, OTSU algorithm for image segmentation was adopted, the optimal threshold of 4 kinds of leaf samples was determined, and finally the sample images containing only leaf lesion regions were extracted. After image segmentation, 3 color feature parameters were extracted from the single-channel first moments, second moments and three-order moments of each feature image based on color moments; and 20 texture parameters were calculated from the 4 directions (0 , 45, 90 and 135°) of energy, contrast, correlation, stability and entropy based on gray level co-occurrence matrix (GLCM); and 3 spectral characteristic parameters of relative spectral reflectance of sensitive bands, including 560, 640 and 780 nm, were obtained. The color feature, texture feature and spectral feature were optimized into 2 feature vectors, and the training set and test set were tested by BP (back propagation) neural network and support vector machine (SVM) respectively. A total of 188 samples, including 50 anthracnose samples, 48 brown spot disease samples, 50 white star disease samples and 40 healthy leaf samples, were randomly selected as the training set, and the remaining 104 samples were used as the test set. The recognition rates of the test set through the feature vector combination of color features and texture features were generally low by BP neural network and SVM, and the recognition rates of the test set through the feature vector combination of color feature, texture feature and spectral feature were higher, which were 89.59% and 86.67% for BP neural network and SVM respectively. In order to further improve the recognition rate and shorten the modeling time, genetic algorithm was used to reduce the dimensionality of the input feature. Through taking selection, crossover and mutation operations, 26-dimensional input features were optimized to 14 dimensions, and then BP neural network was to recognize the tea spots. Finally, the average recognition rate was raised to 94.17%, and the model setup time was also shortened from 6.6 to 1.7 s. The result shows that it is possible to achieve fast and efficient identification of tea diseases by the fusion of spectral information and image information with pattern recognition technique. The neural network identification model based on genetic algorithm optimization has the advantages of short modeling time and high recognition accuracy.