Research On Citrus Huanglongbing Diagnosis System Based On Edge Computing

Huanglongbing is a disease easily infected by citrus plants,which spreads quickly and does great harm.Now,the disease has appeared in many countries,causing a devastating blow to the local citrus industry.However,there is no cure and only prevention can be used.In order to increase the yield of citrus and reduce the economic loss,it is of great significance for the citrus industry in China and even in the world to find the plants infected with Huanglongbing accurately,efficiently and quickly and cut off the source of the disease.In this thesis,the hyperspectral imaging system is used to obtain the data set,and the deep learning method and edge computing technology are introduced to study the orange Huanglongbing diagnosis system.Firstly,the hyperspectral imaging system was used to scan and image the leaf samples of two varieties of Citrus respectively,and the hyperspectral map data were collected;Secondly,the recursive feature elimination algorithm is used to screen the main feature bands and extract the image data under the band,so as to construct the image data set.At the same time,the spectral data are preprocessed by multiple scattering correction(MSc),savitzky Golay revolution smoothing and wavelet transform(WT)respectively,so as to construct the spectral data set;Thirdly,the convolution neural network and recurrent neural network are used to analyze the image data set and spectral data set;Finally,the deep learning model is deployed in the edge computing platform including field programmable gate array(FPGA),graphics processing units(GPU)and application specific integrated circuits(ASIC),so as to realize the construction of Huanglongbing edge computing system.1)In order to detect citrues Huanglongbing,this thesis compares and analyzes the actual performance of cyclic neural network and recurrent neural network in this problem.The results show that the traditional cyclic neural network will show the problem of gradient dispersion,and the improved cyclic neural network: long short term memory(LSTM)and gate recurrent unit(GRU)can make up for this defect.In addition,compared with convolution neural network algorithm,the improved cyclic neural network has faster reasoning speed,higher accuracy,smaller model,stronger model generalization ability,and less hardware resources.It is suitable for deployment in edge computing devices that require power consumption control,and finally realize edge computing.2)Because there is some noise in the collected spectral data,which affects the model training and prediction accuracy,the preprocessing of spectral data is conducive to the improvement of model accuracy.The results show that the reasoning accuracy of spectral model is greatly improved after preprocessing,and SG convolution smoothing algorithm is the best.The comparative test shows that the prediction accuracy of the models trained by LSTM + SG convolution smoothing and GRU + SG convolution smoothing can reach more than 99.66%,followed by the models trained by MSC + LSTM and MSC + GRU,and the prediction accuracy can also reach more than 98.82%..3)Since the model of convolutional neural network is large and the FPGA resources are relatively small,in order to enable the trained model to obtain better speed and accuracy in edge calculation,this thesis adopts model compression algorithm and model optimization strategy to carry out hardware acceleration.The experimental results show that the FPGA platform can achieve better energy efficiency ratio than PC in the process of edge calculation with less loss of reasoning precision and reasoning speed.4)The data precision of dataset will directly affect the edge computing speed,memory consumption and other indicators.Therefore,this paper uses the maximum quantization algorithm to quantize the original spectral data set,converting it from fp32 data type to int8 data type,and test it in the edge computing platform.The results show that under the premise of maintaining the prediction accuracy,the size of the spectral data set is reduced by 67.77%,and the computing speed in the LSTM model and Gru model is increased by 15.26% and 10.13% respectively,which significantly reduces the memory and bandwidth overhead of the edge computing devices and improves the overall energyefficiency ratio of the system.