Identification of rice sheath blight by imaging hyperspectral camera

In this study, a 400-1000nm hyperspectral camera was applied, and FS23, a product of Hangzhou Color Spectrum Technology Co., LTD., could be used for related research. FigSpec® series imaging hyperspectral cameras use a transmission grating beam splitting module with high diffraction efficiency and a high sensitivity surface array camera, combined with built-in scanning imaging and auxiliary camera technology, to solve the traditional hyperspectral cameras require external push-scan imaging mechanism and complex focusing and other difficult problems. It can be directly integrated with the standard C interface imaging lens or microscope to achieve fast acquisition of spectral images.

Precision agriculture is an important way to achieve low consumption, high efficiency, high quality and safety in agriculture. As the largest grain crop in our country, the stable yield and high yield of rice have always been the focus of our agricultural production, and timely and effective disease control is an important guarantee to achieve stable yield and high yield. Rice leaf blight is one of the three major diseases of rice. If the cause and degree of damage of damaged crops can be detected in the early stage of rice disease, combined with variable application in fine agriculture, the disease rate of rice disease infection can be effectively reduced, the harm scope can be narrowed, and the rice yield can be effectively increased. Variable application mainly refers to the timely diagnosis of the cause and degree of damage of the affected crops according to the information of crop pests and diseases, and the application of chemical agents according to the appropriate disease treatment, local conditions and demand, so as to reduce the use of chemical agents and achieve the purpose of timely prevention and control.

In this study, hyperspectral imaging technology was used to recognize rice sheath blight. The PLS-DA discriminant analysis model was established after different pretreatment of the original spectra, and good results were obtained. Under SG, SNV and MSC pretreatment methods, the accuracy of prediction sample discrimination was 82.8%, 92.1% and 89.1%, respectively. The PLS-DA model established by SNV pretreatment spectrum had the highest accuracy, while the PLS-DA model established by SG pretreatment spectrum had the lowest accuracy, but the accuracy was more than 80%. So these three methods are feasible. The accuracy of the prediction set of LDA and BPNN discriminant models based on MNF feature information extraction is 95.3% and 98.4%, respectively, which is better than the PLS-DA model based on all bands. After comprehensive comparison of the three models, the BPNN model based on MNF feature information extraction achieves the optimal discriminant effect, and the accuracy of modeling set and prediction set is 99.1% and 98.4%, respectively. The experimental results show that the hyperspectral imaging technology can be used to identify rice grain wilt, and the MNF algorithm can be used to extract characteristic information to represent the original spectrum, and greatly reduce the calculation amount. The algorithm has a wide application prospect in the process of rapid recognition and modeling of rice disease.