Research On The Detection Method Of Pear Brown Spot Based On Hyperspectral Imaging Technology

Pear is one of the five major fruits in the world.China’s pear cultivation area and total output are both in the forefront of the world.However,pear trees are threatened by various diseases during the growth process.Pear brown spot is a widespread worldwide disease,which can seriously affect the quality and yield of pear fruit.The traditional disease detection technology is highly subjective,time-consuming,and cumbersome,and the detection cost is high.Hyperspectral imaging technology can quickly and non-destructively detect pear tree diseases.In this paper,taking the leaves of pear tree brown spot as the object,hyperspectral imaging technology is used to classify and identify the disease,detect the degree of disease infection,and conduct research on the quantitative estimation of chlorophyll content(SPAD)of pear tree leaves under the stress of brown spot disease.Provide theoretical support for the qualitative and quantitative detection of pear trees under brown spot disease,and provide information support for the precision of disease control.The main research content and results are as follows:(1)Using hyperspectral imaging technology combined with machine learning to classify and detect pear tree diseased leaves,establish a discriminant analysis model for pear tree brown spot,black spot,sunburn and healthy leaves,and realize rapid and non-destructive detection of multiple types of pear tree diseases.Five preprocessing methods were used to preprocess the original spectral data of leaves,combined with support vector machine(SVM)and BP neural network modeling methods respectively,an optimization algorithm was selected to determine the parameters of the discriminant model,and a comprehensive comparison found that the standard normal transformation(SNV)The preprocessing effect of the method is the best,and the accuracy rate is more than 90%under the two types of models.In order to be further applied in practice,20 and 16 characteristic wavelengths were extracted based on the principal component loading method and the continuous projection algorithm(SPA).Model verification showed that the characteristic wavelengths selected by SPA were more effective in distinguishing diseased leaves of pear trees.The SNV-SPA-SVM model obtained the best results,and the recognition accuracy of the model for all kinds of leaves reached more than 90%.(2)Based on hyperspectral imaging technology,disease classification diagnosis of pear brown spots was achieved in both field and laboratory environments.Using the same test samples,the same equipment was used to construct the datasets in the field and laboratory environments respectively.The competitive adaptive weighting algorithm and SPA algorithm were used to extract feature wavelengths and combined with SVM models to diagnose brown spot leaves at all levels,and the results showed that the SVM models were generally accurate in detecting the disease leaves at all levels in the field dataset.Therefore,a combined model using a convolutional neural network and SVM was proposed for the graded diagnosis of brown spot disease in the field.The model led to a significant improvement in the discriminatory effect,and the accuracy of the test set reached 83.75%,which has a certain guiding significance for the control of pear diseases in practical production.(3)A model for the quantitative detection of chlorophyll content in pear leaves under brown spot disease stress was constructed based on hyperspectral imaging in the field.The differences between different disease levels and chlorophyll content were analyzed,and the optimal spectral indices of DI,VI,NDVI,and m NDI were determined using the correlation matrix method and fitted to the chlorophyll content using traditional functions.The hyperspectral estimation model of SPAD values of diseased leaves was further developed and compared with the PLSR,SVR,and CNN algorithms using the combination of optimal spectral indices of leaves,CARS,SPA algorithms,and the key feature wavelengths extracted by combining the two algorithms as input.The results showed that the models constructed with the optimal combination of spectral indices outperformed the traditional functional models,and the test set of the models constructed with the feature variables extracted from SPA and CARS was higher than 0.74.The SVR model with the key wavelengths extracted from CARS-SPA as input variables had the best prediction of chlorophyll content,with a test set coefficient of determination of 0.8029 and an RMSEP of 2.8856.The model was further tested using an external validation set with R~2=0.7366 and RMSE=3.5967,indicating that the CARS-SPA-SVR model is stable and generalizable,which can achieve better chlorophyll content estimation of pear brown spot disease in the field,which can provide information support for monitoring the growth status of diseased plants.