You are on page 1of 9

Clim Dyn (2016) 46:2749–2757

DOI 10.1007/s00382-015-2727-6

Deficiencies in the simulation of the geographic distribution


of climate types by global climate models
Xianliang Zhang1 · Xiaodong Yan2

Received: 23 April 2014 / Accepted: 22 June 2015 / Published online: 8 July 2015
© Springer-Verlag Berlin Heidelberg 2015

Abstract The performances of General Circulation Mod- Keywords Model evaluation · Spatial distribution ·
els (GCMs) when checked with conventional methods (i.e. GCMs · Regional climate pattern
correlation, bias, root-mean-square error) can only be eval-
uated for each variable individually. The geographic distri-
bution of climate type in GCM simulations, which reflects 1 Introduction
the spatial attributes of models and is related closely to
the terrestrial biosphere, has not yet been evaluated. Thus, A General Circulation Model (GCM) is a mathematical
whether the geographic distribution of climate types was model of the general circulation that is established based
well simulated by GCMs was evaluated in this study for on equations that describe the physics of the conserva-
nine GCMs. The results showed that large areas of climate tion of mass, energy and momentum (Hansen et al. 1983).
zones classified by the GCMs were allocated incorrectly Due to their solid physics-based foundations, GCMs are
when compared to the basic climate zones established by widely accepted as a useful tool to apply in weather fore-
observed data. The percentages of wrong areas covered casting, understanding the climate system, and projecting
approximately 30–50 % of the total land area for most climate change (Räisänen 2007). GCMs have evolved and
models. In addition, the temporal shift in the distribution of improved into the currently used coupled atmosphere-
climate zones according to the GCMs was found to be inac- ocean GCMs (AOGCMs) (Taylor et al. 2012). Using a
curate. Not only were the locations of shifts poorly simu- number of AOGCMs, historical climate change from 1850
lated, but also the areas of shift in climate zones. Overall, to 2005 and the future climate change under different sce-
the geographic distribution of climate types was not simu- narios have been simulated (IPCC AR5 2013). The simu-
lated well by the GCMs, nor was the temporal shift in the lated results are important for understanding the climate
distribution of climate zones. Thus, a new method on how variability of the past and for predicting changes in the
to evaluate the simulated distribution of climate types for future. However, there are more than 40 AOGCMs cur-
GCMs was provided in this study. rently used to simulate the world’s climate, each with its
own strengths and weaknesses. The skill of these models
in simulating present-day and past climate changes affects
their reliability in future predictions (Whetton et al. 2007;
* Xianliang Zhang Reichler and Kim 2008; Reifen and Toumi 2009). There-
zhangxianliang@syau.edu.cn fore, the performance of each AOGCM should be evaluated
* Xiaodong Yan before using it to predict future climate changes.
yxd@bnu.edu.cn In general, model performance is measured by three
1 statistics; namely, correlation, bias, and root-mean-square
College of Forestry, Shenyang Agriculture University,
Shenyang 110086, China error (RMSE) (Gates et al. 1999; Taylor 2001). Correla-
2 tion evaluates the association between observed data and
State Key Laboratory of Earth Surface Processes
and Resource Ecology, Beijing Normal University, model simulations, while bias and RMSE indicate the dis-
Beijing 100087, China tances between observed and simulated data. However,

