In order to improve the accuracy and efficiency of predictive modeling of the starch content of Huangshui, a byproduct of Baijiu production by solid-state fermentation, spectral information of Huangshui was collected using a Fourier transform near-infrared (FTIR) spectrometer and preprocessed by first derivative. Based on the preprocessed spectra, a predictive model for the starch content of Huangshui was developed using partial least squares regression (PLSR), and its performance was evaluated by determination coefficient (R²) and root mean square error of prediction (RMSEP). As the original spectra contained a lot of redundant information, in order to effectively improve the detection accuracy and to optimize the modeling efficiency, the advantages of different feature extraction methods were combined. Finally, it was found that the PLSR model established by using the spectral features extracted by competitive adaptive reweighted sampling (CARS) combined with the successive projections algorithm (SPA) was significantly better than the model built without feature extraction or using single feature extraction. The results showed that the R² and RMSEP of the model established using CARS were 0.9654 and 0.2012%, while those obtained using CARS-SPA were 0.9738 and 0.1748%, respectively. The spectral dimension reduced from 2203 to 126 after the combination of CARS with SPA, which improved both the prediction accuracy and the modeling efficiency. The method proposed in this study provides an effective means to optimize near-infrared spectral quantitative modeling of starch in Huangshui.