Synoptic-Statistical Approach to Regional1
Downscaling of IPCC 21st Century Climate2
Projections: Seasonal Rainfall over the3
Hawaiian Islands4
Oliver Timm∗
IPRC, SOEST & University of Hawai’i
1680 East West Road, Honolulu, HI 96822, USA
5
Henry F. Diaz
NOAA-CIRES Climate Diagnostics Center
325 Broadway, Boulder, CO 80305
6
February 25, 20097
revised version8
9
Abstract10
A linear statistical downscaling technique is applied to the pro-11
jection of the Intergovernmental Panel on Climate Change (IPCC)12
fourth assessment report (AR4) climate change scenarios onto Hawai-13
ian rainfall for the late 21st century. Hawaii’s regional rainfall is14
largely controlled by the strength of the trade winds. During the win-15
ter months, disturbances in the westerlies can produce heavy rainfall16
throughout the Islands. A diagnostic analysis of sea level pressure17
(SLP), near-surface winds, and rainfall measurements at 134 weather18
observing stations around the islands characterize the correlations19
between the circulation and rainfall during the nominal wetseason20
(November–April) and dry season (May–October). A comparison of21
the base climate 20th century AR4 model simulations with reanalysis22
data for the period 1970–2000 is used to define objective selection cri-23
terion for the AR4 models. Six out of 21 available models werecho-24
1
sen for the statistical downscaling. These were chosen on the basis25
of their ability to more realistically simulate the modern large-scale26
circulation fields in the Hawaiian Islands region.27
For the AR4 A1B emission scenario, the six analyzed models28
show important changes in the wind fields around Hawaii by the29
late 21st century. Two models clearly indicate opposite signs in the30
anomalies. One model projects 20–30% rainfall increase over the is-31
lands; the other model suggests a rainfall decrease of about10–20%32
during the wet season. It is concluded from the 6-model ensemble33
that the most likely scenario for Hawaii is a 5–10% reductionof the34
wet-season precipitation and a 5% increase during the dry season, as35
a result of changes in the wind field. We discuss the sources ofun-36
certainties in the projected rainfall changes, and consider future im-37
provements of the statistical downscaling work, and implications for38
dynamical downscaling methods.39
2
1. Introduction40
The Intergovernmental Panel for Climate Change (IPCC) fourth assessment re-41
port (AR4) leaves no doubt that the increase in atmospheric CO2 concentration42
will lead to significant environmental changes in all regions of the globe (IPCC43
2007). However, the degree of uncertainty of the projected climate changes in-44
creases from global to regional scale. Small-scale topographic features such as45
the Hawaiian Islands are below the typical horizontal and vertical resolution of46
the GCMs and cannot be represented. In order to project the large-scale climate47
changes onto regional scales, nested regional models or statistical methods have48
been adopted (von Storch et al. 1993; Wilby and Wigley 1997; Christensen et al.49
2007; Schmidli et al. 2007; Wang and Zhang 2008). Here, we will apply statisti-50
cal downscaling (SD) to the rainfall of the Hawaiian Islandsunder the assumption51
that GCMs best simulate the large-scale atmospheric circulation patterns, and that52
future changes in those principal climate modes would be themost likely to affect53
the climate of Hawaii.54
The Hawaiian Islands are located in the trade wind zone southwest of the sub-55
tropical high-pressure system of the North Pacific Ocean. Asnoted above, the56
small-scale topographic features of the islands are not represented in the suite of57
3
IPCC (2007) GCMs, and therefore the interaction between thelarge-scale circu-58
lation and high topography of the islands must be consideredin order to obtain a59
regionally consistent estimate of projected rainfall changes over Hawaii.60
The climate of Hawaii is characterized by a wet (during the Northern Hemisphere61
winter half-year) and a dry (summer) season. Four differentmechanisms for rain-62
fall production have been identified in previous meteorological studies (Schroeder63
1993; Chu et al. 1993). Lyons (1982) found that trade-wind-induced rainfall along64
the windward facing sides of the mountain chains is the most significant contrib-65
utor to the statewide rainfall budget. The strong covariance between precipitation66
on Hawaii and the trade wind circulation that was found in earlier studies suggests67
that the SD of large-scale circulation pattern onto Hawaiian rainfall stations is68
feasible, despite the complexity of the physical processesthat actually control the69
formation of precipitation (Woodcock 1975; Ramage and Schroeder 1999; Yang70
and Chen 2003; Cao et al. 2007). In addition, frontal rain associated with extrat-71
ropical storms over the North Pacific is observed during the wet season. Cold-air72
troughs dipping far to the south sometimes become detached from the main west-73
erly circulation and develop a cutoff low pressure system known as a Kona low.74
This leads to advection of moist warm air masses from the equatorial Pacific to75
the islands and the cool air aloft provide the thermodynamicinstability for the76
4
development of intense rainfall events. The island topographic can locally force77
extreme rainfall events of 10–12 inches in 24 hours. These extreme rainfall events78
contribute up to more than 50% of the annual rainfall total ondrier leeward sides79
of the islands (except for leeward side of the Big Island withthe two big volca-80
noes that reach to∼4100 m in elevation). Therefore, much of the interannual to81
decadal variability is determined by relatively few events(see also Chu and Chen82
2005), which will make the task for statistical climate change projections more83
difficult than in regions with higher frequency of rainfall events, such as along the84
windward facing mountain chains.85
The present study aims for a first quantitative assessment ofthe expected mean86
seasonal rainfall changes given the projected circulationchanges under the A1B87
emission scenario. Regionalized precipitation change information can serve many88
purposes. Among others, Hawaii’s ecosystems are vulnerable to the disturbances89
in the climatological pattern under global warming. Changes in the rainfall could90
impose a severe stress onto endemic species (Loope 1995; Loope and Giambel-91
luca 1998). On the stakeholders site, water resource management plans would92
need to be revised if Hawaii was expected to experience significant changes in the93
rainfall under global warming (Oki 2004). We emphasize thatthe present study is94
a first attempt to downscale the large-scale circulation changes of the 21st century95
5
from the AR4 report onto individual rainfall stations on theHawaiian Islands. In96
section two, the data and methods are described. Section three presents the results97
from diagnostic studies and the calibration and validationof the statistical transfer98
model. Further, the evaluation of the various GCMs from the IPCC AR4 report are99
presented and finally the results from the 21st century climate change downscal-100
ing are presented. In section four, a discussion of the associated uncertainties is101
given. Section five will summarize the results and closes with concluding remarks102
on future improvements for statistical downscaling.103
2. Data and Methods104
SD involves three essential steps. First, the establishment of a statistical rela-105
tionship between the large-scale predictors and the local predictands. Second, an106
assessment of the predictive skill of the statistical modeland third the application107
of the statistical model (Wilby and Wigley 1997). For the application to the 21st108
century climate change simulations, the SD method must further include an anal-109
ysis and selection of the GCM scenario runs, with the goal to select those models110
that compare best to the observed climate during the 20th century. The chosen111
models will provide the predictor information for the thirdstep of the SD.112
6
Daily and monthly rainfall data were obtained from the National Climatic Data113
Center in Asheville, NC. Originally, stations with less than 10% missing data114
were used to form monthly mean rainfall values. In this study, the majority of the115
records are available for the period 1950 to present. The climatological annual116
cycle in Hawaiian rainfall has been previously studied (Giambelluca et al. 1986).117
