Flat black tea (FBT) has been innovatively developed to alleviate homogenisation competition, but the dynamic changes in aroma components during the process remain unclear. This study employed HS-SPME-GC-MS to analyse the aroma components of tea samples from various processing stages of FBT, and to make a comparative assessment with conventional strip-like Congou black tea (SBT). Additionally, a proteomic analysis was conducted on fresh leaves, withered leaves, and frozen–thawed leaves. Significant changes were observed in the aroma components and proteins during the processing. The results of the multivariate and odour activity value analysis demonstrated that the principal aroma components present during the processing of FBT were linalool, (E)-2-hexen-1-al, methyl salicylate, geraniol, hexanal, benzeneacetaldehyde, (Z)-3-hexenyl butyrate, dimethyl sulphide, 2-methylbutanal, 2-ethylfuran, nonanal, nonanol, 3-methylbutanal, (Z)-3-hexen-1-ol, 2-pentylfuran, linalool oxide I, and β-myrcene. Freezing–thawing and final roasting are the key processing steps for forming the aroma quality of FBT. The final roasting yielded a considerable quantity of pyrazines and pyrroles, resulting in a high-fried aroma, but caused a significant reduction in linalool, geraniol, β-myrcene, and esters, which led to a loss of floral and fruity aromas. The freezing–thawing treatment resulted in an accelerated loss of aroma substances, accompanied by a decrease in the expression level of lipoxygenase and 2-C-methyl-D-erythritol 2,4-cyclodiphosphate synthase. The formation of aroma substances in the linoleic acid metabolic pathway and terpenoid metabolic process was hindered, which had a negative impact on tea aroma. This study elucidates the causes of unsatisfactory aroma quality in tea products made from frozen tea leaves, providing theoretical support for the utilisation of frostbitten tea leaves, and helps us to understand the mechanism of aroma formation in black tea.

The sensory evaluation of Longjing tea quality relies predominantly on subjective human judgment. Straightness is a critical sensory descriptor in the morphological quality assessment of Longjing tea, which lacks objective and quantifiable criteria. To objectively quantify this descriptor from a morphological perspective, this study developed an interpretable machine learning framework that integrates computer vision and SHAP analysis to digitally characterize the sensory attribute of straightness. Nine morphological features were extracted from Longjing tea samples collected across major production regions, including six contour-based and three skeleton-based descriptors. Three classification models, namely random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM), were constructed and compared based on morphological features. SHAP analysis was employed to interpret model decisions, identifying three core morphological determinants, length ratio (lr), standard deviation (sd) and rectangularity (R), which consistently ranked highest in both univariate effect sizes and cross-model feature importance, with Cliff’s δ of 0.93, 0.81 and −0.78, respectively. By establishing a quantitative link between morphological features and sensory terminology, this study offers an interpretable tool for automated tea quality evaluation, with broader implications for digital standardization in quality evaluation of agricultural products.