GCMs are known to have coarse horizontal resolutions with feedback and simulation biases. RCMs, on the other hand, are more sophisticated and have high horizontal resolutions, yet are not free from biases since both regional downscaling simulation and the parent driving GCM induce spatial and temporal biases (see e.g., Burhan et al., 2014). For the purpose of projecting robust climate change extremes and to have high confidence in the results, model selection is made by gauging its ability to emulate observed climatology both for temperature and precipitation. Moreover, Taylor diagrams for GCMs and RCMs are made to obtain a concise statistical summary of how well the simulated climatology match the observed climatology in terms of their correlation, their root-mean-square difference and the ratio of their variances (Taylor, 2001).
GCMs emulating observed climatology
Area averaged mean annual maximum and minimum temperatures, and area averaged total annual rainfall is extracted from the 7 GCMs and put to comparison with the observed climatology of maximum and minimum temperature, and rainfall. Province based climatologies are constructed (Figure 3.1). It is seen that in all provinces, maximum and minimum temperature climatologies are well represented by all GCMs except for Had-GEM2-AO (which is not in fact the actual Hadley center simulated model, but its modeling center is NIMR, South Korea). In terms of precipitation, generally models tend to have high scatter, so the precipitation results for GCMs thereafter tend to remain in low confidence.
GCMs emulating observed climatology over Baluchistan
In terms of rainfall in Baluchistan, quite haphazard climatological patterns displayed by different GCMs suggests that the coarse resolution of the GCMs are unable to resolve rugged terrain and barren climate of the province. However, CCSM4 to some extent is able to capture the peaks of annual rainfall cycle, yet with considerable wet biases yielding over-estimated results for the past climate of the province.
In Baluchistan province, GCMs correlate well (95-99%), root-mean-square difference is small (less than 0.5 units), and the models lie within ±0.5 standard deviation of the observed maximum and minimum temperatures. Nevertheless, no good correlation is seen to be established between any of the models representing precipitation statistics of the Baluchistan province (Figure 3.2).
GCMs emulating observed climatology over GB-AJK
Both maximum and minimum temperatures in GB-AJK are well represented by simulated GCMs climatologies, however, with considerable cold biases. Moreover, GCMs tend to pick up the winter (DJFM) precipitation climatological cycle comparatively better than that in the summer (JJAS), yet, once again, with huge biases in the province. CCSM4 is seen as better representing the annual precipitation cycle, as compared to the rest of the models.
In GB-AJK, all the models except HadGEM2-AO are giving 95-99% correlation with the observed in terms of maximum temperature. All the models except HadGEM2-AO are within the half root mean square difference, as well as within the +0.5 standard deviation of the observed maximum temperature. Same results hold for the minimum temperature in GB-AJK. In terms of precipitation, there is both large variability and error, along with weak correlations (some models even giving negative correlations) in the GB-AJK Province. CCSM4 gives a comparatively better estimation with 60% correlation, however, both the root-mean-square difference and the ratio of the variances tends to remain on the higher side.
GCMs emulating observed climatology over KPK
Summer (JJAS) maximum and minimum temperature in KPK is better resolved as compared to that in the winter (DJFM) by GCMs. Precipitation climatologies represented by GCMs are poorly resolved for summer (JJAS) and partially better resolved in winters (DJFM), however, with both wet and dry biases. CCSM4, nevertheless, captures both winter (DJFM) and summer (JJAS) precipitation peaks, yet, once again suffers with dry biases.
Both simulated maximum and minimum temperatures show good correlations (95-99%), small root-mean-square differences, and within +0.5 standard deviation existence in the KPK province. There is however, the same exception for the HadGEM2-AO model which does not emulate the observed statistics at all. In terms of precipitation, the scatter of different GCMs is large with most of the models bearing poor Taylor Statistics. Nevertheless, CCSM4 model has a comparatively better correlation (nearly 80%), and within -0.5 standard deviation of the observed, yet the root-mean-square difference remains above 0.5 units in the KPK province.
GCMs emulating observed climatology over Punjab
Unlike, Baluchistan and GB-AJK, Punjab is a plane area with vast gradually altering slopes, hence resolving terrain borne climatic features of this province is comparatively less challenging for the GCMs. However, Punjab at the same time is the province that receives maximum amount of monsoon rainfall in JAS, and hence monsoon dynamics play an important role in determining the precipitation patterns and cycles of this region. Owing to this probable limitation, the GCMs have not been able to successfully emulate precipitation peaks of monsoon in JAS. Nevertheless, CCSM4 is the only model in our analysis that has shown comparatively better annual cycle representation of precipitation with the observed in the Punjab province.
The simulated maximum temperature in Punjab displays good correlations (90-99%) with the observed. The root-mean-square difference is small and the models tend to remain within an acceptable range of +0.5 standard deviation of the observed. This holds same for the minimum temperature. However, in terms of precipitation, only CCSM4 model displays high correlation (nearly 90%), a smaller root-mean-square difference and an acceptable standard deviation ratio with the observed climate of Punjab.
GCMs emulating observed climatology over Sindh
Maximum and minimum temperature climatology of Sindh is resolved within justifiable ranges. However, similar to other provinces, precipitation cycle remains a challenge to get resolved with currently analyzed GCMs’ monsoon dynamics. CCSM4 captures the monsoon rainfall peaks better than the remaining models. However, quite large over-estimation of precipitation in CCSM4 in monsoon rains is also observed.
Simulated maximum temperature displays a high correlation (90-99%), and remains within an acceptable range of root-mean-square difference of the observed climate in Sindh province. In terms of minimum temperature, the correlation is even higher (models tending to be at 99% correlation). The ratio of the variances displayed by simulated minimum temperature is within ±0.5 units and the root-mean-square difference remains less than 0.5 units for the minimum temperature. In terms of precipitation bcc-CSM1.1 model displays a comparatively better correlation (nearly 99%), however, at the same time it suffers with high magnitudes of variance ratio, and high magnitudes of root-mean-square differences in Sindh province.
STATISTICALLY DOWNSCALED GCMS EMULATING OBSERVED CLIMATOLOGY
Statistically downscaled Nex-NASA GCMs output is area averaged for maximum temperature, minimum temperature, and precipitation independently over all the provinces of the country. Climatologies are extracted and annual cycles are constructed to make a comparison with the observed (Figure 3.3). Moreover, these climatologies are further put into Taylor Statistics to obtain high confidence over model selection (Figure 3.4). The results show good climatological aspects of Nex-NASA product both in terms of temperature and precipitation.
Statistically downscaled GCMs emulating observed climatology over KPK
Annual cycles of maximum and minimum temperatures, as well as of precipitation are well captured, however the temperatures suffer from cold biases whereas the precipitation suffers from dry biases. The winter peaks of precipitation are well overlapped with the observed, yet the summer peaks tend to remain below the observed, indicating under-estimation in the KPK province.
All parameters (TMAX, TMIN, and precipitation) display satisfying values for Taylor Statistics in KPK province. The correlations are high in precipitation (95-99%) whereas they are higher in TMAX and TMIN (> 99%) in the KPK region. The root mean square is less than 0.5 units for all the analyzed models and the parameters for the KPK region. Moreover, the root mean square difference is small (< 0.5 units) in precipitation, whilst it is further smaller in TMAX and TMIN.
Statistically downscaled GCMs emulating observed climatology over Punjab
The climatology of Punjab is brilliantly captured by the Nex-NASA product, for both maximum and minimum temperature, as well as for precipitation, for all the analyzed models. Monsoon peaks in precipitation cycle show small under-estimation by Can-ESM2 and CNRM-CM5, whereas they show small over-estimation by MRI-CGCM3, when compared with the observed. Nevertheless, the magnitude of under-estimation and over-estimation is quite small as compared to raw GCM products.
The temperature parameters display very small values for root-mean-square differences, as well as for the ratio of the variances, along with high correlations (> 99%) with the observed climate in Punjab. In terms of precipitation, the Nex-NASA product tends to display small ratios of variances (within ±0.5 standard deviation of observed), small root mean square differences (< 0.5 units), and simultaneous high correlations (> 95%) with the observed.
Statistically downscaled GCMs emulating observed climatology over Sindh
Both maximum and minimum temperatures display a nice overlapping of the Nex-NASA and the observed data over the Sindh province. Climatology of precipitation is also well captured, however, small under-estimation of precipitation in monsoon season is observed. Nevertheless, the biases are systematic and may be removed by simple bias correction techniques for the purpose of impact studies.
Temperature and precipitation climatologies of the Nex-NASA product are well coordinated with the observed, in terms of all three Taylor statistics over the Sindh region. The TMAX and TMIN have virtually no root mean square differences, and the ratios of the model variances to the observed is 1. Moreover the correlations are high (> 99%) in the region. In terms of precipitation, the root mean square differences are small (< 0.5 units). The ratios of variances to the observed tends to remain between 0.5 and 1. Furthermore, the correlations are high with values exceeding 99% for the region.
