Nonlin. Processes Geophys.

, 18, 10131028, 2011

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlinear Processes
doi:10.5194/npg-18-1013-2011 in Geophysics
Author(s) 2011. CC Attribution 3.0 License.

An artificial neural network technique for downscaling GCM

outputs to RCM spatial scale
R. Chadwick1,* , E. Coppola2 , and F. Giorgi2
1 Department of Meteorology, University of Reading, UK
2 TheAbdus Salaam International Centre for Theoretical Physics, Trieste, Italy
* now at: Met Office Hadley Centre, Fitzroy rd, Exeter, UK

Received: 23 September 2011 Revised: 7 December 2011 Accepted: 12 December 2011 Published: 22 December 2011

Abstract. An Artificial Neural Network (ANN) approach 1999). However, current RCMs, which have a resolution of
is used to downscale ECHAM5 GCM temperature (T ) and around 1050 km, are still computationally expensive to run,
rainfall (R) fields to RegCM3 regional model scale over even for only a fraction of the globe.
Europe. The main inputs to the neural network were the Uncertainties in future climate projections derive from var-
ECHAM5 fields and topography, and RegCM3 topography. ious sources, such as future emission scenarios, internal vari-
An ANN trained for the period 19601980 was able to recre- ability of the climate system and the accuracy and configu-
ate the RegCM3 19812000 mean T and R fields with rea- ration of the climate models themselves, both global and re-
sonable accuracy. The ANN showed an improvement over gional (e.g. Giorgi and Francisco, 2000; Deque et al., 2005).
a simple lapse-rate correction method for T , although the These uncertainties can be quantified to some extent by ex-
ANN R field did not capture all the fine-scale detail of the amining the climate projections for a range of different sce-
RCM field. An ANN trained over a smaller area of Southern narios, global climate models, climate realizations, and re-
Europe was able to capture this detail with more precision. gional downscaling tools (Deque et al., 2005). This process
The ANN was unable to accurately recreate the RCM climate might require the completion of a very large number of sim-
change (CC) signal between 19812000 and 20812100, and ulations (Giorgi et al., 2008), which may not currently be
it is suggested that this is because the relationship between feasible due to the computational expense of each GCM and
the GCM fields, RCM fields and topography is not constant RCM model run. Therefore there is interest in possible al-
with time and changing climate. An ANN trained with three ternative, less expensive approaches to fill the GCM-RCM
ten-year time-slices was able to better reproduce the RCM climate change simulation matrix.
CC signal, particularly for the full European domain. This Statistical and statistical/dynamical GCM downscaling
approach shows encouraging results but will need further re- techniques require much less computational time than a pure
finement before becoming a viable supplement to dynamical dynamical approach. Statistical downscaling methods use
regional climate modelling of temperature and rainfall. empirical relationships established between large-scale and
fine-scale variables using historical data, and a number of
these techniques are described in Giorgi et al. (2001) and
Maraun et al. (2010). These methods rely on these statis-
1 Introduction
tical relationships remaining constant with future climate, an
Studies of climate impacts usually require estimates of future assumption that is open to question (Wilby et al., 1998).
climate parameters at a higher spatial scale than can be pro- Statistical/dynamical approaches construct a statistical re-
vided by the current generation of General Circulation Mod- lationship between RCM and GCM outputs over a limited
els (GCMs). One method of producing higher resolution es- RCM run, then extrapolate this relationship to produce fine-
timates is to run a Regional Climate Model (RCM) for a par- scale estimates for other periods of the GCM run (e.g. Busch
ticular sub-region of the globe, forced with lateral and sur- and Heimann, 2001). This type of approach is computa-
face boundary conditions from a GCM (Giorgi and Mearns, tionally less expensive than running the RCM for the whole
time period of interest, but relies on the relationship between
GCM and RCM output parameters remaining constant with
Correspondence to: R. Chadwick time and changing climate.
([email protected])

Published by Copernicus Publications on behalf of the European Geosciences Union and the American Geophysical Union.
1014 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

In this paper we explore an approach similar to statisti- Table 1. ANN inputs.

