The RTFDDA "station nudging" approach, in combination with diverse data sources having good coverage in time and space, can greatly alleviate the well-known problem in mesoscale data assimilation and NWP where observations are excessively sparse and the dynamics and cloud/precipitation "spin-ups" associated with "cold-starts" can severely affect the accuracy of the short-term (0 - 12 hour) forecasts. Nevertheless, data from different platforms have different instrument, sampling and processing errors and the error in the observations from one platform may vary in time and space. Furthermore, when these data are applied to a common data assimilation model grid, the representativeness error relative to the grid resolution can be a crucial factor affecting the analyses and forecasts. How to estimate and make use of the integral of these errors (referred as to total observation error) is a critical and challenging problem in data assimilation.
In conjunction with the "spin-up-free" FDDA analyses and forecasts, the high accuracy 0 - 3 hour model forecasts are available at all times, in real-time. These forecasts provide an ideal three-dimensional weather state that can be used to estimate the total observation error of each observation, without regard to the platform from which it was observed. In this paper, statistical comparison studies were conducted to confirm the accuracy of these 0 - 3 hour model forecasts. An advanced and efficient data quality control (QC) procedure is developed by making use the 0 - 3 hour forecasts and the error information of two reliable high-quality observation platforms (rawinsondes and METAR). The QC procedure not only defines appropriate error tolerance criteria to eliminates bad observations, but also calculate and assign a generalized quality flag to each observation. By scaling the quality flag between 1 (the best) to 0 (the worst), the QC flags of the observations are applied in the nudging process as a confidence weight that is unique to each observation. Numerical experiments were conducted, and the results show that the use of the observation error estimate leads to significant improvement in the RTFDDA analyses and short-term forecasts.