| Wheat is one of the three major grain crops in China,and wheat yield is related to national food security,which is of vital importance to social stability and national economic development.Diseases can be a major threat to wheat yield,and rust is one of the most serious diseases of wheat and a major control target in wheat field management.In the current field management of wheat disease,there are problems such as "inaccurate identification,untimely discovery and imprecise prevention and control".This study explored the use of modern information technology to transform the traditional means of disease identification and monitoring to help farmers prevent and control wheat rust in a timely and precise manner.In this study,leaf level images and canopy level UAV remote sensing images of wheat rust were collected respectively,and deep learning-based disease(remote sensing)image identification and classification,detection of onset centres and segmentation of onset areas and their severity grading assessment were carried out.The main research results are as follows:(1)Confounding factor analysis and attention mechanism help to improve the recognition accuracy of wheat rust images.Phenotypic morphology is the main factor causing confusion among the three types of rusts.In this study,we used a confusion matrix to find out the main misclassification categories and their numbers,and comprehensively analysed the specific causes of misclassification.Accordingly,the dataset,model structure and learning strategy were improved,and three datasets,WRD(Wheat Rusts Dataset),WDD(Wheat Disease Dataset)and PDD(Poaceae crops Disease Dataset),were created respectively,and a new model,Imp-Dense Net,was proposed for wheat rust image recognition,which finally achieved the accurate recognition of the three types of rust diseases in wheat,with the Top-1 accuracy =98.32%(WRD).The results of the ablation experiments show that the image processing method named PPRN,which is proposed for the preliminary classification results and the phenotypic and morphological features of the three types of rust diseases,contributes the most to the improvement of the classification accuracy,which reaches 5.19%,and it proves that it can guarantee the feature stability of the image when it flows between the convolutional layers within the model.The comprehensive modification of the Dense Net model structure(incorporation of SE attention mechanism,SPPNet and factorisation ideas)is also more effective,and its contribution to the classification accuracy improvement reaches 4.68%.It is demonstrated that a suitable Convolutional Neural Networks(CNN)model can achieve effective self-iteration in learning rust image features,and thus identify the disease type quickly and accurately.At the same time,the effect of model modification was visualized by analyzing the changes in convolution calculation focus trajectory through Grad-CAM.The superior performance of Imp-Dense Net is demonstrated through performance comparison experiments with emerging models.Through the gradual expansion of data categories and quantities,high-precision and automated identification and classification of common diseases of cereal grain crops such as wheat,maize and rice were achieved,with accuracies of Top-3 accuracy = 97.30%(WDD)and Top-5 accuracy = 96.60%(PDD),respectively,which verified Imp-Dense Net’s robustness.(2)Propose an optimized object detection algorithm that can provide information on the distribution of stripe rust disease centers and achieve early warning from a whole-field perspective.The multiple spectral information of the experimental field was extracted by UAV and various on-board sensors such as multispectral and visible light.Aiming at the characteristics of small and numerous early stripe rust incidence centres in remote sensing images,a small target feature enhancement algorithm and its functional modules(ESCD)was designed and incorporated into the Yolov5 target detection model for structural modification.The combination of Vari Focal Loss function and Swish activation function further enhances the detection of small incidence centres.The results show that the newly proposed Yolo-EVS model achieves a detection accuracy of 95.53%(m AP)and an image inference time of about 15.9 ms/frame.With the prediction time only 0.4 ms/frame slower than the original model,the detection accuracy was improved by 10.02%.The total number of onset centres in the experimental field was counted as 1628 by manual survey and visual interpretation,while the number of onset centres predicted by the Yolo-EVS model was1607,and the recall of the number of onset centres reached 98.71%(Recall).The results demonstrated that the improved algorithm and model effectively improved the leakage and misdetection of small sized incidence centres in the field,and achieved high-precision and fast target detection of wheat stripe rust incidence centres from a whole-field perspective.Meanwhile,it also proves the feasibility of obtaining RGB images through UAV for early warning of wheat stripe rust centers,and has the advantages of simple process and operation,flexibility,and low cost,which has the value of promotion and application.(3)A semantic segmentation algorithm is proposed,which can be combined with geographic information to realise the segmentation of wheat stripe rust incidence areas,and with the Visible Atmospheric Resistance Index(VARI)to realise the severity assessment of each incidence area.During the experiment,it is verified that the multimodal data can be used as an effective reference for the production of semantic segmentation labels of disease-affected areas,and it is also verified that the multibranch binary classification and the integrated learning framework have a positive effect on the improvement of semantic segmentation accuracy.:the overall instance segmentation accuracy of the onset region reached 94.13%(mean F1 score),improved by 10.72% over single-branch multiclassification,and the overall segmentation accuracy was improved by 6.39%(mean F1-score).Precise instance segmentation of wheat stripe rust onset regions was achieved,while the recall of the number of onset regions reached 96.03%(Recall),the accuracy of the pixel-by-pixel detection of onset region contours reached 93.00%(F1-score),and the intersection ratio of onset region and non-onset region contour segmentation reached 91.02%(FWIo U).Meanwhile,the reliability of the results of instance segmentation of wheat stripe rust onset areas and canopy level severity grading assessment was verified by leaf level severity manual survey,in which the accuracy of severity assessment matched with the results of 316 ground survey sites reached 94.62%.Combined with the results of the aforementioned target detection of onset centre,the quantitative data such as the centre coordinates of wheat stripe rust onset area,onset area,onset area contour and the severity of each onset area were automatically obtained,and the automated mapping of the onset severity grading results with obvious colour gradient differentiation was also achieved.The final results proved the feasibility of using the VARI index to calculate the severity of onset areas,at the same time,the feasibility and reliability of using modified algorithms(SUNet)for instance segmentation of wheat stripe rust disease incidence areas in UAV remote sensing images have been demonstrated in terms of technology and performance,which can provide quantitative data support for the subsequent precision application operations.This study used deep learning techniques to identify and monitor wheat rust at the leaf and canopy levels,respectively.In the experiment,the three major image processing tasks in the current computer vision field,including image classification,object detection,and image segmentation,were integrated from simple to complex.Three major current image processing tasks in the field of computer vision,such as image classification,target detection and image segmentation,were incorporated in the experiment from simple to complex.Firstly,the Imp-Dense Net model for wheat rust image classification task was constructed,mainly solving the crop disease recogniting problem of "what is it?".Then,the Yolo-EVS model was constructed for the remote sensing target detection task of wheat stripe rust,which can accurately locate and statistically count the early onset centres in the field,reduce the leakage rate and false detection rate of small onset centres,and mainly solve the onset centres localizating problem of "where is it?".Finally,the SUNet model for instance segmentation of remote sensing images of wheat stripe rust was constructed to achieve the fine instance segmentation and severity grading assessment of the onset area,which provides more accurate and objective parameters for the assessment of the onset of the disease in the whole field and the formulation of prevention and control strategies,and mainly solves the crop diseases severity assessment problem of "how is it?".This study not only improves the automation and intelligence level of wheat rust identification and monitoring,but also provides an effective reference and methodology for the identification and monitoring of other crop diseases. |
