444 DOI: 10.5675/ICWRER_2013 ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors for the Upper Blue Nile River Basin, Ethiopia Netsanet Cherie 1 · Manfred Koch 1 1 Department of Geohydraulics and Engineering Hydrology, Kassel University Abstract High population increase and poor land- and water-management have led to a decline in recent years of the already low agricultural productivity in the Upper Blue Nile river basin (UBNRB), Ethiopia, where more than 85% of the population depends entirely on rain-fed agriculture. This adverse situation is further exacerbated by imminent climate change in the region as a whole. To analyze the possible impacts of future climate change on, particularly, the water resources in the UBNRB, climate predictions for the basin using downscaled predictor from various GCM’s have been carried out. The two statistical downscaling models used are SDSM and LARS-WG, whereby SDSM is used in a mono- model manner, i.e. employing predictors from only one GCM (ECHAM 5) and LARS-WG is used both in mono-model- (ECHAM 5) and multi-model- (ECHAM5, GFLD21 and CSIRO- MK3) mode. These three models are selected in a pre-screening analysis from about 17 GCMs in the MAGICC/SCENGEN database environment and found to be the most suitable for describing the observed historical (baseline) data (1970-2000 time period) in the UBNRB. The future climate predictors for the GCM’s selected cover the two time periods 2046-2065 (2050s) and 2081-2100 (2090s), respectively, and are available for the two SRES-scenarios A1B and A2. The future climate predictions for the UBNRB using the SDSM and LARS-WG tools indicate an increase of the seasonal temperatures for both downscaling tools and both SRES scenarios. Thus, for the 2050s time period the seasonal maximum temperatures T max rise between 0.6°C to 2.7°C and the minimum ones T min by 2.44°C. Similarly, during the 2090s the seasonal T max increase by 0.9°C to 4.63°C and T min by 1°C to 4.5°C, whereby these increases are generally higher for the A2 than for the A1B scenario. For most sub-basins, the predicted changes of T min are larger than those for T max . Moreover, for both SRES- scenarios and both simulated future time periods, the T max and T min for spring and summer seasons are found to be warmer than for autumn and winter . In general, the two LARS- WG downscaling approaches forecast a warmer future climate than SDSM. As for the precipitation, the predictions of the three downscaling methodologies are more variant across the sub-basins and the seasons of the year, however, all of them predict overall decreasing trends for most of the seasons of the year, but the autumn, particularly, for the later 2090s period. The SDSM- predicted negative precipitation changes are generally
44
Embed
Mono- and multi-Model statistical Downscaling of GCM ... · ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 446 DOI: 10.5675/ICWRER_2013
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
444
DOI: 10.5675/ICWRER_2013
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors …
Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors for the Upper Blue Nile River Basin, Ethiopia
Netsanet Cherie1 · Manfred Koch1 1 Department of Geohydraulics and Engineering Hydrology, Kassel University
AbstractHigh population increase and poor land- and water-management have led to a decline in
recent years of the already low agricultural productivity in the Upper Blue Nile river basin
(UBNRB), Ethiopia, where more than 85% of the population depends entirely on rain-fed
agriculture. This adverse situation is further exacerbated by imminent climate change
in the region as a whole. To analyze the possible impacts of future climate change on,
particularly, the water resources in the UBNRB, climate predictions for the basin using
downscaled predictor from various GCM’s have been carried out. The two statistical
downscaling models used are SDSM and LARS-WG, whereby SDSM is used in a mono-
model manner, i.e. employing predictors from only one GCM (ECHAM 5) and LARS-WG is
used both in mono-model- (ECHAM 5) and multi-model- (ECHAM5, GFLD21 and CSIRO-
MK3) mode. These three models are selected in a pre-screening analysis from about
17 GCMs in the MAGICC/SCENGEN database environment and found to be the most
suitable for describing the observed historical (baseline) data (1970-2000 time period)
in the UBNRB. The future climate predictors for the GCM’s selected cover the two time
periods 2046-2065 (2050s) and 2081-2100 (2090s), respectively, and are available for the
two SRES-scenarios A1B and A2.
