Climate impact assessments require information about climate change at
regional and ideally also local scales. In dendroecological studies, this
information has traditionally been obtained using statistical methods, which
preclude the linkage of local climate changes to large-scale drivers in a
process-based way. As part of recent efforts to investigate the impact of
climate change on forest ecosystems in Bavaria, Germany, we developed a
high-resolution atmospheric modelling dataset, BAYWRF, for this region over
the thirty-year period of September 1987 to August 2018. The atmospheric
model employed in this study, the Weather Research and Forecasting (WRF) model, was configured with two nested domains of
7.5 and 1.5 km grid spacing centred over Bavaria and forced at the outer
lateral boundaries by ERA5 reanalysis data. Using an extensive network of
observational data, we evaluate (i) the impact of using grid analysis
nudging for a single-year simulation of the period of September 2017 to
August 2018 and (ii) the full BAYWRF dataset generated using nudging. The
evaluation shows that the model represents variability in near-surface
meteorological conditions generally well, although there are both seasonal
and spatial biases in the dataset that interested users should take into
account. BAYWRF provides a unique and valuable tool for
investigating climate change in Bavaria with high interdisciplinary
relevance. Data from the finest-resolution WRF domain are available for
download at daily temporal resolution from a public repository at the Open
Science Framework (Collier, 2020;
The forcing of climate change in modern times is clearly of global nature, and many important scientific problems can be understood at the global scale as well (e.g. Held and Soden, 2006). Climate impact assessments, however, must also understand the effects at regional and even local scales in order to develop appropriate adaptation and mitigation measures. Although local phenomena such as glaciers, lakes, vegetation patterns or stream flow show a strong dependence on large-scale climate dynamics, these proxies experience further variability when the large-scale signal is transferred to their location (e.g. Mölg et al., 2014). In order to contextualize local changes, there is a need to link local climate to the large-scale climate, ideally in a process-based way.
In dendroclimatological studies, the traditional approach is to compute a
calibration function between local or regional tree-ring parameters and
climatic variables. Typically, such a statistical relationship would try to
utilize local station data (which are generally sparse), gridded
observations (which tend to be of coarse resolution) or indices of large-scale
climate dynamics (which describe coupled atmosphere–ocean modes) as the
climatic influence (e.g. Hochreuther et
al., 2016). Besides known problems like stationarity
(e.g. Frías et al., 2006),
statistical approaches also limit the possibilities to explain the
influences at the various scales on a process-resolving level. Dynamical
downscaling with a full numerical atmospheric model provides a physical
answer (Giorgi and Mearns, 1991), yet
the disadvantage is the high computational cost. Hence, dynamical
downscaling at near-kilometre resolution has traditionally been performed on
a case-study basis for weather events
(e.g. Gohm et al., 2008).
Multi-decadal simulations, on the other hand, were typically limited to
resolutions of tens of kilometres (e.g. Di
Luca et al., 2016). With the progress of computational resources, dynamical
downscaling is becoming a candidate for climate impact studies that require
local-scale information, and the first decadal simulations at
The management of forests is a classical impact study where adaptation and
mitigation measures meet the heterogeneous effects of climate change at
local scales (e.g. Lindner et al.,
2014). With this background, the project BayTreeNet was started recently
under the umbrella of the interdisciplinary climatological research network
Bayklif (
Previous regional climate simulations including Bavaria over continuous multi-decadal periods were performed with model resolutions as high as 5–7 km and up to the year 2009 (e.g. Berg et al., 2013; Warscher et al., 2019). However, to the best of our knowledge, such datasets at the kilometre scale and up to the near present do not yet exist, despite previous research highlighting the importance of convection-permitting resolution in this region (Fosser et al., 2014). We address this data gap by performing simulations with an atmospheric model, configured with a convection-permitting spatial resolution in a nested domain over Bavaria, for the recent climatological period of 1987 to 2018. These data have the potential to find multidisciplinary interest among researchers assessing ecological and human dependencies on the climate for scientific and practical questions.