cal/dynamical downscaling to provide a means to populate a
downscaled climate projection matrix without having to run
Input Input
RCMs for the full range of GCM simulations. Instead of Number
a statistical method, an Artificial Neural Network (ANN) is
used to construct a relationship between a GCM and corre- 1 NE GCM temperature/rainfall
sponding nested RCM fields. If such a relationship, which 2 NW GCM temperature/rainfall
likely depends on the given GCM-RCM pair considered, can 3 SW GCM temperature/rainfall
4 SE GCM temperature/rainfall
be shown to be accurate, it can be used to downscale the
5 NE GCM elevation
information of a given GCM to the grid of a given RCM with- 6 NW GCM elevation
out having to re-run the RCM for each GCM experiment. For 7 SW GCM elevation
example this could be used to downscale a given GCM onto 8 SE GCM elevation
a given RCM grid for different greenhouse gas emission sce- 9 NE GCM gridpoint to RCM gridpoint distance
narios or for different realisations of the same scenario. This 10 NW GCM gridpoint to RCM gridpoint distance
approach of course relies on the ability of the derived ANN 11 SW GCM gridpoint to RCM gridpoint distance
relationship to be valid in future climate conditions as well 12 SE GCM gridpoint to RCM gridpoint distance
as the present. The concept can be seen as a type of emula- 13 RCM elevation
tor (Bastos and OHagan, 2008; Conti and OHagan, 2010), 14 RCM landuse
where an ANN is used to emulate RCM output for future 15 Month
16 Year
scenarios without needing to run the RCM.
We develop the ANN model and test its accuracy for
present and future climate conditions using a transient cli-
mate projection (19602100) over the European region European region at a grid spacing of 25 km. RegCM3 was
completed with the Abdus Salam International Centre for integrated for the period 19602100. ECHAM5 was inte-
Theoretical Physics (ICTP) RCM, RegCM3 (Pal et al., grated for the same period at a resolution of 1.89 degrees,
2007), nested within the Max Plank Institute global model using the A1B scenario of greenhouse gas forcing (Solomon
ECHAM5 (Roeckner et al., 2003). We test the accuracy of et al., 2007). The lateral buffer zone is 15 grid points in
the ANN model for different historical periods and future width, and uses an exponential relaxation term (Giorgi et al.,
time slices in the simulation, and suggest optimal approaches 1993). The size of the buffer zone is small compared to the
to maximize the model reliability. The paper begins in the RCM domain, and RegCM3 is therefore free to produce its
next section with a description of the climate and ANN mod- own internal circulations and climatology.
els and of the experiment design.
3 ANN methodology
2 Model and simulation description Artificial neural networks (ANNs) are a pattern recognition
tool capable of reproducing empirical, possibly non-linear
The regional climate model used here is RegCM3, described relationships between a set of input variables and some
by Pal et al. (2007) and references therein. RegCM3 is corresponding output variables. ANNs are based on the
based on the hydrostatic primitive equation dynamical core physiology of the brain, with a series of nodes able to pass
of the NCAR mesoscale 95 meteorological model MM5 information between one another in a similar way to cells in
(Grell et al., 1994). Resolvable precipitation processes are the brain. For a full description of ANN theory, see Bishop
treated with a sub-grid explicit moisture scheme (Pal et al., (2000) or Picton (2000). ANNs have been applied on several
2000), and the convection scheme used here is that of Grell previous occasions to problems within climate science (see
et al. (1994), with the Fritsch and Chappell (1980) clo- Hsieh, 2009; Haupt et al., 2009 for recent reviews of these
sure assumption. The Kiehl et al. (1996) CCM3 radiation applications), including to the field of global and regional at-
scheme, biosphere-atmosphere transfer scheme (Dickinson tribution of climate change (Pasini et al., 2006; Pasini and
et al., 1993) and Zeng et al. (1998) ocean flux scheme are Langone, 2010).
also used. In this case, the interest was in establishing the relation-
Initial and 6-hourly boundary conditions and SSTs are ob- ship between GCM and RCM output parameter fields, specif-
tained from the ECHAM5 AOGCM (Roeckner et al., 2003). ically temperature and rainfall (T and R). This was achieved
This parallel ECHAM5 run also provides the GCM data used by presenting properties of the GCM and RCM fields to an
to develop and test the artificial neural network in this study. ANN, and training it to be able to replicate the relation-
The RCM domain is that used for the EU ENSEM- ship between the two. The method used is described in detail
BLES project (Van Der Linden and Mitchell, 2009), cover- below.
ing the entire European region. The domain covers the entire

Nonlin. Processes Geophys., 18, 10131028, 2011 www.nonlin-processes-geophys.net/18/1013/2011/

R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale 1015

Fig. 1. DJF climatological mean T ( C) for Europe, 19812000. From RCM, GCM with lapse-rate correction, EurANN (ANN trained over
the full European domain), and EurANNTS (Time-slice
ANN trained over the full European domain).
Fig. 1. DJF climatological mean T ( C) for Europe, DJF 1981-2000. From RCM, GCM with lapse-rate correc-
tion, EurANN (ANN trained over the full European domain), and EurANNTS N
(Time-slice ANN trained over
j 1
The particular type of ANN used here is known as a feed- X
the full
forward European
multi-layer domain).(MLP). MLPs consist of a
perceptron Inputjy = W(j 1)xy Output(j 1)x , (1)
x=1
number of layers, each containing several nodes. Each node
in a particular layer is linked to every node in the previous where Output(j 1)x is the Output of node x in the (j 1)-
and subsequent layers. For a feed-forward network such as th layer, W(j 1)xy is the weight connecting node x in the
the one used here, nodes can only pass information one way, (j 1)-th layer and node y in the j -th layer, and Nj 1 is the
in the direction leading from the first, input layer to the final, total number of nodes in the (j 1)-th layer.
output layer. The number of layers, combined with the num- This input is mapped to an output for that node by use of
ber of nodes in each layer, is collectively known as the ANN a sigmoid function, where:
architecture.
Each connection between nodes has an associated 1
Outputjy = . (2)
weight W , the value of which can be modified to 1 + exp(Inputjy )
strengthen or weaken the connection between nodes. It is
the value of these weights, together with the ANN architec- This sigmoid function is similar to a simple threshold func-
ture, that largely determine the properties of an MLP. The tion but is also continuous and differentiable, which is neces-
input to a particular node consists of the weighted sum of the sary for the use of a back-propagation algorithm. The output
outputs from all nodes in the previous layer. So the input to 13 Outputjy is then passed on to each node in the j +1-th layer,
a particular node y in the j -th layer of an ANN is given by: and so on until the final layer of output nodes is reached.
A network is prepared for a certain purpose by means of
a calibration, or training process. A large number of train-
ing patterns are used, each pattern consisting of a set of input
variables and the corresponding target output variable. In

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlin. Processes Geophys., 18, 10131028, 2011

