3A.1 NN Technique for Producing Consistent Ocean Color Data for Assimilation in Ocean Models

Tuesday, 14 January 2020: 8:30 AM
156A (Boston Convention and Exhibition Center)
Vladimir Krasnopolsky, NOAA, College Park, MD

Satellite remote-sensing of ocean color (OC) parameters provides the only means for broadly observing the biological component of the world’s oceans. Consequently, this capability must be exploited for analysis and prediction of ocean bio-physical processes and establishing a linkage to biological components of ocean ecological forecasts. Operational integration/assimilation of ocean OC field (chlorophyll, Kd490, KdPAR) into NOAA’s operational ocean models has three fundamental requirements/conditions: 1) gaps in the observations need to be addressed, both in the current instance and for extended gaps; 2) the data being assimilated must have a long data record to facilitate compilation of a robust statistical database spanning multiple seasons; and 3) the data being assimilated must be for a predicted parameter.

In our previous work [1] we demonstrated that neural network (NN) technique [1] can be successfully used to fill both short and small (several days and several grid points), and extended (several months and global) gaps in OC satellite measurements. In this work, we show that the other two principal requirements can also be satisfied using NN technique.

Consistent Ocean Color

There exist three major OC data sets produced by SeaWiFS (09/1997 – 12/2010), MODIS (07/2002 – present), and VIIRS (1/2011 – present) sensors. These three data sets have different error statistics, and it is not a simple task to integrate them for producing one consistent long-term data set. In this study, we examined one of possible approached. We used three years of VIIRS data (the mostly accurate resent measurement) to train an ensemble of NNs. Each of the NN ensemble members performs a mapping of relevant ocean variables (SST, SSH, and upper portions of temperature and salinity profiles) to the logarithm of OC (chlorophyll-a concentration, C). Logarithm of OC, C, was selected to guarantee a higher accuracy of approximation and extrapolation provided by NNs trained on VIIRS data. For NN training the mean square error function is used. This error function is optimal for normally distributed data. OC data have almost log-normal distribution (see Fig. 1); thus, is almost normally distributed. Using tree years of VIIRS data for training and as NN output allowed as developing an ensemble of NNs capable of stable far extrapolation of OC data.

In this study, satellite-derived surface variables — sea-surface temperature (SST), sea-surface height (SSH), and gridded ARGO salinity and temperature profiles from 0 – 75m depth 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. NNs are trained using data for three years (2012 to 2014) and tested for period of 10 years (2005 to 2014). To reduce the impact of the noise in the data and to obtain a stable computation of the NN Jacobian for sensitivity studies and data assimilation [1], an ensemble of NNs with different weights is constructed. Results are assessed using the root-mean-square error (RMSE) metric and cross-correlations between observed OC fields and NN output. Measurements from different OC sensors (VIIRS, MODIS, and SeaWiFS) available during the validation period were use. Correlations between NN generated and observed OC, C, slightly decreases (from ~0.85 to ~ 0.75) with moving away from the training interval (2012 to 2014). However, RMS differences do not change significantly. Results for all three used satellite sensors are very consistent. It means that NN generated C can serve as a consistent long-term OC data for different use, including assimilating in ocean models.

Assimilating OC in ocean models.

OC, C, is not a prognostic variable in current oceanic models. Therefore, OC assimilating in the model requires a coupling of a biochemical component to the ocean model or introducing into data assimilation system (DAS) an observation operator relating C to ocean prognostic variables. NN (1) can serve as such an operator. Jacobian of NNs (1) can relate innovations in C to innovations in ocean prognostic variables in DAS[1].

References

[1] Krasnopolsky V., 2013, "The Application of Neural Networks in the Earth System Sciences. Neural Network Emulations for Complex Multidimensional Mappings", Springer, 200 pp.

[2] Krasnopolsky V., S. Nadiga, A. Mehra, E. Bayler, and D. Behringer, 2015. “Neural Network Technique for Filling Gaps in Satellite Measurements: Application to Satellite Ocean Color Observations”, Computational Intelligence and Neuroscience, Article ID 923230, December, http://www.hindawi.com/journals/cin/aa/923230/

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner