Evaluation and projection of a snow coverage rate over the Upper Yarkant River Basin, China

Among the five major spheres of the climate system, namely cryosphere, atmosphere, hydrosphere, biosphere and lithosphere in the Earth system, the cryosphere is an important one, having the second-largest heat capacity next to the marine (Zhang 2005). Snow cover is a key component of the cryosphere – the low thermal conductivity of snow insulates the underlying ground from atmospheric temperatures and the relatively high albedo of snow alters the ground energy fluxes (Armstrong & Brun 2008; Christian et al. 2013). Consequently, these properties have important implications for the energy balance and water balance between the world's land and atmosphere (Zhu & Dong 2013; Xia & Wang 2015). For example, snowmelt can result in surface radiation budget reduction, lower surface air temperature and the effects on local water cycle, which leads to the anomalies of atmospheric circulation and water resource balance in different periods (Zhang 2005; Vavrus 2007). The annual and seasonal variations of snow cover are significant in the cryosphere, being sensitive to changes in temperature. The snow cover responds to anytime- and space-scale climate change. Snow cover not only responds passively to climate change, but also has significant and lasting effects on physical, biological and social systems. The high reflectivity and hydrological effect of snow cover affect the climate by changing the surface energy balance, water cycle and atmospheric circulation (Kong & Wang 2017). With the climate change, climate variability modeling and analysis have always been a research hotspot, and there are some successful works. For example, the hydrologic cycle is accelerating due to global warming. These changes can affect the water resource system, which is especially fragile in the arid areas of northwestern China, where rivers are mainly fed by meltwater from snow and ice (Chen et al. 2014). Total precipitation is a key factor in climatological studies. Different research methods were used to analyze the spatial and temporal variabilities of precipitation in Iran. The combination of principal component analysis and cluster analysis was used for homogeneous precipitation area detection, which reveal the importance of time scale in the detection of homogeneous precipitation sub-zones. The correlation decay distance and the coefficient of variation of annual precipitation of 42 stations were computed in Chaharmahal and Bakhtiari province to analyze the precipitation variability (Arab Amiri & Mesgari 2017, 2019; Arab Amiri et al. 2017; Gocic & Arab Amiri 2021). Recently, studies on snow cover change appear to be particularly crucial with the global climate warming. Some scholars have carried out the study of snow cover change under the background of climate change. The mean snow cover in the Swiss Alps showed a gradual increase before the 1980s, then was followed by a significant decrease toward the end of the century (Laternser & Schneebeli 2003). In Switzerland, the strongest relative reduction of mean snow water equivalent (SWE) in winter <1,500 m amounts to 40–80% by mid-century relative to 1971–2000, and at an elevation of 2,000–2,500 m, the reduction amounts to 10–60% by mid-century (Christian et al. 2013). Ma et al. (2011) used the Coupled Model Intercomparison Project 3 (CMIP3) to predict the decreasing trend of the SWE in Eurasia over the next 50 years (Ma et al. 2011). Xia & Wang (2015) selected 13 CMIP5 models to predict the future snow cover fraction in Eurasia based on the simulation by remote sensing data, which showed a decline of the snow coverage rate in the future (Xia & Wang 2015). Nevertheless, previous studies mainly focused on the regions of Eurasia, the Northern Hemisphere, East Asia and so on (Wang & Ding 2011), while few research methods have been performed in alpine mountain areas, which are more sensitive to climate change.

The Yarkant River that originates from the North Karakoram on the Tibetan Plateau has inadequate water resources due to its intense evaporation and dry climate. It is a major tributary of the longest continental river – the Tarim River of north-western China, whose water source comes from snowmelt and glacial meltwater in upstream mountains (Zhang & Zhou 1991). In this region, glacier and perennial snow are widely distributed; in addition, the snow cover has clear seasonal variation characteristics and interannual and interdecadal changing trends (Wei et al. 2018). The snow hydrology of the basin has unique mountain regional characteristics; besides, the anomaly of this snow cover is closely related to floods and drought in the downstream. According to the statistics, snow cover in the Northern Hemisphere decreased at a rate of 1.4% (10a) from the 1950s to the 21st century (IPCC 2013). Consequently, future changes and anomalies of the snow cover in the Upper Yarkant River Basin are of major concern owing to global warming.

