Model Output Statistics QUalified by Intercomparison with True Observations

Project aims

The first phase of MiKlip has shown that MiKlip model simulations generally produce lower than observed temperatures in the troposphere. The model atmosphere is also too moist and vertical stability is less than observed. Depending on the model initialisation, the simulated temperatures also drift over some of the simulation years. The aims of project MOSQUITO are a comprehensive characterisation of these biases and drifts, using high quality radiosonde data as a reference. In addition, representation errors of large scale model simulations and meteorological reanalysis fields will be investigated by comparison with the small scale structure present in the radiosonde data.

Project Structure

MOSQUITO belongs to Module E and is a work-package of the Module-E DWD project. MOSQUITO will be carried out at DWD’s Meteorological Observatory Hohenpeissenberg

Tasks of the project

MOSQUITO will extend the high quality observational database from radiosondes. The observations will be used to estimate systematic errors and uncertainties of meteorological re-analyses and MiKlip global and regional simulations. After characterisation, calibration procedures for simulated temperature and humidity profiles will be implemented. Among other things, this should improve probability forecasts for severe weather events.

Fig 1. Systematic temperature bias of MiKlip simulated temperature against radiosonde observations over Germany. B0-lR, B1-LR, B1-MR and P-LR are baseline 0, baseline 1 and prototype low resolution (LR) and mixed resolution (MR) experiments. B1-LR-reg is the COSMO regional simulations for B1-LR.
Fig 2. Same as in 1. but showing the change of temperature bias at 500 hPa as a function of forecast year. The full-field initialisation in P-LR results in a large drift. (Click to enlarge)


The project will deliver data-sets to the central MiKlip server and the Central Evaluation System. Calibration procedures will be integrated into the general model post-processing.

Progress so far

Validation of the MiKlip prediction system with radiosonde observations

The MOSQUITO2 work package started in January 2017. Initial validation of the pre-operational MiKlip II model runs indicates that the previous substantial temperature and humidity biases of the model (see Figs. 1 and 2 above) have been reduced substantially. As an example, Fig. 3 now shows good agreement between observed and modelled probability density functions for upper air temperature (left) and relative humidity (right) above Central Europe. In previous model versions, the simulated temperatures were often too low, and the simulated humidity distribution showed an unrealistic large peak at high humidities. In these areas, the MiKlip2 MPIESM 1.2 pre-operational hindcasts give a large improvement.

Fig. 3: Probability density functions for modelled and observed 500 hPa temperature (left) and 700 hPa relative humidity near the Payerne radiosonde station. Model results are for MiKlip II MPIESM 1.2 pre-operational hindcasts.

Comparison between reanalyses and radiosonde observations

Another question to be addressed was how smaller scale variability observed by the radiosondes is captured by global reanalyses. As an example, Fig. 4 shows the standard deviation between radiosonde-measured temperatures and two reanalyses. Generally, the reanalyses (NCEP, MERRA, ERA-Interim, JRA25, 20CR) capture the radiosonde observed temperatures quite well. Over the European continent, the standard deviation of differences, e.g. at the 300 hPa level and for ERA-Interim, is generally smaller than 0.3 K – comparable to the accuracy of the temperature sensors. Near the coast, however, larger differences are seen, with standard deviations reaching or exceeding 1 K. For the 20th century reanalysis, however, larger differences, of the order of 1 K, are seen everywhere. This is not surprising, because the 20th century reanalysis is constrained only by surface pressure observations, whereas all the other reanalyses also consider upper-air information.

Fig. 4: Standard deviation between radiosonde measured temperatures and reanalysis temperatures at 300 hPa, for European stations, and for ERA-Interim (left) and 20th century reanalysis (right).
Evaluation of severe weather probability in radiosonde observations and model hindcasts (based on K-Index)

In the second phase of the project, the quality of MiKlip hindcasts for the estimation of extreme weather probability is being evaluated. As shown in the first phase of the project, the K-Index is a viable indicator for e.g. the probability of thunderstorms.

The K-Index is given by: K = (T850 – T500) + TD850 – (T700 – TD700), where T is air temperature, TD is dew point temperature and the assigned values mean pressure level in hPa. K examines the vertical temperature lapse rate, the moisture content of the lower atmosphere, and the vertical extent of the moist layer.

Fig.5 shows examples of the probability distribution of K-Index values calculated from radiosonde data (Obs.) and various MiKlip experiments: Baseline1-LR (b1-LR), Baseline1-MR (b1-MR), Preoperational-LR (PreopLR) and Preoperational-HR (PreopHR). K-index values below 15 (marked with cyan vertical line in Fig. 5) indicate a stable atmosphere, where thunderstorms are not likely to occur. K-index values above 35 (marked with purple line in Fig. 5) indicate a high potential for thunderstorms. This applies in particular for the thunderstorm season. In winter, because of the lack of moisture, even fairly large values do not always mean that conditions are favourable for thunderstorms.

PreopHR hindcasts give the best agreement with observations for the majority of the European stations taken into account. In most of the investigated cases, PreopHR does not overestimate thunderstorm probability. It is definitely better not only than baseline 1 (both low and middle resolution) but also than PreopLR – which already gives better results than b1. The only exception is the Mediterranean Sea region, where, for stations located on the coastline/islands, both PreopHR and PreopLR underestimate K-Index values during the warm season, whereas Baseline1 hindcasts are in good agreement with the corresponding radiosonde data.

Fig.5 Observed and hindcast modelled probability distribution of the K-index during the warm (left) and cold (right) seasons, for model lead years 2-9 (1996-2016, 12 UTC data). Cyan vertical line: threshold below which thunderstorms are very unlikely; purple line: threshold above which thunderstorms are very likely.
Correction of Model Bias using radiosonde data

Bias correction tools (Recalibration, DriftCorrection Plugin) already developed within MiKlip, are evaluated using radiosonde data as the reference observations. In addition, a new bias correction through quantile mapping is tested (based on empirical cumulative distribution functions).

Comparing the methods to each other, they all perform very similarly when annual mean values are considered (Fig.6).

For monthly data, and for the data distribution, however, quantile mapping shows the best performance, particularly at high quantiles. This is advantageous for applications related to extreme events. The improvements are obtained for all regions (Fig. 7). Note that the Recalibration plugin currently does not support a correction of monthly biases.

Fig.6 Annual mean relative humidity at pressure level 850 hPa (2002-2011) derived from MiKlip baseline1-LR (init. 2001, R1i1p1,) raw and after application of three different bias correction methods, in comparison to the radiosonde observations.
Fig.7 Probability distribution of monthly mean relative humidity at pressure level 850 hPa (2002-2011) derived from MiKlip baseline1-LR (init. 2001, R1i1p1,) ) raw and after application of two different bias correction methods, in comparison to radiosonde observations. The bias corrections are quantile mapping and the Generalised Linear Model (GLM) method from the DriftCorrection Plugin.