| Late blight caused by pathogenic blight infestation is a devastating disease in potato production.The disease can be harmful throughout the potato reproductive period,and in severe cases can lead to potato crop failure.To achieve the rapid detection and control of potato late blight,this paper realizes the rapid classification and early prediction of potato late blight disease based on deep learning and spectral technology,respectively.A small,simple,and adaptable potato late blight detector was developed using hardware including a Raspberry Pi,a camera,a fiber optic spectrometer,and a humidity sensor for secondary development.It can classify disease levels based on the external morphological characteristics of leaves and also can realize the early discrimination of disease and the prediction of the epidemic period based on leaf spectral information when no symptoms appear.In addition,early warning and control information can be issued based on the measurement results.The main research and conclusions of this paper are as follows:（1）Potato late blight detection and discrimination based on deep learning.Firstly,a dataset of potato leaf diseases in single and complex contexts was constructed,which contained seven types of potato leaf images.It was augmented using data enhancement methods with an improvement in class imbalance.Thereafter,the performance of seven pre-trained models for fine-grained classification of potato leaf diseases was evaluated comprehensively in terms of accuracy,inference speed,and number of parameters.The Shuffle Net V2 2×model with better generalization ability and faster inference was selected from among them and modified.Three improvement strategies were proposed:introducing an attention mechanism module,reducing the depth of the network,and reducing the number of1×1 convolutions.Their effects on the performance of the underlying model were explored through ablation experiments,and the best form of improvement was determined.The improved model reduced the number of parameters and computation by about 23%,increased the classification accuracy by 0.14%,improved the CPU inference speed to 30.2 frames s-1,and achieved an accuracy of over 85%and a recall of over 90%for all 7 types of images.Finally,a lightweight segmentation network S-UNet suitable for embedded devices was proposed,which used Shuffle Net V2 with ECA module as the backbone network of U-Net for feature extraction.The m Io U,m PA,m Precision and m Recall for the segmentation of four types of pixels:background,leaf,early blight and late blight were 84.62%,91.23%,91.84%and 91.23%,respectively.The S-UNet model was further optimized using pruning and quantization methods.The S-UNet model was further optimized using pruning and quantization methods.The compressed model model maintained a high segmentation accuracy,the model size was reduced to 21.56 M,and the inference speed on GPU and CPU was improved to 26.5 FPS and 21.4 FPS,respectively.Two deep learning models were established to discriminate the level of potato late blight and segment the disease spot area,respectively,with a high recognition accuracy and a fast inference speed,which provided important technical support for the automatic identification of potato late blight.（2）Early discrimination and epidemic prediction of potato late blight based on spectroscopic techniques.Firstly,the response of spectral transmittance and physicochemical values at different disease stages were analyzed and found that spectral transmittance decreased with the aggravation of disease,POD activity decreased slightly and then increased significantly,and SPAD values decreased continuously,indicating that the changes of transmittance and physicochemical values could reflect the degree of disease.Secondly,a systematic combination of preprocessing,downscaling and classification methods was implemented to discriminate disease classes based on spectral transmittance and physicochemical values,respectively.For the transmittance-based classification model,the choice of classification method was the most important,followed by the dimensionality reduction method,and the established MC-NCA-KNN classification model had the best classification effect with an accuracy of 96.89%,and the model transfer using the direct correction algorithm helped to establish a robust spectral diagnosis model.The establishment of a high-performance physicochemical value prediction model is a prerequisite for classifying disease levels based on physicochemical values.SG was used for pretreatment and combined with UVE for wavelength selection to improve the prediction of both physicochemical values,and then GBDT was used to classify the predicted physicochemical values with an accuracy of 95.35%.In addition,there were significant differences in reflectance at sensitive wavelengths that were extremely relevant to both physicochemical values,confirming the feasibility of rapid non-destructive determination of leaf physicochemical values based on spectroscopic techniques for potato late blight disease classification.Finally,the feasibility of predicting the epidemic period based on the changes of physicochemical values was further investigated,and it was found that humidity and incubation time were two important factors affecting the changes of physicochemical values.The Rp2 of the fitted POD activity and SPAD value regression models based on these two factors were 0.9432 and 0.8754,respectively,and the final combination of regression and classification models based on physical and chemical values achieved the prediction of the late blight epidemic period with 92.8%accuracy.The study showed that the rapid non-destructive determination of leaf physicochemical values based on spectroscopic techniques for qualitative discrimination and epidemic prediction of late blight diseases was feasible and provided a theoretical reference for early detection and control of potato late blight.（3）The potato late blight detector was developed and tested for its performance.Firstly,the overall design of the hardware system of the detector was conducted based on the functional requirements,the functions and hardware composition of each module were clarified.The designed detector includes four parts:an acquisition module,a control module,a display module,and a power module,where the acquisition module includes a temperature and humidity sensor,a camera,a spectrometer,and accessories to obtain phenotypic information about the potato plant and its environmental temperature and humidity;the control module uses an ARM control board to communicate with and power the hardware of the acquisition module;the display module uses a touch screen to facilitate user interaction;and the power module uses an uninterruptible power supply to meet the requirements of field use.Subsequently,the hardware of each functional module was selected,and the structural design of the detector housing was implemented based on the connection relationship of the hardware system.After that,the overall design of the software was carried out based on Python language and using layered architecture to clarify the correspondence between the human-computer interaction layer,functional module layer,and device driver layer,and the corresponding program was designed and written according to each software function,and the interactive interface was designed using Qtdesigner and Py Qt5 to complete the development of portable GUI application.Finally,the performance of the detector was tested from three aspects:stability,accuracy and real-time.During the test,the CPU occupancy and memory occupancy were kept below 50%.The accuracy of classifying late blight based on images and spectra was94%and 92%,respectively,and the accuracy of predicting the epidemic period of late blight based on spectra was 87%.The average time required for image detection and spectral detection was 10.9 s and 6.2 s,respectively.The results showed that the detector software and hardware operation was stable and reliable,the detection accuracy was high,and the device had practical application value. |
PDF Full Text Request