The scanning weather radar is still the preferred instrument to observe mesoscale and convective phenomena because of its high spatial and temporal resolution over a large domain. In addition, the computational capacity of modern computers allows fine resolution forecasts. This sets the stage for high resolution data assimilation, a key technique for improving quantitative precipitation forecasts (QPF).

The McGill data assimilation is based on the variational formalism. The cost function includes a background term, observations and the cloud-resolving model as constraints. To take into account the model errors, and to reduce the large amount of computations, a weak constraint is applied in the McGill data assimilation system. The system assimilates both reflectivity and Doppler velocity from the single Doppler radar (McGill S-band) observations. Interested readers are referred Chung et al. 2009 for further details.

Recently, we have included the McGill data assimilation algorithm into a semi-operational system called Mesoscale Analysis System (MAS), and coupled it with a high resolution (2.5 km) regional prediction GEM-LAM model output. We have routinely obtained the analyses field within the radar observation coverage. Two features of the current system must be noticed: first, since the GEM-LAM forecast is used as background and also as the initial guess in the assimilation system, if the background field is unsuitable, then it is hardly useful for obtaining the optimal analyses in one assimilation window. By comparing the reflectivity and Doppler component between observation and background, the position correction (with displacement and rotation) is applied to obtain a better background. Second, the interpolation of data across the no data region very near the radar creates a spurious convergence and divergence. To avoid this problem, we do not assimilate the observations within a certain range near radar site.

In a separate effort, we have also investigated the sensitivity of mesoscale forecasts to different initial condition errors and tried to establish the suitability of different instruments in detecting the growing errors between model and reality (Fabry and Sun 2010).