Most regions in Hawaii experience a pronounced seasonal cycle of precipitation118
(Fig. 1), with the bulk of the precipitation falling during the months of November119
through April (the wet season) and less frequent rainfall events during the months120
May through October (the dry season). Hence, we will developSD projection121
scenarios for the dry and wet season. 134 stations that coverthe period 1958–122
2000 are used for the downscaling purpose, in order to allow for an independent123
validation of the SD models.124
[Figure 1 about here.]125
Large-scale climate information was obtained for the period 1958–2000 from the126
ERA-40 reanalysis products (Uppala et al. 2005). The data were downloaded127
from the APDRC server (http://apdrc.soest.hawaii.edu/).The AR4 20th century128
climate model simulations and the 21st century A1B emissions scenario data were129
obtained from the FTP server (ftp-esg.ucllnl.org) that is maintained by the Earth130
7
System Grid II (ESG) research project sponsored by the U.S. DOE Office of Sci-131
ence. We worked with a total of 21 models shown in Table 1.132
[Table 1 about here.]133
2a. Diagnostic analysis: calibration and validation of the SD model134
In this study a composite analysis is used for the synoptic classification of the135
large-scale circulations that are associated with either very dry or very wet months136
at individual stations in Hawaii. For each station the monthly mean rainfall data137
1958–2000 were divided into two seasonal subsets (n=258), the wet (Nov.–Apr.)138
and the dry (May–Oct.) seasons, and sorted by increasing precipitation amounts.139
For each station the lower and upper 5% of monthly values wereidentified and140
the associated dates (i.e. month and years) provide the basis to form composite141
maps of the large-scale circulation.142
We concentrate our analysis on a key region around Hawaii (180◦E – 120◦W, 10◦S143
– 40◦N). This region is assumed to be wide enough to contain sufficient large-scale144
climate information relevant to Hawaii’s rainfall. A thorough objective criterion145
that would identify the optimal domain size for the downscaling purpose has not146
been attempted in this study. A smaller domain would give more weight to the147
8
regional aspects of the interacting large and local-scale processes. This results in148
limited degrees of freedom for the SD. A limited domain size,however, reduces149
the risk of incorporating model biases from remote regions during the downscaling150
process. Thus, a compromise has to be made between the inclusion of biases151
and the loss of valuable predictor information (Kang et al. 2007). We note that152
the extension of the region across the equator helps to include the atmospheric153
imprints of El Nino and La Nina events over the eastern tropical Pacific. The154
northern latitudes in our domain allow us to include information related to the155
Pacific Decadal Oscillation.156
With regards to the dynamics of the large-scale circulation, we concentrate the157
diagnostic analysis onto variables that can serve as predictors for the 21st cen-158
tury downscaling pursuit. Based on our understanding that the low-level wind159
regime has a profound effect on mean rainfall over the islands, and that much of160
the monthly, seasonal, and interannual climate variability in the upper atmosphere161
is linked to changes in the low-level circulation (Sanderson 1993; Held and Soden162
2006), we concentrate on near surface variables in this study. The close connec-163
tion between station rainfall, low-level winds and upper level circulation were164
initially tested with multiple linear regression, canonical correlation analysis and165
maximum covariance analysis [not shown]. All these resultsdemonstrated that the166
9
near surface (1000 hPa) meridional wind component can explain a significant frac-167
tion of the precipitation anomalies throughout the islands. It was also found that168
additional low-level variables (zonal wind, SLP) add little independent informa-169
tion in addition to the meridional wind field. Other variables such as precipitable170
water and 500 hPa geopotential are highly correlated with the meridional wind171
field. Therefore they bring limited independent information into the SD model.172
Other factors are known to have a strong control on the rainfall over Hawaii: the173
strength and height of the trade-wind inversion layer (Cao et al. 2007) and the174
vertical velocity in the 850 hPa to 700 hPa level. The former diagnostic is not175
immediately available from the IPCC AR4 database. Limited data were available176
for the vertical winds and we therefore decided to focus on the meridional wind177
field component in this study.178
The composite analysis identifies the typical circulation pattern that occur during179
the driest and wettest months at one station. The differencebetween the associated180
meridional wind fields of the lower and upper 5% rainfall months is used as a pro-181
jection pattern for the 1958–2000 (n=258) monthly mean v-wind fields from the182
ERA-40 reanlysis data. Note that in this step we take advantage of the larger sam-183
ple size and work with monthly mean data instead of seasonal mean data (n=43).184
The resulting index time series measures the similarity of the individual monthly185
10
mean v-wind field with the composite pattern. The station rainfall data are then186
converted into relative rainfall values ((p-pmean)/pmean*100 [%]) at each station.187
The relationship between the station rainfall and the projection indices from the188
v-wind is estimated with ordinary linear regression. For the calibration of the SD189
model, we decided to estimate the linear regression betweenseasonal means of190
the station rainfall data and the v-wind projection indices. The use of seasonal191
data was favored because of our basic interest in seasonal mean rainfall changes192
during the 21st century. The confidence intervals (von Storch and Zwiers 1999,193
p.154) of linear regression are calculated during the calibration step. These take194
the uncertainty of the regression coefficients and the uncertainty of the residual195
error into account. Since the regression line will be applied to 30-yr averages of196
the AR1 scenarios, the uncertainty of the residual varianceis scaled with a factor197
of 1/30.198
The validation of the regression models follows the idea of dividing the obser-199
vational data into a calibration and validation period (Michaelsen 1987). In this200
study, a moving window with a width of 21 years is applied starting with the cali-201
bration interval 1958–1978. The window is shifted by one year until it covers the202
period 1980–2000. The years outside the calibration windoware used to compare203
the station rainfall with the estimates of the linear regressions. Explained vari-204
11
ances (R2cal, R
2val) were calculated to test the robustness of the linear relationships.205
After validation of the linear regressions, the IPCC’s 20th-21st century simu-206
lations are projected onto the diagnosed difference pattern (from the composite207
analysis) and the linear regressions are used to estimate the corresponding rainfall208
anomalies that are expected from the wind changes at the end of the 21st century209
(i.e. difference mean 2070–2099 minus mean of 1970–2000).210
2b. Analysis of model skills211
In order to find an quantitative measure regarding which models should be in-212
cluded in our study of the 21st century rainfall projectionsfor Hawaii, we an-213
alyzed the fields of the 20th century climate simulations (20c3m) to select the214
“best” models. This analysis was carried out over the region(180◦ E – 120◦ W, 10215
◦ S – 40◦ N). The SLP field was used. The decision process is guided by a visual216
selection process. We refrain from a more formal development of an objective217
metric for our specific SD pursuit.218
Three different criteria were tested for each model for the wet and dry season:219
• How well is the SLP field represented in the models compared with the220
12
ERA-40 climatology (1970–2000)?221
• How well is the spatiotemporal variability reproduced in the models?222
• Does the variance of the meridional wind component match theobserved223
ERA-40 variance?224
The differences in the mean are measured in terms of the mean absolute error
(MAE)
MAE(k) =1
n
∑
ij
|xk(i, j) − xr(i, j)| , (1)
wherei,j are indices for the longitude and latitude of then grid points, andxk(i, j)225
is the SLP of model numberk, xr(i, j) is the ERA-40 SLP (see Fig. 2) . Note that226