13
2750 X. Zhang, X. Yan

model performance can only be evaluated for single vari- Climate zones, as a representation of the spatial distri-
ables individually (Gates et al. 1999; Pincus et al. 2008; bution of regional climate patterns, were classified in the
Reichler and Kim 2008). Thus, the spatial distribution of present study based on observational data and model simu-
climate could be reflected by the distribution of model lations. Model performance in terms of the spatial attrib-
mean bias or the RMSE for climate variables (Reichler utes of climate was then evaluated by comparing the dis-
and Kim 2008; Knutti and Sedlacek 2013). These methods, tribution of climate zones obtained by the models with
however, cannot evaluate model performance with respect that obtained from the observational data. Models showing
to the geographic distribution of climate types, which is an better performance could then be used to predict the dis-
indicator of regional climate characteristics. For instance, tribution of climate zones in the future with greater confi-
whether or not a model correctly represents the location of dence. Accordingly, the main goals of this study were to (a)
arid climate zones could not be evaluated using the above- evaluate model performances in terms of their representa-
mentioned statistics. The successful simulation of the dis- tion of the geographic distribution of climate types using
tribution of climate zones by climate models is crucial to climate classification; and (b) investigate whether temporal
test model predictions of the future distribution of climate shifts in climate zones could be detected using the model
zones. Future shifts in the distribution of climate zones can simulations.
be predicted using model simulations under different sce-
narios, and it is not only important for climate research (De
Castro et al. 2007; Kim et al. 2008; Feng et al. 2012; Hanf 2 Data and method
et al. 2012), but also for other disciplines, such as agricul-
ture (Papadakis 1975; Cramer and Solomon 1993) and ter- 2.1 Climate data
restrial ecology (Duchesne and Magnan 1997). Given the
importance of temporal shifts in the distribution of climate Global monthly mean temperature and precipitation data
types in related areas, it is clear that an accurate representa- simulated by nine AOGCMs (BCC-CSM1-1, CAN-ESM2,
tion of the geographic distribution of climate types is cru- CNRM-CM5, GISS-E2-H, INMCM4, IPSL-CM5A-LR,
cial to both climate and bio-climatic research. MIROC-ESM, MPI-ESM-P and NOR-ESM1-M) from
The geographic distribution of climate types can be 1901 to 2005 were retrieved from the CMIP5 website
obtained by a climate classification process (Kottek et al. (http://www-pcmdi.llnl.gov/ipcc/about_ipcc.php). Detailed
2006; Mahlstein and Knutti 2010). Generally, the Koeppen descriptions of the models and their data can be found
classification method and its subsequent revisions are the on the website. The monthly observed mean temperature
most widely used approaches to classify the climate of the and precipitation data over land (excluding the Antarctic)
world (Guetter and Kutzbach 1990; Fraedrich et al. 2001; were obtained from the Climate Research Unit (CRU) TS
Beck et al. 2005; Peel et al. 2007; Baker et al. 2010). The 3.0 dataset (www.cru.uea.ac.uk). The resolution of the
errors in model simulations have been evaluated using the observed data is 1° × 1°, so the model simulations were
Koeppen climate classification (Gnanadesikan and Stouffer interpolated to the same resolution using a bilinear method
2006). The biases of threshold values can be reflected by and masked with the observed data.
the Koeppen climate classification method because the
boundaries are defined by threshold values. Thus, Koep- 2.2 K‑means cluster method
pen climate classification has advantages in evaluating the
spatial distribution of model bias, as many models produce The K-means cluster method is an unsupervised method
a cold bias in most grid cells of the world (Cordero and widely used in cluster analysis to classify multivariable
Forster 2006). However, classifying regions with similar data into a given number of clusters. Generally, there are
climate characteristics to the same climate zone is difficult four steps involved in the K-means method. First, K cen-
to achieve based on the climate zones used in the Koeppen troids are randomly selected as the initial centroids for K
climate classification method. As pointed out by Gnana- groups. Second, each point in the dataset is assigned to its
desikan and Stouffer (2006), the more recently developed nearest centroid, so the dataset is allocated to K groups.
schemes should instead be used to investigate these issues. Third, the centroid in every group is recalculated and
As an alternative approach, the K-means cluster method is assigned as a new centroid. Then, the K new centroids are
an effective method that can classify multiple variables to obtained. Finally, steps 2 and 3 are repeated until the cen-
a certain number of climate zones (Mahlstein and Knutti troids no longer move. More details of the K-means algo-
2010; Zhang and Yan 2014a, b). Regions with similar cli- rithm can be found in Zhang and Yan (2014a).
mate attributes can be classified to one climate zone. The The initial centroids used in the K-means method have a
K-means cluster method is thus suitable for detecting the significant influence on the location of climate zones. The
geographic distribution of climate types. initial centroids were generated using the observed climate

13
Deficiencies in the simulation of the geographic distribution of climate types by global… 2751

data and used as the initial centroids in the classification 3 Results


for every GCM. However, the initial centroids did not influ-
ence the final results. The boundaries of climate zones were 3.1 Advantages and disadvantages of different metrics
only sensitive to the bias in temperature and precipitation,
and small differences in climate variables could not affect In order to illustrate the advantages and disadvantages of
the classification in boundary regions significantly. different metrics, an example is first provided. Consider the
The number of clusters should be assigned prior to the following temperature distributions as a function of y over
cluster analysis. Following the climate classification by the the interval (0, L):
K-means method in Zhang and Yan (2014a), the number of
clusters was assigned to 14 in the present study. Moreover, 1. T1 = 20 × cos(pi × y/L)
the influences of units should be eliminated for temperature 2. T2 = 18 × cos(pi × y/L)
and precipitation. Thus, monthly temperature and precipita- 3. T3 = 20 × cos(pi × y/L) − 2
tion were standardized using the following method: 4. T4 = 20 × (1 − 2* abs(y)/L)
x − min(X) 5. T5 = 20 × (1 − 2* abs(y)/L) − 2
Z= (1)
max(X) − min(X)
If we try to cluster these temperature distributions
where Z is the standard index, the scale is 0–1, X is a vari- into “warm” and “cold” regimes, we obtain the same
able, x is every value in X, max(X) is the maximum value in two clusters with the same boundary (y = L/2) for all of
X, and min(X) is the minimum value in X. these spatial distributions. If, however, we look at corre-
lations, we find that the linear ones are the same and the
2.3 Kappa test sinusoids are the same; while if we look at a threshold
of 10, we obtain different amounts classified in different
The observed data and model simulations were classi- clusters (T1 and T4 will form one cluster, T2 another,
fied using the K-means method. The agreement between and T3 and T5 a third). Using different measures, we
the climate zones obtained using observed data and those find:
obtained using model-simulated data were checked using
the Kappa test. Kappa statistics were first introduced by 1. RMSE: Sensitive to patterns within climate zones,
Cohen (1960) to measure the agreement between observ- global biases; insensitive to errors at zone boundaries,
ers of psychological behavior. The pattern of agreement values of thresholds.
between two climate zones could be evaluated by the 2. Correlation: Sensitive to patterns within climate zones;
Kappa value, and was calculated as: insensitive to global biases and values of thresholds,
po − p e errors at zone boundaries.
Kappa = (2) 3. Koeppen: Insensitive to patterns within climate zones;
1 − pe
sensitive to global biases, errors at zone boundaries,
Suppose i represents a climate zone for the observation and value of thresholds.
j represents a climate zone for the model simulation. Pij is 4. K-means: Less sensitive to patterns within climate
the proportion of the total number of grids assigned to cli- zones (some sensitivity as an extreme value will
mate zones i and j. These proportions can be formed into a shift the boundary a little) and to mean errors at zone
square matrix. The main diagonal consists of the individual boundaries; sensitive to regional bias, and picks thresh-
proportions Pii of grids where both the climate zones agree. olds more objectively.
Po is the sum of the Pii values. Pe is the expected propor-
tion of agreement for the case in which the simulation and A major aim of the present study is to reflect the spatial
observations are uncorrelated. distribution of regional climate patterns. From the above
The range of the Kappa value is −1.0 to 1.0. A value descriptions, we can see that the K-means method can clas-
of zero indicates that the agreement is no better than that sify climate types objectively and reflect the regional bias.
expected by chance, a negative value indicates less agree- Besides, the K-means method carries more advantages than
ment than would be expected by chance, and a value of 1.0 the other three metrics in classifying climate types. First,
represents complete agreement. “Agreement” can be “poor systematic bias does not affect the clustering results. All
to very poor” when Kappa <0.4, while 0.40–0.55 is “fair”, model simulations contain a systematic bias, which can be
0.55–0.70 is “good”, 0.70–0.85 is “very good to excellent”, removed by a bias-removal method (e.g. Kharin and Zwi-
and Kappa >0.99 is “perfect” agreement (Landis and Koch ers 2002; Acharya et al. 2013), and the anomalies of model
1977; Monserud and Leemans 1992). simulations can be obtained after the systematic bias has

