Identification Of Cotton Leaf Diseases And Pests Based On Machine Learning And Android

Cotton is the major cash crop in China.Cotton will be constantly attacked by pests and diseases during the growth process.About 15% of the annual economic loss of cotton in China comes from pests and diseases.Existing identification methods for cotton diseases and pests mostly utilize traditional artificial identification techniques.The automatic identification of cotton leaf diseases and pests is highly desired.In order to meet the needs of the development of smart agriculture,to solve the problems of low detection and recognition efficiency and poor reliability of cotton leaf diseases and insect pests in my country,combining Machine learning methods and the occurrence characteristics of cotton pests and diseases,the following work mainly:(1)Collected 975 images of healthy,red leaf blight,red spider,Fusarium wilt,Verticillium wilt,and double-spotted leaf beetle cotton leaf as sample sets.Trained different machine learning models: support vector machine models,extreme learning machine models,and convolutional neural network models.(2)The comparative study and analysis of the three machine learning models shows that the support vector machine model when the kernel function is a polynomial kernel function,the training accuracy rate of the model for the fusion features(gray level co-occurrence matrix + Hu7 + color histogram)of cotton leaf images of pests and diseases is 99.99%,and the test accuracy rate is 96.11%;The extreme learning machine whose activation function is Sigmoid has a training accuracy rate of 98.33% for the feature of cotton leaf color histograms when the number of hidden layer neurons is 200,and the test accuracy rate is 95.55%;In the transfer learning mode,the fine-tuning CaffeNet model with a learning rate of 0.001 has a training accuracy rate of 100.00% for cotton leaves of pests and diseases,and a test accuracy rate of 99.91%.(3)Using the confusion matrix to perform performance test analysis on the optimal model(fine-tuning CaffeNet model with a learning rate of 0.005 in the transfer learning mode)in this study,the average precision of the model is 98.24%,and the average recall is 98.09% The average F1 score is 98.12%,which has a higher recognition accuracy rate than traditional machine learning methods.(4)Finally,Using the Android Studio platform combined with JNI technology to complete the design of the App program framework and main functional interface and model migration.The App mainly includes the functions of image acquisition,feature extraction and detection and identification of cotton leaf diseases and insect pests.