Probabilistic nowcasting of PBL profiles with surface observations and an ensemble filter

A long-term goal of this work is to find an efficient system for probabilistic planetary boundary layer (PBL) nowcasting that can be employed wherever surface observations are present. One approach showing promise is the use of a single column model (SCM) and ensemble filter (EF) data assimilation techniques.

Earlier work showed that surface observations can be an important source of information with an SCM and an EF. Here we extend that work to quantify the probabilistic skill of ensemble SCM predictions with added complexity. Although it is appealing to add additional physics and dynamics to the SCM model it is not immediately clear that additional complexity will improve the performance of a PBL nowcasting system based on a simple model. We address this question with regard to treatment of surface assimilation, radiation in the column, and also advection to account for realistic 3D dynamics (a timely WRF prediction). We adopt a factor separation analysis to quantify the individual contribution of each model component to the probabilistic skill of the system, as well as any beneficial or detrimental interactions between the different factors.

Probabilistic skill of the system is evaluated through the Brier Skill Score (BSS) and the area under the relative operating characteristic (ROC) curve (AUR). The BSS is further decomposed into both a reliability and resolution term to understand the trade-offs in different components of probabilistic skill.

We compare the skill of the flow dependent probabilistic forecast obtained with the SCM-EF system to that of climatological probabilistic information that could be obtained from deterministic 3D mesoscale forecasts at sites where surface observations are available. Such climatological uncertainty is derived following a "dressing" technique by vertically projecting the error distribution at the surface with weights that are proportional to the climatological covariance of the surface model variables with the profile above. In addition, the deterministic profile forecasts are adjusted by linearly regressing the latest error at the surface with these climatological covariances.

Results show that assimilation of surface observations can improve deterministic and probabilistic predictions more significantly than major model improvements. The SCM-EF mean forecasts perform in most cases better than the climatological-adjusted profiles derived using the dressing technique, and the flow-dependent uncertainty is in general sharper than the climatological error variance. Results' analysis still in progress suggests that the dressing technique may lead to improved results whenever PBL parameterizations show too excessive vertical mixing.