At present, the detection of dry rot and potato scab was completed by manual visual inspection, and the detection results were subjective. This experiment studied the spectral detection method for classification and recognition of normal, dry rot and scab of potato. 116 potato samples were collected in the experiment, and the spectrum collection range was 860~1 745 nm. After the first derivative (FD) processing, the principal component analysis (PCA) classification recognition effect was better, and FD was used as the spectral preprocessing method. The shape of the spectral curve was determined by the extreme points on the spectral curve, the midpoint between the extreme points and the slope line between the extreme points. The shape change of the spectral curve represented the change of the internal substance and had fingerprint characteristics. The mode eigenvector was composed of the spectrum corresponding to the key points or the line slope between the extreme points. The average spectra of the key points of the three samples were used to form the standard pattern feature vectors. By calculating the Mahalanobis distance between the feature vectors composed of the key points of the tested samples and the standard pattern feature vectors, the minimum Mahalanobis distance was used to determine the attribution of the samples, and the error recognition rate tested the recognition performance of the model. There were 13, 12 and 15 key points in normal, dry rot and scab samples, respectively. The pattern feature vector was composed of the reflectance corresponding to each key point, and the error recognition rate of the three types of samples was zero. By removing redundant key points and integrating them into a standard pattern feature vector, the error recognition rate of normal and scab samples was zero, that of dry rot samples was 14.3%, and all were scab samples. The feature vector data points increase the fit degree between disease samples and reduces the discrimination between two types of disease samples. Using the slope between two points at the wavelength of 911, 1 269 and 1 455 nm to form the pattern feature vector, the error recognition rate of normal and scab samples was zero, and the error recognition rate of dry rot samples was 2.4%. Linear discriminant analysis (LDA) and Bayesian classifier (BC) were used to build the classification model by using the scores of the first two principal components as the parameters. Different classification models were provided. The effectiveness of the classification model based on the pattern feature vector was compared and verified. The error recognition rate of the two recognition methods was zero. The experimental results show that the pattern feature vectors representing the structural features of spectral curves could be used as the classification parameters, and the distance method could be used for modeling, which had the same recognition accuracy as the standard recognition methods.