1016 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

Fig. 2. DJF climatological mean R (mm day1 ) for Europe, 19812000. From RCM, GCM bilinearly interpolated to RCM scale, EurANN
(ANN Fig.
trained
2. over
DJFthe full European domain),
climatological mean R and EurANNTS
(mm/day) (Time-slice
for Europe, ANN trained
1981-2000. overRCM,
From the fullGCM
European domain).interpolated
bilinearly
to RCM scale, EurANN (ANN trained over the full European domain), and EurANNTS (Time-slice ANN
this case it was desired to produce an output of T or R for practise it can take several thousand epochs before a true
trained over the full European domain).
each point in the RCM grid. ANNs were trained separately minimum is reached. A momentum term is also included
for T and R, so each network produced only one output vari- in the training process. This reduces oscillations of the so-
able, corresponding to a single node in the output layer of the lution around the absolute minimum of E and reduces the
ANN. number of training epochs needed.
The ANN weights are randomly initialised, and the inputs A crucial property of ANNs is that they should be able
for the first pattern are fed into the ANN, producing a cor- to generalise, meaning that after training they can process a
responding output. This output is compared with the corre- set of inputs not used in the training process (and hence not
sponding target output for the pattern, resulting in an error seen before by the network) into a reasonable output value.
value for that pattern. This process is repeated for all remain- In some cases when too many epochs of training are used, an
ing patterns, and the total error, E is taken to be the sum of ANN can become overtrained to its training dataset and
the squared errors from all patterns. One cycle of this process loses the ability to generalise. In order to avoid this, the
involving all training patterns is known as an epoch. available input/output patterns were separated into a training
The goal of ANN training is to minimise E by adjust- dataset and a smaller test dataset. After each epoch, E was
ing the ANN weights. This is achieved by use of a back- calculated for both the training and test datasets. If E contin-
propagation algorithm (Rumelhart and McClelland, 1986). ued to fall or stabilised for the training dataset but began to
A relationship E/W can be analytically established for ev- rise for the test dataset, then overtraining was diagnosed and
ery weight W in the ANN. The value of each weight is then the training process was stopped.
modified by an amount proportional to E/W . The process The ANN code used here was based on a modified ver-
is then repeated, with all training patterns processed through sion of that developed by Lonnblad et al. (1991) for pattern
the modified network, E calculated, and weights modified, recognition problems in particle physics.
until the value of E is regarded to be at a minimum. In
Fig. 3. DJF standard deviation of DJF seasonal mean T ( C) for Europe, 1981-2000. From EurANN (ANN
Nonlin. Processes
trained over Geophys., 18, 10131028,
the full European domain)2011
and RCM. www.nonlin-processes-geophys.net/18/1013/2011/
R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale 1017

Fig. 3. DJF standard deviation of seasonal mean T ( C) for Europe, 19812000. From EurANN (ANN trained over the full European
domain) and RCM.

Fig. 4. DJF standard deviation of seasonal mean R (mm day1 ) for Europe, 19812000. From EurANN (ANN trained over the full European
domain) and RCM.

An MLP with at least 3 layers should be able to reproduce surrounding 4 GCM gridpoints are used as inputs. So one
any mapping between input and output variables (Hornick GCM gridpoint will be to the NE of the RCM gridpoint, one
et al., 1989), but 4 layers are used here as it has been found to the NW, one to the SW and one to the SE. The inputs used
previously that an extra layer leads to more efficient ANN are shown in Table 1. For RCM-scale T as output, GCM T
training for problems similar to this one (Tomassetti et al., input values were used, while GCM R values were used for
2009; Chadwick and Grimes, 2011; Coppola, 2006). RCM-scale R output.
As inputs it was possible to use any information from the To improve the efficiency of ANN training, inputs are stan-
GCM output field, and non-time-variant properties of the dardised to values in the range [0 1] by the function:
RCM field such as gridpoint elevation and land-use. As x xmin
GCM and RCM gridpoints are not in general spatially co- xfinal = , (3)
xmax xmin
incident, and the RCM field has a higher spatial resolution, a
decision had to be taken on which GCM gridpoints to use as where xmax and xmin are the min/max values for a particular
input data for each RCM gridpoint. input and xfinal is the final standardised input for use in the
ANN.
Tomassetti et al. (2009) used an ANN approach to down-
GCM and RCM T /R fields were used at monthly time-
scale precipitation fields from 27 km to 3 km, and their
resolution. The available coincident RCM and GCM dataset
method of matching the two fields of differing spatial res-
was for the period 19602100.
olution has been adopted here. For each RCM gridpoint, the

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlin. Processes Geophys., 18, 10131028, 2011

1018 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

Fig. 5. DJF mean T ( C) for the Mediterranean region, 19812000. From RCM, GCM with lapse-rate correction, EurANN, and MedANN.

4 Results and discussion bilinearly interpolated on to the RCM grid. The differ-
ence 4h in km between the interpolated GCM elevation and
The ANN was trained initially with RCM data from 1960 the RCM elevation is then calculated at each RCM grid-
1980 (separate randomly selected subsets of this dataset were point. Finally a standard atmospheric lapse-rate correction of
used for ANN training and testing) and validated with RCM 6.5 K km1 is applied to each point in the interpolated GCM
data from 19812000. Due to the large number of RCM grid- T field:
points available in the European domain over the 21 yr train-
ing period, and in order to reduce processing time, only a ran- Tlapserate = TGCMinterp 6.5 4h (4)
domly selected selection of 20 % of the total available grid- This lapse-rate corrected field is shown in Fig. 1. As no
points for each month in the training period were used for similar lapse-rate correction is appropriate for rainfall, Fig. 2
training. The particular gridpoints used were selected sepa- instead shows the GCM R field bilinearly interpolated on to
rately for each month of the training period, in order to fully the RCM grid.
sample the RCM domain during training. Sensitivity tests es- As the purpose here was to reproduce RCM T and R
tablished that increasing this proportion above 20 % made no fields, ANN performance was measured by how well it could
noticeable difference to ANN performance, so 20 % of the reproduce the large and small-scale features of these fields.
data was considered to be a sufficient amount to properly Although RCM fields may not necessarily reflect reality, if
train the ANN. ANN training proceeded with the training the ANN reproduced these fields accurately then it was con-
dataset until E was minimised, with the subset of test data sidered to be performing well. As a quantitiative measure
used to prevent overtraining. Once trained, the ANN was ap- of model performance we use the area averaged bias, root
plied to the validation dataset and ANN estimates produced mean square error (RMSE) and spatial coefficient of corre-
for each validation month and RCM gridpoint. lation between the ANN or GCM models against the RCM
The time-mean 19812000 ANN DJF output using the en- results. These are shown in Tables 2 and 3.
tire European RCM domain (excluding buffer zones for both It was found that the ANN was able to successfully re-
training and validation) are shown in Figs. 1 and 2. The RCM produce the large scale features of the climatological mean
fields are also shown. RCM T and R fields (see Figs. 1 and 2). The representa-
The ANN downscaling method described here was also tion of detailed features in the mean T field is good in the
compared against a simple lapse-rate downscaling method ANN, but there is no obvious qualitative improvement over
for temperature fields similar to the one described in Engen- the lapse-rate down-scaling method here. The ANN R field
Skaugen (2007). The GCM T and elevation fields are first does represent several features of the RCM field better than

Nonlin. Processes Geophys., 18, 10131028, 2011 www.nonlin-processes-geophys.net/18/1013/2011/

R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale 1019

Table 2. Area averaged bias, RMSE and spatial correlation coefficient of temperature, for EurANN (ANN trained over the full European
domain), EurANNTS (Time-slice ANN trained over the full European domain), MedANN (ANN trained over the Mediterranean region),
MedANNTS (Time-slice ANN trained over the Mediterranean region), lapse-rate corrected ECHAM5 and ECHAM5, taking RegCM3 fields
as truth, for the period 19812000 over the Mediterranean region.

