Phytoplankton is crucial for marine ecosystems and serves as an essential indicator of marine quality. However, identifying species is challenging due to their variability, and traditional methods like expert microscopic counting or high-throughput flow cytometry are often time-consuming or inaccurate. This study investigates the application of embedded feature reduction techniques to improve phytoplankton species classification. Using the Random Forest (RF) algorithm, we reduce the initial set of 72 features to the 30 most significant ones, resulting in a notable increase in classification accuracy to 99.47%, compared to 98.24% with the full feature set. Our findings emphasize that not all features contribute equally to classification performance, and removing irrelevant or noisy features can significantly enhance model effectiveness. Additionally, the RF-selected features demonstrate superior class separation compared to those chosen by XGBoost, emphasizing the effectiveness of RF for developing robust phytoplankton classification models.