The Good the Bad and the Ugly of Single Sensor Error Statistics for Sea Surface Temperature: What Do Spaghetti Westerns and Quality Levels Have in Common?

Friday, 19 December 2014: 3:10 PM
Guillermo P Podesta1, Katherine Ann Kilpatrick2, Robert Evans1 and Peter J Minnett3, (1)Univ Miami / RSMAS, Miami, FL, United States, (2)University of Miami, Miami, FL, United States, (3)Univ Miami, Miami, FL, United States
Global High Resolution Sea Surface Temperature (GHRSST) L2p products are available from many different satellite-based instruments. Assimilation and fusion of SSTs from multiple sources requires knowledge of the expected accuracy and uncertainly of a retrieval and confidence that a pixel is within the uncertainty required by the application. The GHRSST L2p files provide quality levels with labels such as “worst”, “acceptable”, or “best.” But what do these levels actually mean to a user in terms of uncertainty and bias? Does “acceptable” have the same meaning for data providers and users? Currently there is no standard or consensus for the accuracy requirement of a quality level. GHRSST providers rely on matchup databases of satellite and in situ SSTs to provide Single Sensor Error Statistics (SSES). The GHRSST L2P MODIS SSES values currently are stored in a 6-dimensional Look Up Table (LUT) – often referred to as a hypercube. This LUT lists uncertainties stratified by quality level, season, latitude, viewing geometry, surface temperature, and “wet” or “dry” atmospheres. While this approach is more useful than a single aggregate SSES estimate, the coarse nature of the hypercube bins produces obvious discontinuities in uncertainty fields at some bin boundaries when geographically mapped. In reality, the SSES should vary more smoothly as a function of different combinations of factors influencing the accuracy of retrievals. We will present ongoing efforts to model the SST bias and uncertainty as a smooth function of some of the dimensions of the SSES hypercube. Additionally, we aim to classify pixels into quality categories with explicitly defined ranges of bias and uncertainty. The ultimate goal is to develop methods that could be deployed across multiple sensors to establish a standard for objective, quantitative definitions of SST retrieval quality.