Research On Tea Disease Recognition Based On Machine Learning And Hyperspectral Imaging

In order to achieve fast and efficient identification of tea diseases and provide important theoretical support and reference frane for plant protection unmanned aerial vehicle disease monitoring and early warning,the common tea diseases such as anthracnose,brown leaf spot and white star disease were used as the research objects in this paper,and the method of identifying tea diseases based on hyperspectral imaging technology combined with machine learning was put forward.The main contents and conclusions of this research were as follows.(1)Study the method of hyperspectral data collection and spectral feature selection for tea diseases.For tea disease samples and healthy leaf samples,the hyperspectral images at 616 bands were obtained at 358-1021 nm were collected through hyperspectral inaging system with appropriate parameter settings.The black-white correction and smoothing were used for hyperspectral images to eliminate the noise interference in the collection process.A region of 200 pixels square near the tip of the main vein was selected as the region of interest,and the average spectral reflectance of all the pixels in the region of interest was extracted as the spectral reflectance of the band.Obtain the overall changes and differences of spectral reflectance of all the samples according to the spectral reflectance curve in the whole range of the sample,select three bands 560,6405 780 nm that make the maximum separation between various samples as sensitive band based on caleulation of spectral reflectance separation,the spectral reflectance of the sensitive band is extracted as the spectral feature of the sample.(2)Study the method of image feature selection for hyperspectral data of tea diseases.According to the strong correlation of images in the adjacent band of hyperspectral image data,use the first principal component of each sample to find the characteristic wavelengths after principal component analysis,which are 762,700,721 and 719 nm.In order to reduce the computational complexity,the second principal component analysis was carried out with 4 characteristic wavelengths and the second main components were identified as the feature images.According to the Otsu segmentation algorithm,accurately isolate leaf lesion area and finally obtain the inage only contains the lesion.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.(3)Obtain different combination of feature vectors for saliency test and algorithm verification through feature combination and feature dimensionality reduction.For the spectral features,color features and texture features,Color feature and texture feature were combined as combination of feature vector 1;Color feature and texture feature and spectral feature were combined as combination of feature vector 2;And the combination of feature vector 3 is the output of the combination of feature vector 2 after reducing the dimension by genetic algorithm.(4)Establish machine learnilg model to test three combination of feature vectors.The model of support vector machine,random forest and BP neural network were respectively constructed for sample training and testing,Three models apparently improved the recognition rate of the samples by the combination of feature vectors,1,2 and 3 respectively.Compared with the combination of feature vector 1,the combination of feature vector 2 which include spectral information had greatly improved the recognition rate of tea on the three models,and the recognition rate for the four samples was above 80%,indicated that the spectral characteristics have important contributions and notable effects in the identification of tea diseases.The combination of feature vectors 3 not only reduced the computation complexity and shortened the modeling time,but also had high accuracy on the three classification models,and the recognition rate for the four kinds of samples was all above 90%.The research results showed that the method of hyperspectral imaging technology combined with machine learning can achieve fast and accurate classification of tea disease,provide valuable reference for plant protection unmanned aerial vehicle disease monitoring and early warning at low altitude remote sensing.