EurANN EurANNTS MedANN MedANNTS Lapse-rate GCM

T Bias
( C)
DJF 0.041 0.234 0.186 0.045 1.244 1.141
MAM 0.263 0.116 0.012 0.107 1.364 1.262
JJA 0.198 0.091 0.064 0.073 1.042 0.941
SON 0.359 0.377 0.027 0.088 1.960 1.857
T RMSE
( C)
DJF 0.730 0.722 0.427 0.411 1.467 1.875
MAM 0.362 0.319 0.205 0.219 1.364 1.448
JJA 0.746 0.652 0.461 0.475 1.343 1.296
SON 0.627 0.562 0.355 0.303 1.974 2.114
T corr.
DJF 0.970 0.973 0.991 0.990 0.953 0.852
MAM 0.990 0.989 0.995 0.995 0.982 0.760
JJA 0.872 0.896 0.946 0.945 0.796 0.478
SON 0.979 0.986 0.991 0.994 0.970 0.834

Table 3. Area averaged bias, RMSE and spatial correlation coefficient of rainfall, for EurANN (ANN trained over the full European domain),
EurANNTS (Time-slice ANN trained over the full European domain), MedANN (ANN trained over the Mediterranean region), MedANNTS
(Time-slice ANN trained over the Mediterranean region) and ECHAM5, taking RegCM3 fields as truth, for the period 19812000 over the
Mediterranean region.

EurANN EurANNTS MedANN MedANNTS GCM

R Bias
(mm day1 )
DJF 0.158 0.117 0.078 0.095 0.615
MAM 0.262 0.153 0.021 0.074 0.742
JJA 0.049 0.082 0.154 0.163 0.409
SON 0.160 0.074 0.003 0.050 0.743
R RMSE
(mm day1 )
DJF 0.634 0.630 0.472 0.526 0.801
MAM 0.387 0.338 0.282 0.297 0.763
JJA 0.195 0.203 0.226 0.214 0.411
SON 0.445 0.433 0.348 0.373 0.777
R Corr
DJF 0.591 0.601 0.853 0.831 0.327
MAM 0.814 0.823 0.906 0.902 0.595
JJA 0.877 0.883 0.848 0.869 0.818
SON 0.648 0.656 0.836 0.806 0.461

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlin. Processes Geophys., 18, 10131028, 2011

1020 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

Fig. 6. DJF mean R (mm day1 ) for the Mediterranean region, 19812000. From RCM, GCM bilinearly interpolated to RCM scale,
EurANN, and MedANN.

the GCM, such as the high precipitation over North-West apparent. The MedANN temperature field also shows some
Spain and Portugal (see Fig. 2). However some of the more improvement over the lapse-rate method when observed at
detailed features such as enhanced rainfall over the moun- this level of detail, with better agreement with the RCM
tainous coast of Croatia, and the extent of enhanced rain- in Southern Italy, Sicily and Sardinia. Area-averaged bias,
fall over the Alps, are not captured by the ANN. Quantita- RMSE and spatial correlation statistics of T and R, for Eu-
tively, it can be seen from the bias, RMSE and spatial cor- rANN, MedANN, ECHAM5 lapse-rate corrected ECHAM5
relation shown in period 1 (19812000) of Fig. 12 that the (for T only) and ECHAM5 with respect to RegCM3 fields,
ANN shows an improvement over the both the GCM and the for the period 19812000 over the Mediterranean region are
lapse-rate correction, for both T and R. shown in Tables 2 and 3. It can be seen that in general
Figures 3 and 4 show the standard deviation of the RCM MedANN performs better than EurANN over the Mediter-
and ANN fields over the period DJF 19812000. In general ranean region (with the exception of JJA rainfall).
the two fields have the same pattern, with the ANN slightly These results indicate that the performance of the ANN is
underestimating the magnitude of the RCM standard devia- sensitive to the choice of region of calibration. The more lo-
tion, particularly for rainfall. This is consistent with previous calised the region of ANN calibration, the more it is able to
work showing that ANNs often fail to capture the full range reproduce the specific relationship between orography, GCM
of values that they are attempting to simulate (Chadwick and variables and RCM variables found in that region. So the
Grimes, 2011). MedANN is more appropriate for estimating rainfall and
In order to try and improve ANN performance for smaller temperature over the Mediterranean than the EurANN, but
regional rainfall features, a European sub-region covering equally the MedANN would be expected to perform poorly
the Mediterranean was focused on (defined as 3644 E, 3 if applied over the European region as a whole.
25 N). The ANN was retrained specifically for this reduced Our results concerning the performance of the ANN for
area in the same way as before, but using only data from the historical period show encouraging indication towards
within the region. T and R output from this more localised the use of the ANN to downscale GCM information at the
ANN (referred to as MedANN) are shown in Figs. 5 and 6. RCM level. It then remains to be seen whether the ANN can
It can be seen that MedANN captures the high rainfall correctly reproduce the climate-change (CC) signal shown
along the Croatian coast far more effectively than either by the RCM over Europe. Towards this goal, estimates were
the original European ANN (EurANN) or the GCM, with produced with the same ANN as above for the future time
some enhanced rainfall over mountainous areas of Italy also period 20812100, and the differences of the mean T and

Nonlin. Processes Geophys., 18, 10131028, 2011 www.nonlin-processes-geophys.net/18/1013/2011/

R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale 1021

Fig. 7. JJA T ( C) climate change signal, mean 20812100 mean 19812000 for Europe. From RCM, GCM with lapse-rate correction,
EurANN, and EurANNTS.

