<p>The aroma type and quality grade of Tieguanyin tea directly influence its market value and consumer acceptance, but conventional sensory evaluation and laboratory chemical analyses are often limited by subjectivity, operational complexity, or low throughput. This study aimed to develop a rapid and non-destructive method for the simultaneous discrimination of Tieguanyin aroma types and quality grades by integrating hyperspectral reflectance and transmittance imaging with machine learning. Reflectance and transmittance hyperspectral images were acquired from 72 independent sample units covering 12 categories across three aroma types and different quality grades, and 1728 ROI-level mean spectra were extracted. Raw and preprocessed spectra were used to establish single-spectrum classification models with convolutional neural network (CNN), support vector machine (SVM), random forest (RF), and multilayer perceptron (MLP) classifiers. Based on the optimal single-spectrum configurations, data-level fusion and feature-level fusion models were further constructed, with competitive adaptive reweighted sampling (CARS) and uninformative variable elimination (UVE) used for characteristic wavelength selection. The results showed that reflectance spectra were more effective for aroma type discrimination, whereas transmittance spectra provided complementary information for quality-grade identification. Data fusion further improved classification performance. The UVE-CNN feature-level fusion model achieved the best result, with a test accuracy of 96.53% for the 12-category task and 100% accuracy for aroma type identification, while also reducing training and prediction time after wavelength selection. These findings provide a useful reference for objective grading, intelligent quality assessment, and industrial discrimination of Tieguanyin aroma types and quality grades.</p> Graphical Abstract <p></p>