Based on comprehensively evaluating the ability of snow cover simulation in the Yarkant River Basin using the CMIP5 model, the main objective of this research is to project future changes in the snow coverage rate over the Yarkant River Basin under the three different representative concentration pathways (RCP2.6, RCP4.5 and RCP8.5). Previous studies on the variation of snow cover under the climate change background mainly focused on the regions of Eurasia, the Northern Hemisphere, East Asia, Qinghai-Tibet Plateau and so on (Zhu & Dong 2013; Liu et al. 2014; Xia & Wang 2015; Yang et al. 2017), while few research works have been performed in small-scale alpine mountain areas, which are more sensitive to climate change. Research works on snow cover in the Yarkant River Basin mainly focused on the vertical distribution and seasonal changes (Jia et al. 2014; Wei et al. 2018). The paper first focused on the changes of the future snow cover rate under the three different RCPs (RCP2.6, RCP4.5 and RCP8.5) using CMIP5 model data, in order to provide a beneficial guidance for flood control and drought relief in the downstream of the alpine basin.

The Upper Yarkant River Basin is located in Xinjiang, China, at 35°27′–38°20′N and 74°27′–78°25′E (Figure 1). Shaksgam River and Taxkorgan River are two main tributaries of the Yarkant River, originating from the Karakoram glacier on the western side of the Karakoram mountains. The river is 660 km long from the source to the Kaqun hydrological station with a drainage area of 50,248 km² which is between 1,448 and 8,569 m above sea level (a.s.l.) (Kan et al. 2018). The Yarkant River Basin lies in the hinterland of Eurasia, far from the ocean (Wang 2010), on the edge of the Tarim Basin, where it is surrounded by Karokoram mountains, Kunlun mountains and Pamir mountains. Because of its peculiar position, the climate is characterized by low precipitation, high evapotranspiration and long daylight hours. The basin contains almost 3,000 glaciers with an estimated ice volume of 660 km³. To be specific, the glacier area of Karakoram mountain accounts for 83.5% as the main distribution area of the glacier; besides, Kunlun mountains occupy 11.5% and Pamir mountains occupy 5.0%. Furthermore, the Yarkant River is a typical snow- and ice-supplied river, as snow and ice meltwater make up approximately 64% of runoff. Obviously, glaciers are crucial water resources in the Yarkant Basin, spreading over almost all the sources of tributaries. Based on the climate characteristics, the seasons can be defined as follows: March to May is considered as spring, June to August is summer, September to November is autumn and December to February of the following year is winter (Wei et al. 2019).

• 1.

  NOAA visible satellite snow product

The NOAA visible satellite snow product, which has been re-gridded to the equal area EASE-grid, is used as the observed data to evaluate the simulation capacity of the CMIP5 snow cover data (Yang et al. 2017). Weekly snow cover data that the satellite of the Northern Hemisphere retrieved were distributed by the National Snow and Ice Data Center, and they provided an extremely useful means of assessing hemispheric snow extent (Li & Wang 2011). The snow cover data have a spatial resolution of 25 km, and the data series are from 1966 to present. Nevertheless, the data in the first few years were abandoned due to large observation errors, and only data from 1993 to 2005 were utilized in the paper.

• 2.

  CMIP5 snow cover data

The CMIP5 models are available for download from the Program for Climate Model Diagnosis and Intercomparison (PCMDI, http://cmip-pcmdi.llnl.gov/index.html) at the Lawrence Livermore National Laboratory. In this paper, we select the CMIP5 models which all have the monthly average snow coverage output under both historical simulation and different future emission scenarios of RCP2.6, RCP4.5 and RCP8.5. Each model includes multiple products: the differences among them are the resolution and whether to consider the carbon cycle, etc. Moreover, the earth system simulation product with high resolution is selected (David & Allan 2000; Masson & Knutti 2011; Zhao et al. 2013). Before analyzing a future snow cover change with simulation results in detail, it is necessary to utilize observational data in order to assess historical simulation results and then select the best model to be used. Eight models that meet this criterion are listed in Table 1.

The key of bilinear interpolation algorithm (BIA) is to carry out a linear interpolation along the x-axis and y-axis separately, and obtain the function value of the awaiting intercalation point.

