Global estimation of phytoplankton functional types using PACE OCI hyperspectral data and a novel deep ensemble learning approach

• First global-scale monitoring of phytoplankton functional types using NASA’s PACE OCI imaging spectrometer. • Creation of the largest synthetic hyperspectral-HPLC dataset for enhanced phytoplankton functional types modeling. • Development of HyperPFT: an innovative deep ensemble learning framework for precise phytoplankton functional types retrieval. • HyperPFT outperforms conventional machine learning models, enhancing phytoplankton functional types inversion accuracy. Hyperspectral remote sensing is revolutionizing the monitoring of marine ecosystems by enabling finer discrimination of phytoplankton functional types (PFTs). However, a persistent limitation to global-scale hyperspectral PFT modeling is the scarcity of co-located field samples that simultaneously contain high-performance liquid chromatography (HPLC) pigment data and hyperspectral measurements. This study addresses this critical gap by introducing a novel two-stage framework that synthesizes the largest HPLC–hyperspectral training dataset and deploys an advanced deep ensemble learning model for global PFT retrieval from NASA’s PACE Ocean Color Instrument (OCI) data. In Stage 1, we develop an extensive synthetic training dataset by combining real HPLC–hyperspectral pairs with reconstructed hyperspectral data generated from historical multispectral ocean color observations using a generative adversarial network (GAN). This innovative spectral reconstruction strategy enables the integration of over 4,000 previously unusable HPLC samples into the hyperspectral modeling process. In Stage 2, we propose HyperPFT, a lightweight Transformer-based deep ensemble learning model specifically tailored to hyperspectral data. HyperPFT simultaneously delivers high-accuracy PFT estimations and pixel-wise uncertainty quantification, leveraging 100 sub-models trained via bootstrap resampling.Comprehensive validation against in situ datasets demonstrates that HyperPFT consistently outperforms traditional machine learning models in pigment and PFT prediction, particularly under conditions of spectral redundancy and noise. Applied to 8-day composite PACE OCI data, HyperPFT successfully maps global distributions of eight PFTs, revealing ecologically coherent biogeographical patterns and identifying uncertainty hotspots. This work marks the first global-scale application of PACE hyperspectral data for operational PFT inversion and represents a critical step toward scalable, uncertainty-aware remote sensing of marine phytoplankton diversity. The proposed approach lays a robust foundation for future hyperspectral ocean color missions and ecosystem monitoring applications.