13
2752 X. Zhang, X. Yan

been removed. However, Koeppen classification methods zones. Moreover, large areas were classified to the wrong
could not process the anomalies of the model simulations. climate zone for every model. The climate zone DF was
Second, the K-means method carries advantages in cluster- classified to a totally wrong climate zone (ET) in models
ing multivariable data into several groups according to their CAN-ESM2, MIROC-ESM and MPI-ESM-P. However, the
consistencies. The level of consistency in multivariable model MPI-ESM-P obtained the most similar spatial distri-
data is important in climate research. For instance, in some bution pattern of climate zones.
climate types, the temperature is low when there is a lot of The total area classified to the wrong climate zone
precipitation. The consistency in multivariable data could was calculated (Table 2). The “wrong” areas amounted
not be reflected by the other clustering methods. Thus, the to >50 million km2, and the percentages of wrong areas
K-means method is superior to the other methods in clus- covered approximately 30–50 % of the total land area for
tering climate types and was used to classify the climate most models. The areal coverage of every cluster obtained
zones in this study. using observations and the model simulations were also
calculated (Table 3). Large areas were classified to the
3.2 Evaluation of the distribution of climate zones wrong climate zones for more than half of all climate
in model simulations types.

Climate zones, which reflect the spatial distribution of 3.3 Temporal shift in the distribution of climate zones
regional climate patterns, were classified for every model in the model simulations
to evaluate whether the distributions of climate zones were
well simulated. The basic climate zones classified accord- The temporal shift in the distribution of climate zones
ing to observations and the climate zones classified accord- was studied by comparing the climate zones in the period
ing to the nine models are shown in Fig. 1. The distribution 1981–2005 with those in the period 1901–1925. Those
of the simulated climate zones was similar to the distri- obtained by the nine climate models were compared with
bution of the basic climate zones. However, the climate that obtained by the observational data (Fig. 3). There
types ET (tundra climate) and DF (fully humid continen- was a clear temporal shift in the boundaries of climate
tal climate) were classified as one climate type in models zones in the observation; specifically, in the boundaries
CAN-ESM2, MIROC-ESM and MPI-ESM-P. Moreover, of AF and AM (equatorial monsoon climate), DF and
certain climate types had no corresponding basic climate DW (continental climate with dry winter), BS (steppe cli-
zone because one basic climate type was classified into mate) and BW (desert climate). However, these changes
two climate types in the model simulations. For instance, were not reproduced by the models. In fact, the bounda-
AF (fully humid equatorial rainforest) of the basic climate ries of climate zones that did not shift in the basic cli-
zones was classified to two climate zones in model BCC- mate zones, did shift in many of the model results. For
CSM1-1 (AF and a “New” climate zone). Moreover, the instance, a larger than observed degree of boundary shift
RMSE between the centroids of the clusters from the mod- was simulated by the model BCC-CSM1-1. Moreover,
els and the centroids of the clusters from the observation the area of shift was not precisely captured by the mod-
was small (Table 1). However, differences in RMSE were els; a larger than observed area was detected in some
large in the clusters of some models when the “New” cli- models, such as BCC-CSM1-1and GISS-E2-H, while a
mate type occurred for these models. smaller than observed area was detected in others, such
The level of agreement between the distribution of the as INMCM4. The results of model MPI-ESM-P were the
modeled climate zones and the basic climate zones was closest to observations; however, it still showed less shift
evaluated by Kappa values (Table 2). The model MPI- (in terms of area) than was actually the case, according
ESM-P had the highest Kappa value (0.6481), suggesting to the observed data. Overall, the shift in the spatial dis-
“good” agreement with the basic climate zones. “Good” tribution of climate zones was not simulated well by the
agreement was also found for the models BCC-CSM1-1, models.
CAN-ESM2, CNRM-CM5, GISS-E2-H and INMCM4.
The Kappa values of models IPSL-CM5A-LR, MIROC-
ESM and NOR-ESM1-M were lower than 0.55, suggesting 4 Discussion
“fair” agreement with the basic climate zones.
The differences between the distribution of climate Conventionally, model performance tends to be evaluated
zones obtained by the model simulations and the basic cli- with respect to a single time series, such as the regional
mate zones are shown in Fig. 2. Only the boundaries of a mean, using measures such as correlation and RMSE.
few climate zones were similar to those of the basic climate In the present study, we took a different approach to

