Hyperspectral scattering is a promising technique for noninvasive measurement of apple mealiness. An uninformative variable elimination (UVE) coupled with locally linear embedding (LLE) algorithm was proposed for assessing apple mealiness. After the algorithm, the number of effective wavelengths decreased to 23.5% of full wavelengths of hyperspectral scattering images. LLE was utilized to reduce the dimensionality of images composed of effective wavelengths. Partial least squares discriminant analysis was applied to develop classification model. Compared with mean reflectance (75.8%) and UVE coupled with mean reflectance algorithm (77.4%), LLE and UVE coupled with LLE model yielded better results (79.0%). UVE coupled with LLE model with the presevation of classification accuracy only used 23.5% wavelength of LLE model. Therefore, it provides a useful algorithm for online classification and data saving.