The IMD dataset is produced by the India Meteorological Department for the whole India. The time period is 1951–2007 and the spatial resolution is 1∘×1∘. The data are interpolated from gauge measurements by using the Shepard method . compare the IMD dataset with the Variability Analysis of Surface Climate Observations (VASClimo) dataset and conclude that the IMD dataset is more accurate in terms of spatial variation. The IMD dataset has been extensively used in climate related research and applications, such as validation of climate models and monsoon variability and predictions .

The number of used gauges varies during the period as well as spatially across the region. The average number of gauges per grid point is 2.99 ranging from 0.2 to 4.4 . Spatially, more gauges are used in the central south; less gauges near the borders of India and in the northern part. No gauge measurements are available near the latitude of 35.5∘ N and northward.

The APHRODITE (Asian Precipitation – Highly Resolved Observational Data Integration Towards Evaluation of Water Resources) dataset is interpolated by the Sphere map method based on data collected at 5000–12 000 gauges . The interpolated parameter is the precipitation anomaly or ratio, instead of the precipitation amount . Elevation corrections are considered by a weighting function, which is based on the angular distance when considering topography . The dataset covers Asia over the period of 1951–2007. Different versions of the APHRODITE dataset have been used to determine Asian monsoon precipitation change, hydrological modeling , verification of high-resolution model simulations and satellite precipitation estimates . In this research, we use the latest version (V1101) for monsoon Asia at a spatial resolution of 0.25∘×0.25∘ . The APHRODITE dataset uses the largest number of gauge observations among interpolated products, and is believed to be one of the most realistic precipitation datasets for Asia .

The ERA-Interim dataset is the precipitation product of ERA-Interim , which is a spatially and temporally complete dataset of multiple climate variables at high spatial and temporal resolution. The data we use here are on a Gaussian grid (with a resolution of 0.7×0.7∘ at the Equator) with a 3 h time resolution, and aggregated to daily time step. ERA-Interim is a global atmospheric reanalysis dataset produced by the ECMWF (European Center for Medium-Range Weather Forecasts)

The next generation reanalysis, ERA5, featuring a higher horizontal resolution (∼31 km) and a 10-member ensemble approach for uncertainty estimates, is released by the end of 2017. See, e.g., http://www.ecmwf.int/en/newsletter/147/news/era5-reanalysis-production (last access: 1 May 2015).

The WRF dataset is generated by using a regional climate model, the Weather Research & Forecasting Model (v3.7.1). The climate model is a limited-area, non-hydrostatic, primitive-equation model with multiple options for various physical parameterization schemes. The model has been used in climate simulation in Asia and other areas . Here we use the Thompson scheme for microphysics, CAM for short- and long-wave radiation, the Noah Land-Surface scheme, Mellor–Yamada–Janjic TKE for the planetary boundary layer and Kain–Fritsch (new Eta) for convection. The model is forced by 6-hourly ERA-Interim reanalysis data. To avoid error at boundary edges and to facilitate further hydrological modeling work, we set up the model at a very large domain (59–91∘ E, 9–46∘ N). The spatial resolution is around 16 km, where topography and land use are aggregated from data with an accuracy of 10 m. They are preprocessed by using the WRF Preprocessing System (WPS). We divide the atmosphere into 30 vertical layers with model top pressure 50 hPa. The height of the lowest model level varies between 15 and 27 m depending on the surface pressure. The whole simulation period is from 1979 to 2007, and the period 1979–1980 is used as model spinup. Due to the long model running time, we restart the model around every 5 years. Model setup is summarized in Table  and the whole setting is a file in the Supplement.

Rain gauge Bhuntar lies in a valley at a small town in the state of Himachal Pradesh, India (Fig. ). The Bhuntar gauge is only 400 m down the confluence of the Parvati River with the Beas River. The altitude of the gauge is 1080 m a.s.l., and both precipitation and discharge data are used in this study. Annual precipitation is 921 mm year-1 based on data from 1981 to 2007, with most rainfall in July and August. Temperature is rarely below 0 ∘C, and only minimum temperature is occasionally below 0 ∘C in winters. The precipitation data have been used in hydrological modeling research for the Beas Basin .

Three discharge series are selected to cross validate water balance. They are respectively Pandoh (downstream), Bhuntar (middle stream) and Manali (upstream). These stations are operated by Central Water Commission regional office in India. The catchments are located in the Beas Basin, which is a main tributary of the Indus River in northern India (Fig. ). The catchments are nested from upstream to downstream. The purpose is to reflect precipitation data at various elevations within a hydrological scale. Runoff is considerably influenced from glacier melting . According to the 0.5 km MODIS-based Global Land Cover Climatology by the USGS Land Cover Institute (https://landcover.usgs.gov/global_climatology.php, last access: 1 May 2017), coverage of snow and ice is 16 % in the Pandoh catchment, 24 % in the Bhuntar catchment and 21 % in the Manali catchment. The discharge data have been manually quality controlled and missing data are filled by discharge anomaly. Discharge measurements are more qualified than precipitation in the snow and ice dominated area . Therefore, the quality of runoff simulation can infer by the forcing precipitation data. use the WRF-Hydro (v3.5.1) modeling system in the Beas Basin, and they find that the distribution of simulated daily discharge values agrees well with observations, which reversely confirms the precipitation simulations.