The general principles of the BIA are as follows (Liu 2013; Bai 2014):

To evaluate the simulation capacity of the CMIP5 models, spatial bias (B), SDR, correlation analysis (R) and the S index between the observation data and the CMIP5 model data are analyzed.

• 1.

  Bias (B)

The seasonal snow coverage rate spatial distribution of the NOAA observed data in 1993–2005 is illustrated in Figure 2. The snow coverage rate of the Yarkant River Basin in winter is highest ranging from 55.16 to 95.05%, and it is about the same in spring and autumn, ranging from 34.03 to 80.80% and 31.05 to 75.31%, respectively; instead, it is lowest in summer ranging from 13.56 to 46.37%. Additionally, the snow coverage rate drops dramatically from southwest to northeast, which is in accordance with the characteristics of the topography of the Yarkant River Basin. The higher the elevation, the greater is the snow coverage rate. The snow coverage rate is the highest in the Shaksgam Valley at the source of the Yarkant River because of the highest altitude and decreases sharply in the basin export. The change rule of NOAA snow cover data is consistent with that using MOD10A2 snow cover data in other research works (Jia et al. 2014; Wei et al. 2018).

Figure 3 shows the snow coverage rate spatial distribution in spring, summer, autumn and winter of the CMIP5 model data, and only the CNRM-CM5 model, GISS-E2-H model, NorESM1-M model and multi-model ensemble mean are listed. Also, the snow coverage rate is listed in Table 2, which demonstrates that the snow coverage rate in spring of the CanESM2 model, Inmcm4 model and NorESM1-M model is closest to that of the observed data, ranging from 49.16–85.07%, 30.82–89.98% and 31.51–80.77%, respectively. In brief, the three models can basically outline a sharply decreasing trend of spatial distribution from southwest to northeast. To be specific, the simulation of Inmcm4 model for a snow coverage rate over the Karakoram main mountain range is slightly smaller than the observed data, and CanESM2 and NorESM-1 models’ snow coverage rate simulation results for the Karakoram mountains are basically consistent with the observed data, which are larger than the observed data of the northwestern Kunlun mountain. In addition, all models simulate poorly in summer and the results are less than the observed data. In summer, the higher temperature leads to a lower snow coverage rate, so the simulation errors in summer are relatively large. It is also proven that the snow cover schemes of CMIP5 models were related to temperature and precipitation, and the simulation of precipitation and temperature of each model all showed a positive deviation (Wu & Wu 2004; Li et al. 2009; Xia & Wang 2015). However, in high-altitude mountainous areas like the Yarkant River Basin, the high simulated temperature is not conducive to the accumulation of snow.

The simulated ranges of a snow coverage rate in autumn of the CanESM2 model, CNRM-CM5 model, GISS-E2-H model, Inmcm4 model and NorESM1-M model are close to the observed range. In autumn, all other models can simulate the falling trend of spatial distribution from southwest to northeast, except the CanESM2 model. Besides, the simulation effect of all models on winter is better than the one of spring, summer and autumn. Among all models, the CanESM2 model, CNRM-CM5 model, Inmcm4 model, MIROC-ESM model and NorESM1-M model simulate best for the snow coverage rate in winter, with simulation ranges of 58.10–90.73%, 16.15–94.62%, 72.48–98.35%, 76.30–93.31% and 82.31–100%, respectively.

In terms of spatial distribution, the simulation of each model basically coincides with the observed data. However, there is a certain error between the simulated value and the observed value, and the simulation results of each model are different. The observed data NOAA, which are visible-light remote sensing snow cover data, are greatly affected by clouds (Zhu & Ding 2007). The Yarkant River Basin is a cloudy alpine area where the influence of clouds leads to observation errors. Second, the topography of the study region is extremely complex, and there are some deviations in the simulation of complex terrain areas for each CMIP5 model (Xia & Wang 2015). As for the multi-model set, it can essentially reproduce the spatial distribution law of the snow coverage rate in all seasons, and the variation range of a snow coverage rate in each season is closer to the observed data than all the other models.

To quantify the ability of each model to simulate a snow cover rate, the monthly correlation coefficients between the snow coverage rate of each model and the observed data were calculated, as shown in Figure 4.

