| Citrus is grown in 135 countries in the world today,so it is closely related to the development of global economy and trade.There are many varieties of citrus.With the change of temperature and humidity,the sugar content and soluble solids content of citrus changed,which caused the quality difference of citrus.The classification of citrus leaves can monitor the growth of citrus,so as to formulate the optimal planting programs such as weeding,pest control and spraying of pesticide,which have important practical significance for the cultivation and management of crops.Conventional classification methods of leaves such as chemical detection,image features classification have disadvantages such as high cost and unstable models.This study used spectroscopy and digital image processing techniques to identify citrus leaf varieties.The main research contents are as follows:(1)The varieties of citrus leaves were identified based on hyperspectral technology.A total of 2799 citrus leaves from 13 varieties were collected.Thirteen varieties of citrus leaves were identified based on hyperspectral technology.Four models of SVM,PCA-SVM,PCA-RBF-NN and PCA-LDA were established for the original spectrum.Among them,the identification accuracy of the SVM model reached 92.53%,which was the highest.The original spectrum was pre-processed by Detrend,and the above four models were also established.The identification accuracy of the PCA-SVM model reached 95.40%,which was the highest.The SPA algorithm was used to extract the characteristic wavelengths of the preprocessed spectra of 5 varieties of citrus leaves.Extracted the features of hyperspectral images and fused them with feature wavelengths of spectral.A total of 52 characteristics of each leaf of the 5 citrus varieties were modeled.The identification accuracy of the SVM model reached 95.00%,and the identification accuracy of the LDA model reached 97.00%.The results showed that hyperspectral technology can be used to identify varieties of citrus leaves.(2)Based on the feature extraction method of digital images,the varieties of citrus leaves were identified.Feature extraction method was used to identify 5 varieties of citrus leaves.Extracted 5 geometric features,7 Hu invariant moments and 4 texture features for calculation.The SVM models were established by three methods:morphological features,texture features,and fusion data of morphology and texture features.The classification accuracy based on fusion data was the highest,which was75.44%.The classification effect of morphological features of citrus leaves was better than texture features.Feature extraction method was used to identify 13 varieties of citrus leaves.The SVM models were established by morphological features,texture features and fusion data of morphology and texture features.None of the three models could make a good distinction between 13 varieties of citrus leaves.Even the classification accuracy of fusion data was only 56.42%.(3)The convolutional neural network models based on digital images were used to identify the varieties of citrus leaves.A classification model of citrus leaf varieties based on Alex Net was built.When the initial learning rate was 0.0001,the accuracy of the model validation set was reaching 99.37% and the accuracy of the test set was 99.10%.A Res Net18 model for classification of citrus leaf varieties based on residual blocks was established.When the initial learning rate was 0.001,the accuracy of the model validation set was reaching 99.73% and the accuracy of the test set was 99.64%.It can be seen that the CNN models richly expressed the differences between varieties through the features automatically extracted by the deep network,which could accurately identify and classify the varieties of citrus leaves.This study have shown that both hyperspectral technology and digital image technology can achieve the classification of multiple varieties of citrus leaves,and the accuracy rates are more than 95%.However,the time cost of acquiring images by hyperspectral technology is relatively high.Digital images using convolutional neural network algorithm is better than traditional machine learning algorithm for the classification of citrus leaf varieties. |
