1 Introduction

Himalayan watersheds of the Indus basin (Fig. 1), namely, Jhelum, Kabul and upper Indus basin (UIB) accumulate around 80% of the total surface water available in Pakistan that largely contributes to the socio-economic wellbeing of the country. The exposure of mountainous environments to climate change (e.g. Huber et al. 2006; Grabherr et al. 2010) and recently established elevation dependent warming (MRI 2015), raise serious concerns about the future hydroclimatology of Himalayan watersheds as well as their contributions to the downstream water availability (Hasson 2016a; Hasson et al. 2017). Thus, knowledge-based advice on the potential impacts of climate change on the water resources of Himalayan watersheds is a timely need for informed decisions in order to avoid, adapt or mitigate climate change adversities thereby ensuring sustainable agrarian economy downstream. However, addressing this concern requires, before all, robust information on the present climate and on its future changes as projected under a variety of greenhouse gases (GHG) emission scenarios.

Fig. 1
figure 1

The Himalayan watersheds of Indus basin, namely, Kabul, Jhelum and upper Indus basin (UIB)

Full size image
Table 1 List of CORDEX South Asia experiments and their driving CMIP5 models
Full size table

Presently, state-of-the-art climate models serve as primary tools for emulating Earth’s climate, and to project its changes under a variety of anthropogenic or natural forcings (Hasson et al. 2016b). The rapid development of these climate models undoubtedly raised expectations of their robust support towards devising relevant climate change policy and development of national action plans. However, obtaining reliable and valid future climate change information from these climate models depends highly upon their skill in reproducing the observed hydroclimatic phenomena. This, in turn, requires a near accurate representation of the region specific geophysical features, their associated processes and their sophisticated feedback on a variety of representative spatiotemporal scales. Given that the study region features a tremendous diversity in such fine scale features and processes, it becomes challenging for the climate models to prove their fidelity, and thus to earn reliability for their projected future climatic changes.

For instance, the Himalayan watersheds, aboding a highly concentrated cryosphere and complex terrain of western Himalaya, Karakoram, and eastern Hindukush (HKH) mountain ranges as well as of western Tibetan Plateau, are situated at the borders of two large-scale circulation modes, the extra tropical westerlies and the south Asian summer monsoon. The extra tropical westerlies (summer monsoon) bring moisture mainly during winter and spring (July–September) and largely over the high (low) altitude regions (Wake 1989; Ali et al. 2009; Hewitt 2011; Ridley et al. 2013; Hasson et al. 2016b; Janowiak and Xie 2003; Krishnamurti et al. 2012). These distinct large-scale circulation modes and their associated precipitation regimes are controlled by both global scale phenomena and local forcing. Precipitation is anomalously high during the warm phase of the El Niño-Southern Oscillation (ENSO) (Shaman and Tziperman 2005) and during the positive phase of the North Atlantic Oscillation (NAO) (Syed et al. 2006). On a local scale, the complex HKH terrain largely influences the spatiotemporal precipitation distribution by modulating the moisture-laden winds and contributes in altering the local climate (Dimri and Niyogi 2013). The modulated behavior of large-scale circulations and their associated precipitation regimes adds to the seasonality of available moisture and its form (rain or snow), and subsequently to the seasonality of water availability, defining the overall hydrological regimes of the Himalayan watersheds. Due to importance of the abundant frozen water resources of these watersheds for ensuring food and energy security downstream, their future water availability assessment largely depends upon valid climate change projections from the climate models. Hence, the fidelity of climate models needs to be critically assessed for the Himalayan watersheds.

A relatively few number of studies have been performed in this regard over the study region, focusing on global and regional climate models (GCMs and RCMs). For instance, Syed et al. (2009) and Islam et al. (2009a) have analyzed the performance of 06 and 17 GCMs participating in the third and fourth assessment report (TAR and AR4) of the IPCC (2001 and 2007), respectively, and found that the majority of models feature substantial cold and wet biases against the CRU (Climate Research Unit) observational dataset. Their findings suggest that the performance of typically coarse resolution GCMs is limited over the complex terrain of the study area. In contrast, RCMs are generally expected to outperform the coarse resolution GCMs, due to their relatively detailed representation of the geophysical features and associated processes. Using the PRECIS (Providing Regional Climates for Impact Studies) RCM, Islam et al. (2009b) have downscaled the ECMWF ERA40 reanalysis and HadAM3P GCM datasets on 0.44\(^{\circ }\) resolution for the historical period (1961–1990). Besides achieving fine resolution, they have consistently reported substantial cold and wet biases against the CRU observations. Kulkarni et al. (2013) have downscaled three perturbed physics ensembles from HadCM3 GCM on 50 x 50 km resolution using the PRECIS RCM and validated the outputs against the APHRODITE observed rainfall and NCEP-NCAR reanalysis temperatures over eastern, western and central Himalayan regions. They have reported marked cold and wet biases for the western Himalayas, partially covering the study area. Validating the same experiment over the whole Indus basin using APHRODITE rainfall and CRU temperature observations, Rajbhandari et al. (2015) have also shown cold and wet biases over the UIB. Similar findings of substantial cold and wet biases have been reported by Ali et al. (2015), who have validated the fine scale derivatives of two GCMs participating in the Coupled Model Intercomparison Project Phase-5 (CMIP5) obtained by CCAM (Conformal-Cubic Atmospheric) and RegCM (ICTP Regional Climate Model) models. It is to note that the magnitudes of qualitatively agreed cold and wet biases among the above mentioned studies vary mainly due to: (1) the integration over different study domains, as none of these studies particularly focus on the Himalayan watersheds of Indus Basin; (2) the observational datasets considered as a truth for validation, emphasizing on the need to consider a broader database for climate model validation; (3) the datasets used to drive the RCM, implying that the limited performance of driving datasets further limits the skill of downscaled experiments.