R fields between 20812100 and 19812000 were calcu- successful application of the ANN to other time periods re-
lated for EurANN and MedANN, as shown in Figs. 79. lies on the relationship between input and output data re-
Bias, RMSE and spatial correlation statistics for these cli- maining constant with time and changing climate. This as-
mate change estimates are shown in Tables 4 and 5. sumption appears to hold for the validation period of 1981
Figure 7 shows that EurANN reproduces the large-scale 2000, which is close to the training period, but may break
RCM CC signal for temperature relatively well, and is much down for time periods further in the future such as 2081
more similar to the RCM field than the GCM is. However 2100.
smaller scale T CC patterns are not so well represented. The In order to try and improve the CC signal of the ANN,
ability of the ANN to reproduce the fine-scale RCM signal is retraining was performed using a time-slice approach. In
variable between seasons, with Fig. 8 showing a reasonably this case, training data was taken from 3 separate 10 yr time-
accurate ANN estimate and Fig. 9 a less accurate example. periods (19611970, 20112020 and 20612070), and this
time-slice ANN (ANNTS) was then used to produce T and R
Figure 10 and Table 5 shows that the EurANN CC signal estimates for 19812000 and 20812100. The ANNTS was
does not reproduce the RCM rainfall field well. In particu- trained both for the whole European domain (EurANNTS)
lar, the ANN shows a general drying over Southern Europe and for the Mediterranean (MedANNTS) in a similar way to
which is present in the GCM but not the RCM. The MedANN the original ANN.
also performs relatively poorly, as can be seen in Fig. 11 and The ANNTS output for 19812000 is shown in Figs. 1
Table 5. 6, and it can be seen that this new approach produces results
As the ANN is capable of reproducing the RCM 1981 similar to the ones obtained from the original ANN. The AN-
2000 mean fields reasonably accurately and consistently, this NTS seems able to replicate the 19812000 RCM output as
failure to consistently reproduce the CC signal must be due well as the original non-timeslice ANN, and this can also be
to an inability to correctly reproduce the 20812100 fields. seen in the statistics in Tables 2 and 3 and Figs. 12 and 13.
The ANN was trained with data from 19601980, and the

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlin. Processes Geophys., 18, 10131028, 2011

1022 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

Fig. 8. MAM T ( C) climate change signal, mean 20812100 mean 19812000 for the Mediterranean region. From RCM, GCM with
lapse-rate correction, MedANN, and MedANNTS.

Fig. 9. DJF T ( C) climate change signal, mean 20812100 mean 19812000 for the Mediterranean region. From RCM, GCM with
lapse-rate correction, MedANN, and MedANNTS.

The ANNTS CC signal is shown in Figs. 711, and here CC signal is well reproduced for both T and R (see Figs. 7,
there is some improvement over the original ANN (see Ta- 10 and Tables 4 and 5), with the rainfall field in particular
bles 4 and 5 for a comparison of skill statistics). The Eu- much improved over the original ANN and a slight improve-
rANNTS CC signal shows that on the large scale the RCM ment over the GCM. For the MedANNTS, the detail of the

Nonlin. Processes Geophys., 18, 10131028, 2011 www.nonlin-processes-geophys.net/18/1013/2011/

R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale 1023

Fig. 10. JJA R (%) climate change signal, mean 20812100 mean 19812000 for Europe. From RCM, GCM with lapse-rate correction,
EurANN, and EurANNTS.

RCM CC signal is often reproduced well (and slightly bet- This drift in accuracy explains the poor replication of the
ter than the GCM), but in some cases is not an improvement CC trend by the ANN, despite the accuracy of the ANN
over the GCM CC signal (e.g. DJF T and P ). In most cases, for any single time period remaining better than that of the
the MedANNTS is an improvement over the original non- GCM. The stationarity in bias and RMSE of ANNTS should
timeslice ANN. Although the ANNTS is able to correctly re- make the CC trend obtained by this method more reliable
produce the RCM CC signal in many cases, this is not always than that of the original ANN.
true.
To examine why a timeslice training approach produces a
more accurate CC signal than training with a single period,
5 Conclusions
ANN and ANNTS estimates were produced for a series of
six different 20 yr periods (shown in Fig. 12). Area-averaged
Bias, RMSE and spatial correlation values of EurANN, Eu- In this paper we present an ANN approach for the downscal-
rANNTS, MedANN and MedANNTS (taking RegCM3 as ing of GCM temperature and precipitation fields to reproduce
truth) for the time-mean output of each of these periods corresponding fields obtained with a nested RCM, and apply
are shown in Figs. 12 and 13. Figure 12 shows that while this approach to a nested RCM transient climate change sim-
the GCM bias and RMSE remain approximately constant ulation over Europe. The ANN appears able to realistically
with time, the EurANN estimates degrade as the time period recreate RCM output fields from GCM input data for both T
gets further from the ANN calibration period. In contrast and R. An ANN trained with the full European domain rep-
the EurANNTS bias and RMSE do not drift with time. For resents large scale patterns well, but misses smaller rainfall
MedANN and MedANNTS similar behaviour can be seen features such as high rainfall over the mountains of Croa-
for the RMSE in Fig. 13, though there is no drift in bias of tia. The EurANN appears to represent the RCM rainfall field
the MedANN. more closely than the GCM rainfall field does (see Fig. 2).
For temperature the lapse-rate corrected GCM, the EurANN
and the RCM fields are all qualitatively similar (Fig. 1), but

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlin. Processes Geophys., 18, 10131028, 2011

1024 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

Fig. 11. DJF R (%) climate change signal, mean 20812100 mean 19812000 for the Mediterranean region. From RCM, GCM bilinearly
interpolated to RCM scale, MedANN, and MedANNTS.

0.1
0.5 0.0
R Eur Bias
T Eur Bias

-0.1
0.0 -0.2
ANN
ANNTS -0.3
-0.5 Lapse-rate -0.4
GCM
-0.5
1981 2001 2021 2041 2061 2081 1981 2001 2021 2041 2061 2081

1.4 0.6
1.2
R Eur RMSE
T Eur RMSE

0.5
1.0
0.8 0.4
0.6
0.3
0.4
0.2 0.2
1981 2001 2021 2041 2061 2081 1981 2001 2021 2041 2061 2081

1.000 1.00
0.995
R Eur Corr.