It can be seen from Figure 4 that the correlation coefficient (R) in spring, autumn and winter is basically >0.8, whereas the simulation effect is poor in summer with the minimum correlation coefficient of 0.02. In summer, the CNRM-CM5 model, GISS-E2-H model and multi-model are better than other models, and the R ranges are 0.73–0.86, 0.71–0.83 and 0.73–0.75, respectively. The variation of R between the MIROC-ESM model and the NOAA data is different from other models, which shows abnormal phenomenon in May. The anomalies are related to the snow cover schemes of the MIROC-ESM model. Generally, the correlation between the multi-model set and the observed data is good, which is shown with the red line in Figure 4. Even from July to September, R varies between 0.73–0.76, but the simulation of other models is worse at the same time. The monthly correlation coefficients between models and the observed data are listed in Table 3.

Figure 5 presents the annual change of the SDR between the model simulated values and the observed values. The closer the SDR value is to 1, the closer the spatial change of model simulation is to the observed data. Moreover, Figure 5 proves that the SDR value of each model in summer deviates furthest from 1, and the simulation effect is poorer than the one in the other seasons. In the other seasons, the spatial variation simulated by the CNRM-CM5 model deviates furthest from the observed data. From January to April, the spatial variation simulation of the CSIRO-MK3.6.0 model, GISS-E2-H model and CanESM2 model is closest to the spatial variation of the observed values. From October to December, the SDR of the GISS-E2-H model and Inmcm4 model is closest to 1, and the spatial variation is similar to that of the observed data. In addition, the SDR between the multi-model set data and the observed data ranges from 0.44 to 1.17. Despite that the SDR in August and September is low, at 0.59 and 0.44, respectively, and the SDR in October is 0.71, the SDR in other months is all >0.84. As the red line expresses in Figure 5, the SDR of the multi-model is closer to 1 every month. The monthly SDR between models and the observed data are listed in Table 4.

The S index is a comprehensive index describing the ability of a simulated snow coverage rate in each model, and its monthly variation rule is illustrated in Figure 6.

The S index is generally low in summer, at close to 0, especially in July, August and September. In detail, the simulation capability of the MIROC-ESM model is the worst compared with the other models each month and the variations are different from other models which are related to the snow cover schemes of the MIROC-ESM model. In contrast, the other models have the best simulation capability in March, April, November and December. The simulation ability of the multi-model set is strong each month as the S index reaches 0.91, 0.94 and 0.91, respectively, in February, March and April. Even in January, November and December, it is >0.80, while in September alone, the minimum simulation ability is 0.32.

According to the comparison work carried out by Wu & Wu (2004) and Li et al. (2009) for different snow cover schemes of CMIP5 models, it can be seen that the simulated snow coverage rate of different models, which have different snow cover schemes under the same physical framework of atmospheric forcing and land surface process, varied greatly (Wu & Wu 2004; Li et al. 2009). For example, the NorESM1-M model uses the scheme in CLM4.0, which considers the influence of snow depth and snow density on snow cover. The scheme in CanESM2 and GISS-E2-H models is only related to the SWE which can be converted into snow depth. It is proven that schemes of most models are related to snow depth (Xia & Wang 2015). Moreover, the quality of snow depth simulation mainly depends on temperature and precipitation. The high-temperature simulation of models in high-altitude areas is the reason for the lack of snow cover simulation in summer. The use of different snow cover schemes in CMIP5 models causes simulation deviation.

To sum up, considering the four aspects of spatial analysis, correlation coefficient, standard deviation variability rate and S index, all models have certain simulation abilities for the snow coverage rate in the upper reaches of the Yarkant River Basin. Furthermore, each model has a weak simulation ability in summer, and the multi-model set simulates best in all seasons of the basin. Hence, the snow coverage rate variation of the future in upper reaches of the Yarkant River Basin will be predicted by adopting the average of the multi-model set.

The average of multi-model sets simulated the snow coverage rate of the Yarkant River Basin best. Thus, in order to estimate the future snow coverage rate changes under three different greenhouse gas emission scenarios (RCP2.6, RCP4.5 and RCP8.5), the snow coverage rate of the future (2006–2100) was calculated by the arithmetic mean method of various models.

The deviations of a snow coverage rate during 2006–2100 and 1993–2005 under RCP2.6, RCP4.5 and RCP8.5 scenarios in spring, summer, autumn and winter were calculated. Then, the spatial distributions of deviations in four seasons are outlined in Figure 7(a)–(c).

