Monday, August 7, 2017 \ 12pm, noon, 401 Akasofu
Max Yaremchuk’s talk will discuss the following:
A modification of the gaussianization technique is proposed for assimilating observations whose errors cannot be described by the Gaussian statistics due to the natural bounds of their variability, such as ice concentration (IC). The method exploits similarity between the univariate PDFs of IC innovations and increments when observations are abundant, such as SSMI/IMS data from satellites. In this case, utilization of a single gaussianization transform in both observation and state spaces provides a significant improvement of the forecast skill compared to non-gaussianized assiimilation.
To account for the errors in the PDF estimation, degaussianization of the innovations is augmented by an additional algorithm optimizing the estimate of the background error variance field with respect to the innovations obtained after the analysis. Performance of the technique has been tested with a regional configuration of the CICE model in the Beaufort Sea at 2 km resolution driven by the operational pan-Arctic run of the ANCFS system in September-December 2015. Assimilation experiments have shown faster convergence of the iterative solver in the gaussianized case and 5-10% better 24-hour forecast skill compared to the operational system, which currently employs an isotropic correlation model without gaussianization.