The future climate predictions for the UBNRB using the SDSM and LARS-WG tools indicate
an increase of the seasonal temperatures for both downscaling tools and both SRES
scenarios. Thus, for the 2050s time period the seasonal maximum temperatures Tmax rise
between 0.6°C to 2.7°C and the minimum ones Tmin by 2.44°C. Similarly, during the 2090s
the seasonal Tmax increase by 0.9°C to 4.63°C and Tmin by 1°C to 4.5°C, whereby these
increases are generally higher for the A2 than for the A1B scenario. For most sub-basins,
the predicted changes of Tmin are larger than those for Tmax. Moreover, for both SRES-
scenarios and both simulated future time periods, the Tmax and Tmin for spring and summer
seasons are found to be warmer than for autumn and winter . In general, the two LARS-
WG downscaling approaches forecast a warmer future climate than SDSM. As for the
precipitation, the predictions of the three downscaling methodologies are more variant
across the sub-basins and the seasons of the year, however, all of them predict overall
decreasing trends for most of the seasons of the year, but the autumn, particularly, for
the later 2090s period. The SDSM- predicted negative precipitation changes are generally
GFDL-CM2.1 USA 2.0 × 2.5 A1B, A2 1.43 / 3.19 0.86 / -2.69 * Numbers left and right of the slash denote precipitation and mean temperature, respectively.
3.3. Preparation and clean-up of observed climate input data
The observed meteorological data used in this study includes spatially and temporally
limited records of rainfall, maximum and minimum temperatures. Although there are
many meteorological gauging stations in the study area, most of the gauges have either
short record periods, or have plenty of missing and erroneous data. Among the stations
in the UBNRB, 53 rainfall- and 33 temperature stations are considered here (see Figure 1).
The data from these stations are collected by the Ethiopian National Meteorological
Service Agency (ENMSA) in the time period 1970-2000. In spite of this 31-years long
record length, most of the stations suffer from a variety of errors which include different
coordinate units in a gauge, coordinate shifts, typing errors, significant missing records
and recording errors. All errors, except those due to missing records, are corrected by
a thorough investigation using different sources such as the literature, UBNRB master
plans and by referring to the original data.
The complex treatment of missing data, i.e. the filling a record of a particular station in this
study area is done, using LARS-WG (Long Ashton Research Station Weather Generator),
developed by Semenov et al. [1998]. Hashmi et al. [2011] show the capability of LARS-WG
to synthesize and to fill in missing values in daily climate time series ( precipitation and
temperature).
LARS-WG utilizes semi-empirical distributions (SED), i.e. histograms, to represent series
of wet and dry days, daily precipitation, minimum and maximum temperature and solar
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 458
DOI: 10.5675/ICWRER_2013
is referred to Cherie [2013]. One may notice from the figure that basically the same
dominant NCEP- predictor variables of the list in Table 3 are obtained for the different
sub-basins, and the former are often correlated with precipitation and temperature in an
opposite way which, from a meteorological point of view, appears to make some sense.
Figure 6 Correlation coefficients between NCEP- predictors and average sub-basin precipitation (top) and maximum temperature (bottom) for a few UBNRB sub-basins
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 461
DOI: 10.5675/ICWRER_2013
Figure 7 Annual 1970-2000 observed time series for precipitation (top) maximum (middle) and minimum temperature (bottom) for the various sub-basins of the UBRNB
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 462
DOI: 10.5675/ICWRER_2013
Table 4: Seasonal Mann-Kendall test for a trend (hypothesis H1) in precipitation, maximum and minimum temperature for the various sub-basins of the UBNRB.
SN Sub-basinSeasonal Mann-Kendall trend: H1
PCP Tmax Tmin1 Anger No No Sign(+)2 Belles No No No3 Beshilo No Sign(+) No4 Dabus No No Sign(-)5 Didessa No No Sign(+)6 Fincha No No No7 Guder Sign(+) No No8 Jimma No No Sign(+)9 Muger Sign(+) No No10 N_Gojjam No Sign(+) Sign(+)11 S_Gojjam Sign(-) No No12 Tana No No No13 Welaka No No No14 Wenbera Sign(-) No Sign(-)# UBNRB Sign(+) Sign(+) No
"No" for hypothesis H1 implies that there is no statistically significant trend, "Sign" represents the presence of a significant trend , with (+) and (-) indicated increasing and decreasing trend, respectively.
Figure 8 1970-2000 decadely averaged values for precipitation/year, maximum and minimum temperatures for the various sub-basins of the UBNRB
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 464
DOI: 10.5675/ICWRER_2013
4.2.2. 21th - century future climate predictions
Although the performance of the SDSM- mono model downscaling to simulate the past
20th -century climate in the UBNRB appears not to be so good, for a latter comparison with
the other downscaling methods, we present a few results of this approach method with
regard to the prediction of 21th - century future climate variations in the basin. Figures 11
and 12 shows these future predictions of the monthly averages of the maximum (Tmax)
and minimum (Tmin) temperatures and the rainfall, respectively, for the sub-basin Tana.