Extent and modelled topographic height in WRF D1
The atmospheric simulations were performed using the advanced research
version of the Weather Research and Forecasting (WRF) model version 4.1
(Skamarock and Klemp, 2008)
configured with two one-way nested domains of 7.5 and 1.5 km grid spacing
situated over Bavaria (Fig. 1), hereafter referred to as D1 and D2. Terrain
data were taken from NASA Shuttle Radar Topographic Mission data re-sampled
to 1 km and 500 m grids (Jarvis et al.,
2008;
Summary of the WRF configuration used for BAYWRF. A full sample namelist is provided in Appendix A.
Forcing data at the lateral boundaries of D1 and bottom boundaries of both
domains was taken from the ERA5 reanalysis
(Copernicus Climate Change Service (C3S), 2017) at
3-hourly temporal resolution. The 30-year simulation was divided into 30
annual simulations that were run continuously from 15 August of year
Each run required 12 d of wall time with 320 processors on the Meggie
compute cluster at the Erlangen Regional Computing Center, for a total of
2.86 million core hours. The model was compiled using Intel 17.0 compilers
and run using distributed-memory parallelization. Model output was written
at 2-hourly intervals, amounting to more than 55 TB of data, in addition
to
For the period of 00:00 UTC, 1 September 2017, to 00:00 UTC, 1 September 2018, we
compared two simulations with different forcing approaches: one excluding
and one including grid-analysis nudging to constrain drift in the
large-scale circulation (e.g. Bowden et
al., 2013). This period was selected due to the higher availability of
observational data closer to present day and because the summer of 2018
contained a record heatwave with drought conditions (Beyer,
2018), permitting evaluation of an extreme event. We refer to these
simulations as WRF_NO_NUDGE and
WRF_NUDGE, respectively. For the WRF_NUDGE
simulation, analysis nudging was applied in D1 outside of the planetary
boundary layer and above the lowest 10 model levels using the default
strength (
The location of the stations used for model evaluation during the
most recent simulation year (September 2017 to August 2018) for each dataset
listed in Table 2. Datasets labelled in black are shown by filled black
circles, while datasets labelled in pink are shown by open pink circles,
illustrating that locations for measurements of air temperature and humidity
(
For model evaluation, we used data from the German Weather Service (DWD)
Climate Data Center for all stations in Bavaria with hourly temporal
resolution available, which provide good spatial coverage of our study area
(Table 2; Fig. 2). To evaluate the forcing approach, we compared the
following near-surface atmospheric variables at the highest temporal
resolution available in the simulations, which is 2-hourly: air
temperature and relative humidity at 2 m (
A summary of data used for model evaluation.
For statistical analysis, we computed the mean deviation (MD), mean absolute
deviation (MAD) and the coefficient of determination (
Box-percentile plots (Esty and Banfield,
2003) of mean deviation (MD), mean absolute deviation (MAD), and coefficient
of determination (
A summary of the statistical evaluation of the
WRF_NO_NUDGE (italics) and WRF_NUDGE (bold italics) simulations, considering the evaluation period of
1 September 2017 to 1 September 2018. The table presents the mean deviation
(MD), the mean absolute deviation (MAD) and the coefficient of determination
(
Finally, we also compared night-time land surface temperature (LST) from the MODIS MYD11A1 dataset (Table 2) at 1 km spatial and daily temporal resolution with simulated skin temperature in D2 for the period of 1 June to 31 August 2018. The night view time ranged from 1.2 to 2.8 h in local solar time, with a domain and time averaged value of 2.2 h. As WRF data were only available at two-hourly time steps, we averaged 00:00 and 02:00 UTC (01:00 and 03:00 local time) data from D2 for comparison with MODIS. In our comparison, we excluded nights when MODIS had more than 50 % missing data over D2, leaving a sample size of 52.