13
Deficiencies in the simulation of the geographic distribution of climate types by global… 2753

Fig. 1  Climate zones classified according to observations (CRU) and ter; DF, fully humid continental climate; DS, continental climate with
the nine AOGCMs using the K-means method. Climate zone abbre- dry summer; DW, continental climate with dry winter; ET, tundra cli-
viations: AF, fully humid equatorial rainforest; AM, equatorial mon- mate; EF, frost climate. Climate zone “NEW” in the models BCC-
soon climate; AW, equatorial savannah with dry winter; AS, equatorial M1-1, CNRM-CM5, GISS-E2-H, NOR-ESM1-M, MIROC-ESM and
savannah with dry summer; BW, desert climate; BS, steppe climate; MPI-ESM-P was assigned because it had no corresponding basic cli-
CF, fully humid warm temperate climate; CS, warm temperate cli- mate zone
mate with dry summer; CW, warm temperate climate with dry win-

evaluating the geographic distribution of climate types by reasons for the climate pattern bias could be inaccuracies
using the K-means cluster method. in the simulation of the large-scale atmospheric circulation
The geographic distribution of climate types was not and/or deficiencies in land surface modeling or in processes
well simulated by the models, with large areas classified to associated with the hydrological cycle. The areas classi-
incorrect climate zones—a finding that is consistent with fied to incorrect climate zones were mainly two climate
the research of Gnanadesikan and Stouffer (2006). Possible zones classified to one climate zone, or one climate zone

13
2754 X. Zhang, X. Yan

Table 1  The root mean square error (RMSE) between the clusters from the models and the clusters from the observation, calculated based on
the standardized data to avoid the influences of units
BCC-CSM1-1 CAN-ESM2 CNRM-CM5 GISS-E2-H INMCM4 IPSL-CM5A-LR MIROC-ESM MPI-ESM-P NOR-ESM1-M Smallest
RMSE

Af 0.88 0.31 0.45 0.36 0.47 0.76 0.77 0.43 0.21 0.21
Am 0.58 0.39 0.32 0.83 0.29 – 0.64 0.42 0.20 0.20
As 0.11 0.32 0.14 0.38 0.23 0.35 0.42 0.35 0.16 0.11
Aw 0.37 0.27 0.24 0.27 0.24 0.36 0.60 0.31 0.30 0.24
BS 0.27 0.38 0.19 0.40 0.19 0.25 0.52 0.42 0.13 0.13
BW 0.19 0.25 0.15 0.18 0.42 0.20 0.48 0.25 0.17 0.15
Cs 0.35 0.28 0.26 0.50 0.21 0.31 0.22 0.31 0.17 0.17
Cw 0.45 0.27 0.34 0.51 0.15 0.31 0.28 0.45 0.17 0.17
Cf 0.58 0.39 0.61 0.40 0.28 0.44 0.78 0.63 0.18 0.18
Ds 0.28 0.08 0.19 0.29 0.14 0.39 0.25 0.31 0.28 0.14
Dw 0.27 0.17 0.21 0.31 0.15 0.17 0.32 0.35 0.34 0.17
Df 0.27 0.53 0.20 0.32 0.12 0.16 – – 0.39 0.12
ET – 0.95 – – 0.59 0.82 1.07 1.06 – 0.59
EF 0.42 0.43 0.44 0.40 0.31 0.34 0.44 0.50 0.43 0.31

