Rapid identification of marine microorganisms is critical in marine ecology, and Raman spectroscopy is a promising means to achieve this. Single cell Raman spectra contain the biochemical profile of a cell, which can be used to identify cell phenotype through classification models. However, traditional classification methods require a substantial reference database, which is highly challenging when sampling at difficult-to-access locations. In this scenario, only a few spectra are available to create a taxonomy model, making qualitative analysis difficult. And the accuracy of classification is reduced when the signal-to-noise ratio of a spectrum is low. Here, we describe a novel method for categorizing microorganisms that combines optical tweezers Raman spectroscopy, Progressive Growing of Generative Adversarial Nets (PGGAN), and Residual network (ResNet) analysis. Using the optical Raman tweezers, we acquired single cell Raman spectra from five deep-sea bacterial strains. We randomly selected 300 spectra from each strain as the database for training a PGGAN model. PGGAN generates a large number of high-resolution spectra similar to the real data for the training of the residual neural network. Experimental validations show that the method enhances machine learning classification accuracy while also reducing the demand for a considerable amount of training data, both of which are advantageous for analyzing Raman spectra of low signal-to-noise ratios. A classification model was built with this method, which reduces the spectra collection time to 1/3 without compromising the classification accuracy.