0.95
T Eur Corr.

0.990
0.985 0.90
0.980
0.85
0.975
0.80
1981 2001 2021 2041 2061 2081 1981 2001 2021 2041 2061 2081
Time Period Time Period

Fig. 12. Area averaged bias (top), RMSE (middle) and spatial coefficient of correlation (bottom) of EurANN (ANN trained over the full
European domain), EurANNTS (Time-slice ANN trained over the full European domain), ECHAM5 lapse-rate corrected ECHAM5 (for T
Arearespect
Fig. 12.with
only) and ECHAM5 averaged bias (top),
to RegCM3 RMSE
fields for 6(middle)
differentand spatial
time coefficient
periods of correlation
over Europe. Statistics(bottom) of EurANN
are averaged over all four seasons.
Left-hand plots showtrained
(ANN temperature statistics,
over the right-hand
full European plots show
domain), rainfall. Time
EurANNTS periods ANN
(Time-slice are 20trained
yr means starting
over from
the full the year shown on the
European
x axis.
domain), ECHAM5 lapse-rate corrected ECHAM5 (for T only) and ECHAM5 with respect to RegCM3 fields
for 6 different time periods over Europe. Statistics are averaged over all four seasons. Left-hand plots show
temperature
Nonlin. Processes statistics,
Geophys., 18, right-hand
10131028,plots show rainfall. Time periods www.nonlin-processes-geophys.net/18/1013/2011/
2011 are 20 year means starting from the year
shown on the x axis.
R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale 1025

1.5 0.2
0.0

R Med Bias
ANN

T Med Bias
1.0 ANNTS
Lapse-rate -0.2
GCM
0.5 -0.4
0.0 -0.6
-0.8
1981 2001 2021 2041 2061 2081 1981 2001 2021 2041 2061 2081

2.0 0.8
0.7

R Med RMSE
T Med RMSE

1.5
0.6
0.5
1.0
0.4
0.5 0.3
0.2
1981 2001 2021 2041 2061 2081 1981 2001 2021 2041 2061 2081

1.0 0.9

R Med Corr.
T Med Corr.

0.9 0.8

0.8 0.7

0.7 0.6
0.6 0.5
1981 2001 2021 2041 2061 2081 1981 2001 2021 2041 2061 2081
Time Period Time Period

Fig. 13. Area averaged bias (top), RMSE (middle) and spatial coefficient of correlation (bottom) of MedANN, MedANNTS, ECHAM5
lapse-rate corrected
Fig. 13. ECHAM5 (for Tbias
Area averaged only) andRMSE
(top), ECHAM5 with and
(middle) respect to RegCM3
spatial fields
coefficient for time-means
of correlation of 6 of
(bottom) different
MedANN,20-yr periods over
the Mediterranean region. Statistics are averaged over all four seasons. Left-hand plots show temperature statistics, right-hand plots show
MedANNTS, ECHAM5 lapse-rate corrected ECHAM5 (for T only) and ECHAM5 with respect to RegCM3
rainfall. Time periods are 20 yr means starting from the year shown on the x axis.
fields for time-means of 6 different 20-year periods over the Mediterranean region. Statistics are averaged over
the EurANN all skill scores are
four seasons. an improvement
Left-hand on the lapse-
plots show temperature land-atmosphere
statistics, right-hand plotsfeedbacks (e.g.Time
show rainfall. soil periods
moisture-precipitation)
are
rate method (see period 1 (19812000) of Fig. 12). occur within the RCM during climate change that are not
20 year means starting from the year shown on the x axis. also present in the GCM. This could alter the relationship
Training the ANN over a smaller regional domain im-
proves the level of detail that it can capture. In this case, between RCM and GCM T and R values at coincident grid-
the MedANN shows an improvement over the lapse-rate points, and render the 19601980 calibration unreliable for
method for recreating RCM temperature fields (Fig. 5), and later time periods.
the MedANN rainfall fields represent RCM rainfall much We attempted to mitigate this problem by training the
more closely than the GCM rainfall does (Fig. 6). MedANN ANN model using different time slices in the future climate
also shows an improvement over EurANN for this region. simulation (ANNTS model). With this approach the error re-
Although the ability of ANNs to reproduce RCM mean mains stationary throughout the range of time periods. The
fields for a certain time period is interesting, the real prop- improvement could be either because the ANNTS has access
erty of note here is whether an ANN approach can correctly to a wider range of T and R data during training, because
replicate the climate change signal of an RCM. The origi- year is included in the ANN inputs and the ANN is able to
nal ANN, trained only with data from 19601980, does not apply a time-variable relationship between GCM input, ele-
reproduce the 20812100 19812000 CC signal well for23 vation and output, or a combination of the two. In this way
either R or T , and for either the full European domain or the the ANNTS might be able to capture time-varying non-linear
Mediterranean sub-region . dynamical processes in the RCM.
Figures 12 and 13 show that this is due to a drift with time The non-stationarity of ANN error shown here has im-
in the accuracy of the ANN. As the validation time period plications for statistical-dynamical (and possibly statistical)
gets further from the ANN training period the accuracy of GCM downscaling methods. In particular, any method which
the ANN decreases. This may be because the relationship relies on the relationship between GCM variables, orogra-
between input and output data changes with time and chang- phy and RCM variables remaining constant with changing
ing climate, meaning that the patterns established during climate might also be expected to be subject to the same drift
ANN training for 19601980 are not suitable for 20812100. in accuracy as seen here. Training using future time slices
One possibility is that regional circulation changes and/or improves the performance and applicability of the ANN

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlin. Processes Geophys., 18, 10131028, 2011

1026 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

Table 4. Area averaged bias, RMSE and spatial correlation coefficient of temperature climate change signal between 19812000 and 2081
2100, for ANN, ANNTS, and lapse-rate corrected ECHAM5, taking RegCM3 fields as truth. First three columns show results for EurANN
and EurANNTS validated over the European region, and the remaining columns show results for MedANN and MedANNTS validated over
the Mediterannean region.

