In the present work,“Fuji” apples from Shandong Yantai were used to take the diffuse reflection spectra by FT-NIR.PLS components (i.e.,factors) were computed by nonlinear iterative partial least squares (NIPALS) and the number of latent factors (LV) was optimized by a leave-one-out cross-validation procedure on the calibration set.On the basis of partial least square (PLS) regression,the models for apples’ firmness before and after peeling were compared.In order to eliminate the effect of apple peel on prediction,spectral pretreatments such as multiplicative scatter correction (MSC),derivative,direct orthogonal signal correction (DOSC) and wavelengths selection based on genetic algorithms (GA) were used.Finally,the results of different spectral treatments were compared.In conclusion,the RSDp of models for apples before and after peeling was 16.71％ and 12.36%,respectively,suggesting that the apple peel played a negative role in constructing good predictive models.Moreover,the traditional spectral pretreatments (such as MSC,derivative) can hardly resolve the problem.In this research,GA-DOSC played an important role in reducing the interference of apple peel.It not only reduced the wavelength variables from 1480 to 36,but also reduced the latent variables from 5 to 1.The correlation coefficient (r) was improved from 0.753 to 0.805,and the RMSECV and RMESP were reduced from 1.019 kgf·cm-2 and 1.197 kgf·cm-2 to 0.919 kgf·cm-2 and 0.924 kgf·cm-2,respectively.Especially,the RSDp was decreased remarkably from 16.71% to 12.89%.The performance of the model after GA-DOSC treatment was similar to the model using spectra of apple flesh (12.36%).It was concluded that the prediction precision based on GA-DOSC satisfied the requirement of NIR non-destruction determination of apples firmness.