Using Neural Networks to Fill Gaps in Satellite Measurements

Integrating/assimilating satellite data into operational atmospheric and ocean models requires scientifically consistent and robust techniques to address data gaps. In this work we introduce one possible approach based on a Neural Network (NN) gap-filling technique, which derives an empirical transfer function linking satellite derived fields of interest with other assimilated satellite and in situ observations. As a testbed for development and validation of this method, satellite-derived Ocean Color (OC), the chlorophyll-a concentration in the upper ocean in particular, was used. The OC variability is primarily driven by biological processes that are related and correlated with the physical processes of the upper ocean. Satellite-derived surface variables (sea-surface temperature (SST), sea-surface height (SSH), and sea-surface salinity (SSS) fields) and upper layers of Argo salinity and temperature profiles are employed as signatures of upper-ocean dynamics. OC fields from NOAA's operational Visible Imaging Infrared Radiometer Suite (VIIRS) are used, as well as NOAA SSH and SST fields and NASA Aquarius mission SSS fields. The OC data correlations with the SSH/SST/SSS fields are spatially and temporally dependent. The NN transfer function is trained, using global data for two years (2012 and 2013), and tested on independent data for 2014. Results are assessed using different statistical metrics applied to the difference between observed global OC fields and NN generated fields. To reduce the impact of the noise in the data and calculate a stable NN Jacobian for sensitivity studies, an ensemble of NNs with different weights is constructed. The results for the ensemble mean are compared with those for a single NN. Finally, the impact of the NN training period on the generalization ability of the NN has been evaluated by comparing performances of NNs trained on one- vs, two-year training period.