Optimization of a neural network-based biological model for chlorophyll-a concentration in the upper ocean

Integrating/assimilating satellite ocean color (OC) fields (chlorophyll - a, Kd 490, Kd PAR) in NOAA's operational ocean models requires scientifically consistent and robust techniques to obtain long (several years long) time series of OC fields. In addition, in order to dynamically take into account the OC signal in ocean and coupled climate models, a biological model for components of OC is required. In this work, we introduce one possible approach based on a Neural Network (NN) technique, linking Chl-a variability -- which is primarily driven by biological processes -- with the physical processes of the upper ocean, using NN-based biological model for Chl-a. A NN method for correlating satellite OC fields with other assimilated satellite and in situ observations: a) instigates fewer assimilation errors (since the inputs to the NN are already being assimilated), b) reduces reliance on sparse in situ OC observations, and c) provides a dynamical feedback between biological and physical processes in the upper ocean in ocean and coupled climate models. In this study, satellite-derived surface variables -- sea-surface temperature (SST), sea-surface height (SSH), and sea-surface salinity (SSS) fields -- and gridded ARGO salinity and temperature profiles from 0-75m depth are employed as signatures of upper-ocean dynamics. OC Chl-a fields from NOAA's operational Visible Imaging Infrared Radiometer Suite (VIIRS) are used, as well as MODIS and SeaWiFS Chl-a concentrations. Different methods of optimization of the NN technique are investigated. Results are assessed using the root-mean-square error (RMSE) metric and cross-correlations between observed OC fields and NN output. To reduce the impact of the noise in the data and to obtain a stable computation of the NN Jacobian, an ensemble of NN with different weights is constructed. The results for the ensemble mean are compared with those for a single NN. This study demonstrated that the NN technique provides an accurate, computationally cheap method to generate long (up to10 yearlong) time series of consistent Chl-a concentration, which are in good agreement with Chl-a data observed by different OC sensors during this period. It is noteworthy that a single NN (or a single NN ensemble) is capable of generating OC fields all over the globe (at all grid points of the global grid). Also, the accuracy of NN prediction does not deteriorate during the validation period: NN trained on three years of data (2012 and 2013) performs well during the ten years (2005-2014) validation period. These results demonstrate a very good generalization ability of the NN both in terms of spatial and temporal generalization. It means that the NN-based empirical biological model for Chl-a can be used in the oceanic models and coupled climate prediction system to dynamically take into account biological processes in the upper ocean.