The MODIS Global Evapotranspiration Project (MOD16) (http://www.ntsg.umt.edu/project/modis/mod16.php, last access: 1 May 2017) is selected to reveal actual evaporation. The MODIS project is started in 2000, and has a short overlap period with the study period. Additionally, part of the catchments is covered by permanent snow and ice and the sensors cannot work well on this type surface. Therefore, we use annual mean amounts of 2000 to 2013 to reduce uncertainties. The missing ratios of annual mean actual evaporation are 22 % for the Pandoh catchment, 32 % for the Bhuntar catchment and 31 % for the Manali catchment.

The four datasets show similar spatial pattern of mean annual precipitation (Fig. ). The highest precipitation is located at the foothill of the mountains and stretched from southeast to northwest. Visually, the high precipitation belt (the foothills of the mountains and the southeastern corner) is most clearly shown by the WRF dataset. The spatial variability increases from the IMD dataset to the WRF dataset. Their coefficients of variation are respectively 0.5 for the IMD data, 0.6 for the ERA-Interim data, 0.7 for the APHRODITE data and 1.1 for the WRF data. The density curves of mean annual precipitation values in all grid points (Fig. ) and the statistics of the Kolmogorov–Smirnov test (Table ) show the variabilities and the differences among the datasets more clearly.

Both the IMD and APHRODITE datasets are interpolated from observations at gauges. However, the APHRODITE dataset shows a rain belt at the mountains' foothills much better. Additionally, the APHRODITE dataset shows much lower estimates (less than 300 mm year-1) in the northeastern corner. This area is quite high, with mean elevation at 4650 m a.s.l. and elevation ranges from 906 to 7677 m a.s.l. The temperature is -2.35 ∘C of annual mean and as low as -16.81 ∘C in January AphroTemp,. The reason for this low-precipitation area is that the APHRODITE dataset uses more gauges, particularly also observations from Nepal, Bhutan and China . These gauges have undercatch problems, which means rain gauges could only catch part of snowfall due to wind and disturbance. In contrast, the IMD dataset uses only the gauges in the low-valley area of India and extends north by interpolation . Eventually, the APHRODITE dataset has the lowest annual amount, only 61 % of the IMD dataset.

The ERA-Interim and WRF datasets are products with different dynamical models. The ERA-Interim data and the WRF data are similar in terms of annual total amount (ERA-Interim: 718 mm year-1, WRF: 688 mm year-1) and spatial pattern, partially due to the fact that in this area the observations that are assimilated into the data assimilation system are sparse and unevenly distributed. The WRF data are more realistic than the ERA-Interim data due to finer spatial resolution, especially in complex topography areas .

The effects of location and topography are shown in Fig. . The summer precipitation changes dramatically. Over the high flat plateau, precipitation decreases with latitude since the strength of the monsoon decreases with distance from its source. As the monsoon gets closer to the mountains, precipitation starts to increase. As the air parcel is lifted to high elevation, climate gets dry and cold. The winter precipitation occurs mainly along the upslope. The magnitude is also small and decreased along the path of the winter monsoon. The highest precipitation occurs in the windward of the upslope region, but it is 0.5 or 1.5∘ (around 55–110 km) far away from the mountains in summer. analyze a decade of TRMM data and also find the highest annual precipitation is offset by a few 10 s of km south of either high topography or relief. This offset has been found only over tall and broad mountain regions rather than narrow mountain peaks .

The differences among the datasets are more obvious in summer at the mountain foot. The WRF dataset gives much more precipitation (700 mm month-1) in July and August at the mountain foot, almost 2 times that of other datasets (300 mm month-1). This is reported as a moisture bias in summer . It is often cited as orographic bias which describes as strong overprediction of precipitation rates along windward slopes while predicted snowfall lies under measured values along leeward slopes .

The inter-annual patterns are very similar as indicated by high correlations between pairs of datasets, shown in Table . The correlation between the IMD and APHRODITE datasets is the highest, reaching 0.91. The WRF dataset has low correlation with all other datasets. Spatially, the four datasets show a similar seasonal distribution, and the WRF dataset has the highest variability (Fig. ). The intra-annual cycle is also similar as shown in Fig. . The WRF and APHRODITE datasets have respectively the highest and lowest precipitation in summer.

To look at changes over time, we select the Theil–Sen median method to calculate trends due to its robustness and the non-parametric Mann–Kendall test for the significance test. The trend analysis and significance test are done for the areal mean of each month (Fig. ), and every individual grid for summer (Figs.  and ) and winter (Figs.  and ). The figures show an increase in summer precipitation and a decrease in winter precipitation, although both increase and decrease exist in each dataset. Three of the areal mean trends (May by the WRF dataset; June by the IMD and ERA-Interim datasets) are statistically significant at the 95 % confidence level. The spatial distribution of trends in summer precipitation varies a lot. Most decreasing trends of winter precipitation occur in the northern part. Approximately 10 % of grids are significant at the 10 % confidence level.

