Nondestructive detection of apple mealiness was investigated by using neighborhood rough set theory and hyperspectral scattering image technology. Spectral scattering profiles between 600nm and 1000nm were acquired by hyperspectral scattering image system for 576 apple samples. The optimal wavelength sets were chosen from 81 raw characteristic attributes by neighborhood rough set. 526 samples were selected randomly for calibration set and 50 samples were selected for validation set to develop classification model using optimal wavelengths coupled with support vector machine (SVM). Simulation was repeated 10 times to investigate the ability of classification model. Results showed that neighborhood rough set could select 14 optimal wavelengths effectively. The validation model using 14 optimal wavelengths yielded better result (classification accuracy 75%) than the full spectrum model (classification accuracy 71%) and the principle component analysis algorithm (classification accuracy 74%).