A Self-Attention-Based Deep Learning Model for Estimating Global Phytoplankton Pigment Profiles

Characterizing global phytoplankton pigment profiles in the ocean is crucial for comprehending phytoplankton dynamics. This study develops a self-attention-based deep learning model (Pigmentsformer) to estimate nine types of global phytoplankton pigment profiles in the top 300-m ocean. Pigmentsformer employed 15-day sequences of ocean-color data to capture the dynamic changes in phytoplankton growth. The model was trained on an extensive collection of matchups, including 33 975 samples between global high-performance liquid chromatography (HPLC) in situ measurements, ocean-color satellite measurements, and the environment fields from the reanalysis dataset. Validation of the model employs a tenfold leave-one-out cross-validation (LOOCV) approach. The coefficient of determination ( ${R}^{2}$ ) between in situ HPLC and Pigmentsformer-estimated concentrations ranges from 0.67 to 0.87, with the mean absolute error (MAE) ranging from 0.01 to $0.33~{\text {mg}\cdot }{\text {m}}^{-3}$ for nine types of pigments. The backpropagation technique reveals that among all predictors, optical properties are paramount when estimating total chlorophyll-a (TChla) at the surface, with ocean current being the most influential environmental factor. However, as depth increases, the effect of environmental variables SSH and temperature exceed that of optical properties and ocean current. An examination of 20 years of model-generated phytoplankton size classes (PSCs) was conducted to explore the correlation between changes in phytoplankton communities and the El Niño-Southern Oscillation (ENSO) in the Equatorial Pacific. The location of the maximum phytoplankton layer has a positive relationship with Niño 3.4 index ( $R =0.70$ for micro-phytoplankton, $R =0.68$ for nano-phytoplankton, and $R =0.45$ for pico-phytoplankton) within the Equatorial Pacific of the Niño 3.4 region.