For potato early blight at different infection periods (DPP), the spectral data is interfered with by factors such as stray light, noise, etc. In addition, a large number of bands, a large amount of data and complex bands will adversely affect the quantitative and qualitative analysis of the spectrum. 9 kinds of spectral preprocessing methods are studied, combined with the experimental results, the preprocessing methods are arranged and combined, and the 16 spectral preprocessing methods are extended and improved to combine with the continuous projection algorithm, the competitive adaptive weighting algorithm and the genetic algorithm. The band extraction methods are combined to obtain 64 spectral processing methods to optimize the original spectral data. In the convolutional neural network (CNN) classification model, most of the classification accuracy after spectral processing is significantly improved compared to the unprocessed prediction classification accuracy of 86.67%, and the classification accuracy of 12 spectral processing methods is 100%, the ideal classification of potato early blight at different disease stages can be achieved. In order to further quantitatively analyze the different infection stages of potato early blight, the spectral data processed by the spectral processing method were quantitatively analyzed using the constructed CNN quantitative estimation model. The spectral information useful to the target variable in the analysis method will lead to the result that the R2 and RMSE of the data processed by the spectral analysis method will decrease compared with the original spectral data. The fusion spectral processing method used in the study can further optimize the original spectral data. Improve the performance of the model. Among them, the CNN quantitative estimation model based on the spectral processing method combined with mean centering, multivariate scattering correction, and moving average smoothing has achieved the best results. The fitting degree of the value is 100%, and its RMSE is only 0.001 1, indicating that the deviation between the estimated value and the actual value of potato early blight at different disease stages is close to 0, indicating that the model can perfectly predict potato early blight in different disease stages. The results show that the proposed CNN can perform effective classification detection and quantitative analysis of different infection periods of potato early blight, and an effective combination of various preprocessing and feature band extraction methods according to the optimization purpose can effectively improve the modeling effect, Provide theoretical and technical support for non-destructive, accurate and intelligent detection of crop diseases.