This study introduces a novel machine learning-based framework, termed the in-situ multi-excitation fluorescence machine learning (mf-ML) method, for the high-resolution quantification of phytoplankton communities. By integrating in-situ multi-excitation chlorophyll fluorescence data (collected across nine wavelengths: 375–590 nm) with traditional high-performance liquid chromatography (HPLC) and CHEMTAX analyses, the framework overcomes the spatial and temporal limitations of discrete sampling and satellite remote sensing. Field validation was conducted in the coastal waters of Yeosu, South Korea, during a harmful algal bloom (HAB) warning period. Feature engineering—including band ratios, continuum-removed spectra, and recursive feature elimination—was combined with Bayesian hyperparameter optimization to train an ensemble XGBoost model. The model demonstrated exceptional predictive accuracy across eight major phytoplankton groups, achieving an average coefficient of determination (R^2) of 0.977. Notably, the mf-ML framework successfully reconstructed the continuous, high-resolution vertical profiles of the phytoplankton community, accurately capturing the diel vertical migration (DVM) of dinoflagellates down to the pycnocline (8–10 m) at night and their ascent by day. It also resolved the subtle vertical distribution shifts of non-migrating groups in response to solar irradiance. This study highlights multi-wavelength excited fluorescence spectrometry combined with machine learning as a cost-effective, robust, and high-resolution tool for in-situ marine monitoring. By providing continuous insights into phytoplankton behavior and community structure, this approach offers a significant advancement for marine ecosystem management and proactive HAB early-warning systems.