Research On Wheat Diseases Monitoring And Automatic Identification Technology Based On Field Images

In recent years,with the climate change,wheat diseases have shown a trend of multiple,recurring and overlapping occurrences,resulting in a large reduction in yield or even no harvest in some wheat fields,which seriously restricts the continued high yield of grain in my country.Traditional wheat disease identification and monitoring mainly relies on long-term observation and judgment of wheat disease types and disease severity by plant protection personnel.Accurate disease identification and estimation is the basis and premise of accurate prevention and control.It is of great significance to develop non-destructive,efficient and intelligent wheat disease monitoring methods.Aiming at the lack of information technology in current crop disease monitoring,this paper focuses on the segmentation and identification of wheat leaf diseases and remote sensing monitoring of wheat powdery mildew images.Taking wheat powdery mildew,stripe rust and leaf blight as the research objects,the wheat leaf disease images were obtained through a smartphone,and the pure disease pixels were extracted based on the Grabcut segmentation algorithm and adaptive threshold algorithm,and a segmentation model of wheat disease parts under complex background was constructed.And the convolutional neural network algorithm was used to construct the wheat disease identification and diagnosis model.At the same time,a field with severe wheat powdery mildew was selected,and a six-rotor UAV was used as a remote sensing platform to obtain images of wheat powdery mildew through the on-board multispectral and thermal infrared cameras.Based on multi-data source fusion and machine learning algorithms,wheat powdery mildew was detected monitor.This provides key technical support for accurate diagnosis of wheat diseases,and is of great significance for accurate prevention and control of crop diseases and sustainable agricultural development.The main results are as follows:(1)In view of the difficulty of segmentation of wheat disease images under field conditions,this paper optimizes the conventional Grabcut algorithm according to the characteristics of the disease and the characteristics of green pixels,and replaces the generated ROI markers with the human-computer interaction marker frame to enhance the automation level of disease segmentation.and efficiency.At the same time,the adaptive threshold algorithm is used to weaken the over-segmentation of the edge disease area by the Grabcut algorithm.The Precision,Recall,MPA and MIOU of the improved algorithm are 0.854,0.917,0.912 and 0.776,respectively.Compared with the conventional Grabcut algorithm,the Precision,Recall and MIOU increased by 4.27%,8%,and 8.38%,respectively,while MPA increased by a smaller margin(1.11%).(2)Extract color and texture features from the segmented lesion pixels and input them as model parameters,using the nearest neighbor node algorithm(K-Nearest Neighbor,KNN),Support Vector Machines(Support Vector Machines,SVM)and random forests(Random Forest,RF)for disease classification and identification.Combining color and texture features through the RF model,the disease classification results are the best with an accuracy rate of 97.59%,followed by SVM(97.21%)and KNN(95.44%).(3)Using the segmented field image as a dataset,five convolutional neural networks(VGGNet,Goog Le Net,Res Net,Dense Net and Mobile Net)are used for training and validation,and the network dataset is used for testing.The Dense Net model has the highest accuracy in both validation and test sets,with an accuracy of 99.47% and 94.14%,followed by Res Net(98.98%,90.92%),Goog Le Net(98.94%,89.75%),Mobile Net(98.22%,85.36%)and VGGNet(97.64%,81.26%).(4)Wheat fields with severe powdery mildew were selected,and three remote sensing data sources were extracted from UAV images.Multiple Linear Regression(MLR),Back Propagation(BP),Extreme Learning Machine(Extreme Learning Machine)Learning Machine,ELM)and Random Forest(RF)modeling.Among them,vegetation index(VI)is the most suitable for estimating the severity of wheat powdery mildew,followed by temperature(TP)information,and texture feature(TF)is the worst.Integrating different information sources,the three information source fusion models have the highest accuracy(R~2=0.836),followed by the vegetation index and temperature fusion model(R~2=0.758),while the vegetation index and texture fusion model(R~2=0.676),temperature and temperature The texture fusion model(R~2=0.647)was poor;whether using single data or multi-source information fusion,the monitoring accuracy of the RF model was higher than that of other algorithm models,and it was more suitable for UAV remote sensing to monitor the incidence of wheat powdery mildew.