The work aims to realize rapid nondestructive detection of firmness of apples based on hyperspectral imaging technology and error back propagation (BP) network. The hyperspectral imaging acquisition system was used to acquire the hyperspectral images of postharvest 'Fuji' apples, and then the average reflectance spectra in the whole region of apple samples was extracted. The successive projection algorithm (SPA) and competitive adaptive reweighted sampling (CARS) method were used to conduct data mining of spectral data preprocessed by the standard normal variation. A BP network model for predicting firmness of apples based on full spectra and characteristic spectra was studied. The results showed that, 18 and 16 characteristic wavelengths were extracted by SPA and CARS from 256 full wavelengths, which obviously improved the working efficiency of prediction model. Moreover, SPA+BP network model had a relatively good prediction ability for firmness of apples (rp=0.728, RPm=0.282 kg/cm2). This study indicates that the prediction model based on hyperspectral imaging technology and BP network can be applied in the rapid nondestructive prediction of firmness of apples.