DEEP LEARNING-BASED ESTIMATION OF PHYTOPLANKTON PIGMENTS DISTRIBUTION FROM SATELLITE REMOTE SENSING

A novel deep learning approach based on self-attention was developed to estimate global pigment profiles in the ocean using a comprehensive dataset of HPLC measurements. The model successfully handled time-series satellite data even when some values were missing, providing accurate estimations. By analyzing the model's gradients, visual attention results were obtained, revealing the significance of different satellite data or products in estimating phytoplankton pigments across various regions and depths. Remote sensing reflectance (Rrs) was found to be the most important parameter for estimating total chlorophyll a (Tchla) on the ocean's surface and subsurface. The trained model was employed to generate global concentration profiles of pigments. Using the diagnostic pigment analysis (DPA) method, the distribution of phytoplankton size classes (PSCs) at different depths was calculated to analyze phytoplankton communities. This research highlights the potential of deep learning methods in advancing marine ecology studies.