With the aim to understand the relationship between internal properties and nearinfrared (NIR) characteristics of apples during late developmental period, and provide a basis for field management and harvest in time, NIR diffuse reflection technology was used to measure the absorbance of ‘Fuji’ apples over the wavelength range of 833~2500Nm during the last three months of fruits’ late developmental period. Then, the internal qualities (soluble solids content (SSC), firmness (F), pH value and moisture content (MC)) of apples were measured. The linear correlations between each internal quality and the light absorption intensity at a single wavelength were analyzed. The results showed that there were weak linear correlations between the internal quality and the light absorption intensity at a single wavelength. It was difficult to predict the internal qualities of apples based on the intensity of light absorption at a given wavelength. Therefore, combined with chemometrics, the least squares support vector machine and extreme learning machine (ELM) models were established for predicting SSC, F, pH value and MC, and the effect of three data reduction methods (principal component analysis (PCA), successive projection algorithm (SPA) and uninformative variable elimination (UVE)) on the prediction performance of models was analyzed. Modeling results revealed that the optimal models for predicting SSC and pH value were SPA-ELM, whose RMSEPwas 0.4435°Brix and 0.0068, respectively;the optimal models for F and MC were PCA-ELM, whoseRMSEP was 0.2612kg/cm2 and 0.6235%, respectively. Comparing three kinds of data reduction methods, SPA had better data reduction effect than those of PCA and UVE, which not only could make the model have better prediction performance and robustness, but also have obvious data reduction effect. The number of characteristic wavelength extracted by SPA was only 0.29%~0.53% of the original data.