Table 2  Kappa values and the area/proportion of incorrect climate Indeed, the distribution of all climate zones would have
zones (excluding the Antarctic) for the nine AOGCMs been similar between them and observations if the climate
Model Kappa value Area (million Proportion of zone DF had been properly represented. Moreover, the
km2) global land area distribution of certain climate zones was simulated well
(%) by some models. The Kappa value is a useful measure for
BCC-CSM1-1 0.5726 57.2 39.06 ranking the order of agreement between two patterns (Mon-
CAN-ESM2 0.6213 50.4 34.43 serud and Leemans 1992), and was adopted in the present
CNRM-CM5 0.5451 56.6 38.67 study to rank the performance of the models in terms of
GISS-E2-H 0.5718 58.5 39.98 their representation of the spatial distribution of climate.
INMCM4 0.5740 56.8 38.83 We recommend that the Kappa value is used more widely
IPSL-CM5A-LR 0.4455 75.4 51.48 in the future as a tool to evaluate model performance with
MIROC-ESM 0.4800 74.0 50.56 respect to spatial attributes.
MPI-ESM-P 0.6481 48.1 32.88
The temporal shift of climate zones was also not simu-
NOR-ESM1-M 0.4785 72.0 49.19
lated well by the models. The boundaries of climate zones
were inaccurately positioned, meaning the shifts of bound-
aries in the models’ results were also inaccurate. Moreo-
ver, not only were the locations of shifts poorly simulated,
classified to two climate zones. When two climate zones but also the areas of shift in climate zones. Thus, cau-
were classified to one climate zone, one of the two climate tion should be applied when using model simulations to
zones was assigned as the basic climate zone because it examine the temporal shifts in the distribution of climate
had a larger crossover of similarity with the basic climate zones. This is also true of the Koeppen climate classifica-
zone than the other. The other of the two climate zones was tion method, under which approach it has also been found
then assigned as a new climate zone, despite having simi- that the boundaries of climate zones are poorly simulated
lar climate characteristics as the basic climate zone. In the (Gnanadesikan and Stouffer 2006). Therefore, future shifts
other situation, the climate attributes of one basic climate in the distribution of climate zones under different scenar-
zone were simulated by some models as being quite simi- ios reported in the literature (Rubel and Kottek 2010; Hanf
lar to those of another basic climate zone, and so the two et al. 2012; Mahlstein et al. 2013) should also be regarded
basic climate zones were classified to one climate zone in with caution. In conclusion, the spatial distribution of
these models’ results. Among these models, CAN-ESM2 regional climate patterns cannot be simulated well by the
and MPI-ESM-P produced climate zone distribution pat- models examined in this study, nor can the temporal shift in
terns that were the most similar to the basic climate zones. the distribution of climate zones.

13
Deficiencies in the simulation of the geographic distribution of climate types by global… 2755

Fig. 2  The areas classified to incorrect climate zones by the nine AOGCMs

Table 3  Areal coverage (million km2) of every cluster from the model simulations and the observation
CRU BCC-CSM1-1 CAN-ESM2 CNRM-CM5 GISS-E2-H INMCM4 IPSL-CM5A-LR MIROC-ESM MPI-ESM-P NOR-ESM1-M