Europe Med
EurANN EurANNTS Lapse-rate/GCM MedANN MedANNTS Lapse-rate
T Bias
( C)
DJF 0.692 0.112 0.309 0.359 0.268 0.129
MAM 0.524 0.152 0.002 0.122 0.093 0.228
JJA 0.673 0.057 0.151 0.176 0.224 0.320
SON 0.672 0.033 0.235 0.299 0.137 0.247
T RMSE
( C)
DJF 0.697 0.206 0.413 0.695 0.274 0.220
MAM 0.562 0.253 0.325 0.410 0.160 0.239
JJA 0.887 0.365 0.457 0.719 0.346 0.613
SON 0.678 0.213 0.377 0.507 0.241 0.315
T corr.
DJF 0.936 0.955 0.956 0.359 0.654 0.847
MAM 0.932 0.944 0.947 0.209 0.718 0.788
JJA 0.894 0.917 0.908 0.722 0.580 0.562
SON 0.952 0.946 0.917 0.546 0.782 0.797

Table 5. Area averaged bias, RMSE and spatial correlation coefficient of temperature and rainfall climate change signal between 1981
2000 and 20812100, for ANN, ANNTS, and ECHAM5 interpolated to RegCM3 spatial scale, taking RegCM3 fields as truth. First
three columns show results for EurANN and EurANNTS validated over the European region, and the remaining columns show results for
MedANN and MedANNTS validated over the Mediterannean region.

Europe Med
EurANN EurANNTS Lapse-rate/GCM MedANN MedANNTS GCM
R Bias
(mm day1 )
DJF 0.247 0.068 0.023 0.316 0.023 0.037
MAM 0.255 0.065 0.035 0.071 0.087 0.131
JJA 0.358 0.055 0.073 0.319 0.105 0.053
SON 0.300 0.005 0.016 0.278 0.599 0.517
R RMSE
(mm day1 )
DJF 0.292 0.161 0.169 0.417 0.187 0.152
MAM 0.286 0.162 0.163 0.245 0.173 0.207
JJA 0.367 0.149 0.169 0.320 0.124 0.091
SON 0.328 0.147 0.173 0.287 0.231 0.247
R Corr
DJF 0.750 0.887 0.803 0.395 0.696 0.831
MAM 0.793 0.889 0.850 0.072 0.562 0.207
JJA 0.728 0.764 0.654 0.561 0.636 0.567
SON 0.556 0.756 0.622 0.278 0.599 0.517

Nonlin. Processes Geophys., 18, 10131028, 2011 www.nonlin-processes-geophys.net/18/1013/2011/

R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale 1027

approach. However our results do point to the importance Dickinson, R., Henderson-Sellers, A., and Kennedy, P.: Biosphere-
of the issue of stationarity in statistical downscaling models, Atmosphere Transfer Scheme, BATS: version 1E as coupled to
which needs to be addressed in further research. the NCAR Community Climate Model, Tech. Rep. NCAR/TN-
The use of local inputs could in principle reduce the abil- 387+STR, NCAR, 1993.
ity of the ANN to reproduce precipitation and temperature Engen-Skaugen, T.: Refinement of dynamically downscaled pre-
cipitation and temperature scenarios, Climatic Change, 84, 365
changes associated with changes in synoptic scale circula-
382, 2007.
tion. As the ANN performs well for the validation period Fritsch, J. and Chappell, C.: Numerical prediction of convectively
19812000, the use of local inputs does not appear to be a driven mesoscale pressure systems, J. Atmos. Sci., 37, 1722
problem when the climate change signal between training 1733, 1980.
and validation period is small. However, as noted above, Giorgi, F. and Francisco, R.: Evaluating uncertainties in the predic-
one possible cause of the poor performance of the original tion of regional climate change, Geophys. Res. Lett., 27, 1295
non-timeslice ANN in the 20812100 period is that regional 1298, 2000.
circulation changes occur in the RCM that are not present Giorgi, F. and Mearns, L.: Introduction to special section: Regional
in the GCM. It is possible that the introduction of non-local climate modeling revisited, J. Geophys. Res.-Atmos., 104, 6335
inputs into the time-slice ANN could provide information to 6352, 1999.
the ANN about any divergence between the RCM and GCM Giorgi, F., Marinucci, M., Bates, G., and DeCanio, G.: Develop-
ment of a second generation regional climate model (RegCM2)
circulations with time, and this is something that merits fur-
2: Convective processes and assimilation of lateral boundary
ther investigation. conditions, Mon. Weather Rev., 121, 28142832, 1993.
The time-slice ANN was in general able to reproduce the Giorgi, F., Hewitson, B., Christenson, J., Fu, C., Jones, R., Hulme,
RCM climate-change signal more accurately than the GCM, M., Mearns, L., Storch, H. V., and Whetton, P.: Regional climate
but not in all cases. We plan to further refine the ANNTS ap- information evaluation and projections., In IPCC WG1 TAR,
proach in order to use it as a tool to fill the GCM-RCM sim- 2001.
ulation matrix necessary to assess uncertainties in regional Giorgi, F., Diffenbaugh, N., Gao, X., Coppola, E., Dash, S., Fru-
climate change projections. mento, O., Seidou Sanda, I., Rauscher, S., Remedio, A., Steiner,
A., Sylla, B., and Zakey, A.: Exploring uncertainties in regional
Acknowledgements. This study was performed while RC was a climate change: The Regional Climate Change Hyper-Matrix
visiting scientist in the Earth Systems Physics group of the ICTP, Framework., Eos, 89, 445446, 2008.
and was made possible by the kind invitation for him to spend two Grell, G., Dudhia, J. and Stauffer, D.: Description of the fifth gen-
months working there. Helpful comments from two anonymous eration Penn State/NCAR Mesoscale Model, MM5, Tech. Rep.
reviewers also helped to improve the manuscript. NCAR/TN-398+STR, NCAR, 1994.
Haupt, S., Pasini, A., and Marzban, C. (Eds.): Artificial Intelligence
Edited by: S. Vannitsem Methods in the Environmental Sciences, Springer, 2009.
Reviewed by: two anonymous referees Hornick, K., Stinchcombe, M., and White, H.: Multilayer feedfor-
ward networks are universal approximators, Neural Networks, 2,
359366, 1989.
References Hsieh, W.: Machine learning methods in the environmental sci-
ences, Cambridge University Press, Cambridge, UK, 2009.
Bastos, L. and OHagan, A.: Diagnostics for Gaussian process em- Kiehl, J., Hack, J., Bonan, G., Boville, B., Briegleb, B., Williamson,
ulators, Technometrics, 51, 425438, 2008. D., and Rasch, P.: Desciption of the NCAR Community Climate
Bishop, C.: Neural Networks for Pattern Recognition, Clarendon Model, CCM3, Tech. Rep. NCAR/TN-420+STR, NCAR, 1996.
Press, UK, 2000. Lonnblad, L., Peterson, C., and Rognvaldsson, T.: Using neural
Busch, U. and Heimann, D.: Statistical-dynamical extrapolation of networks to identify jets, Nuclear Phys. B, 349, 675702, 1991.
a nested regional climate simulation, Clim. Res., 19, 113, 2001. Maraun, D., Wetterhall, F., Ireson, A., Chandler, R., Kendon, E.,
Chadwick, R. and Grimes, D.: An artificial neural network ap- Widmann, M., Brienen, S., Rust, H., Sauter, T., Themesl, M.,
proach to multi-spectral rainfall estimation over Africa., J. Hy- Venema, V., Chun, K., Goodess, C., Jones, R., Onof, C., Vrac,
drometeor., submitted, 2011. M., and Thiele-Eich, I.: Precipitation downscaling under climate
Conti, S. and OHagan, A.: Bayesian emulation of complex multi- change: Recent developments to bridge the gap between dy-
output and dynamic computer models, J. Stat. Plan. Inf., 140, namical models and the end user, Rev. Geophys., 48, RG3003,
640651, 2010. doi:10.1029/2009RG000314, 2010.
Coppola, E.: Validation of improved TAMANN neural network for Pal, J., Small, E., and Eltahir, E.: Simulation of regional-scale water
operational satellite rainfall estimation in Africa, J. Appl. Mete- and energy budgets: Representation of subgrid cloud and precip-
orol., 45, 15571572, 2006. itation processes within RegCM, J. Geophys. Res., 105, 29579
Deque, M., Jones, R., Wild, M., Giorgi, F., Christensen, J., Hassell, 29594, 2000.
D., Vidale, P., Rockel, B., Jacob, D., Kjellstrom, E., de Castro, Pal, J., Giorgi, F., Bi, X., Elguindi, F., Solomon, F., Gao, X., Fran-
M., Kucharski, F., and van den Hurk, B.: Global high resolu- cisco, R., Zakey, A., Winter, J., Ashfaq, M., Syed, F., Bell,
tion versus Limited Area Model climate change projections over J., Diffanbaugh, N., Kamacharya, J., Konare, A., Martinez, D.,
Europe: quantifying confidence level from PRUDENCE results, da Rocha, R., Sloan, L., and Steiner, A.: The ICTP RegCM3
Clim. Dynam., 25, 653670, 2005.