one could also work with the root mean square error (RMS) (e.g. Taylor 2001) but227
the MAE gives less weight to outliers.228
[Figure 2 about here.]229
For the spatiotemporal variability comparison, the SLP fields 1970–2000 from the
ERA-40 and the models were subject to an EOF analysis. The leading m = 10
spatial eigenvectors and the explained variances were usedto define a EOF skill
13
score (ESS) according to:
ESSk =
∑m
i
∑m
j w(i, j) |rk(i, j)|∑m
i
∑m
j w(i, j). (2)
In Eq. 2 the correlationsrk(i, j) betweeni-th spatial EOF pattern of model k and
thej-th EOF of the ERA-40 reanalysis are summed over a limited range of EOF
combinations. Since EOF analysis often result in pairs of equally important modes
(North et al. 1982), the skill score gives weight to spatially correlated modes that
are offset by one rank using the weight function
w(i, j) =
0.5√
λk(i) λr(j) : i = j − 1
1.0√
λk(i) λr(j) : i = j
0.5√
λk(i) λr(j) : i = j + 1
0 : elsewhere
, (3)
with i,j representing the EOF modes1 . . .m . The weights in the EOF skill score230
ESSk also account for the explained varianceλk(i) andλr(j) of the eigenmodes231
of the model and the reanalysis, respectively. This heuristically derived measure-232
ment summarizes the information that is shown for illustrative purposes in next233
section (Fig. 10). A perfect correspondence of the spatial EOF modes in the234
14
model compared with the ERA-40 reanalysis would result in a skill score of 1235
(note that the minimum skill score is 0).236
Finally, it is crucially important for the statistical projection purpose that the vari-237
ability in the meridional wind field of the ERA-40 reanalysisand the models are238
locally of similar magnitude during the 20th century. A biasin the interannual-239
to-decadal variance may also distort the projected changesin the wind fields and,240
ultimately, the projected rainfall anomalies.241
One way to measure the model skill of reproducing the observed variability in
the wind is the average over the logarithmic ratios between the reanalysis and the
modeled variance at each grid point in the study area:
MLVk =1
n
∑
i,j
log
{
max
[
vk(i, j)
vr(i, j),vr(i, j)
vk(i, j)
]}
, (4)
where the indicesi, j correspond to then spatial grid points. The mean logarith-242
mic variance (MLV) is a non-negative value. A perfect match between modeled243
and reanalysis variability would result in a MLVk of zero. A value of 1 would244
indicate that the variance ratio is on average an order of onemagnitude different.245
The maximum function avoids cancellation effects if modelsshow regions of both246
underestimated and overestimated variance.247
15
3. Results248
3a. Observed relationship between large-scale circulation and rainfall249
This section addresses the question of how the large-scale circulation is related250
to the station-based precipitation on the Islands of Hawaii. The aim is to identify251
which locations on the islands are predominantly controlled by large-scale circu-252
lation anomalies. The dynamical structure of circulation modes are analyzed and253
serve as a ’proof of concept’ for the SD approach.254
After applying the composite analysis (see methods) to the 1000 hPa winds for255
the individual station rainfall data, the v-wind from the ERA-40 reanalysis is256
projected onto the associated anomaly composite pattern. The resulting index257
time series are regressed onto the associated relative rainfall anomalies. Here, the258
wet/dry seasons of the years 1958–1988 are used for the calibration of the linear259
regression parameters. The correlations between the individual rainfall time series260
and the v-wind indices show that a significant amount of the rainfall variability is261
controlled by the winds, especially during the wet season (Fig. 3b). The spatial262
distribution further indicates that the SD can be applied toboth the windward and263
leeward sides of the islands. However, the statistical relation between the wind264
field and the rainfall is weaker in the dry season (Fig. 3a). Weaker correlations265
16
during the dry summer month were expected, since few sporadic and localized266
heavy rainfall events have more influence on the seasonal mean precipitation.267
[Figure 3 about here.]268
In order to test the predictive skill of the calibrated linear regressions, the ob-269
servational period 1958–2000 was divided into non-overlapping calibration and a270
validation intervals (see section 2). Although individualstations can show very271
different validation results, the average over all stations indicate reasonableR2val272
values (Fig 4). It must be noted that the validation of the regression models serves273
as a means to lend credence to the application of the SD to future climate change.274
Aside from low amounts of explained variability, large discrepancies betweenR2cal275
andR2val are indicators of nonstationarity in the case of ordinary linear regressions.276
The climate shift in the 1970s that was identified in many climate records over the277
Pacific (Trenberth 1990; Miller et al. 1994; Gedalof and Smith 2001; Meehl et al.278
2008) could have served as an ideal test environment for the SD method (Wang279
and Zhang 2008). However, the validation is sensitive to changes in the data qual-280
ity of the reanalysis products and the station data. Advances in remote sensing281
significantly improved the reanalysis products beginning in the 1970s. System-282
atic increase in data gaps in the station network during the late 1970s and 1980s283
17
change the number of sample in validation and calibration windows. All these284
factors contribute to differences between the explained variances of the calibra-285
tion and validation. Overall, the validation shows that thecalibration statistics286
provide a reasonable estimate of the predictive skill. Onlystations with sufficient287
sample size (mimimum of 10 seasonal rainfall estimates) andwith a significant288
(p=5%) statistical relationship were used for the final downscaling of the scenar-289
ios onto the station rainfall (Section 3c). This resulted ina reduction to 97 (57)290
out of 134 rainfall stations suitable for the downscaling inthe wet (dry) season.291
In the following, we decided to use 1958–1988 as the calibration period for the292
linear regression models.293
[Figure 4 about here.]294
We also applied the vertical velocity fields in 700 hPa (ω700) as a predictor for the295
dry season rain. The station-averaged correlation betweendry season rainfall and296
ω700 is similar to the one obtained with the v-wind. In case of v-wind andω700 as297
predictors, the overall correlation was not increased (Fig. 5). The latter result is an298
indicator of multicollinearity in the predictor variables. Physically, the statistical299
results suggest a close connection between the horizontal winds (i.e. convergence)300
and the vertical motions above.301
18
[Figure 5 about here.]302
The composite analysis of the 134 stations showed two dominant modes in the303
v-wind field. To illustrate the dynamical structure of thesemodes, composites of304
zonal winds and SLP were calculated in addition to the v-wind. Figs. 6, 7 depict305
the composites for Hilo on the Big Island and Waiawa on Kauai,respectively.306
The wet season rainfall at Hilo has a correlation (R2cal = 0.29) with the v-wind307
index 1958–1988 (n=31). The composite maps of the ERA-40 SLPand 1000 hPa308
wind field show that a more pronounced subtropical high northeast of the islands309
produces stronger NE trade winds that favors the formation of rainfall.310
[Figure 6 about here.]311
Negative precipitation anomalies are observed during seasons when the high pres-312
sure cell and the trade winds are weaker than average. Despite the stronger low313
pressure systems in the North Pacific, the contribution fromfrontal rain systems314
cannot compensate for the reduction in trade-wind induced rainfall. The majority315
of the windward sites on the islands fall into this trade-wind related rainfall regime316
(Woodcock 1975; Lyons 1982; Sanderson 1993; Chu and Chen 2005).317
At Waiawa, a station on the leeward side of Kauai, the v-wind field explains 17%318
of the wet season rainfall variability (Fig. 7). The rainfall is associated with Kona319
19
lows W–NW of the islands. The resulting southerly winds advect warm moist air320
masses towards the islands. This synoptic weather pattern occurs less frequently321
than the trade-wind regime but it can cause intense rainfallevents over the islands.322