Af 9.99 3.25 6.93 4.61 8.34 8.04 5.04 6.73 4.95 9.24
Am 4.09 5.50 12.52 8.36 0.65 0.00 5.01 6.20 6.57 9.06
As 13.48 13.13 14.20 14.23 8.24 14.28 9.75 9.78 11.81 10.69
Aw 16.00 17.17 13.24 18.77 21.73 13.12 14.22 11.21 15.88 9.69
BS 12.27 14.17 9.23 12.53 14.23 16.30 14.71 8.71 13.94 16.37
BW 17.31 16.25 15.59 17.30 11.92 15.14 14.11 15.34 15.19 10.76
Cs 12.45 13.20 12.97 12.51 13.47 19.41 15.56 10.73 12.48 11.41
Cw 15.15 14.04 13.81 7.49 15.14 14.70 14.15 13.84 15.15 15.03
Cf 4.39 8.51 3.90 3.43 3.86 6.30 4.55 6.88 5.85 6.55
Ds 16.35 14.17 14.21 14.69 14.98 5.29 15.96 15.69 15.70 14.02
Dw 12.72 10.63 13.66 13.99 11.69 17.86 12.27 15.20 11.78 9.70
Df 4.45 7.42 5.13 6.65 8.76 9.67 10.39 – – 6.85
ET 5.78 – 8.52 – – 4.35 8.22 10.06 8.30 –
EF 1.97 2.59 2.46 1.24 3.57 1.92 2.44 2.86 2.38 2.88
New – 6.35 – 10.57 9.32 – – 13.15 6.38 14.11

13
2756 X. Zhang, X. Yan

Fig. 3  Comparison of the temporal shift in the distribution of climate zones between the periods 1981–2005 and 1901–1925 obtained by the
nine AOGCMs (black) with that obtained by the observational data (green)

Acknowledgments This work was funded by a National Key Sci- Cohen J (1960) A coefficient of agreement for nominal scales. Educ
entific Project of China (2012CB95570000) and the National Natural Psychol Meas 20:37–46
Science Foundation of China (41271066 and 41330527). The authors Cordero EC, Forster PDF (2006) Stratospheric variability and
acknowledge the World Climate Research Program’s Working Group trends in models used for the IPCC AR4. Atmos Chem Phys
on Coupled Modeling, which is responsible for CMIP, and we thank 6:5369–5380
the climate modeling groups for producing and making available their Cramer WP, Solomon AM (1993) Climatic classification and future
model output. We also thank the two anonymous reviewers for their global redistribution of agricultural land. Clim Res 3:97–110
constructive comments. De Castro M, Gallardo C, Jylha K, Tuomenvirta H (2007) The use of
a climate-type classification for assessing climate change effects
in Europe from an ensemble of nine regional climate models.
Clim Change 81:329–341
References Duchesne J, Magnan P (1997) The use of climate classification
parameters to investigate geographical variations in the life his-
Acharya N, Chattopadhyay S, Mohanty UC, Dash SK, Sahoo LN tory traits of ectotherms, with special reference to the white
(2013) On the bias correction of general circulation model out- sucker (Catostomus commersoni). Ecoscience 4:140–150
put for Indian summer monsoon. Meteorol Appl 20:349–356 (Sainte-Foy)
Baker B, Diaz H, Hargrove W, Hoffman F (2010) Use of the Kop- Feng S, Ho C, Hu Q, Oglesby RJ, Jeong S, Kim B (2012) Evaluat-
pen Trewartha climate classification to evaluate climatic refugia ing observed and projected future climate changes for the Arctic
in statistically derived ecoregions for the People’s Republic of using the Köppen-Trewartha climate classification. Clim Dyn
China. Clim Change 98:113–131 38:1359–1373
Beck C, Grieser J, Kottek M, Rubel F, Rudolf B (2005) Character- Fraedrich K, Gerstengarbe FW, Werner PC (2001) Climate shifts dur-
izing global climate change by means of Koeppen climate clas- ing the last century. Clim Change 50:405–417
sification Klimastatusbericht. Dtsch. Wetterdienst, Berlin, pp Gates WL, Boyle JS, Covey C, Dease CG, Doutriaux CM, Drach
139–149 RS, Fiorino M, Gleckler PJ, Hnilo JJ, Marlais SM (1999) An

