Sorghum is a primary raw material for Chinese Baijiu production. Rapid and accurate identification of sorghum varieties is crucial for quality control in the Baijiu industry. Near-infrared spectroscopy (NIRS) offers a promising nondestructive solution, though conventional discriminant models-relying solely on category labels-lack interpretability. This study proposes a novel classification method integrating NIRS with decision-level fusion (DLF). The approach involves developing quantitative NIRS models for sorghum tannin and amylopectin contents, building a label-based sorghum variety discriminant model, and fusing predictions from these three models via DLF to calculate the new category prediction values. Experiments on 81 sorghum flour samples from diverse sources demonstrated that DLF generally improved model discrimination performance. Specifically, partial least squares discriminant analysis (PLS-DA) model preprocessed with standard normal variate (SNV) + competitive adaptive reweighted sampling (CARS) and optimized by DLF achieved an accuracy of 0.938, while the CARS-PLS-DA model with DLF reached an accuracy of 0.951. The proposed method effectively bridges quantitative and qualitative analysis, enhancing model interpretability while improving discriminatory power.