Fig. 8 gives two rather typical examples of the contributionby different climato-323
logical quantiles of the observed daily rainfall to the annual totals for Wainanae324
on the leeward side of Oahu and Haleakala, Maui. The plots show for each year325
of record the amount of annual rainfall associated with different quantiles from326
the daily rain probability density function (PDF) (Fig. 8b,d) and the percentage327
contributions to each annual total (Fig. 8 a,c). The upper 10% of all daily rainfall328
events account for 50% or more of the annual totals. Therefore, the lower corre-329
lation observed in Fig. 7d with the large-scale wind field is consistent with the330
physical mechanisms of rainfall generation.331
[Figure 7 about here.]332
[Figure 8 about here.]333
20
3b. GCM evaluation334
The model evaluation is based on the comparison between the ERA-40 reanal-335
ysis climate and individual AR4 20th century scenario runs (years 1970–2000).336
The objective quantification of the model skills is evaluated in the region around337
Hawaii (180◦ E – 120◦ W, 10 ◦ S – 40◦ N). SLP and the v-wind in 1000 hPa are338
analyzed. Based on the three test criteria (see section 2) and data availability this339
screening procedure reduces the number of GCMs from 21 to six, which are used340
in the final SD step.341
First, the differences in the climatological mean SLP are analyzed for the wet and342
dry season. The MAE (Eq. 1) ranges from 0.5 to 3.0 hPa, withoutany seasonal343
dependence (Fig. 9). Second, the spatial correlation between the dominant EOF344
modes in the SLP field gives a quantitative measure of the models’ ability to re-345
produce the interannual to decadal variability around the Hawaiian Islands and346
extratropical/tropical teleconnection regions (see Fig.9). To illustrate the mean-347
ing of the EOF skill score (ESS, Eq. 2), Fig. 10 represents twoexamples of the348
spatial correlation matrix derived from the leading 10 EOF modes. The better the349
spatial agreement in the EOF modes and the better the agreement in the ranks of350
these EOFs, the closer are the values with high correlation aligned along the di-351
21
agonal. The two given examples correspond to models (’q’,’u’ in Tab. 1) with a352
good correlation with the ERA-40 EOF modes.353
[Figure 9 about here.]354
The information contained in Fig. 10 is represented in the ESS (Eq. 2) and Fig. 9355
shows the results for the AR4 models. The models are generally of equal quality in356
terms of EOF modes during the summer season (gray symbols in Fig. 9). During357
the winter months, a group of models reveal markedly lower skills. Together358
with the differences in the mean SLP fields and the requirement of equally good359
performance in the wet and dry seasons, we used Fig. 9 as a guidance and selected360
six out of the 21 AR4 models for the rainfall projection purpose. It must be noted361
that for model ’o’ (see Fig. 9) the meridional wind field was not available at the362
time of our analysis. Furthermore, the third criterion, which measures the local363
variance ratios of the v-wind between models and reanalysis, showed that the364
selected models are in a reasonable variance range (Fig. 11).365
[Figure 10 about here.]366
[Figure 11 about here.]367
22
3c. Projected climate change and rainfall change368
In this section, the projected changes in the v-wind component and the regressed369
Hawaiian rainfall anomalies are presented. The question will be addressed to what370
extend the models project the same large-scale circulationchanges and what are371
the expected rainfall anomalies from these wind changes. The projected rainfall372
changes are based on the linear regressions between the v-wind field and the sea-373
sonal rainfall data that were developed in the previous section (Section 3a). Only374
stations that passed the statistical significance test are used for purposes of down-375
scaling the AR4 projections.376
In Fig. 12 the differences between the end 21st century (2070–2099) and the late377
20th century (1970–1999) averages for the wet season circulation are depicted for378
the selected models. The largest differences are projectedin simulation ’q’ and379
’t’ (Fig. 12 e,f). For model ’t’, the v-wind anomalies indicate a stronger northerly380
wind NE of the islands and anomalous southerly winds in the equatorial region.381
The slight positive wind anomalies above the islands and northerly anomalies west382
of the islands are the result of SLP anomaly that resembles a cyclonic circulation383
NW of the islands. The v-wind pattern resembles both a strengthened trade wind384
regime and an enhanced Kona Low activity. In contrast, the change in the v-385
23
wind simulated by model ’q’ shows almost the inverse anomalypattern. The386
most pronounced changes appear in the equatorial region. Weaker trade winds387
and an anomalous northerly component in the equatorial regions that extend over388
the Hawaiian Islands are in clear contrast to the results of model ’t’. The other389
model simulations show a smaller amplitude in the change of the v-wind field.390
In the ensemble average, the large amplitude changes canceleach other and the391
resulting wind anomalies are smaller, which makes a dynamical interpretation of392
the ensemble mean changes more ambiguous, despite their estimated significance393
(p=0.1). The results further highlight that the different AR4 models locally project394
different wind changes over Hawaii and their teleconnection to the tropics and395
extratropical centers of action are not robust among the models. This imposes a396
fundamental problem on the objective definition of the proper domain size for the397
SD.398
[Figure 12 about here.]399
Without discussing the details for the circulation changesduring the dry season, it400
is noteworthy that model ’q’ and model ’t’ project similar wind anomalies during401
the dry season compared with the wet season anomalies (Fig. 13). Largest wind402
anomalies occur over the equatorial regions in these individual models, but the403
24
ensemble mean shows only weak anomalies in the equatorial regions and south of404
the islands.405
[Figure 13 about here.]406
The projection of these simulated v-wind changes onto the station-related com-407
posite pattern provide the index values that are translatedinto projected rainfall408
anomalies at each station by application of the establishedregression lines. These409
downscaled scenarios of the late 21st century are plotted inFigs. 14 - 16 in form410
of relative precipitation change at each station. The estimated 2-sigma standard411
deviation (i.e. approx 95% confidence intervals ) are also provided. As expected412
from the weak circulation changes depicted for the 6-memberensemble mean in413
Figs. 12(g) and 13(g), the projected rainfall changes are low and not exceed-414
ing 20% at most stations. In the dry season (Fig. 14(a–c)), the ensemble mean415
projects slightly increased rainfall for northwest Maui, but the statistical uncer-416
tainties of downscaled rainfall anomalies are too large to draw affirmative con-417
clusions about the projected regional rainfall changes. For the same reason, no418
important precipitation changes are found in the 6-model ensemble mean during419
the wet season (d–f). An increase along the west-facing coast of the Big Island is420
projected, but the statistical uncertainty (which includes the unresolved variabil-421
25
ity) is rather large. Since the individual models in the 6-member ensemble reveal422
some striking variations in their projected wind anomalies, Fig. 15 and 16 present423
the downscaled precipitation changes for the two models with the largest anomaly424
amplitudes, model ’q’ and ’t’, respectively. In the dry season the local character of425
the precipitation is noticeable. On the northernmost island (Kauai), the two down-426
scaling scenarios show pronounced regional gradients fromwet to dry anomalies,427
however, with opposing signs between the two models. Over the Big Island, an428
east-west gradient is noticeable, but again the signs oppose each other in model429
’q’ and ’t’. During the wet season the effect of the large-scale circulation changes430
translates into a spatially more homogeneous pattern. Yet again, the two solutions431
derived from the wind anomalies provide two opposing scenarios for Hawaii: a432
dryer climate or a wetter climate. It should be noted that forthe majority of the433
stations, the percentage rainfall changes are below 30% andonly a few locations434
indicate robust changes with regards to the statistical confidence ranges.435
