Neural Network Technique For: (a) Gap-Filling Of Satellite Ocean Color Observations, And (b) Bridging Multiple Satellite Ocean Color Missions

Abstract:

A Neural Network (NN) technique is used for gap-filling of satellite-derived ocean color fields and for bridging multiple satellite ocean color missions by linking ocean color variability, primarily driven by biological processes, with the physical processes of the upper ocean. Satellite-derived surface variable fields, sea-surface temperature (SST), sea-surface height (SSH) and sea-surface salinity (SSS), along with gridded vertical profiles of temperature (T) and salinity (S) from ARGO, are used as signatures of upper-ocean dynamics. The NN technique employs adaptive weights that are tuned using statistical learning (training) algorithms applied to past data sets, providing robustness with respect to random noise, accuracy, fast emulations, and fault-tolerance. This study uses Visible Imaging Infrared Radiometer Suite (VIIRS) ocean color fields, satellite SSS/SSH/SST fields, and gridded vertical profiles of temperature (T) and salinity (S) from ARGO. All data sets were interpolated to the same spatial (one-degree latitude-longitude) grid and temporal resolution (daily) for 2012-2014. The NN technique is trained for two years and tested on the remaining year; however, by rotating the time series, we are able to cover all three years. The NN output are assessed for bias, root-mean-square error (RMSE), and cross-correlations; and a Jacobian is evaluated to estimate the impact of each input (SSH, SSS, SST, T and S) on the NN ocean color estimates. The differences between results from an ensemble of NNs versus a single NN are examined. After training the NN for the VIIRS period, the NN is retrospectively applied to 2005-2012 data, a period covered by other satellite ocean color missions — the Moderate Resolution Imaging Spectroradiometer (MODIS AQUA) and the Sea-viewing Wide Field-of-View Sensor (SeaWiFS).