More specifically, likewise to all following figures, results for the two future time-slices
2050s and 2090s and the two SRES scenarios A1B and A2 (see Table 2). In addition to the
absolute values of the corresponding climate variable, relative changes with respect to
the observations of the 20th - century reference period - after additive and multiplicative
Figure 9 SDSM- 20th - century simulation results for precipitation, maximum and minimum temperatures for two UBNRB sub-basins. Shown are monthly averaged observed, SDSM- NCEP- reanalysis- and SDSM-ECHAM-5- predictions of the three climate variables
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 465
DOI: 10.5675/ICWRER_2013
Figure 10 Average monthly bias between 1970-2000 observed climate data (RF, Tmax and Tmin) and SDSM- NCEP- reanalysis- and SDSM-ECHAM-5 predictions, respectively, for two sub-basins of the UBNRB (see also Figure 9)
"delta"- corrections for the past biases of temperatures and precipitation, respectively,
as mentioned earlier, e.g. Cherie [2013] - and the trend-lines over the whole 1970-2000
simulation period (broken-up in three intervals) are shown.
The future SDSM- predicted temperatures, including the changes, shown in Figure 11
for the sub-basin Tana, are pretty much representative for the whole UBNRN, namely,
systematic temperatures increases for both Tmax and Tmin which, depending on the sub-
basin and the season, range between 1°C and 4°C. Generally, the increase of Tmax is higher
for the wet season months (June to August) than for the dry season months (December
to February). Also, expectedly, Tmax is higher for the 2090- than for the 2050 decade,
whereby, particularly, for the 2090 time period, the impact of the (more extreme) SRES-
scenario A2 is slightly stronger than that of A1B.
For the minimum temperatures Tmin the trends are similar, if not even more consistent
across the various sub-basins of the UBNRB. Compared with the 20th –century reference
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 466
DOI: 10.5675/ICWRER_2013
Figure 11 SDSM- mono model predictions of future maximum (upper panels) and minimum (bottom panels) temperatures in sub-basin Tana for the two time-slices 2050s and 2090s and the two SRES-scenarios A1B and A2
period, warmer temperatures are observed for the two future time periods considered,
whereby, expectedly, those for the 2090- decade are higher than those of the 2050-decade.
Moreover, for most sub-basins the warming trend is higher for SRES- scenario A2 than
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 469
DOI: 10.5675/ICWRER_2013
Table 5: LARS-WG statistical performance to simulate the distribution (KS-test) and the mean (t-test) of the various aggregated observed climate variables for each the sub-basin. Numbers indicate how many times the hypothesis H0,(simulated and observed variables are the same) has been rejected in favor of the alternative hypothesis H1
more specifically changes, relative to the 20th - century reference period - of the maximum
and minimum temperatures and the precipitation for the sub-basins South Gojjam and
Tana. Results for other sub-basins are presented in Cherie [2013] and are rather similar,
particularly as far as temperatures are concerned, whereas seasonal precipitation is
somewhat more variable across the various sub-basins.
The most noteworthy feature of the various panels of Figure 14 is the systematic future
increase of both the maximum and minimum temperature for all seasons, of up to
4 °C for the 2090s decade. This temperature increase is particularly pronounced for Tmax
Figure 13 LARS-WG- mono-model monthly calibration (left column) and validation (right column) results for Tmin, Tmax and Rain in the UBNRB- sub-basins Fincha, Guder and Jimma
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 471
DOI: 10.5675/ICWRER_2013
during the future summer seasons. Moreover, similar to the SDSM-downscaling results
of the previous section, the LARS-WG- predicted future warm-up across the UBNRB
is also consistently higher for SRES- scenario A2 than for A1B, which should be of no
surprise, as the same large-scale GCM (ECHAM5) predictors have been used in the two
downscaling methods.
In contrast, the LARS-WG- mono-model future precipitation changes exhibit stronger
seasonal variations. As shown in the two corresponding panels of Figure 14 for the two
sub-basins - likewise to all other sub-basins - for the 2050s- period the winter precipitation
will go up tremendously, but this increase will have tapered down again for the 2090‘s.
Figure 14 LARS-WG- mono-model future predicted seasonally averaged changes, relative to the 20th - century reference period, of maximum (Tmax ) (top), minimum (Tmin ) (middle) temperature and precipitation (pcp) (bottom), for sub-basins South Gojjam and Tana for the two time-slices 2050s and 2090s and the two SRES-scenarios A1B and A2.