[Figure 14 about here.]436
[Figure 15 about here.]437
[Figure 16 about here.]438
26
[Figure 17 about here.]439
A more comprehensive description of the island-wide rainfall changes is given440
in form of a histogram. Taking the percentage changes at all stations and for441
all six model scenarios together the number of stations falling into 5%-bins are442
counted. The bimodality in Fig. 17(b) reflects the two opposing scenarios during443
the wet season. The maximum likelihood value suggests an island-wide 5%–444
10% rainfall decrease in the wet season. The histogram highlights the differences445
between the mean, median and maximum likelihood estimate. Because of the446
bimodality, we prefer the value of 5%– 10% rainfall decreaseover the ensemble447
mean value (which is close to 0%). The histogram of the dry season is closer to a448
unimodal distribtution. A 5% increase in rainfall over the islands is indicated as449
the maximum likelihood value.450
4. Discussion and Summary451
The SD presented here is an attempt to exploit the dynamical linkage between452
the large-scale circulation and individual station rainfall on the Hawaiian Islands.453
Prior to this study very limited information was available for this remote region in454
the Pacific. The current state-of-the-art GCM models that have been used for the455
27
IPCC AR4 scenario runs have a coarse grid resolution of about200× 200 km.456
In these models, the islands with their topographic features are not represented.457
The underlying work hypothesis for the SD procedure is that the missing physical458
processes (such as the topographic effect of the mountains on the trade winds) can459
be represented by statistical relationships.460
In this case study only linear methods have been applied. We restricted our predic-461
tor information to the meridional wind in the lower atmosphere. Especially for the462
dry season and the dry regions of the islands, this approach has limited success.463
Future improvements will require additional information in the form of multivari-464
ate predictor climate fields. The height and strength of the trade wind inversion,465
vertical motions on the 700 hPa level, transient eddy activity and low level mois-466
ture convergence are fundamental control factors (Lyons 1982; Chu et al. 1993;467
Schroeder 1993; Cao et al. 2007). Therefore multivariate regression methods with468
these additional large-scale circulation characteristics are promising to improve469
the overall statistical projection skill. However, our first tests with the inclusion of470
the vertical winds at 700 hPa did not improve the overall predictive skills, though471
some regressions of individual stations profit from this extra information. A care-472
ful case-by-case analysis will be required in future.473
The linear regression technique are less suited in regions with long dry spells and474
28
few rain events in the seasons. Here, a few extreme synoptic weather events can475
bring the bulk of rainfall during the year (Fig. 8). A strong linear connection with476
seasonally averaged large-scale circulation fields cannotbe expected to explain a477
large percentage of the local rainfall variability. Futuredownscaling methods must478
try to derive optimal non-linear transformation functionsbetween the predictors479
and the extreme rainfall (Wang and Zhang 2008). Since this study was the first480
attempt to downscale the projected 21st century climate change scenarios onto481
Hawaiian rainfall, it was important to understand the linear statistical relations.482
The 95% confidence ranges that were estimated for the projected rainfall changes483
serve as a guideline for the uncertainty associated with theestimate. However,484
the true uncertainty range cannot be estimated. One major problem with the con-485
fidence ranges is the implicitly assumed stationarity of thePDFs of rainfall and486
of the covariability between predictands and predictors. Under global warming487
it is likely that the statistical properties will change. The general assumption is488
that precipitation on a global scale will shift towards moreextreme events on489
the one side and more severe droughts on the other side (Held and Soden 2006).490
Whether rain over Hawaii will follow this general scenario is not known yet, but491
it is our contention that higher SSTs will lead to higher amounts of precipitable492
water and thus increases the likelihood of extreme events. Therefore, it remains493
29
a major challenge for SD methods to incorporate this type of uncertainty into the494
confidence range. Dynamical downscaling methods with regional climate models495
are ideal tools to take these aspects into account (Schmidliet al. 2007), but the496
extremely high computational costs of such models will require a well-justified497
selection of the atmospheric boundary conditions.498
For the 21st century climate change projection purpose, theinter-model compari-499
son revealed that the basic dynamic features of the circulation changes vary dras-500
tically among the models. The objective quality assessmentof the models could501
not narrow down the ensemble spread. The difficulties of estimating the rainfall502
changes in the area around Hawaii were already present in theIPPC report in Fig.503
11.25 (IPCC 2007). Part of the spread in the model ensemble isthe result of dif-504
ferent dynamical responses over the Pacific region (Barsugli et al. 2006; IPCC505
2007; Vecchi et al. 2008). In fact, this region is under the direct control of the506
atmospheric response to ENSO and the PDO (Chu 1995; Chu and Chen 2005).507
ENSO’s response to increased 21st century CO2 levels is not well defined in the508
ensemble and differences in the atmospheric teleconnection pattern result in am-509
biguous circulation response pattern. Therefore, it is notsurprising that the region510
around Hawaii exhibits low coherency in the climate change scenarios among the511
models. In this context, it must also be mentioned that the global scaling argu-512
30
ments of Held and Soden (2006) (see also Vecchi et al. 2006; Mitas and Clement513
2005) cannot immediately be projected onto the regional scales of Hawaii. As514
discussed in Held and Soden (2006), the trade winds are expected to weaken on515
a zonal average, but for Hawaii the changes in the stationaryeddies (i.e. the sub-516
tropical cell) are of crucial importance (Barsugli et al. 2006). The application of517
the theory and the statistical results presented here for the AR4 scenarios are not518
contradicting each other.519
The uncertainty inherent in the ensemble of the GCM models isdirectly passed520
through the statistical transfer model onto the estimated station rainfall anoma-521
lies. The transfer function cannot reduce this type of climate change uncertainty.522
Neither can regional models overcome this uncertainty. Future generations of523
GCMs are expected to provide more consistent circulation scenarios over the Pa-524
cific. Lastly, the emission scenarios themselves are highlydisputed. In addition525
to the statistical uncertainties and the model differences, the wide range of likely526
emission scenarios also add a significant amount of uncertainty to the Hawaiian527
rainfall change scenarios.528
Based on the IPCC AR4 A1B scenarios we find that the potential exists for the529
Hawaiian Islands to experience significant changes in the circulation pattern. Our530
SD projects moderate rainfall changes for Hawaii by the end of the 21st century531
31
as a consequence of the circulation changes. In one of the sixanalyzed model532
scenarios, an increase of 20–30% in the winter-time rainfall is projected. Taken533
all six models together, the most likely projection is a 5–10% decrease during the534
wet winter season by the end of the 21st century. For the dry season the maximum535
likelihood value has a rather broad distribution. A modest (5%) shift to increased536
rainfall is indicated. We notice that this study only investigated the mean seasonal537
rainfall changes and not the extreme events. The results from this SD method538
are collected on-line athttp://www.———.edu/——-/——and future results with539
optimized preditors and other climate change scenarios will be added to the web-540
pages.541
5. Acknowledgments542
The authors are grateful to the anonymous reviewers for their constructive crit-543
icism. H.F. Diaz was supported by NOAA and the U.S. Department of Energy.544
O. Timm has been supported by the Japan Agency for Marine-Earth Science and545
