Hyperspectral scattering image technology is an effective method for nondestructive measurement of internal qualities of agricultural products. However, hyperspectral scattering images contain a large number of redundant data that affect the detection performance and efficiency. A new semi-supervised affinity propagation (AP) (NSAP) algorithm coupled with partial least square regression was proposed to select the feature wavelengths from the hyperspectral scattering profiles of “Golden Delicious” apples for predicting apple firmness and soluble solid content (SSC). Six hundred apples were analyzed in the experiment, 400 of which were used for the calibration model and the remaining 200 apples were used for the prediction model. Compared with full wavelengths, the number of effective wavelengths for apple firmness and SSC prediction selected by NSAP, respectively, decreased to 28 and 40 %. The root mean square error of prediction decreased from 6.6 to 6.1 N and from 0.66 to 0.63 %, respectively, whereas the correlation coefficient increased from 0.840 to 0.862 and from 0.876 to 0.890, respectively. Better prediction accuracy was achieved by the prediction model using selected wavelengths by NSAP than that by traditional AP, SAP, and genetic algorithm. The NSAP approach provided an effective means of wavelength selection using hyperspectral scattering image technique.