Statistically downscaled GCMs emulating observed climatology over Baluchistan
The Nex-NASA product in Baluchistan is well overlapped over observed climatology of the maximum and the minimum temperature. The precipitation cycle, though well coordinated, suffers from under-estimation in the summer season.
The TMAX and TMIN emulation of Nex-NASA product displays high correlations (99%) in the Baluchistan province. The root mean square difference is small (< 0.5 units), and the ratios of variances to the observed are between 0.5 and 1. In terms of precipitation, the correlations are high (80-95%), with ratios of variances to the observed ranging between 0.5 and 1 for the Baluchistan region.
Statistically downscaled GCMs emulating observed climatology over GB-AJK
Both maximum and minimum temperature as well as precipitation products of Nex-NASA display comparative annual cycles with the observed, yet the temperature is over-estimated whilst the precipitation is under-estimated in the GB-AJK region. Peaks for both winter and summer precipitation are captured nicely over the region, however, the peak of the winter precipitation of the Nex-NASA data lags behind by one month leading to some possible time oriented biases in the region.
Taylor Statistics of maximum and minimum temperatures of GB-AJK display very good correlations (> 99%), whereas those of precipitation of GB-AJK display good correlations (90-95%) in the region. The ratios of variances in maximum and minimum temperatures are quite close to 1 (meaning almost identical variances), however, those in precipitation are between 1.5 and 2 (meaning large variations in variances) in the GB-AJK region. Moreover, root-mean-square differences in maximum and minimum temperatures are quite small, whereas those in precipitation are greater than 0.5 units.
Dynamically downscaled GCMs emulating observed climatology
Dynamical downscaling, in addition to its non-trivial initial and boundary conditions, features correction of terrain and relief borne climatic biases in the models. The CORDEX RCMs used in the study offers dynamically downscaled GCMs product over South Asian region. Annual climatologies from the six RCMs are extracted and plotted over observed to analyze credibility of the RCMs to emulate annual cycle over each province (Figure 3.5). It is important to note that RCMs are downscaling tools that may induce systematic or non-systematic errors in the output product, based on the driving GCM. The output product from an RCM therefore inherits two bias sources – one from the driving GCM, and the other from the RCM simulation.
Dynamically downscaled GCMs emulating observed climatology over GB-AJK
The CORDEX climatology for GB-AJK maximum and minimum temperature emulates the annual cycle in fact with huge systematic biases. Relatively better representation of the annual cycle is seen in SMHI RCA4 ICHEC-EC-EARTH RCM, yet suffers from small under-estimation in temperature output. The precipitation climatology of CORDEX RCMs in GB-AJK region displays large non-systematic errors that restrict the RCMs ability to correctly represent summer and winter precipitation peaks in the region. Nevertheless SMHI RCA4 ICHEC-EC-EARTH RCM is able to capture the winter precipitation with good ability, yet summer precipitation is quite over-estimated in the GB-AJK region.
Taylor Statistics of GB-AJK region displays all the CORDEX RCMs output for maximum and minimum temperatures to remain within 0.5-1.5 units of ratio of the variances, as well as within 0.5 units of root mean square differences to the observed. The correlations for TMAX and TMIN is > 99% and > 95% respectively. However, none of the CORDEX RCMs successfully imitate Taylor Statistics for observed precipitation in the GB-AJK region (Figure 3.6).
Dynamically downscaled GCMs emulating observed climatology over KPK
Maximum and minimum temperatures from CORDEX models display fairly good climatological patterns with the observed, yet with both over-estimated and under-estimated systematic biases present. In terms of precipitation, all CSIRO modeling center RCMs have fairly captured summer precipitation peaks (though with small under-estimation), while they have been un-successful in capturing the winter precipitation peak (rather showing a trough instead of a peak) for the KPK region. Winter precipitation peak is fairly well captured by REMO2009 MPI-ESM-LR RCM (with small under-estimation) over the KPK province.
The Taylor Statistics for KPK province displays good correlations (> 95 % for TMAX and > 99 % for TMIN) for all the CORDEX analyzed models. The root mean square differences of all the models are within 0.5 units of the observed, whilst the ratio of variances to the observed lie within 1-1.5 units for all the analyzed CORDEX models. For precipitation, REMO2009 MPI-ESM-LR and SMHI RCA4 ICHEC EC EARTH RCMs tend to have relatively high correlations of 50 % and 60 % respectively, with the observed. However, both models have high root mean square differences with the observed (> 0.5 units), in terms of KPK precipitation statistics.