Technology (JAMSTAC) through its sponsorship of the International Pacific Re-546
search Center. We thank K. Hamilton for his support and encouragement in this547
work and L. Mehrhoff from the USGS Pacific Island Ecosystems Research Center,548
32
Honolulu, for his support of this project. We are grateful toThomas Giambelluca549
for the stimulating discussions. This manuscript is IPRC Contribution No. XXX550
and SOEST Contribution No XXXX.551
References552
Barsugli, J. J., S.-I. Shin, and P. D. Sardeshmukh, 2006: Sensitivity of global553
warming to the pattern of tropical ocean warming.Climate Dynamics, 27, 483–554
492, doi:10.1007/s00382-006-0143-7.555
Cao, G., T. W. Giambelluca, D. E. Stevens, and T. A. Schroeder, 2007: Inver-556
sion variability in the Hawaiian trade wind regime.J. Climate, 20, 1 145–1 160,557
doi:10.1175/JCLI4033.1.558
Christensen, J. et al., 2007:Climate Change 2007 — The Physical Science Ba-559
sis. Contribution of Working Group I to the Fourth Assessment Report of the560
IPCC, Cambridge University Press, Cambridge, United Kingdom, chapter 11:561
Regional Climate Projections. 849–926.562
Chu, P.-S., 1995: Hawaii rainfall anomalies and El Nino.J. Climate, 8, 1 697–563
1 703, doi:10.1175/1520-0442(1995)008<1697:HRAAEN>2.0.CO;2.564
33
Chu, P.-S. and H. Chen, 2005: Interannual and interdecadal rainfall variations in565
the Hawaiian islands.J. Climate, 18, 4 796–4 813, doi:10.1175/JCLI3578.1.566
Chu, P.-S., A. J. Nash, and F.-Y. Porter, 1993: Diagnostic studies of two con-567
trasting rainfall episodes in Hawaii: Dry 1981 and wet 1982.J. Climate, 6,568
1 457–1 462, doi:0.1175/1520-0442(1993)006<1457:DSOTCR>2.0.CO;2.569
Gedalof, Z. and D. Smith, 2001: Interdecadal climate variability and regime-scale570
shifts in Pacific North America.Geophys. Res. Lett., 28, 1 515–1 518.571
Giambelluca, T., M. Nullet, and T. Schroeder, 1986: Hawaii Rainfall Atlas. Tech-572
nical Report R76, Hawai’i Division of Water and Land Development, Depart-573
ment of Land and Natural Resources, Honolulu, Hawaii, vi + 267pp.574
Held, I. M. and B. J. Soden, 2006: Robust responses of the hydrological cycle to575
global warming.J. Climate, 19, 5 686–5 699, doi:10.1175/JCLI3990.1.576
IPCC, 2007:Climate Change 2007 — The Physical Science Basis. Contribution577
of Working Group I to the Fourth Assessment Report of the IPCC. Cambridge578
University Press, Cambridge, United Kingdom, 996 pp.579
Kang, K.-H., H.and An, C.-K. Park, A. L. S. Solis, and K. Stitthichivapak, 2007:580
34
Multimodel output statistical downscaling prediction of precipitation in the581
Philippines and Thailand.Geophys. Res. Lett., 34, doi:10.1029/2007GL030730.582
Loope, L. L., 1995: Climate change and island biological diversity. Islands: Bi-583
ological Diversity and Ecosystem Function on Islands, Springer, New York,584
238.585
Loope, L. L. and T. W. Giambelluca, 1998: Vulnerability of island tropical mon-586
tane cloud forests to climate change, with specific reference to East Mauim587
Hawaii.Climatic Change, 39, 503–517.588
Lyons, S. W., 1982: Empirical orthogonal function analysisof Hawai-589
ian rainfall. J. Appl. Meteor., 21, 1 713–1 729, doi:10.1175/1520-590
0450(1982)021<1713:EOFAOH>2.0.CO;2.591
Meehl, G., A. Hu, and B. Santer, 2008: The mid-1970s climate shift in the Pacific592
and the relative roles of forced versus inherent decadal variability. J. Climate,593
doi:10.1175/2008JCLI2552.1, in press.594
Michaelsen, J., 1987: Cross-validation in statistical climate forecast595
models. J. Climate. Appl. Meteor., 26, 1 589–1 600, doi:10.1175/1520-596
0450(1987)026<1589:CVISCF>2.0.CO;2.597
35
Miller, A., D. Cayan, T. Barnett, N. Graham, and J. Oberhuber, 1994: The 1976-598
1977 climate shift of the Pacific ocean.Oceanography, 7, 21–26.599
Mitas, C. M. and A. Clement, 2005: Has the Hadley cell been strengthening in600
recent decades?Geophys. Res. Lett., 32, doi:10.1029/2004GL021765.601
North, G., T. Bell, R. Cahalan, and F. Moeng, 1982: Sampling errors in estimation602
of empirical orthogonal functions.Mon. Wea. Rev., 110, 699–706.603
Oki, D. S., 2004: Trends in streamflow characteristics at long-term gaging sta-604
tions, Hawaii. Scientific Investigations Report 2004—5080, U.S. Department605
of the Interior, U.S. Geological Survey, U.S. Geological Survey, Information606
Services, Denver, CO 80225.607
Ramage, C. S. and T. A. Schroeder, 1999: Trade wind rainfall atop608
Mount Waialeale, Kauai.J. Climate, 127, 2 217–2 226, doi:10.1175/1520-609
0493(1999)127<2217:TWRAMW>2.0.CO;2.610
Sanderson, M., ed., 1993:Prevailing Trade Winds, Weather and Climate in611
Hawaii. University of Hawai’i, 126 pp.612
Schmidli, J., C. M. Goodess, C. Frei, C. R. Haylock, Y. Hundecha, J. Ribalaygua,613
and T. Schmith, 2007: Statistical and dynamical downscaling of precipitation:614
36
An evaluation and comparison of scenarios for the European Alps.J. Geophys.615
Res., 112, doi:10.1029/2005JD007026.616
Schroeder, T. A., 1993: Climate controls.Prevailing Trade Winds: Weather and617
Climate in Hawai’i, M. Sanderson, ed., University of Hawai’i Press, 126pp.618
Taylor, K.-E., 2001: Summarizing multiple aspects of modelperformance in a619
single diagram.J. Geophys. Res., 106, 7 183–7 192.620
Trenberth, K., 1990: Recent observed interdecadal climatechanges in the north-621
ern hemisphere.Bull. Amer. Meteor. Soc., 71, 988–993, doi:10.1175/1520-622
0477(1990)071<0988:ROICCI>2.0.CO;2.623
Uppala, S. M. et al., 2005: The ERA-40 re-analysis.Quart. J. Roy. Meteor. Soc.,624
131, 2 961–3 012, doi:10.1256/qj.04.176.625
Vecchi, G. A., A. Clement, and B. J. Soden, 2008: Examining the tropical Pacific’s626
response to global warming.EOS, 89, 81,83.627
Vecchi, G. A., B. J. Soden, A. T. Wittenberg, I. M. Held, A. Leetmaa, and M. J.628
Harrison, 2006: Weakening of tropical Pacific atmospheric circulation due to629
anthropogenic forcing.Nature, 441, 73–76, doi:10.1038/nature04744.630
37
von Storch, H., E. Zorita, and C. U., 1993: Downscaling of global cli-631
mate change estimates to regional scales: An application toiberian632
rainfall in wintertime. J. Climate, 6, 1161–1171, doi:10.1175/1520-633
0442(1993)006<1161:DOGCCE>2.0.CO;2.634
von Storch, H. and F. Zwiers, 1999:Statistical Analysis in Climate Research.635
Cambridge Univeristy Press, Cambridge, UK, 484 pp.636
Wang, J. and X. Zhang, 2008: Downscaling and projection of winter ex-637
treme daily precipitation over North America.J. Climate, 21, 923–957,638
doi:10.1175/2007JCLI1671.1.639
Wilby, R. and T. Wigley, 1997: Downscaling general circulation model output:640
A review of methods and limitations.Progress Phys. Geogr., 21, 530–548,641
doi:10.1177/030913339702100403.642
Woodcock, A. H., 1975: Anomalous orographic rains of Hawaii. Mon. Wea. Rev.,643
103, 334–343, doi:10.1175/1520-0493(1975)103<0334:AOROH>2.0.CO;2.644
Yang, Y. and Y.-L. Chen, 2003: Circulation and rainfall on the lee side645
of the island of Hawaii during HaRP.Mon. Wea. Rev., 131, 2 525–2 542,646
doi:10.1175/1520-0493(2003)131<2525:CAROTL>2.0.CO;2.647
38
List of Figures648
1 Average monthly precipitation for the main Hawaiian Islands (in649
inches). Averages calculated for the period 1921–1980 based on650
488 precipitation reporting stations. The state annual average (73651
inches or 1854 mm) is a weighted average of the island means. .. 46652
2 Climatological SLP field and wind field in 1000 hPa for (a) thedry653
season and (b) and wet season averaged over 1970–2000. Data654
from the ERA-40 reanalysis project (Uppala et al. 2005). . . .. . 47655
3 Explained variance [%] for each station-based regressionmodel:656
(a) the dry season, (b) the wet season. Note that stations without657
statistically significant (p=5%) linear correlation are marked with658
black crosses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48659
39
4 Explained variances of the calibration (R2cal, black) and valida-660
tion (sgn(Rval)R2val, green) intervals using a 21-yr moving win-661
dow for calibration and the remaining years for the validation of662
the linear regressions: (a) dry season 1958–2000; (b) wet season663
1958–2000. The box and whiskers plot show the median, standard664
deviation, minimum and maximum values of the squared corre-665
lations between the 134 station rainfall and their corresponding666
v-wind indices. Black (green) crosses denote the station-mean667
R2cal (sgn(Rval)R
2val) values for each data window. Note that all668
stations with sufficient data (10 seasonal rainfall values)were in-669
cluded irrespective of the statistical significance of their Rcal val-670
