This paper aims to select the spectral feature grouping method and establish regression detection models based on feature grouping and successive projection algorithms (SPA) feature optimization. At first, the hyperspectral images and sugar content of 174 apple samples were collected to extract the hyperspectral data using ENVI 4.7 software. After pretreating the spectral data by using multiple scattering correction (MSC), all samples were stored according to sugar content. In the case that the training set and the test set were divided into different groups at the same allocation ratio, the full-band grouping modeling method and the continuous projection extraction feature wavelength grouping modeling method were used to model the hyperspectral data and the sugar content data, respectively. By comparing the prediction effect in terms of full-band group modeling in which we character the grouping method as 1, 2, and 3, respectively. In the case of grouping method 1, i.e., when the test set and training set were randomly assigned according to the set ratio column, the prediction effect of the model established was optimal. Nonetheless, the simulation result was stable because it was the averaging R of 100 running results. In the case of grouping method 2, i.e., when the training set and the test set were divided using interval allocation according to the set ratio, the prediction effect of the model, was slightly worse than that of grouping method 1. Grouping method 3, i.e., the training set and the test set were grouped in series according to the set ratio, has the worst prediction effect of the model. Results indicated that the quality of the established model is not only related to whether the characteristic wavelength is extracted, but also has a great relationship with the grouping method. It is of great significance to select the appropriate grouping method of training set and test set for the research object to improve the accuracy of the model. This provides a theoretical basis for future research.