Dynamically downscaled GCMs emulating observed climatology over Punjab
Both maximum and minimum temperature climatologies of CORDEX RCMs display good overlapping with the observed in Punjab province. In terms of precipitation all CSIRO modeling center RCMs have captured the winter precipitation peaks with good precision (yet they display early onset and gradual withdrawal of monsoon rains, as compared to the observed), however, they are not good at representing winter precipitation in Punjab province with correct figures. Nevertheless, REMO2009 MPI-ESM-LR RCM is able to capture the winter precipitation peak of Punjab province, however, it suffers from dry biases.
The root mean square differences of TMAX and TMIN of Punjab province from CORDEX RCMs are small (< 0.5 units) when compared with the observed. The ratio of the variances to the observed for TMAX and TMIN remains within 1-1.5 from all the analyzed CORDEX RCMs. The correlations are high (> 95 % for TMAX and > 99 % for TMIN) for all the CORDEX analyzed models in Punjab province. In terms of precipitation, correlations for all CSIRO based RCMs are high (> 80 %), yet the root mean square differences are > 0.5 units. Nevertheless, the ratio of the variances to the observed precipitation is quite close to 1 which means that the scatter of the CSIRO RCMs data is close to that of observed over Punjab province.
Dynamically downscaled GCMs emulating observed climatology over Sindh
In Sindh province, both maximum and minimum temperature climatologies are well exhibited when compared with the observed. Precipitation however, is suffering from both systematic and time lead/lag biases from CORDEX RCMs. Relatively better representation of Sindh precipitation climatology is given by REMO2009 MPI-ESM-LR RCM where it captures monsoon peak with good precision but with small over-estimation in the Sindh province.
The Taylor Statistics of Sindh province for TMAX and TMIN from CORDEX RCMs display high correlations (> 95 % and > 99 % respectively), small root mean square differences (< 0.5 units), and the ratio of their variances to the observed are quite close to 1. The REMO2009 MPI-ESM-LR RCM seems to better imitate the observed precipitation statistics with > 90 % correlation, < 0.5 units of root mean square difference, and proximity of the ratio of the variances to one.
Dynamically downscaled GCMs emulating observed climatology over Baluchistan
The CORDEX RCMs are displayed to emulate maximum and minimum temperature climatologies of Baluchistan with good precision, yet there are small over-estimated and under-estimated biases. The precipitation climatologies for summer season are also well captured by the majority of the analyzed models (with systematic biases), however, winter season precipitation climatology is not captured by any of the CORDEX RCMs over Baluchistan.
The Taylor Statistics for TMAX and TMIN in Baluchistan province from all the analyzed CORDEX models display good correlations (> 95 % for TMAX and > 99 % for TMIN), as well as small root mean square differences (< 0.5 units). The ratio of the variances to the observed for TMAX and TMIN is close to 1 which indicates that CORDEX models tend to have nearly same spread in their climatological data as seen in that of observed data for TMAX and TMIN. As far as the precipitation is concerned, no robust Taylor Statistics could be found by any of the CORDEX RCMs for Baluchistan province.
Verdict over GCMs and RCMs subsets selection
Models’ behaviour in emulating climate of different provinces of Pakistan leads us to derive some generalized inferences:
a. Raw GCMs have generally shown only systematic biases in TMAX and TMIN, while precipitation is not resolved by any of the raw GCMs except by CCSM4. It is also worth mentioning that grid resolution alone, is not to be taken as a standard for a GCM to emulate observed climate. As is seen in Figure 3.7, CCSM4 has a coarser grid resolution than both the Australian based and the Japanese based models, yet it serves us well with resolving the regional climate better than any of the remaining analyzed models.
b. Since all Nex-NASA statistically downscaled GCMs are bias corrected, therefore they display highest confidence in emulating all the parameters correctly.
c. Individual representation of CORDEX RCMs in terms of describing the climate over different provinces induces huge systematic and non-systematic errors, to which the solution lies in taking the ensemble average of all the RCMs for further analysis. Moreover, it is seen that all RCMs from CSIRO modeling center have shown quite similar climatologies and statistics which leads to a hypothesis that the CSIRO models have large influence of the RCM (CCAM) from which they are downscaled, rather than that of the driving GCMs.
Based on the ability of models to emulate observed climatologies and statistics (as described in the above discussion), the selected raw GCMs, the statistically downscaled GCMs, and the selected dynamically downscaled GCMs, are summarized in Table 3.1.