ues. Center year refers to the center of the 21-yr calibration window. 49671
5 Explained variances of the calibration (R2cal, black) and validation672
(sgn(Rval)R2val, green) as in Fig.4(a) but with the projectedω700673
indices as an additional predictor to the v-wind indices. . .. . . . 50674
40
6 Composite analysis of the SLP and 1000 hPa wind for the high675
and low precipitation winter months at station Hilo on the Big676
Island during 1958–2000 (n=258): (a) the mean circulation av-677
eraged over the months associated with thep95 quantiles of the678
monthly mean rainfall data; (b) the average circulation forthe679
month with lowest rainfall (p5 quantile); (c) difference of the mean680
circulation high-low composite. In (d) the seasonally averaged681
meridional wind field index (see text for details) associated with682
the difference field in (c) is regressed onto the rainfall data from683
station Hilo, 1958–1988 (n=31, r2=0.29). . . . . . . . . . . . . . 51684
7 Same as Fig. 6 but for the station Waiawa on the leeward side685
of Kauai. Note that the linear regression results in a significant686
correlation (n=31,r2=0.17) . . . . . . . . . . . . . . . . . . . . . 52687
8 Time series of annual rainfall at (a,b) Haleakala, Maui, and (c,d)688
Waianae, leeward coast of Oahu for respective periods of record.689
The panels (b) (d) give the annual rainfall totals as contributed690
by the daily events, in terms of their magnitude (quantile value).691
Panels (a), (c) give the percentage contribution to each annual total692
by these quantiles. . . . . . . . . . . . . . . . . . . . . . . . . . . 53693
41
9 Skill score statistics of the AR4 models: bias in the 1970–2000694
mean SLP field in the region 10◦S–40◦N/180◦E-120◦W is mea-695
sured as the mean absolute error (MAE in hPa); spatial correla-696
tion of the leading EOF modes in the 1970–2000 SLP field over697
the same region (Eq. 2). Reference climatology and EOF modes698
were analyzed from the ERA-40 data 1970–2000. Black (gray)699
letters mark the wet (dry) season. See Tab. 1 for model identi-700
fication. The AR4 models selected for the 21st century rainfall701
projection are marked with solid circles. Note that our selection702
criteria would also favor model ’o’ (dashed circle), but data were703
were not available for the 21st century. . . . . . . . . . . . . . . . 54704
10 Matrix of the spatial correlation among the leading 10 EOFeigen-705
mode pattern of the SLP field from the ERA-40 data and two AR4706
models. Color shading indicate the correlation. The highercon-707
centration along the diagonal axis, the better the agreement in the708
spatial eigenmodes between model and ERA-40. Left, model ’q’;709
right model ’u’ (see Tab. 1). . . . . . . . . . . . . . . . . . . . . 55710
42
11 Model skills in reproducing the observed ERA-40 interannual-711
decadal variability in the meridional wind field (1000 hPa) in the712
region 10◦S–40◦N/180◦E-120◦W during 1970–2000: (a) the wet713
season, (b) the dry season. Note that a perfect grid-by-gridmatch714
between observed and modeled variance would give a value of 0.715
See text for details. Filled circles mark the selected models for the716
downscaling analysis. . . . . . . . . . . . . . . . . . . . . . . . . 56717
12 Differences between the end 21st century (2070–2099 averaged)718
and the late 20th century (1970-1999 averaged) wet season v-wind719
for the selected models: (a) model ’a’, (b) model ’d’, (c) model720
’e’, (d) model ’p’, (e) model ’q’, (f) model ’t’, and (g) the 6-721
model ensemble mean. Black contours depict the v-wind changes722
(contour interval 0.2 ms−1). Gray shadings indicate significant723
differences at the two-sided 10% significance level. . . . . . .. . 57724
13 Same as in Fig. 12 but for the dry season. . . . . . . . . . . . . . 58725
43
14 Projected rainfall changes (anomalies with respect to the 1970–726
1999 climatological mean in %) during the dry (a–c) and wet (d–f)727
season using the 6-model ensemble mean of the projected v-wind728
anomalies. Upper row shows the maximum likelihood estimate.729
Middle row shows the estimated lower margin of the 95% statis-730
tical confidence interval, and the bottom row the upper margin of731
the 95% confidence interval. See text for discussion of the sta-732
tistical confidence interval. Note the varying color scalesof the733
plots. Stations without significant statistical relationship between734
rainfall and v-windfield (crosses) were not used in these estimates. 59735
15 Same as in Fig. 14 but for the projected rainfall anomaliesusing736
the v-wind anomalies from model ’q’. Note the varying color737
scales of the plots. . . . . . . . . . . . . . . . . . . . . . . . . . . 60738
16 Same as in Fig. 14 but for the projected rainfall anomaliesus-739
ing the v-wind anomalies from model ’t’. Note the varying color740
scales of the plots. . . . . . . . . . . . . . . . . . . . . . . . . . . 61741
44
17 Number of stations falling into each 5% bin of downscaled rainfall742
changes. Counts were summed over all six models. Dry season743
(a) and wet season (b). . . . . . . . . . . . . . . . . . . . . . . . 62744
45
Figure 1: Average monthly precipitation for the main Hawaiian Islands (ininches). Averages calculated for the period 1921–1980 based on 488 precipita-tion reporting stations. The state annual average (73 inches or 1854 mm) is aweighted average of the island means.
46
(a) dry season (b) wet season
Figure 2: Climatological SLP field and wind field in 1000 hPa for (a) the dryseason and (b) and wet season averaged over 1970–2000. Data from the ERA-40reanalysis project (Uppala et al. 2005).
47
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
10152025303540455055606570
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
10152025303540455055606570
%
-
(a) (b)
%
Figure 3: Explained variance [%] for each station-based regression model: (a) thedry season, (b) the wet season. Note that stations without statistically significant(p=5%) linear correlation are marked with black crosses.
48
(a) (b)
Figure 4: Explained variances of the calibration (R2cal, black) and validation
(sgn(Rval)R2val, green) intervals using a 21-yr moving window for calibration and
the remaining years for the validation of the linear regressions: (a) dry season1958–2000; (b) wet season 1958–2000. The box and whiskers plot show themedian, standard deviation, minimum and maximum values of the squared corre-lations between the 134 station rainfall and their corresponding v-wind indices.Black (green) crosses denote the station-meanR2
cal (sgn(Rval)R2val) values for
each data window. Note that all stations with sufficient data(10 seasonal rain-fall values) were included irrespective of the statisticalsignificance of theirRcal
values. Center year refers to the center of the 21-yr calibration window.
49
Figure 5: Explained variances of the calibration (R2cal, black) and validation
(sgn(Rval)R2val, green) as in Fig.4(a) but with the projectedω700 indices as an
additional predictor to the v-wind indices.
50
−20 −15 −10 −5 0 5 10 15
−60
−20
020
4060
projection index
stat
ion
data
(a) (b)
(c) (d)
204˚ 00' 204˚ 30' 205˚ 00'
19˚ 00'
19˚ 30'
20˚ 00'
204˚ 00' 204˚ 30' 205˚ 00'
19˚ 00'
19˚ 30'
20˚ 00'
Hilo
Figure 6: Composite analysis of the SLP and 1000 hPa wind for the high and lowprecipitation winter months at station Hilo on the Big Island during 1958–2000(n=258): (a) the mean circulation averaged over the months associated with thep95 quantiles of the monthly mean rainfall data; (b) the averagecirculation forthe month with lowest rainfall (p5 quantile); (c) difference of the mean circulationhigh-low composite. In (d) the seasonally averaged meridional wind field index(see text for details) associated with the difference field in (c) is regressed onto therainfall data from station Hilo, 1958–1988 (n=31, r2=0.29).
51
−15 −5 0 5 10 15 20
−50
050
100
projection index
stat
ion
data 200˚ 00' 200˚ 30' 201˚ 00'
22˚ 00'
200˚ 00' 200˚ 30' 201˚ 00'
22˚ 00' Waiawa
(a)
(c) (d)
(b)
Figure 7: Same as Fig. 6 but for the station Waiawa on the leeward side of Kauai.Note that the linear regression results in a significant correlation (n=31,r2=0.17)
52
Figure 8: Time series of annual rainfall at (a,b) Haleakala,Maui, and (c,d) Wa-ianae, leeward coast of Oahu for respective periods of record. The panels (b) (d)give the annual rainfall totals as contributed by the daily events, in terms of theirmagnitude (quantile value). Panels (a), (c) give the percentage contribution toeach annual total by these quantiles.