It is difficult to conclude why northern India of the Western Himalayas shows an increase in summer precipitation. However, find the same increasing monsoon precipitation in northern India but decreasing monsoon precipitation in central Asia. They use a series of climate model experiments, and conclude that such pattern is a robust outcome of a slowdown of the tropical meridional overturning circulation, which could be attributed mainly to human-influenced aerosol emissions. The trends will continue and become more significant with time if greenhouse gas emission continues as usual. Such trends would lead to strong negative mass balance conditions of glaciers, which is discussed in the next section.

To compare the gridded datasets with measurements at the Bhuntar gauge, we extract the time series at the nearest point to the Bhuntar gauge. We look at annual precipitation (Fig. ), monthly anomaly (Fig. ) as well as extreme precipitation, i.e., annual maximum daily precipitation (Table ). As shown in Fig. , all gridded datasets are comparable with the Bhuntar gauge data. The interpolated datasets, IMD and APHRODITE are quantitatively closest to the Bhuntar gauge data. The ERA-Interim and WRF datasets generally give 2 or 3 times higher precipitation than the Bhuntar gauge data. Figure  shows the differences are mainly from March to July in the WRF dataset, and from July and August in the ERA-Interim dataset. In addition, the WRF dataset shows large variations from February to June, and the ERA-Interim dataset shows large variations in July and August. Table  shows the statistics of annual maximum daily precipitation. Notably, the WRF dataset gives the closest estimate to the Bhuntar data in five quantiles, and the APHRODITE dataset gives the best estimate of the maximum precipitation over the whole period.

The annual actual evaporation from MODIS data is 614 mm year-1 at the Pandoh catchment, 639 mm year-1 at the Bhuntar catchment and 649 mm year-1 at the Manali catchment. The values are too high compared with 64 mm year-1 at the Pandoh catchment for the period from 1990 to 2004 calculated by using potential evaporation, mean and maximum temperature. The Pandoh catchment covers the lower and middle parts, and should have the highest evaporation due to warm climate among three catchments. The MODIS data are not qualified at the catchments and at small catchment scales for the study period.

The precipitation and runoff relationship is shown in Fig.  as accumulation of monthly precipitation and runoff. Though the lines have different slopes, but they share very similar linear relationships. They are consistent in terms of temporal changes. Errors are systematic within each dataset. Runoff is generally less than precipitation due to evaporation loss. However, runoff could possibly exceed precipitation at glacierized catchments due to glacier melting. In the Manali catchment, runoff is much more than precipitation. In the Bhuntar catchment, only the ERA-Interim data show less runoff than precipitation. All datasets show less runoff than precipitation in the Pandoh catchment. Precipitation is definitely underestimated in higher-elevation areas, especially in the Manali catchment. reconstruct annual mass balance of Chhota Shigri glacier since 1969. The Chhota Shigri glacier lies in the Western Himalayas, India and it is representative in terms of mass balance for the Western Himalayas glaciers . The mass loss rates are 0.36±0.36 for 1969 to 1985 and 0.57±0.36 m water equivalent per year (m w.e. a-1) for 2001 to 2015. The runoff contribution from glacier melting is only 3306 mm within 29 years with assumptions of 20 % glacier coverage and -0.57 m w.e. a-1.

In the Great Himalayas region, there are many glaciers, and they are key indicators of regional climate change and water resources. Temperature in combination with precipitation controls survival of glaciers. Therefore, we also look at changes in temperature by comparing the temperature results by the same simulation of the WRF precipitation dataset for the first and last 5 years, namely 1981–1985 and 2003–2007. We skip the trend analysis and significance test, because it is already well known that temperature has been increasing quickly in the Great Himalayas region since the 1980s . Temperature is well measured and simulated. Therefore, there is no need to go through many datasets. We are particularly interested in temperature at the equilibrium line altitude (ELA). As the slope of the regression lines shown in Fig. , the WRF model is able to reproduce the lapse rates. Between the two 5-year periods, temperature increases by 0.91 ∘C in winter and by 0.26 ∘C in summer. Such changes lead to an increase in the elevation of the freezing point (0 ∘C) of 125 m in winter and 32 m in summer. As shown in Sect. , precipitation overall decreases in winter. In combination with increasing temperature, this is an unfavourable condition for the glaciers with less accumulation and faster melting. Moreover, the area between 4900 m a.s.l., which is the equilibrium line altitude (ELA) of the Chhota Shigri glacier seeFig. 2, and 5200 m a.s.l. is large. Therefore, as the climate gets warmer, the ELA will further move up. Such a nonlinear characteristic of elevation distribution results in a potential large reduction in the accumulation area and small storage buffer of permanent snow and ice.