www.nonlin-processes-geophys.net/18/1013/2011/ Nonlin. Processes Geophys., 18, 10131028, 2011

1028 R. Chadwick et al.: An ANN technique for downscaling GCM outputs to RCM spatial scale

and RegCNET: Regional climate modelling for the developing Tomassetti, B., Verdecchia, M., and Giorgi, F.: NN5: A neural net-
world, Bull. Amer. Meteorol. Soc., 88, 13951409, 2007. work based approach for the downscaling of precipitation fields
Pasini, A. and Langone, R.: Attribution of precipitation changes Model description and preliminary results, J. Hydrol., 367, 14
on a regional scale by neural netwrok modelling: A case study, 26, 2009.
Water, 2, 321332, 2010. Van Der Linden, P. and Mitchell, J.: ENSEMBLES: Climate
Pasini, A., Lore, M., and Ameli, F.: Neural network modelling for Change and its Impacts: Summary of research and results from
the analysis of forcings/temperatures relationships at different the ENSEMBLES project, Tech. rep., Met Office Hadley Centre,
scales in the climate system, Ecol. Model., 191, 5867, 2006. Fitzroy rd, Exeter, EX1 3PB, UK, 2009.
Picton, P.: Neural Networks, 2nd Edn., Palgrave, Basingstoke, UK, Wilby, R., Wigley, T., Conway, D., Jones, P., Hewitson, B., Main,
2000. J., and Wilks, D.: Statistical downscaling of general circulation
Roeckner, E., Bauml, G., Bonaventura, L., Brokopf, R., Esch, model output: A comparison of methods, Water Resour. Res., 34,
M., Giorgetta, M., Hagemann, S., Kirchner, I., Kornblueh, 29953008, 1998.
L., Manzini, E., Rhodin, A., Schlese, U., Schulzweida, U., Zeng, X., Zhao, M., and Dickinson, R.: Intercomparison of bulk
and Tompkins, A.: The atmosphere general circulation model aerodynamic algorithms for the computation of sea surface fluxes
ECHAM5. Part 1: Model description, Tech. Rep. Rep. 349, Max using TOGA COARE and TAO data, J. Climate, 11, 26282644,
Planck Institute fo Meteorology, 2003. 1998.
Rumelhart, D. and McClelland, J.: Parallel Distributed Processing,
MIT press, Cambridge, MA, 1986.
Solomon, S., Qin, D., Manning, M., Chen, Z., Marquis, M., Av-
eryt, K., Tignor, M., and Miller, H. (Eds.): IPCC, 2007: Cli-
mate Change 2007: The Physical Science Basis. Contribution of
Working Group 1 to the Fourth Assessment Report of the Inter-
governmental Panel on Climate Change, Cambridge University
Press, Cambridge, United Kingdom and New York, NY, USA,
2007.

Nonlin. Processes Geophys., 18, 10131028, 2011 www.nonlin-processes-geophys.net/18/1013/2011/