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 472
DOI: 10.5675/ICWRER_2013
For the future spring and summer seasons the precipitation changes are much smaller
and no clear trend is visible for all sub-basin, i.e. the former may be positive for some and
negative for others. The influence of the SRES-scenario is such that for most sub-basins
the precipitation changes are larger for SRES A1B than for A2. This result is somewhat
surprising, as SRES A1B is more benevolent than A2, and requires further explanations.
4.4. Comparison of SDSM- and LARS-WG- mono model downscaling methodsIn this section the future seasonal UBNRB- climate predictions of the two mono-models
(using the same large-scale GCM, ECHAM5) downscaling methods SDSM and LARS-WG
are compared.
The downscaled seasonally averaged maximum temperatures Tmax shown in Figure 15
(top panel) indicate for both the SDSM- and the LARS-WG- downscaling tool a considerable
increase for the two future time-periods, irrespective of the SRES- scenario (A1B, A2)
assumed.
However, for both SRES, the LARS-WG- Tmax- predictions are consistently higher by 1.1-1.5 °C
for the 2050s and by 2.4-3°C for the 2090s than the SDSM-ones. At this stage one can
only speculate about the reasons for these discrepancies between the two downscaling
methods which, as discussed, are using quite different statistical procedures.
The situation is similar for the predicted minimum temperatures Tmin, shown in the
middle panel of Figure 15, which increase considerably for all seasons, scenarios and
future time periods, particularly, when using LARS-WG downscaling. Thus the latter
predicts an increase of Tmin by 1.1-1.5 °C for the A2_2050- case and and of 3.4-4.3 °C for
the A1B_2090- case. For both downscaling tools the overall rise of Tmin is always larger
than that of Tmax and both are higher in the 2090s than in the 2050s which indicates
clearly that future climate warming in the UBNRB is an unequivocal fact.
The bottom panel of Figure 15 illustrates the predicted seasonal precipitation. One
may notice that for both downscaling approaches and for both future time periods
the precipitation for the SRES scenarios A1B and A2 are more or less similar for all, but
the winter season. Moreover, SDSM shows a seasonal decrease of the precipitation in
the summers and springs of the 2050s, which is accentuated in the 2090s. In contrast,
LARS-WG predicts a minor increase of the precipitation for most seasons, but which is
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 473
DOI: 10.5675/ICWRER_2013
Figure 15 Mono-model- SDMS- and LARS-WG- downscaled average UBNRB future seasonal changes of maximum (top), minimum (middle) temperatures and precipitation (bottom), for the two time-slices 2050s and 2090s and the two SRES-scenarios A1B and A2
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 474
DOI: 10.5675/ICWRER_2013
4.5. LARS-WG- multi-model downscalingSince each global climate model (GCM) has its own weakness and strengths in a typical
climate analysis, it is to be expected that by using predictors from more than one GCM
the prediction bias may be somehow reduced. In the LARS-WG- multi-model downscaling
method proposed here as a novel approach, LARS-WG is forced using averages of the
corresponding time series‘ predictors from the three GCMs GFLD21, SCIRO-MK3 and
ECHAM5 (see Table 2). To that avail the LARS-WG- scenario files have been prepared
with average of the outputs of these three GCMs. All other procedures are more or less
similar to those of the previously used LARS-WG mono-model [e.g. Cherie, 2013].
The results obtained with this methodology reveal unequivocally that for both SRES-
scenarios the predicted future maximum and minimum temperatures will increase
Figure 16 LARS-WG- multi-model future predicted seasonally averaged changes of the precipitation (pcp) for various sub-basins. Other notations are identical to those in Figure 14
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 475
DOI: 10.5675/ICWRER_2013
consistently across all sub-basins. Thus, the average UBNRB areal Tmax will rise by 1.8 °C in
the 2050s and by up to 3.2 °C in the 2090s, whereas the corresponding Tmin increases are,
with 2.0 °C and 3.4 °C, respectively, even higher. For this reason we forgo a sub-basin-
wise presentation of the temperature changes, which are more or less similar to those
obtained with the LARS-WG- mono-model downscaling method previously, and show in
Figure 16 only monthly precipitation predictions for a few sub-basins which, obviously,
are much more variant than the temperatures.
Figure 16 illustrates that for most of the sub-basins shown - which holds also for the
remaining ones (see Cherie [2013]) - namely, those in the highlands of the UBNRB
(Figure 1), precipitation in winter and, to a lesser extent, also in summer is significantly
reduced for the 2090s time period, whereas autumn precipitation is going up slightly.