Further, Syed et al. (2014) have downscaled the ERA40 reanalysis and ECHAM5 GCM output using the PRECIS and ICTP Regional Climate Model version 4 (RegCM4) RCMs. Validating fine-resolution derivatives against the CRU and UDEL (University of Delaware) observations, they suggest that both datasets exhibit substantial cold bias and that besides huge differences in their forcing each RCM improves for biases. This implies that the RCMs are to-some-extent insensitive to their lateral boundary conditions and markedly add either to their skills or to their uncertainties (Gao et al. 2012), which is particularly the case when the downscaling domain is quite large (Karmacharya et al. 2015). This is further confirmed by the findings from Syed et al. (2014) that both RCMs forced with the same GCM feature differences in their simulated climates. Mariotti et al. (2014) have suggested that dissimilarities between the forcing and the fine scale datasets might result from their varying representation of the large-scale circulations.

Nevertheless, differences in driving datasets, relative independence of RCMs to their lateral boundary conditions, and differences in RCMs’ simulated climates, all suggest a need for an ensemble approach applying multiple RCMs driven by multiple datasets, but adopting a common experimental setup as mandated by the coordinated regional climate downscaling experiment (CORDEX) framework of the world climate research program (WCRP). The CORDEX intends to nurture an international collaboration for generating and providing a set of high-resolution retrospective climate simulations as well as of future projections (Fernández et al. 2010) under a variety of GHG emission scenarios. In this regard, output from the number of GCMs participating in the CMIP5 archive (Taylor et al. 2012) have been dynamically downscaled by a suite of RCMs under common downscaling framework for 14 regions across the Globe.

In this study, we assess the fidelity of CORDEX experiments performed over the South Asia domain (CX-SA) and their driving CMIP5 experiments over the Himalayan watersheds of Indus basin during the historical period (1971–2005). The fidelity is assessed against a boarder observational database for the considered skill metrics, which mainly include the first order statistics of the temperature and precipitation climatologies, and the seasonality of precipitation. Such metrics are relevant for the impact assessment studies, particularly to those focusing on the water resources. Performance of the CX-SA experiments is compared to that of their driving experiments, in order to explore whether the dynamical downscaling has brought an added value to the coarse-resolution experiments. Based on the fidelity results for the historical period, we further assess the robustness or uncertainty of the future climate as projected under the high-end emission scenario of representative concentration pathway (RCP) 8.5 by the end of 21st century (2061–2095)—analogous to the signal-to-noise ratio.

2 Data used

2.1 Observations

We have employed a large database of widely used interpolated and remotely-sensed gridded observations in order to estimate the influence of observational uncertainty on the fidelity assessment of climate modeling experiments. Daily mean temperatures at various resolutions were obtained from: (1) the Asian Precipitation-Highly Resolved Observational Data Integration Towards Evaluation (APHRODITE) of Water Resources version V1204R1 (Yasutomi et al. 2011; (2) the Climate Research Unit Time Series version 3.2 (CRU TS3.2) (Harris et al. 2014; (3) the Global Historical Climatology Network version 2 and the Climate Anomaly Monitoring System (GHCN-CAMS) (Fan and Van den Dool 2008), and; (4) the University of Delaware (UDEL)—(Willmott and Matsuura 2001) datasets. Among these temperature observations, APHRO is a daily dataset constructed for the Monsoon Asia region (60–150\(^{\circ }\)E, 15–55\(^{\circ }\)N) and the rest are monthly global datasets, all available at 0.5\(^{\circ }\) resolution. For precipitation, we have obtained a total of eight datasets of which five are interpolated and three are remotely-sensed datasets. The interpolated datasets are: (1) the precipitation construction overland (PREC-L)—(Chen et al. 2002; (2) the global precipitation climatology center (GPCC) version 6 (Schneider et al. 2014; (3) APHRODITE (onward as APHRO) VR1101 (Yatagai et al. 2012; (4) CRU TS3.2 (Harris et al. 2014; (5) UDEL (Willmott and Matsuura 2001), while the remotely-sensed datasets are: (6) the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks—Climate Data Record (PERSIANN-CDR—Ashouri et al. (2015)); (7) the Global Precipitation Climatology Project (GPCP—Xie et al. (2003)), and; (8) the CPC Merged Analysis of Precipitation (CMAP—Xie and Arkin (1997)). Among the considered precipitation observations, PREC/L, GPCC, CRU, UDEL GPCP and CMAP were available in monthly, while the APHRO and PERSIANN-CDR were available in daily temporal resolution.

2.2 Climate modeling experiments

Under the CORDEX south Asia framework, around eleven experiments were performed by dynamically downscaling a set of CMIP5 experiments (Taylor et al. 2012) at 0.44\(^{\circ }\) resolution using several RCMs. The retrospective climate simulations under the present-day forcing were conducted mainly for the period 1950–2005 while the future climate simulations were performed for the period 2006–2100 under RCP 4.5 and 8.5 forcing scenarios (Van Vuuren et al. 2011; Moss et al. 2010). We have obtained temperature and precipitation from seven CORDEX South Asia experiments (onward as CX-SA) for which the daily data were available for the historical period as well as for the future period under the extreme forcing scenario of RCP8.5. The RCP8.5 refers to the highest GHG emissions as a result of higher population but slow income growths, moderate technological development and high-energy demands, all in the absence of climate change policies (Riahi et al. 2011). We have chosen the extreme RCP8.5 scenario as it provides the upper end of the climatic changes over the Himalayan watersheds, so that the uncertainty or the robustness of the future climate changes can be fully appreciated. The obtained CX-SA experiments were performed using three different RCMs (CCAM, RCA4 and REMO) forced with six CMIP5 GCMs (ACCESS1-0, CCSM4, CNRM-CM5, EC-EARTH, GFDL-CM3 and MPI-ESM-LR). In order to assess the added-value of the downscaled CX-SA experiments, we have also obtained the temperature and precipitation data from their six driving CMIP5 experiments. The details of RCMs, their forcing CMIP5 GCMs, and contributing institutions are summarized in Table 1.

3 Methods

The spatial resolution of all interpolated (four temperature and five precipitation) datasets was 0.5\(^{\circ }\) whereas the remotely sensed precipitation of GPCP and CMAP available at 2.5\(^{\circ }\) and of PERSIANN-CDR available at 0.25\(^{\circ }\) were interpolated to 0.5\(^{\circ }\) resolution. It is to mention that constructing observational fields in a highly complex terrain as of the study region is challenging, particularly in the absence of a high-quality dense station network with longterm records (Collins et al. 2013). Few observations incorporated from the study region in the gridded datasets, if at all, are from the low-altitude stations only, which are less representative of the Himalayan topoclimate (Hasson et al. 2014a, 2016a), particularly of the high-altitude solid moisture component (Yatagai et al. 2012; Palazzi et al. 2013; Prakash et al. 2015; Hasson et al. 2016b). Further, these spatially complete observations are subject to discrepancies owing to the limited skill of the interpolation methods and of the remote sensing techniques within the complex Himalayan terrain, thus featuring a varying skill across the study region (Yatagai et al. 2012; Palazzi et al. 2013; Prakash et al. 2015). Since it is difficult to single out any individual dataset as a reference, the mean of the observational ensemble (OBS) was taken as a reference dataset and its robustness was assessed by comparing the pool of 27 meteorological stations against that of the collocating grid cells. For this, nine goodness-of-fit statistics were considered that included Pearson correlation coefficient (r), coefficient of determination (R\(^2\)), index of agreement (d), percent of bias (pbias), Nash-Sutcliffe efficiency (NSE), mean error (ME), mean absolute error (MAE), root mean square distance (RMSD) and its ratio to the standard deviation of station observations (RS). Details on the employed goodness-of-fit statistics can be found in (Legates and McCabe 1999; Nash and Sutcliffe 1970). Before applying the goodness-of-fit statistics, the gridded temperature observations at 0.5\(^{\circ }\) resolution were adjusted for their height difference against the collocating stations using the mean environmental lapse rate.

The CX-SA and CMIP5 datasets were brought to the grid resolution of observations (\(0.5^{\circ }\) × \(0.5^{\circ }\)) using bilinear interpolation (temperatures were adjusted for the elevation differences) for intercomparison. For the observations and retrospective model simulations, a common period of 1971–2005 was taken as the historical period, while a period of 2061–2095 was considered for the future climate simulations under the RCP8.5 scenario. Three seasons of October to February (ONDJF), March to June (MAMJ), and July to September (JAS), along with a complete hydrological year (O–S) were objectively taken as the periods of analyses, which are relevant for the region under study. The ONDJF season refers to the snow accumulation period, MAMJ season spans from the late-snow accumulation to the pre-monsoon rainfall (early melt) period, while JAS season is mainly the monsoon rainfall season over the study region. Since the study area receives its precipitation from the extra-tropical westerly disturbances during winter and spring and from the south Asia summer monsoon during July–September, MAMJ actually refers to a transition period between the active precipitation regimes of two large-scale circulations modes. Therefore, MAMJ is expected to feature prompt changes as suggested by the observations (Hasson et al. 2016a, b), hence, considered here separately.

3.1 Skill metrics

Climatology We have assessed the fidelity of CX-SA and their driving CMIP5 datasets against the OBS ensemble mean. For compactness of our analysis, we have also constructed the ensemble means of seven CX-SA datasets and of their six driving CMIP5 datasets (CMIP5) for each period and variable. The uncertainty or spread of ensembles (OBS, CMIP5 or CX-SA) is taken as the standard deviation among their member datasets. To explore the uncertainty over the elevated regions relative to foothills and plains, we have compared the ensemble spreads against the elevation. For comparing either two different ensembles or the same ensemble between two different periods, we have compared their coefficients of variation (CV), which were calculated as the ratios of ensemble spreads to ensemble means. Ensemble mean biases for the CX-SA and CMIP5 ensembles were calculated as the mean of the biases of their individual ensemble members, estimated as the distance of their historical climatology (1971–2005) from the OBS ensemble mean. Similarly, the ensemble mean future changes for CX-SA and CMIP5 were computed as the mean of the future changes as projected by their individual ensemble members, calculated as the difference between their simulated future climatology (2061–2095) and historical climatology (1971–2005). Spatial performance of each individual member of OBS, CX-SA and CMIP5 ensembles and of their ensemble means has been summarized in Taylor diagrams (Taylor 2001). A Taylor diagram graphically summarizes the relative differences in the spatial patterns amid various simulated and observed datasets in terms of Pearson correlation coefficient, r, centered RMSD and amplitude of variation (represented by the standard deviation). The r is plotted along the radial axis between x-axis and y-axis; the closer the dataset is to x-axis higher the correlation with the reference dataset. The RMSD of each dataset plotted on the diagram refers to its distance from the reference dataset mentioned on the x-axis, where standard deviation is one. The standard deviation of each datasets plotted on the diagram is proportional to its radial distance from the origin.

Precipitation seasonality In addition to mean biases, we have assessed how well the ensemble members and ensemble means reproduce the observed seasonality of precipitation. For this, we applied a dimensionless seasonality index (SI), which was constructed from information theory based relative entropy (RE) or Kullback-Leibler distance (Hao and Singh 2015). The RE index represents a quantitative measure of the degree of concentration of precipitation by computing a relative distance between its actual and uniform distribution (Pascale et al. 2015a). Since biases/changes in the seasonality can be of few days and weeks (Hasson et al. 2014b, 2016b), the quantities have been computed only for GPCP, CMAP and APHRO datasets on pentad temporal resolution.

Considering x as a grid cell location within each gridded datasets over the study region while t representing the total number of pentads within a year, we define \(\pi\) as the precipitation fraction of the pentad i:

$$\begin{aligned} \pi _{i,x}=\frac{p_{i,x}}{p_x}\;\; where\;\;i=1...t\;\; and\;\; P_x=\sum _{i=1}^tp_{i,x} \end{aligned}$$
(1)

then RE is computed for each season as:

$$\begin{aligned} RE_{t,x}=\sum _{i=1}^t\pi _{i,x}log_2(t\pi _{i,x}) \end{aligned}$$
(2)

The maximum value of \(RE(=log_2t)\) suggests that precipitation is concentrated in a single pentad while its minimum value of zero suggests a uniform distribution among all pentads \((\pi _i=1/t, i=1 \ldots t)\). Based on RE, the SI is obtained as:

$$\begin{aligned} SI_{x,t}=RE_{x,t}\frac{P_{x,t}}{P_0} \end{aligned}$$
(3)

where \(P_0\) is the observed maximum of \(P_{x,t}\). The SI estimates approach zero when either precipitation is uniformly distributed or there was total dryness. The SI reaches its maximum \((log_2t)\) when RE estimates for the year of \(P_0\) approaches maximum. It is pertinent to mention that the employed seasonality indicators are better as compared to the annual cycle analysis in a way that these indicators, taking into account both the spatial heterogeneity and temporal variability, yield precipitation seasonality in a quantitative manner. Accordingly, collective statistics of the annual cycle characteristics and seasonality measure can be summarized for each grid location. On the other hand, annual cycle analysis typically integrates the data along the latitudinal/ longitudinal belts or over the whole study domain (Hasson et al. 2014b), thus compromising on the spatial heterogeneity. Further details on SI and RE can be found in Feng et al. (2013), Pascale et al. (2015a, b), Hasson et al. (2016b) and Hasson (2016b).

3.2 Climatic robustness or uncertainty

The robustness or uncertainty of the future projected climatic changes has been assessed in analogy to the signal-to-noise ratio. This comprises calculating the ratios of mean future change ensemble to: (1) the spread of the future climate change ensemble, to assess whether the projected mean future change is greater than the inter-model disagreement on such change; (2) the observational uncertainty, to explore whether disagreement among the observations is higher than the projected mean future change; (3) the ensemble spread for the historical period, to analyze whether the historical inter-model disagreement is larger than their projected mean future change, and most importantly; (4) the ensemble mean bias (offset from observations), to reveal whether the extent of climate model infidelity is higher than the magnitude of their projected mean future change. The climatic robustness or uncertainty of the individual datasets from each ensemble is summarized in a majority agreement, obtained as the number of experiments featuring greater infidelity (offset from the observations) than their projected magnitude of mean future change minus the number of experiments suggesting the opposite.

4 Results and discussion

4.1 Fidelity in mean temperature climatology

The OBS ensemble mean suggests a range of mean temperature from below −20 to above 35 \(^{\circ }\)C from ONDJF to JAS, respectively (Fig. 2). The OBS ensemble members (CRU, GHCN-CAMS, UDEL and APHRO) exhibit a good agreement with each other - the standard deviation is below 1 \(^{\circ }\)C over the foothill plains, and mostly up to 2 \(^{\circ }\)C in elevated regions, though at some places and in some seasons it is slightly higher (Fig. 3). The Taylor diagram also shows a high agreement among the OBS ensemble members, where APHRO and GHCN-CAM are relatively closer to CRU and UDEL datasets, respectively. Such a high agreement amid OBS ensemble members may result from employing a similar database for their construction. Interestingly, although APHRO has reportedly incorporated a relatively large number of stations from the Monsoon Asia region (Yatagai et al. 2012), and is thought to be relatively more robust, its high agreement with the globally constructed gridded observations suggest them equally robust. Besides high inter-dataset agreement, it is notable that the OBS ensemble members are not an exception to typical errors due to the limitations of interpolation schemes and the quality of incorporated gauge data in a complex Himalayan terrain. Thus, these observations exhibit varying strengths and weaknesses over the study region (Suppl. Fig. 1). Therefore, giving each OBS ensemble member an equal weight, we consider their ensemble mean as a reasonable reference dataset for the climate model validation (Fig. 2, Supp. Fig. 1).

Fig. 2
figure 2

Ensemble means of the mean temperature climatology (1971–2005) from the observational datasets (OBS), offsets (or biases) of the CMIP5 and CORDEX ensemble means from the OBS ensemble mean. The last row presents relative differences, where a negative (positive) scale refers to the under- (over-) estimation of CX-SA relative to CMIP5 datasets. Note: stippling indicates where the magnitude of bias is higher than the OBS uncertainty

Full size image
Fig. 3
figure 3

Uncertainty (standard deviation) of the mean temperature for OBS, CMIP5 and CX-SA ensemble datasets. The last row presents difference in the CMIP5 and CX-SA uncertainties, where a negative (positive) scale refers to lower (higher) CX-SA uncertainty relative to CMIP5

Full size image
Table 2 Goodness-of-fit statistics for the validation of pooled OBS ensemble mean grids against 27 collocating meteorological stations
Full size table

Before using the OBS ensemble mean as a reference dataset for validating the climate experiments, it has been validated for the goodness-of-fit against the real observations (Supp. Fig. 2, Table 2). For temperature, the validation statistics suggest ME range of 0.09 to −1.04 \(^{\circ }\)C where MAE is slightly above than 1 \(^{\circ }\)C across the seasons and on annual scale. The RMSD is mostly below 2 \(^{\circ }\)C while its ratio to the standard deviation of station observations is around 0.4 (highly satisfactory). Further, r, R\(^2\) and d are \(\ge\) 0.9 in most cases whereas NSE ranges between 0.85 in JAS and 0.9 in MAMJ. In view of the simple elevation adjustment of the OBS ensemble mean using the environmental lapse rate, the validation results reveal quite a good agreement of the stations observations with the OBS ensemble mean, suggesting it as robust reference datasets for climate model validation.

Validation of the CMIP5 ensemble mean against the OBS ensemble mean suggests a substantial cold bias of at least 6 \(^{\circ }\)C over the whole mountainous region, comprising the Karakoram and western Himalaya ranges and the western Tibetan Plateau. Such cold bias is observed even higher over the cryospheric region and during the cold seasons of ONDJF and MAMJ (Fig. 2). Surprisingly, the pattern of substantial cold bias and its higher magnitude over the cryospheric region is more prominent from the CX-SA ensemble mean, which feature more than 10 \(^{\circ }\)C cold bias over highly glacierized watersheds of Shyok and Hunza and western Himalayas. In contrast to the elevated regions, both CMIP5 and CX-SA ensemble means suggest better agreement with OBS ensemble mean over the foothill plains, where such agreement is higher for CX-SA as compared to CMIP5. Such findings are consistent with the reports from either earlier versions of GCMs, participating in TAR and AR4 of IPCC, or their downscaled derivatives (Syed et al. 2009; Islam et al. 2009a, b; Kulkarni et al. 2013; Rajbhandari et al. 2015). For the latest generation climate models, Mishra (2015) has also reported the substantial cold bias of 6–8 \(^{\circ }\)C over the Himalayan watersheds of Indus, Ganges and Brahmaputra from a subset of CX-SA datasets analyzed here. Ali et al. (2015) have also shown substantial cold bias particularly over the UIB from two downscaled variants of CMIP5 GCMs. Analyzing a CX-SA experiment obtained from RCA4 (Rossby Centre regional atmospheric model), Iqbal et al. (2016) have consistently shown substantial cold bias over the study region. Since we employ a broader observational database – unlike the aforementioned studies – we compare the absolute biases estimated against OBS ensemble mean with the uncertainty existed among the OBS ensemble members. Our results suggest that the absolute biases of CMIP5 and CX-SA are higher than the observational uncertainty over most of the study region, regardless of the altitude (Fig. 2). The area-averaged time series also depict substantial cold biases for CX-SA and their driving datasets and smaller observational uncertainty (Supp. Fig. 3).

In contrast to OBS, CMIP5 ensemble features a larger uncertainty of 2–3 \(^{\circ }\)C for ONDJF, MAMJ and annual scale over most of the study region and above than 4 \(^{\circ }\)C for JAS over the monsoon dominated plains and westerly dominated HKH ranges (Fig. 3). Such higher uncertainty among CMIP5 datasets during JAS may be linked to their varying skills in simulating the prevailing large scale circulations, particularly the south Asian summer monsoon (Mariotti et al. 2014; Hasson et al. 2016b). A Taylor diagram shows that ACCESS1-0 and CNRM-CM5 feature a maximum difference in their spatial variability and RMSD among all CMIP5 datasets for all seasons, though these datasets roughly agree on their spatial patterns (Fig. 4). Similar is the case with CX-SA, which though similar to their driving datasets in terms of spatial patterns, are surprisingly farther from OBS ensemble mean in case of RMSD and spatial variability (Fig. 4). However, CX-SA features higher inter-datasets agreement relative to CMIP5 during the cold seasons and for JAS and on annual scale it is quite comparable to the OBS uncertainty (Fig. 3). The area-averaged time series over the Himalayan watersheds also exhibit huge uncertainty for CX-SA and CMIP5 datasets (Supp. Fig. 3). The elevation dependence of such OBS uncertainty (ensemble spread) reveals that it is minimum over the foothill plains (low elevation regions) but rises with elevation (Suppl. Fig. 4). CX-SA uncertainty for JAS also rises with elevation but for the rest of year such case is revealed only for the region above 1000 m asl. The CMIP5 uncertainty, being higher than that of the CX-SA, reveals negligible variation along the altitude in all seasons.

Fig. 4
figure 4

Taylor diagram showing comparison of the mean temperature climatology of OBS (green), CMIP5 (blue) and CX-SA (red) ensemble datasets and their ensemble means

Full size image

Lower CX-SA ensemble spread might have resulted from downscaling five different CMIP5 GCMs (ACCESS0-1, CCSM4, CNRM-CM5, GFDL-CM3 and MPI-ESM-LR) with a single RCM of CCAM (Fig. 4), implying that the effect of distinct lateral boundary conditions can substantially be modulated by RCMs based upon their chosen physics and experimental setups. Such modulation is further depicted from distinct outputs of CX-SA datasets (MPI-ESM-LR-CCAM, MPI-ESM-LR-REMO) that are obtained by downscaling the same GCM with different RCMs. For instance, the MPI-ESM-LR-REMO simulates higher cold bias during cold seasons and gets closer to OBS ensemble mean during JAS relative to MPI-ESM-LR-CCAM. On the other hand, the effect of lateral boundary conditions, though apparently small and mainly in terms of spatial biases, can be noted from CX-SA datasets that are obtained by downscaling different GCMs with the same RCM. For instance, CNRM-CM5-CCAM and GFDL-CM3-CCAM reduce cold but increase warm biases relative to their driving GCMs as compared to the rest of CCAM-downscaled datasets (Fig. 5). These findings suggest that the skill of fine scale experiments is sensitive to both the correctness of driving datasets as well as the structural setup of RCMs, along with suitability of their internal physics, though the extent of their influence on the downscaled fields may vary. Since RCM resolution and physics can substantially modulate the effect of lateral boundary conditions (Syed et al. 2014), their optimum choice seems to be more important for ensuring the added value in the fine scaled climatic fields.

Fig. 5
figure 5

CX-SA mean temperature biases for each dataset against the OBS ensemble mean. Negative (positive) scale refers to cold (warm) biases. Stippling indicates where CX-SA dataset bias is higher in magnitude than its driving CMIP5 dataset, featuring either the same sign (hollow circle) or the opposite sign (filled circle)

Full size image

4.2 Fidelity in climatology and seasonality of precipitation

Precipitation climatologyThe OBS ensemble mean suggests that precipitation is mainly received in JAS over the Himalayas and plains while during the rest of the year (ONDJF and MAMJ) it mainly occurs over the HKH ranges, where its portion over the Karakoram is higher in MAMJ than in ONDJF and JAS, respectively (Fig. 6). This distribution pattern is qualitatively similar to that of the high-altitude observations (1995–2012) from the UIB, showing that the maximum precipitation occurs during spring season under the westerly disturbances (Hasson et al. 2017). However, a quantitative comparison of the employed globally or regionally constructed observational datasets (Supp. Fig. 1) against 27 stations (122–2200 masl) suggests the underestimation of the south Asian summer monsoon rainfall and of annual precipitation up to 15% (Table 2, Supp. Fig. 2). Such underestimation also holds true for the high-altitude annual precipitation as reported by several field campaigns (Wake 1987; Winiger et al. 2005; Hewitt 2011; Soncini et al. 2015). We can clearly see from Fig. 6 that the northward extension of maximum precipitation band is constrained to over the lower HKH watersheds in MAMJ due to underestimation of high-altitude precipitation over Karakoram and eastern Hindukush. Overall, ME is negligible for ONDJF and MAMJ and around -50 mm for JAS whereas MAE is \(\le\) 100 mm across the seasons. The goodness-of-fit indices of r, R\(^2\), d mostly range between 0.7 and 0.9 while NSE and RS range between 0.51 and 0.71. The RMSD is around 100 mm for the cold seasons while it is above 150 mm for JAS. It is notable that the general underestimation of observed precipitation by the OBS ensemble mean and their medium-level agreement revealed by goodness-of-fit statistics may improve given the interpolated precipitation dataset are adjusted for elevation.

Fig. 6
figure 6

Same as Fig. 2, but for precipitation

Full size image

Both CMIP5 and CX-SA ensemble means apparently overestimate OBS ensemble mean precipitation mainly over the Himalayan watersheds, however such overestimation is not fully valid in view of the underestimation of high-altitude precipitation in OBS datasets (Fig. 6, Supp. Fig. 3). Over the elevated regions of HKH, wet bias of the CX-SA ensemble mean (against OBS ensemble mean) is higher than wet bias of the CMIP5 ensemble mean in all seasons. On an annual scale, however, such wet bias is equal or higher in magnitude to the OBS ensemble mean. Higher precipitation in CX-SA than in CMIP5 most probably results from enhanced orographically-induced precipitation, owing to relatively finer topographic representation in RCMs. Analyzing the resolution sensitivity on the south Asian summer monsoon region, studies (Wehner et al. 2014; Johnson et al. 2015) have also suggested that increasing resolution has resulted in a slight increase in precipitation. Along the foothills, CX-SA and CMIP5 ensemble means underestimate monsoonal (JAS) precipitation. In contrast to CMIP5, CX-SA ensemble mean underestimates precipitation over plains during ONDJF and MAMJ.

The observational uncertainty is 10–20% of the OBS ensemble mean over the plains, which is enhanced along the foothill belts and then further up over the high-altitude regions, with its maximum over the western Tibetan Plateau in ONDJF (Fig. 7). In contrast, CMIP5 uncertainty approaches twice the OBS ensemble mean over the Indus plains during the cold seasons and on annual scale. CX-SA uncertainty is comparable to that of CMIP5 over high-altitude regions but to OBS uncertainty over the Indus plains. Overall, from the area-averaged time series over the Himalayan watersheds it is shown that CX-SA uncertainty is relatively lower than that of CMIP5 (Supp. Fig. 3). While comparing uncertainty against the elevation, we note that OBS uncertainty is minimum at lower elevations that increases with altitude (Suppl. Fig. 4). Same is true for CX-SA, which is further consistent with the findings of Ghimire et al. (2015) over the whole HKH. In contrast, CMIP5 uncertainty is highest at the lowest elevations and decreases at high-altitudes. Overall, CX-SA feature up to 50-100% higher uncertainty over the elevated regions but lower uncertainty over the Indus plains relative to their driving datasets.

Fig. 7
figure 7

Same as of Fig. 3, but for precipitation

Full size image

Individual OBS datasets exhibit a wider spatial variability in JAS relative to other seasons. However, all OBS datasets feature similar spatial pattern (r is around and above 0.9) and RMSD (roughly equi-distant) against the OBS ensemble mean, except CMAP for the cold seasons and subsequently on annual scale (Fig. 8). Interestingly, the remotely sensed datasets are not significantly different than the interpolated datasets. It is clear from Fig. 8 that CMIP5 ensemble though features a larger spatial variability as compared to OBS and CX-SA but such spread is mostly within the OBS uncertainty for JAS followed by on annual scale and cold seasons. In terms of RMSD, CMIP5 ensemble members and their mean are closer to OBS as compared to CX-SA where ACCESS1-0 and EC-EARTH are within the OBS uncertainty for all seasons while CNRM-CM5 and GFDL-CM3 for JAS. Surprisingly only in few instances, CX-SA datasets improve the spatial correlation of their driving datasets with OBS. Further, five CX-SA datasets that are downscaled with the same RCM (CCAM) closely agree with each other during all seasons, besides the fact that their driving GCMs feature huge differences. Such a close agreement indicates overthrowing the influence of lateral boundary conditions by the employed RCM. Moreover, the only added value in these CCAM downscaled datasets relative to their driving datasets is a slight improvement in the spatial correlation against the OBS ensemble mean in few instances; contrarily the spatial variability and RMSD became even worst. Similar case is true for the MPI-ESM-LR GCM, for which the spatial correlation is improved in the CX-SA datasets, which are downscaled by two different RCMs (REMO and CCAM). On the other hand, EC-EARTH-RCA4 dataset exhibit the only added value of improved spatial variability during JAS, performing lesser in the rest of aspects and seasons relative to the EC-EARTH driving GCM. These examples suggest that RCMs add their own uncertainties to those associated with the lateral boundary conditions.

Fig. 8
figure 8

Same as Fig. 4, but for precipitation

Full size image

The improved pattern in CX-SA relative to CMIP5 suggest a wet bias over the elevated regions and a dry bias over plain, where such behavior is common among all CX-SA datasets, owing to similar but detailed topographic representation in employed RCMs (Fig. 9). Further, CX-SA datasets are farther from OBS ensemble mean in absolute terms than their CMIP5 driving datasets over high-altitude regions.

Seasonality of precipitation Our results suggest that RE is generally underestimated by CMIP5 ensemble mean relative to OBS ensemble mean (here calculated for APHRO, CMAP and GPCP only) and by CX-SA relative to both CMIP5 and OBS ensemble means (Fig. 10). This underestimation suggests the inability of climate model experiments in adequately simulating the overall statistics of the annual cycle of precipitation and thus its seasonality over the Himalayan watersheds. Nevertheless, CX-SA and CMIP5 generally feature a low level agreement with the OBS ensemble mean on a spatial pattern of highly concentrated precipitation regime (higher RE) over the monsoon-dominated and lower Indus plains while a less concentrated/intermittent (lower RE) precipitation regime over elevated regions. Subsequently, the climate model experiments also cannot reproduce the higher SI over the western Himalayas and over the belt along its foothills. For the most of domain, climate model experiments underestimate SI against OBS ensemble mean. Looking into the seasonality indicators from each individual dataset (Fig. 11), we note that CMIP5 datasets overall perform better for RE relative to CX-SA datasets. For SI, CX-SA mostly perform similar to that of their driving datasets, featuring lower spatial variability and correlation but high RMSD than OBS (Fig. 11). These findings suggest that the overall ability of CX-SA in reproducing the seasonal cycle of precipitation is relatively poor than of their driving GCMs.

Fig. 9
figure 9

Same as of Fig. 5, but for precipitation

Full size image
Fig. 10
figure 10

Annual precipitation, relative entropy (RE) and seasonality index (SI). Note: here OBS refers to ensemble mean of APHRO, CMAP and GPCP observational datasets

Full size image
Fig. 11
figure 11

Same as Fig. 4, but for RE (left) and SI (right)

Full size image

4.3 Future climatic uncertainty

Both CX-SA and CMIP5 ensembles means uniformly suggest warming (4–6 \(^{\circ }\)C) over most of the study domain under RCP8.5 by the end of 21st century (2061–2095), where the CX-SA ensemble mean suggests roughly higher JAS warming (by 1–3 \(^{\circ }\)C) mainly over elevated regions (Fig. 12). For precipitation, CMIP5 projects a decrease in ONDJF and MAMJ precipitation over the western Himalaya and foothill plains and in monsoonal precipitation over the western Karakoram and eastern Hindukush. CX-SA, however, projects a decrease in monsoonal precipitation over Jhelum and UIB but for the rest of the year an increase over the western Tibetan Plateau, Karakoram and western Himalaya (Fig. 12). The CX-SA generally suggests higher projected increase in temperature and precipitation than CMIP5. Nevertheless, the projected changes in precipitation where significant are quite uncertain as the inter-model disagreement for the projected change is equal or higher than the magnitude of change.

Fig. 12
figure 12

Mean future change from CX-SA and CMIP5 ensembles projected under RCP8.5 scenario during the period 2061–2095. Stippling indicates when CV of the future change ensemble is above 0.5 for Tavg and above 1 for P

Full size image

We note that the inter-model disagreement of CMIP5 for the simulated historical temperature (1971–2005) is equal or higher in magnitude than their projected warming over most of the study region (Fig. 13). In contrast, CX-SA projected warming is higher in magnitude than their inter-model disagreement during historical period. This implies that the varying performance of CMIP5 datasets hinders the robustness of their projected warming, which is not the case with CX-SA. For precipitation, the inter-model disagreement for the historical period is higher than the mean future changes projected by both CMIP5 and CX-SA, implying uncertainty of future changes in precipitation over the study region—regardless of the observational uncertainty. Further, comparison of mean future change ensemble against the OBS uncertainty suggest that it does not effect the robustness of projected warming for both CMIP5 and CX-SA but certainly for their projected precipitation changes over the most of study region (Fig. 14).

Fig. 13
figure 13

Ratio of the mean ensemble future change to the intermodel spread during the historical period. Stippling shows where CV of the historical climate ensemble exceeds the CV of future change ensemble

Full size image
Fig. 14
figure 14

Ratio of the mean ensemble future change to the OBS uncertainty. Stippling shows where CV of OBS ensemble exceeds the CV of future change ensemble

Full size image

Most importantly, we have found that the ensemble mean future changes over the Himalayan watersheds as projected by the CMIP5 and CX-SA are considerably smaller than their ensemble mean offsets (calculated against OBS ensemble mean) during the historical period for both temperature and precipitation (Fig. 15). However, the opposite is true over the Indus plains. The negative ratios for temperature over the Himalayan watersheds in Fig. 15 indicate that both CMIP5 and CX-SA simulate cold bias during the historical period but project warming for the future period. Similarly, negative ratio for the precipitation suggests that the sign of historical bias differs from that of the future change.

Fig. 15
figure 15

Ratio of mean future change to the mean bias against OBS ensemble mean for CX-SA and CMIP5 datasets. Here, negative values indicate where the sign of future change is opposite to the sign of historical bias while positive values indicate that future change and historical bias observe the same sign. Stippling indicates where CV of the historical bias ensemble is higher than the CV of future change ensemble

Full size image

The behavior of greater offset than that of projected changes from the ensemble means is also well agreed among the individual datasets from CMIP5 and CX-SA ensembles. Figure 16 shows that the majority of individual datasets feature greater biases than their projected changes in temperature over the HKH watersheds while around all datasets suggest the same for precipitation roughly over the whole domain. The area-averaged time series over the Himalayan watersheds also reinforce such findings (Supp. Fig. 3). Hence, low fidelity of these climate modeling experiments earns less reliability for their simulated future changes projected under extreme-end warming scenario of RCP8.5 by the end of 21st century (and thus for changes projected under RCP6.0, RCP4.5 and RCP2.6 within 21st century) over the Himalayan watersheds. These findings raise a caution for those impact assessment studies deriving climate change information from such low fidelity experiments, emphasizing on improving their fidelity.

Fig. 16
figure 16

majority model agreement for the greater historical bias (offset from OBS ensemble mean) than the magnitude of projected change for CX-SA and CMIP5

Full size image

5 Summary and conclusions

Our fidelity assessment metrics are not exhaustive and merely focus on mean climatology and seasonality, though the results from impact assessment models can also be largely influenced by extremes (Dosio et al. 2015). However, we speculate that proving fidelity of the analyzed modeling experiments for extremes will be more challenging. Therefore, we have confined the fidelity assessment to the considered skill metrics.