53
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0.55
0.60
0.65
0.70
0.75
MAE [hPa]
ES
S [u
nitle
ss]
aa
bb
c
c
d
d e
e f
f
h
h
i
i
j
j
k
k
l
l
m
m
n
n
o
o
pp
q q
r
r
ss
tt
u
u
good
poor
poorgood
Figure 9: Skill score statistics of the AR4 models: bias in the 1970–2000 meanSLP field in the region 10◦S–40◦N/180◦E-120◦W is measured as the mean ab-solute error (MAE in hPa); spatial correlation of the leading EOF modes in the1970–2000 SLP field over the same region (Eq. 2). Reference climatology andEOF modes were analyzed from the ERA-40 data 1970–2000. Black (gray) let-ters mark the wet (dry) season. See Tab. 1 for model identification. The AR4models selected for the 21st century rainfall projection are marked with solid cir-cles. Note that our selection criteria would also favor model ’o’ (dashed circle),but data were were not available for the 21st century.
54
2 4 6 8 10
24
68
10
20c3m ndjfma #17
EOF # (sim)
EO
F #
(obs)
2 4 6 8 10
24
68
10
20c3m ndjfma #21
EOF # (sim)
EO
F #
(obs)
12345678910
12345678910
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
spatial correlation
0 0.5 1
simulated EOF #simulated EOF #
ER
A−
40 E
OF
#
ER
A−
40 E
OF
#
Figure 10: Matrix of the spatial correlation among the leading 10 EOF eigenmodepattern of the SLP field from the ERA-40 data and two AR4 models. Color shad-ing indicate the correlation. The higher concentration along the diagonal axis, thebetter the agreement in the spatial eigenmodes between model and ERA-40. Left,model ’q’; right model ’u’ (see Tab. 1).
55
5 10 15 20
0.0
1.0
2.0
3.0
model #
varia
nce
skill
sco
re [u
nitle
ss]
5 10 15 20
0.0
1.0
2.0
3.0
model #
varia
nce
skill
sco
re [u
nitle
ss]
pa tjepa tje
(a) (b)
Figure 11: Model skills in reproducing the observed ERA-40 interannual-decadal variability in the meridional wind field (1000 hPa) in the region 10◦S–40◦N/180◦E-120◦W during 1970–2000: (a) the wet season, (b) the dry season.Note that a perfect grid-by-grid match between observed andmodeled variancewould give a value of 0. See text for details. Filled circles mark the selectedmodels for the downscaling analysis.
56
(a) (b) (c)
(d) (e) (f)
(g)
Figure 12: Differences between the end 21st century (2070–2099 averaged) andthe late 20th century (1970-1999 averaged) wet season v-wind for the selectedmodels: (a) model ’a’, (b) model ’d’, (c) model ’e’, (d) model’p’, (e) model’q’, (f) model ’t’, and (g) the 6-model ensemble mean. Black contours depict thev-wind changes (contour interval 0.2 ms−1). Gray shadings indicate significantdifferences at the two-sided 10% significance level.
57
(a) (b) (c)
(d) (e) (f)
(g)
Figure 13: Same as in Fig. 12 but for the dry season.
58
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10
-505
1015202530
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10
-505
1015202530
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10
-505
1015202530
%
(c) dry season (2070−2099) upper 95% confidence
(b) dry season (2070−2099) lower 95% confidence
(a) dry season (2070−2099) estimated change
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10-505
1015202530
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10-505
1015202530
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-20
-15
-10
-5
0
5
10
15
20%
(d) wet season (2070−2099) estimated change
(e) wet season (2070−2099) lower 95% confidence
(f) wet season (2070−2099) upper 95% confidence
Figure 14: Projected rainfall changes (anomalies with respect to the 1970–1999climatological mean in %) during the dry (a–c) and wet (d–f) season using the6-model ensemble mean of the projected v-wind anomalies. Upper row shows themaximum likelihood estimate. Middle row shows the estimated lower margin ofthe 95% statistical confidence interval, and the bottom row the upper margin ofthe 95% confidence interval. See text for discussion of the statistical confidenceinterval. Note the varying color scales of the plots. Stations without significantstatistical relationship between rainfall and v-windfield(crosses) were not used inthese estimates.
59
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-50
-40
-30
-20
-10
0
10
20
30
40
50%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-50
-40
-30
-20
-10
0
10
20
30
40
50%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10
-505
1015202530
%(a) dry season (2070−2099) estimated change
(b) dry season (2070−2099) lower 95% confidence
(c) dry season (2070−2099) upper 95% confidence
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10
-505
1015202530
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-50
-40
-30
-20
-10
0
10
20
30
40
50%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10
-505
1015202530
%(d) wet season (2070−2099) estimated change
(e) wet season (2070−2099) lower 95% confidence
(f) wet season (2070−2099) upper 95% confidence
Figure 15: Same as in Fig. 14 but for the projected rainfall anomalies using thev-wind anomalies from model ’q’. Note the varying color scales of the plots.
60
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-80-70-60-50-40-30-20-10
01020304050607080
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-80-70-60-50-40-30-20-10
01020304050607080
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-50
-40
-30
-20
-10
0
10
20
30
40
50%
(a) dry season (2070−2099) estimated change
(b) dry season (2070−2099) lower 95% confidence
(c) dry season (2070−2099) upper 95% confidence
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-80-70-60-50-40-30-20-10
01020304050607080
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-30-25-20-15-10-505
1015202530
%
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
200˚ 201˚ 202˚ 203˚ 204˚ 205˚
19˚
20˚
21˚
22˚
-60-50-40-30-20-10
0102030405060
%(d) wet season (2070−2099) estimated change
(e) wet season (2070−2099) lower 95% confidence
(f) wet season (2070−2099) upper 95% confidence
Figure 16: Same as in Fig. 14 but for the projected rainfall anomalies using thev-wind anomalies from model ’t’. Note the varying color scales of the plots.
61
−50 0 50
050
100
150
rainfall change [%]
coun
ts
−50 0 50
010
2030
4050
6070
rainfall change [%]
coun
ts
(b)(a)
Figure 17: Number of stations falling into each 5% bin of downscaled rainfallchanges. Counts were summed over all six models. Dry season (a) and wet season(b).
62
List of Tables745
1 List of IPCC AR4 models that are analyzed in this study. Models746
that are selected for the SD of the 21st century A1B scenariosare747
denoted with+. Note that the wind field data used in the down-748
scaling was not available from the marked with∗. Equilibrium749
climate sensitivity values are from Tab.8.2 in IPCC (2007).. . . . 64750
63
No ref. letter Model Equil. sensitivity [K]1 a CCCMA CGCM3 1+ 3.42 b CCCMA CGCM3 1 T63 3.43 c CSIRO MK3 0 3.14 d GFDL CM2 0+ 2.95 e GFDL CM2 1+ 3.46 f GISSAOM n.a.7 g GISSMODEL E H 2.78 h GISS MODEL E R 2.79 i IAP FGOALS1 0 G 2.310 j INGV ECHAM4∗ n.a.11 k INMCM3 0 2.112 l IPSL CM4 4.413 m MIROC3 2 HIRES 4.314 n MIROC3 2 MEDRES 4.015 o MIUB ECHO G∗ 3.216 p MPI ECHAM5+ 3.417 q MRI CGCM2 3 2A+ 3.218 r NCAR CCSM3 0 2.719 s NCAR PCM1 2.120 t UKMO HADCM3+ 3.321 u UKMO HADGEM1∗ 4.4
Table 1: List of IPCC AR4 models that are analyzed in this study. Models that areselected for the SD of the 21st century A1B scenarios are denoted with+. Notethat the wind field data used in the downscaling was not available from the markedwith ∗. Equilibrium climate sensitivity values are from Tab.8.2 in IPCC (2007).
64