Regional Climate Variaiblity and Historical Extreme Events

Partly following on to the initial evaluation of the 1993 Midwest Flooding, and also in working toward evaluation of MERRA and reanalyses for the National Climate Assessment, we have looked closer at US regional climate variability in reanalyses. While the Northwestern US summer precipitation  is handled quite well in all reanalyses (specifically NCEP CFSR and ERA Interim), owing to influence from ENSO teleconnections, the Midwestern region summer precipitation has substantial uncertainty across all reanalyses. In MERRA, for example, the interannual variance is noticeably low, so that droughts are not as dry and pluvial periods not as wet (see 1988 and 1993 respectively in the following figure).

The extreme summers of 1988 and 1993 have been tied to both large scale ENSO teleconnections and local land-atmosphere feedback processes. Given that the reanalyses data assimilation provides a strong reference for the large scale meteorology, the land atmosphere interactions would be a likely weak point in the models that may affect this uncertainty.

These results are discussed in further detail http://dx.doi.org/10.1175/JAMC-D-12-0291.1.

Extreme Precipitation

Some time ago, I saw a poster that showed observed extreme precipitation increasing in time along the east coast and Gulf coast of the US, suggesting increasing extremes due to land falling hurricanes (Ashouri et al, 2012). There is also some supporting analysis of increasing precipitation trends and extremes in the recent National Climate Assessment report (Figures 2.15 and 2.16). To narrow the results to potential hurricane sources, Figure 1 here evaluates the trend of maximum daily precipitation, each season from 1979-2012, where hurricane season is defined as June through November.  Significant trends are seen along the northeast US track as well as some trends along the Gulf coast track in the south east US.

Figure 1 Trend of maximum daily precipitation in each hurricane season from 1979-2012. Trends significantly different from zero at 90% confidence are outlined in white contours.

The MERRA reanalysis is able to reproduce, generally this pattern of increasing extreme precipitation (Figure 2). MERRA's increasing trends in the southeast have a wider area, and in the northeast, the strongest trends  do not extend through the New England states, as observed. Still the reproduction of the trends of such a specialized diagnostic in a relatively coarse reanalysis is noteworthy.


Figure 2 As in Figure 1, except for the precipitation produced in the MERRA reanalysis.

As a further test of these trends, we area average the observed hurricane season maximum precipitation for the North Atlantic states in MERRA and the CPC observations. The interannual variability of the extreme precipitation is well reproduced, though, MERRA's mean value tends to be less than observed. Figure 3 shows increases in time for the northeast, and not just some end point variation caused be recent very large storms (e.g. Irene). though, low anomalies can occur in the recent few years, as well.

Remnants of Tropical Storm Karen produced heavy precipitation over a substantial portion of the Northeast, so that the 2013 season in the northeast will likely also be a positive anomaly (here is some result of that storm). The southeast may not have comparable extreme precipitation in 2013, at least related to tropical storms and hurricanes. We will come back to this as the 2013 hurricane season closes and MERRA is extended through it.


Figure 3 Time series of area averaged extreme precipitation anomalies from CPC gauge observations and MERRA reanalysis. The mean value removed for comparing anomalies is presented in the legend.

Utilizing Increments in Model Development

MERRA data has complete budgets for water and energy, including the incremental analysis updates (IAU) that constrain the model forecast to the analyzed observations. The increments can be interpreted instantaneously (at the six hour analysis) as a representation of the forecast error, or for longer terms as the mean model bias. The magnitude of the IAU terms are not trivial, and should be accounted for, certainly in budget studies, but can also be useful in understanding the representation of weather and climate phenomena in reanalyses.

As an example of utilizing the increments to evaluate the background model, Mapes and Bacmeister (2012) have evaluated MERRA's tropical climate and convection, relating significant IAU values to weaknesses in the representation of physical processes. They suggest diagnostics and potential areas for model development.

Aircraft Temperatures

Early in the MERRA reanalysis period, aircraft observations are sparse, but increase in time, eventually providing a significant amount of conventional observations. Cardinali et al. (2003) identified biases in aircraft temperature observations, and Ballish and Kumar (2008) further examined the biases in each type of commercial aircraft. Figure 1 shows an 2001-2009 mean bias between collocated aircraft and radiosonde 200mb temperature, over the United  States. Almost everywhere, aircraft are warmer than the radiosonde observations.

Figure 1 Collocated aircraft/RAOB at 200mb temperature (K) differences assimilated in MERRA  averaged from 2001-2012. (computed from the differences of each observaitons background departure)

Figure 2 shows the monthly mean difference between collocated radiosonde and aircraft observations assimilated in MERRA over the U. S., while the dots show the number of collocations (thou/yr).  While the data for these figures are binned and gridded, area and monthly averaging include weighting for the number of observations. Early in the reanalysis, there are lower numbers of aircraft observations, and the differences reflect that with more monthly variability. In 1990-1991, increasing number of observations increase the distribution of data, and the warm bias converges. After 1996, there is an exaggeration in the annual cycle, where the summer aircraft observations get even warmer. However, every month is a positive difference.

Figure 2 Time series of monthly mean differences of 200mb temperature collocations over the United States (Aircraft minus RAOB OmF, in red, K). The black dots indicate number of collocations each year (in thousands, right axis).

The increasing number of collocations reflects the increase in availability of aircraft observations. There are many more aircraft observations being assimilated away from the vicinity of the radiosondes. The number of observations then influences the data assimilation, where the analysis is drawn toward the aircraft data. Figure 3 shows the time series of background departure for collocated radiosonde and aircraft 200mb temperature. As the aircraft observations increase in number, their background departure decreases (this also holds for the RMS of the background departure).

Figure 3 Time series of monthly mean background departure (OmF) of the collocated RAOB (black) and Aircraft (red) 200mb temperatures (K, left axis). The black dots indicate number of collocations each year (in thousands, right axis).

Both Cardinali et al (2003) and Ballish and Kumar (2008) have suggested bias corrections for commercial aircraft temperature data, using more limited comparisons than these. ECMWF has implemented a bias correction in their forecast system.

Reanalyses trends

One of the most important topics and calculations in climate science is trend, aiming to determine long term changes. Significant issues exist in the observational record, and methods correcting the problems themselves need to be explained and verified. In a recent update to the U.S. Historical Climate Network (HCN) station data, Vose et al. compare the observational record against several reanalyses near surface air temperature and their ensemble. In looking at the continental United States, their Figure 1 shows the revised HCN trend is larger than the uncorrected data, but also remarkably close to the ensemble of the reanalyses. Also, despite the differences in trends of the reanalyses, there is very good agreement in the reanalyses interannual variability around the trends (their figure 3). The bottom line is that the corrections to HCN are in agreement with reanalyses (all are statistically significant warm trends), but it is noted that this is not a validation of the corrections.

The spatial distribution of reanalysis trend relative to the HCN trend shows substantial local variations among the reanalyses (their figure 4). So, while the observational forcing imposed on reanalyses can influence the large scale features, the model predictions used to make the analyses impart  some uncertainty related to the model physical parameterizations. If the model errors are random, the ensemble should then minimize the error. Errors that are systematic among all reanalyses would persist in the ensemble.

This paper demonstrates some important points about reanalyses. Any one reanalysis may have uncertainty in any given research project. Multiple reanalyses can help identify these uncertainties and perhaps the background model biases in the reanalysis. However, the reanalyses output variables being compared must have equivalent formulations to take advantage of the availability of the current modern reanalyses through such intercomparisons. Likewise, on hourly surface output in MERRA and CFSR were useful in this study.