Frederic Fabry
McGill University
The 8th EnKF Data Assimilation Workshop
Reflection on the challenges of radar data assimilation at meso-convective scales in an ensemble context
Talk:
Fabry_ChallengesRadarDataAssimilation.pptx
Radar is our best instrument to observe storms; assimilation is our best technique to merge information; and numerical modelling is our best forecasting tool. So why are a) simple nowcasting approaches beating numerical forecasting for a couple of hours, b) the skill of numerical forecasting aided by radar data assimilation dropping so rapidly in the first forecast hour, and c) assimilated radar data showing positive effect for much shorter periods (e.g., about two hours) than other instruments (e.g., six hours for surface observations)? And how do we change this state of affairs?
In an attempt to answer these questions, a reflection on the challenges of radar data assimilation in the context of mesoscale and convective scale forecasting was undertaken. In particular, we wanted to understand what made the assimilation of radar data assimilation more challenging and limiting than that of other data sources. The rapid drop in the skill of radar data assimilation with forecast time suggests that radar data assimilation has difficulties affecting the fields that will shape storm development in the future. This arises because:
- Over 90+% of the troposphere, the only information available from radar is “no precipitation”, largely limiting innovation to areas where either the analysis or the observations have precipitation;
- The air feeding the storms and controlling their environment in a few hours largely come from regions of no precipitation; hence its properties are largely unconstrained directly by radar. Any innovation must hence come from distant correlations with precipitation areas.
- Though there is significant correlation between overall storm intensity and thermodynamic properties, the correlations computed at gridpoint scales are limited because the field of precipitation and its errors have much more energy at smaller unpredictable scales compared to other fields. Limited ensemble sizes and localization then prevent such weak long-distance correlations to be fully utilized. And as model resolution increases, this problem becomes worse.
As a result, the most successful methods for assimilating radar data are those that purposefully or accidently take advantage of the larger-scale relationships between storminess and thermodynamics such as LETKF, while those that use gridpoint-scale correlations are the least successful.
What this suggests is that special efforts should be made to take advantage of larger-scale correlations when assimilating radar data. And since many remote sensors so critical to forecast error reduction are largely blind in regions of critical meteorological activity that we want to constrain better (e.g., the lower troposphere under a cloudy developing storm), a case could be made that such an approach should also be used when assimilating other data.
Radar is our best instrument to observe storms; assimilation is our best technique to merge information; and numerical modelling is our best forecasting tool. So why are a) simple nowcasting approaches beating numerical forecasting for a couple of hours, b) the skill of numerical forecasting aided by radar data assimilation dropping so rapidly in the first forecast hour, and c) assimilated radar data showing positive effect for much shorter periods (e.g., about two hours) than other instruments (e.g., six hours for surface observations)? And how do we change this state of affairs?
In an attempt to answer these questions, a reflection on the challenges of radar data assimilation in the context of mesoscale and convective scale forecasting was undertaken. In particular, we wanted to understand what made the assimilation of radar data assimilation more challenging and limiting than that of other data sources. The rapid drop in the skill of radar data assimilation with forecast time suggests that radar data assimilation has difficulties affecting the fields that will shape storm development in the future. This arises because:
- Over 90+% of the troposphere, the only information available from radar is “no precipitation”, largely limiting innovation to areas where either the analysis or the observations have precipitation;
- The air feeding the storms and controlling their environment in a few hours largely come from regions of no precipitation; hence its properties are largely unconstrained directly by radar. Any innovation must hence come from distant correlations with precipitation areas.
- Though there is significant correlation between overall storm intensity and thermodynamic properties, the correlations computed at gridpoint scales are limited because the field of precipitation and its errors have much more energy at smaller unpredictable scales compared to other fields. Limited ensemble sizes and localization then prevent such weak long-distance correlations to be fully utilized. And as model resolution increases, this problem becomes worse.
As a result, the most successful methods for assimilating radar data are those that purposefully or accidently take advantage of the larger-scale relationships between storminess and thermodynamics such as LETKF, while those that use gridpoint-scale correlations are the least successful.
What this suggests is that special efforts should be made to take advantage of larger-scale correlations when assimilating radar data. And since many remote sensors so critical to forecast error reduction are largely blind in regions of critical meteorological activity that we want to constrain better (e.g., the lower troposphere under a cloudy developing storm), a case could be made that such an approach should also be used when assimilating other data.
Thanks to McGill's Atmospheric and Oceanic Sciences department for hosting this web-site.