Under the low-emission scenario of RCP2.6, the snow coverage rate of the whole basin presents a negative difference in summer compared with the average snow coverage rate from 1993 to 2005, and the negative difference in the west of the basin is more obvious than the one in the northeast part. In spring, autumn and winter, the snow coverage rate variation in most areas presents a negative difference. In spring and autumn, it shows a positive difference in some parts of the southeast, while a positive difference appears in some parts of the northwest in winter. Moreover, under the RCP2.6 emission scenario, the negative differences of the whole basin are greater than the positive differences. Some studies showed that precipitation in the arid area of northwest China increased and temperature rose with the climate warming (Chen et al. 2014; Wei et al. 2019) in the northwest and southeast of the study region where the positive difference occurs and the altitude is higher. Under the RCP2.6 emission scenario, although the precipitation increases and temperature also rises, the temperature in the high altitude in spring, autumn and winter is <0°C, the increased precipitation appears in the form of snowfall consequently leading to an increase in the snow coverage rate. Under the RCP4.5 emission scenario, the negative differences of the whole basin are more obvious than that of the RCP2.6 emission scenario; besides, the snow coverage rate in the whole basin presents a negative difference in the spring, summer and autumn. In winter, there are some positive differences in the northwest region, but the positive differences are smaller than those under the RCP2.6 emission scenario. With the increase in greenhouse gas emissions, the snow coverage rate in the whole basin presents negative differences in four seasons under the RCP8.5 scenario and also the negative differences tend to be more significant. Under the same emission scenario, the negative difference in winter is weaker that in the other seasons, as the most prominent negative differences appear in spring and autumn.

The annual trend of the snow coverage rate of the whole basin in all seasons during 2006–2100 is described in Figure 8.

Under the emission scenarios of RCP2.6, RCP4.5 and RCP8.5, the future snow coverage rate predicted by the multi-model set shows a downward tendency each season. In spring, summer and autumn, the rate of snow coverage reduction from 2006 to 2046 is similar; nevertheless, it seems to be clearly different after 2046 due to three greenhouse gas emissions. Under the RCP2.6 emission scenario, the falling trend of snow coverage rate is the gentlest, and it is larger under the RCP4.5 emission scenario. Also, the dropping trend is most obvious under the scenario of RCP8.5. In winter, the snow changing trend of a snow coverage rate in the basin is basically the same under the three gas emission scenarios; additionally, the snow coverage rate in the RCP8.5 emission scenario is the lowest.

The Mann–Kendall (M–K) trend test of a future snow coverage rate from 2006 to 2100 at each grid point of the whole basin was conducted; afterward, we got the Z value of each grid point, which was interpolated to the whole basin by Arcgis10.2. The spatial distribution of a future snow coverage rate Z value calculated by the M–K method under three different emission scenarios is presented in Figure 9.

Under the three emission scenarios of RCP2.6, RCP4.5 and RCP8.5, the snow coverage rate of the whole basin presents a decreasing trend in the future, and the decreasing trend becomes more significant with the rise of greenhouse gas emissions. Under the RCP2.6, the reduction trend is too insignificant to pass the test with a 90% confidence level. Under RCP4.5 and RCP8.5, the reduction trend can pass the trend test with a 99% confidence level.

Through the analysis above, it can be seen that if effective greenhouse gas emission reduction policy is implemented to control future greenhouse emissions at a low level, the snow coverage rate of the study region will not change much in the future. But if the greenhouse gas emissions are not effectively controlled now, the impact on a snow coverage rate may not be significant in the short term (2006–2046), but as time goes by, the impact will become more and more prominent in the second half of the 21th century. The snow cover is an important freshwater supply for the downstream of the study region. Its changes will have a significant impact on the water resource balance and the stability of the ecosystem in the downstream. Therefore, it is necessary to attach great importance to the changes in the snow cover under the greenhouse gas emission scenario and to formulate corresponding greenhouse gas emission reduction and climate adaptation policies in a timely manner.

