Research On Crop Disease Image Recognition Algorithms Based On Ensemble Learning

The current crop disease identification in our country mainly relies on manual approach,causing some problems such as low efficiency,high labor intensity,poor objectivity,and less promotion.Whilst,there are many deficiencies in the existing expert systems for crop disease diagnosis.For these reasons,this dissertation studies the key algorithms and practical applications of crop disease image identification based on ensemble learning.Thereafter,an automatic identification and diagnosis system for crop disease images is established.The main research work and contributions of this dissertation can be summarized as follows.1.The traditional machine learning(ML)methods can be well employed in the crop disease recognition of small sample images,but there are problems such as manual determination of features,difficult interpretation of extracted features,and low accuracy.Hence,this dissertation proposes a new Group Method of Data Handling-Logistic(GMDH-Logistic)model to recognize cucumber leaf disease images.This method overcomes the shortcomings of commonly-used ML methods,automatically determines key features,and has strong interpretability.The experimental results verify the effectiveness of the proposed method,which is superior to other commonly-used ML methods.The average recall rate of the proposed method is 1.67%higher than that of the suboptimal one.2.Manual feature extraction is required for traditional ML methods,while deep learning(DL)methods,especially deep convolution neural network(DCNN),is a type of end-to-end learning approach.The DL methods can automatically extract image features,but need a lot of labeled data to train the models.In view of the research foundation and preparatory conditions required by the DL methods,this dissertation enhances DCGAN and proposes a data augmentation scheme that combines traditional data augmentation methods and enhanced DCGAN.Whilst,this dissertation proposes a two-stage transfer learning approach.3.This dissertation studies the image recognition and diagnosis model of crop diseases based on deep learning.Two DCNN models,including the Inception Visual Geometry Group Network(INC-VGGN)and Densely Connected Inception Convolutional Network(DENS-INCEP),are designed to recognize maize and rice crop diseases.The validation accuracy of the proposed methods on the public maize and UCI rice image datasets is improved by at least 6.83%and 5.55%compared with the existing methods.The test accuracy of the proposed methods reaches 92.00%and 98.63%on the locally collected rice disease image dataset.Experimental findings demonstrate that the proposed methods outperform other traditional DCNN models.4.Because deep learning models are large and require high memory,this dissertation studies the lightweight convolutional neural networks(CNNs)and proposes two lightweight models,including the Squeeze Excitation Mobile Network(SE-MobileNet)and Mobile Dense Attention Network(Mobile-DANet),to recognize crop disease images.Multiple experimental results demonstrate the effectiveness of the proposed methods,which are superior to other lightweight CNNs.On the public PlantVillage dataset,the validation accuracy of the proposed methods is improved by at least 0.07%and 0.26%compared with other state-of-the-art methods.The proposed methods have attained an average recall rate of 92.92%and 83.45%on the local dataset,respectively.5.To improve the stability,accuracy,and robustness of the model,this dissertation studies the ensemble learning.Based on the previous analysis,this dissertation integrates three lightweight CNNs using a two-layer stacking framework.The first-level classification is employed to generate data output values for training the model;the second-level classifier further learns from the output of the first-level classifier and corrects the deviation of each individual classifier in this framework,thereby gaining the final classification results.Extensive experiments are performed on publicly and locally collected plant disease image datasets.On the publicly available rice and potato datasets,the validation accuracy of the proposed Ensemble Mobile Network(Es_MbNet)is improved by at least 0.68%compared with other influential DCNNs.In addition,the average test accuracy of the proposed approach is improved by 7.83%relative to the results reported in the literature.The proposed approach outperforms the single model and achieves an average recall rate of 95.32%on the locally collected crop disease image dataset.