For evaluating the full simulation, we performed a similar analysis with the
aforementioned station datasets for
Time series of monthly mean 2 m
We note that unphysically large sub-surface temperatures were simulated at a
number of glacierized grid points, primarily
during the months of July to September. Considering all of D2, the daily
average number of affected grid cells was 24, compared with 294 glacierized
and 122 500 total cells. The maximum number of affected grid points was 274
on 31 August 2017, corresponding to 0.2 % of D2. In addition, over the
climatological simulation, only one grid point in Bavaria was affected (
Time series of
Averaged over the evaluation year, both WRF simulations capture the
magnitude and variability of sub-diurnal near-surface meteorological
conditions at most sites well (Fig. 3; Table 3). The interquartile range
(IQR; range between upper and lower quartile) of MDs is 1 order of
magnitude smaller than the observed standard deviation for all variables. As
expected, variability is best captured for
Scatter plots of
Shifting to daily timescales, both simulations represent variability in
daily total PR surprisingly well, with the upper quartile of MDs below
Previous studies evaluating WRF over this region have reported root-mean-square deviations (RMSD). For direct comparison, the mean RMSD in
WRF_NUDGE for 2-hourly
Examination of model biases on a monthly basis reveals further insights into
the model performance (Fig. 4). The amplitude of the annual cycle is
overpredicted in WRF, indicating that the good average agreement in
Figure 5 shows a representative time series of
Spatial maps of mean MD
Same as Table 3 but for daily mean variables in WRF_NUDGE only.
In addition to factors internal to WRF, we note that the driving reanalysis
data may also contribute to the warm bias, at least at some locations. From
the available observations, 60 stations have both valid
The inclusion of grid-analysis nudging leads to a small but nearly uniform
improvement in agreement between observed and simulated variables. The
distribution of MDs is closer to zero for all variables except
Averaged over the full simulation period, BAYWRF shows a similar magnitude
of agreement with station
For BAYWRF, we note that in addition to the potential factors contributing
to temperature biases discussed in Sect. 3.1, evaluation of the
climatological simulation is also affected by discontinuities in station
location and instrumentation. One example is Nuremberg, which
moved on 4 December 1995 from (49.4947
Data from BAYWRF are available for download on the Open Science Framework
(OSF; Collier, 2020;
We presented a climatological kilometre-scale simulation with the atmospheric model WRF over Bavaria for the period of September 1987 to August 2018. Comparison of simulations for the period of September 2017 to August 2018 with and without grid-analysis nudging against extensive meteorological measurements across Bavaria showed that nudging decreased the mean deviations and increased the coefficient of determination at the majority of sites for nearly all evaluated atmospheric variables, in particular precipitation. This approach was therefore adopted for generating the full BAYWRF dataset. In general, BAYWRF represents the variability of near-surface meteorological conditions well, albeit with both seasonal and spatial biases that are explored briefly here. Future users of this dataset are encouraged to rigorously evaluate biases for the variables and time periods relevant to their particular study areas and applications. BAYWRF provides a useful database for linking large-scale climate, as represented by the ERA5 reanalysis, to mesoscale climate over Germany and to local conditions in Bavaria in a physically based way. The data are intended for dendroecological research applications but would also provide a valuable tool for investigations of the climate dependence of economic, societal, ecological and agricultural processes in Bavaria.
EC performed the simulations, analysed the data and wrote the manuscript. TM developed the study concept and wrote the manuscript.
The authors declare that they have no conflict of interest.
We thank Michael Warscher and Benjamin Poschlod for their helpful and insightful reviews of our manuscript. We also gratefully acknowledge the compute resources and support provided by the Erlangen Regional Computing Center (RRZE), and we thank Thomas Zeiser for his assistance with the timely completion of the simulations.
This research has been supported by the Bavarian State Ministry of Science and the Arts in the context of the Bavarian Climate Research Network (bayklif).
This paper was edited by David Carlson and reviewed by Michael Warscher and Benjamin Poschlod.