The detection of adulterants in tea using infrared spectroscopy has gained prominence. However, there has yet to be a systematic comparison of the performance of traditional machine learning methods versus deep learning in the context of modelling infrared data for tea quality. This study compares machine learning and deep learning for modeling spectral data. Machine learning methods like Random Forest, K-Nearest Neighbors (KNN), Support Vector Classification, and Gaussian Naive Bayes used the Successive Projections Algorithm (SPA) for feature extraction, while deep learning models automatically extracted features. SPA-KNN showed superior performance with 0.950 accuracy, 0.953 macro-precision, 0.950 macro-recall, and 0.950 macro-F1 score on the test set (n=80), processing in 1.33 seconds. Deep learning models such as ResNet achieved a lower accuracy of 0.688 and required a longer processing time of 335.60 seconds. This may be partly due to the limited generalization ability caused by the small sample size. Additionally, the complex structure of ResNet, such as its depth, may contribute to the longer processing time. This study offers insights for selecting appropriate methods in tea quality assessment. Machine learning methods, especially with spectral preprocessing and SPA-based feature extraction, remain effective, while deep learning may not excel in limited data scenarios due to its higher computational demands.