When using near infrared spectroscopy to predict the moisture content of potato leaves, a large amount of spectral data needs to be processed, resulting in a time-consuming and labor-intensive calculation process. This paper proposes to use a variety of feature wavelength extraction methods to reduce the amount of calculation of near-infrared spectral data, and according to the comparison of prediction results, the feature wavelength extraction method with the best extraction effect is obtained. First, the spectral reflectance information of 110 fresh potato leaves in the 900~2100nm band is collected, and then the Regression Coefficient (RC), Principal Component Analysis (PCA), first-order derivative correlation extraction are used respectively Method, extract the characteristic wavelengths from the full-band spectral data, and finally establish a BP neural network prediction model according to the characteristic wavelengths extracted in three different ways, and compare the prediction results to obtain the optimal characteristic wavelength extraction method. The results show that the BP neural network model established by the characteristic wavelength extracted by the Regression Coefficient (RC) has the best prediction effect, the prediction set decision coefficient (R2) is 0.9698, and the root mean square error (RMSE) is 0.3177. In this experiment, on the basis of reducing the amount of near infrared spectroscopy data by more than 90%, a good prediction effect was achieved, and the purpose of quickly and concisely predicting the moisture content of potato leaves was achieved.
Two non-destructive detection methods for potato blight based on hyperspectral imaging were used: convolutional neural network (CNN) and support vector machine (SVM) to classify potato leaves. By comparing the classification results, the advantages and disadvantages of different methods are analyzed. In the experiment, normal potato leaves and early blight leaves were selected as research objects. Hyperspectral images of samples were obtained by hyperspectral imaging system, and then principal component images were extracted by principal component analysis method. It was found that the principal component images of normal leaves and blight leaves were significantly different, and finally two models of blight detection were established for convolutional neural network and support vector machine. The experimental results showed that the convolutional neural network was better than the support vector function in the detection of potato blight.
Hyperspectral imaging technology is used to study methods for diagnosing and monitoring potato diseases, so as to improve the efficiency and accuracy of disease diagnosis and monitoring, and to solve the information acquisition, processing and analysis of potato diseases. The experiment took potato late blight as the research object, using hyperspectral image acquisition system to collect the hyperspectral image data of potato plants for different days. After disease inoculation, in the hyperspectral image data in the range of 366-976 nm, feature extraction method was used to select the image corresponding to the characteristic wavelength and optimal principal component image. Hyperspectral imagery of potato plants was used to obtain spectral reflectance at different disease stages, and then the disease incidence trend of inoculation of late plague was studied to realize disease monitoring. The research results show that using hyperspectral imaging technology can achieve rapid and accurate diagnosis of potato disease.