A SAR-based NWP error warning product

Verification of numerical weather prediction models at sea is complicated by the relative dearth of in situ observations. Aloft, this issue can be addressed in part by using cloud-track winds while at the surface scatterometer and passive radiometric wind speed estimates are available. Surface verification is equally challenging with satellite borne scatterometers such as ASCAT and passive techniques offering the primary means of mapping the oceanic surface wind field. In both cases the process is somewhat incestuous as these observations are also assimilated into many over-ocean NWP models. Both also suffer from relatively low spatial resolution, limiting verification of mesoscale and sub-mesoscale features. Spaceborne synthetic aperture radar (SAR) offers the potential to address this resolution issue while verifying against independent data. In this study we used Radarsat-1 and Envisat Wide Swath synthetic aperture radar (SAR) observations to detect surface wind field errors in the NOGAPS global model. The study focused on the Gulf of Alaska and adjacent waters because of the broad range of synoptic and mesoscale phenomena present in this region.

Using SAR to verify NWP model analyses (or predictions) is complicated by the physics responsible for microwave backscatter from the sea surface. Because the radar signal is scattered primarily by wind-driven waves of a few centimeter wavelength, the backscatter intensity varies depending on the direction of the incident beam relative to the wave orientation. Thus, the backscatter signal varies with radar look angle relative to surface wind direction. The backscatter also varies with wind speed as faster winds increase wave amplitude. For the SAR wavelengths used in this study the relevant wave age is a on the order of seconds to 10s of seconds so fetch is not an issue on large bodies of water. Therefore, in order to compute a surface wind speed using SAR backscatter from the sea surface, one must first know the surface wind direction. While the wind direction can frequently be deduced from microscale and mesoscale features of the SAR backscatter image, in this study we used the analyzed wind directions from the NWP model. This approach not only provides guarantee of wind direction data being available, but also allows SAR to be used as a check on model wind direction as well as wind speed.

This dual capability arises because there are four possible outcomes of a SAR based NWP model wind verification at a given location. First, the SAR and NWP model surface wind speeds can agree because both the model wind speed and direction are correct. In this situation the SAR wind speed is correct because the model wind direction is correct, and thus the SAR wind speed matches the model wind speed. Second, the SAR and NWP model wind speeds can disagree because although the model wind direction is correct, and thus the SAR wind speed is correct, the model wind speed is incorrect. Third, the SAR and model wind speeds can disagree because although the model wind speed is correct, the model wind direction is incorrect, thus the SAR wind speed is incorrect. Unless independent SAR wind directions are available to check against the model wind directions, there is no way to distinguish between the second and third outcomes so as to determine whether it is the model wind speed or model wind direction that is in error. Finally, under fortuitous combinations of model wind speed and wind direction errors, the SAR wind speed may equal the model wind speed. Thus, in this unlikely circumstance, the incorrect SAR and model wind speeds will match the incorrect model wind speed.

SAR images were analyzed for a number of weather phenomena including prefrontal conveyor belts, gap flow, island wakes, and mountain waves. The results suggest that NOGAPS surface wind errors result from two primary causes: displacement of synoptic scale weather features and failure to resolve mesoscale structures of topographic origin. Moreover, verification against SAR wind speeds which have been spatially filtered to the model resolution reveals that root-mean-square error on model resolved scales is typically larger than that due to the filtering inherent in the model's finite spatial resolution.