To address the demand for precise sorghum tannin control in liquor brewing, and to overcome the inefficiency and high cost of traditional methods, this study developed a non-destructive approach by fusing features from dual hyperspectral sensors. Based on 240 representative sorghum samples covering different varieties and production regions, visible and near-infrared (VNIR) and short-wave infrared (SWIR) hyperspectral data were sequentially collected, and the tannin content was determined using standard chemical methods as reference values. Using the competitive adaptive reweighted sampling (CARS) method, characteristic wavelength bands were extracted and fused feature subsets were constructed. Combined with partial least squares (PLS), support vector machine (SVM), and convolutional neural network (CNN) algorithms, the performance of models built from both full-data concatenation and feature fusion of VNIR and SWIR data was systematically compared. The results demonstrated that the feature-based models exhibited superior performance to the full-spectrum models, while the model incorporating dual-sensor feature fusion achieved the best overall results. The fused-feature-CNN model achieved the optimal prediction performance, with values of 0.83 for coefficient of determination for the prediction set (RP2), 0.29 for root mean squared error for the prediction set (RMSEP), and 2.42 for residual predictive deviation for the prediction set (RPDP). This study confirms that the integration of multi-sensor feature fusion with deep learning strategies can provide an effective technical pathway for the rapid, non-destructive detection of sorghum tannin content and the development of online sorting equipment.