We found that the CX-SA experiments feature substantial cold biases, which are higher than those observed for their driving CMIP5 datasets. Similar case is observed for the wet precipitation biases. Surprisingly, the substantial cold and wet biases are consistent across the GCMs vintages (Syed et al. 2009; Islam et al. 2009a) and their fine-scaled derivatives (Islam et al. 2009b; Kulkarni et al. 2013; Syed et al. 2014; Rajbhandari et al. 2015), regardless of their diversified representations of region-specific geophysical features (from coarser to 0.44\(^{\circ }\) resolutions) and variety of applied physics options. A consistent low fidelity of the climate modeling experiments over the complex terrain of the Himalayan watersheds suggests almost negligible improvement due to the extensive development of the climate models in recent decades, where the high-resolution simulations generally seem to contribute to the uncertainty rather to the skill of their driving datasets. Wehner et al. (2014) have also suggested that the high resolution simulation though generally produce more precipitation, of course with some spatio-temporal exceptions, however, such an increase is neither always realistic nor it always outperforms the well-tuned coarse resolution model simulations. Thus, an extensive model configuration regarding the representation of crucial fine scale features along with the relevant physics options is utmost necessary in downscaling experiments. The RCMs’ ability to modulate the influence of lateral boundary conditions reinforces a well-configured downscaling experiment setup.

The observational uncertainty over the study region challenges the climate models to prove their fidelity for the retrospective climate reconstruction. This is particularly true for the observed precipitation, which is underestimated (Immerzeel et al. 2015). This requires a comprehensive monitoring of high-altitude precipitation and particularly snow within the Himalayan watersheds. These high-altitude point observations can be used to extensively validate the spatially complete satellite based precipitation products, particularly those that yield solid moisture component separately such as the Global Precipitation Measurement (GPM). The validated precipitation products are more suitable reference datasets for the climate model fidelity assessment as well as for direct use in the impact studies. It is to note that although OBS uncertainty in precipitation is generally higher, for most of the region it is still smaller than the magnitude of CMIP5 and CX-SA biases. On the other hand, the observational uncertainty in temperature is so small that it neither influences much the climate model fidelity assessment nor is it highly likely to feature a greater offset of up to 10 \(^{\circ }\)C from the truth, as revealed from the validation of OBS ensemble mean against the station observations. Thus, a consistent behavior of substantial cold biases over the Himalayan region in global to regional scale climate simulations needs to be investigated in detail.

The robustness of mean future warming (4–6\(^{\circ }\)) projected by CMIP5 is not influenced by the observational uncertainty and the inter-dataset disagreement for the projected change (2061–2095) but certainly by higher disagreement of CMIP5 datasets during the historical period (1971–2005), which however is not the case with CX-SA. On the other hand, minute changes projected in precipitation by both CMIP5 and CX-SA are smaller in magnitude than the observational uncertainty, and the inter-dataset disagreements for the historical period as well as for the future change, implying uncertain future wet/dry conditions over the study region. Such finding indicates that regardless of high uncertainty of observed precipitation over the study region, projected wet/dry conditions are still uncertain due to a varying skill of the climate modeling experiments in reproducing the observed climate as well as due to their disagreement on the projected changes. Most importantly, higher infidelity (biases against OBS ensemble mean) of CMIP5 and CX-SA experiments than their projected warming and wetting/drying under extreme-end warming scenario of RCP8.5 by the end of 21st century (and of RCP6.0, RCP4.5 and RCP2.6 within 21st century) make these projections highly uncertain for the Himalayan watersheds of Indus Basin.

We have found no single best experiment in terms of the chosen metrics. Under such a situation, several studies typically attempt to weigh the skills of partially skillful models, in order to present robust findings. However, we postulate that due to structural limitations of climate models, smaller sample size of their experiments and considered non-exhaustive objective/subjective criteria-based skill metrics, it is difficult to accurately weight the models according to their skill/fidelity. As a result, studies merely focusing on the south Asian summer monsoonal precipitation have come up with different sets of ’best’ or ’reasonable’ models, which do not necessarily agree with each other on their projected changes (e.g. Sharmila et al. (2015); Jena et al. (2015); Hasson et al. (2016b)). Similarly, even though the models feature high fidelity in terms of skill metrics, their closeness to the reality during the historical period does not guarantee their convergence on the projected changes for the future period under the same forcing (Tebaldi and Knutti 2007). In this regard, we suggest the assessment of both the fidelity as well as future projections in qualitative terms, for which employing a full set of available models, rather than focusing on few partially skillful models, is a more plausible choice (Chiew et al. 2009). Thus, our presented majority agreement may serve as a best indicator of the robustness or uncertainty for the qualitative projected changes in terms of chosen metric.

We argue that the presented low fidelity of the fine scaled CX-SA and their driving CMIP5 datasets along with the associated future climatic uncertainties earn less reliability for their projected changes, at least in quantitative terms. The low fidelity of CX-SA experiments further indicates that their regional scale resolution is still too coarse to adequately represent the tremendous diversity of the hydro-climatic features over the Himalayan watersheds, collectively determined by prevailing large-scale circulation modes at their extreme margins, their interactions and their modulation by the local physio-geographical features, mainly the extensive cryosphere and the complex HKH terrain. Hence, a paradigm shift of exploiting the potential of meso-to-local scale instead of the regional scale climate models (e.g. weather research forecast WRF etc.) in terms of range of microphysics, resolution and convective closure, is mission-critical in order to improve the fidelity of hydroclimatic simulations over the Himalayan watersheds. However, performing meso-to-local scale climate simulations requires a multi-model chain, which is computationally expensive and may also prone to the structural limitations of the employed models. In this regard, applications of the global-to-local scale solo performance climate models with a capability of either telescopic zooming over a particular region (e.g. Laboratoire de Meterologie Dynamique—LMDz), or of the straightforward regional-to-local refinement over arbitrary locations (e.g. Icosahedral non-hydrostatic ICON) are the best alternates.