When averaging these still for the sub-basins varying precipitation changes over the
whole UBNRB, Figure 17 results. This figure shows that for the 2050s time period the
UBNRB seasonal precipitation will decline by -2% to -3% in the spring, by - 4% to -5%
in the summer and by -6% to -7% in the autumn. During the 2090s period, a further
decrease is obtained for all, but the autumn season (where an increase of 4% -5% is
observed), namely, for winter where a decline by -11% to -15% is predicted.
So far in the paper only the predictions for the three primary meteorological variables
maximum, minimum temperatures and precipitation have been discussed. However,
the primary GCM‘s used in the downscaling processes (see Table 2) provide further
Figure 17 LARS-WG- multi-model downscaled future seasonal predicted average UBNRB- precipitation changes for the two future time slices and the two SRES-scenarios.
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 476
DOI: 10.5675/ICWRER_2013
Figure 18 LARS-WG- multi-model future predicted seasonal lengths of the wet-spells for various sub-basins. Other notations are identical to those in Figure 14
climate output, namely, and the lengths of the wet- and dry-spells, already mentioned
briefly in Table 5. These two variables provide information on how the day-intervals of
precipitation / no precipitation, respectively, are actually distributed over a season or
the year, and are thus of particular interest for agriculture, which in fact, is the dominant
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 477
DOI: 10.5675/ICWRER_2013
Figures 18 and 19 show LARS-WG- multi-model future predicted seasonal lengths of the
wet-spells and dry-spells, respectively, for various sub-basins. For comparison the 20th -
century reference predictions are also included in the various panels.
Figure 18 indicates that, for the summer season, in particular, the future predicted
wet-spell lengths are in general shorter than the recently observed ones. For the other
seasons of the year there are no noticeable differences between future and presence.
For the future dry-spell lengths (Figure 19) the situation is just the opposite, i.e. these will
be longer in the future for most sub-basin. This means, obviously, that drought intervals
across the UBNRB will be longer by the end of this century.
Figure 19 LARS-WG- multi-model future predicted seasonal lengths of the dry-spells for various sub-basins. Other notations are identical to those in Figure 14
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 478
DOI: 10.5675/ICWRER_2013
From the two figures one may also notice that the named changes of both the wet- and
the dry-spells do not depend much on the SRES-scenario used.
4.6. Comparison of LARS-WG- mono-model- and multi-model methodsIn this section the prediction results of the two LARS-WG downscaling variants, i.e using
either mono-model- or multi-model predictor data from the parent GCMs (see Table 2)
are compared. The various panels of Figure 20 show the sub-basin-wise predictions of
Figure 20 Sub-basin-wise predictions of maximum- (top), minimum (middle) temperature- and precipitation- changes (bottom) for the 2050s (left column) and the 2090s (right column) using LARS-WG- mono-modal- (index 1) and LARS-WG- multi-model downscaling (index 2)
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 479
DOI: 10.5675/ICWRER_2013
maximum-, minimum temperature and precipitation for the two future time periods
using the LARS-WG- mono-modal- and the LARS-WG- multi-model downscaling variants.
From the temperature panels of Figure 20 one may notice that with the exception of Belles
sub-basin, for all other sub-basins the mono- and multi-model LARS-WG approaches
deliver almost the same results for the maximum- and minimum temperature- changes
(relative to the 20th - century reference period) in the 2050s future decade. In contrast,
considerable differences are obtained for the 2090s, when the multi-model- predicted
temperature changes are more than 1°C lower than the ones obtained with the mono-
model LARS-WG variant. A more detailed seasonal analysis [Cherie, 2013] shows that
these methodology-originating differences are particularly pronounced for the future
summer seasons, but are only minor for the other seasons of the year.
The two precipitation panels of Figure 20 indicate clearly that for all sub-basins the annual
precipitation changes predicted with the LARS-WG- multi-model approach are also smaller
than those of the mono-model variant. However, on the contrary to the temperatures above,
these precipitation differences are, surprisingly, more apparent for the 2050s than for the
Figure 21 Comparison of aggregated UBNRB climate prediction changes (Tmin, Tmax, precipitation P) obtained with LARS-WG- mono-modal- and LARS-WG- multi-model downscaling methods for the two future time periods and the two SRES scenarios
ICWRER 2013 | Mono- and multi-Model statistical Downscaling of GCM- Climate Predictors … 481
DOI: 10.5675/ICWRER_2013
Figure 22 Summary of future seasonal UBNRB- predictions of maximum- (top), minimum (middle) temperature- and precipitation (bottom) changes for the three downscaling variants SDSM, LARS-WG- mono-mode (index 1) and LARS-WG- multi-model (index3)