13
Deficiencies in the simulation of the geographic distribution of climate types by global… 2757

overview of the results of the Atmospheric Model Intercompari- Peel MC, Finlayson BL, McMahon TA (2007) Updated world map of
son Project (AMIP I). Bull Am Meteorol Soc 80:29–55 the Köppen-Geiger climate classification. Hydrol Earth Syst Sci
Gnanadesikan A, Stouffer RJ (2006) Diagnosing atmosphere-ocean Discuss 4:439–473
general circulation model errors relevant to the terrestrial bio- Pincus R, Batstone CP, Hofmann RJP, Taylor KE, Glecker PJ (2008)
sphere using the Köppen climate classification. Geophys Res Evaluating the present-day simulation of clouds, precipitation,
Lett 33:L22701. doi:10.1029/2006GL028098 and radiation in climate models. J Geophys Res 113:D14209. doi
Guetter PJ, Kutzbach JE (1990) A modified Koppen classification :10.1029/2007JD009334
applied to model simulations of glacial and interglacial climates. Räisänen J (2007) How reliable are climate models? Tellus A 59:2–29
Clim Change 16:193–215 Reichler T, Kim J (2008) How well do coupled models simulate
Hanf F, Korper J, Spangehl T, Cubasch U (2012) Shifts of climate today’s climate? B Am Meteorol Soc 89:303–311
zones in multi-model climate change experiments using the Kop- Reifen C, Toumi R (2009) Climate projections: past performance no
pen climate classification. Meteorol Z 21:111–123 guarantee of future skill. Geophys Res Lett 36:L13704. doi:10.1
Hansen J, Russell G, Rind D, Stone P, Lacis A, Lebedeff S, Ruedy 029/2009GL038082
R, Travis L (1983) Efficient three-dimensional global models for Rubel F, Kottek M (2010) Observed and projected climate shifts
climate studies: models I and II. Mon Weather Rev 111:609–662 19012100 depicted by world maps of the Koppen-Geiger climate
Kharin VV, Zwiers FW (2002) Climate predictions with multimodel classification. Meteorol Z 19:135–141
ensembles. J Climate 15:793–799 Stocker TF, Qin D, Plattner GK, Tignor M, Allen SK, Boschung J,
Kim H, Wang B, Ding Q, Chung I (2008) Changes in arid climate Nauels A, Xia Y, Bex B, Midgley BM (2013) Climate change
over North China detected by the Koppen climate classification. 2013: the physical science basis intergovernmental panel on
J Meteorol Soc Jpn 86:981–990 climate change, working group I contribution to the IPCC fifth
Knutti R, Sedláček J (2013) Robustness and uncertainties in the new assessment report (AR5). Cambridge University Press, New York
CMIP5 climate model projections. Nat Clim Change 3:369–373 Taylor KE (2001) Summarizing multiple aspects of model perfor-
Kottek M, Grieser J, Beck C, Rudolf B, Rubel F (2006) World map of mance in a single diagram. J Geophys Res 106:7183–7192
the Kppen-Geiger climate classification. Meteorol Z 15:259–263 Taylor KE, Stouffer RJ, Meehl GA (2012) An overview of CMIP5
Landis JR, Koch GG (1977) The measurement of observer agreement and the experiment design. B Am Meteorol Soc 93:485–498
for categorical data. Biometrics 33:159–174 Whetton P, Macadam I, Bathols J, O’Grady J (2007) Assessment of
Mahlstein I, Knutti R (2010) Regional climate change patterns identi- the use of current climate patterns to evaluate regional enhanced
fied by cluster analysis. Clim Dyn 35:587–600 greenhouse response patterns of climate models. Geophys Res
Mahlstein I, Daniel JS, Solomon S (2013) Pace of shifts in climate Lett. doi:10.1029/2007GL030025
regions increases with global temperature. Nat Clim Change Zhang X, Yan X (2014a) Spatiotemporal change in geographical dis-
3:739–743 tribution of global climate types in the context of climate warm-
Monserud RA, Leemans R (1992) Comparing global vegetation maps ing. Clim Dyn 43:595–605
with the Kappa statistic. Ecol Model 62:275–293 Zhang X, Yan X (2014b) Temporal change of climate zones in
Papadakis J (1975) Climates of the world and their agricultural poten- China in the context of climate warming. Theor Appl Climatol
tialities. Edición Argentina, Buenos Aires, p 200 115:167–175

13

You might also like