NOAA remote sensing snow coverage data and CMIP5 models’ snow coverage data were processed by MATLAB. Foremost, we evaluated the simulation ability of the eight CMIP5 models by comparing the spatial distribution between the observed data and CMIP5 models’ data, calculating the correlation coefficient (R), the SDR and the S index. Next, the arithmetic mean method was used to collect the eight CMIP5 models. Finally, we selected the multi-model set to conduct deviation analysis with the historical data at each grid point and adopted the M–K trend to analyze the changing trend of a snow coverage rate from 2006 to 2100. The conclusions are as follows:

• All models can basically present a sharp decrease of the spatial distribution from the southwest to the northeast. The simulation ability of the snow coverage rate in spring, autumn and winter is better than that in summer. In addition, the spatial simulation ability using multi-model data is the best.

• The correlation coefficients (R) between the observed data and CMIP5 models’ data in spring, autumn and winter are basically >0.8, while the simulation effect is poor in summer. Moreover, the correlations between the multi-model set and the observed data are generally good in all seasons, and the R ranges between 0.73 and 0.76 even from July to September; in contrast, the simulation of the other models is worse at the same time.

• The SDR between the multi-model set data and the observed data ranges from 0.44 to 1.17.

Despite that, the SDR in August and September is low, at 0.59 and 0.44, respectively, and the SDR in October is 0.71, while the SDR in other months is all >0.84.

• The simulation ability of the multi-model set is strong each month as the S index reaches 0.91, 0.94 and 0.91, respectively, in February, March and April. Even in January, November and December, it is >0.80.

  In brief, considering the four aspects of spatial analysis, correlation coefficient, standard deviation variability rate and S index, the multi-model set simulates best in all seasons of the basin. Therefore, the snow coverage variation of the future in the Upper Yarkant River Basin is supposed to be predicted using the average of the multi-model set.

• Compared with the average snow coverage rate from 1993 to 2005, the snow coverage rate of the whole basin presents negative differences in most areas and positive differences in local areas under the emission scenarios of RCP2.6 and RCP4.5. Under the RCP8.5 emission scenario, the snow coverage rate in the whole basin shows negative differences in the four seasons and the negative differences become more significant. Under the same emission scenario, the negative difference in winter is weaker than the one in other seasons, as the most prominent negative differences appear in spring and autumn.

• Under the three emission scenarios of RCP2.6, RCP4.5 and RCP8.5, the snow coverage rate of the whole basin presents a declining trend in the future, and the decreasing trend gets more significant with the increase of greenhouse gas emissions. Under the RCP2.6, the reduction trend is too insignificant to pass the test with a 90% confidence level. Under RCP4.5 and RCP8.5 scenarios, the reduction trend can pass the trend test with a 99% confidence level.

According to the analysis of the above, the emission of greenhouse gas has a crucial impact on the changes of snow coverage in upper reaches of the Yarkant River Basin, particularly spring and autumn are affected most obviously. The snow coverage in plateaus and the Alpine region is a valuable source of fresh water supply in the arid area of China; therefore, the snow coverage in upper reaches of the Yarkant River Basin has a great influence on the stability of the ecosystem and the balance of water resources in the lower agricultural irrigation district of southern Xinjiang. As a result, it is extremely important to give great importance to greenhouse gas emissions and implement effective policies to reduce greenhouse gas emissions.

The article only presents the obtained results and does not analyze the causes of the results because the causes are comprehensive and complicated. It is worth mentioning that although the CMIP5 has been developed, there are still some errors between the snow cover simulated by the CMIP5 models and the observed data. How to improve the simulation of snow cover in the CMIP5 models is a challenging task. In addition, the future prediction of the CMIP5 models based on the greenhouse gas emission scenario does not take into account the impact of natural changes. Therefore, the simulation has some uncertainty. Also, the prediction results are uncertain through the simple equal-weighted arithmetic average set method. All the above-mentioned problems need to be further studied in the future.

The authors thank the Program for Climate Model Diagnosis and Intercomparison (PCMDI, http://cmip-pcmdi.llnl.gov/index.html) at the Lawrence Livermore National Laboratory for providing the database used in this study. They also thank the anonymous reviewers and editors of this paper for their very helpful comments and suggestions.

All authors were involved in designing and discussing the study. R.W. undertook the data analysis and drafted the manuscript. L.P. collected the required data. Y.W. and L.P. revised the manuscript and edited the language. C.L. and L.Z. contributed to the setup of the simulations and the write-up of the paper. All authors read and approved the final manuscript.

This research was funded by the National Natural Science Foundation of China (Nos 51569031 and 51469034).

The authors declare no conflicts of interest.

All relevant data are available from http://cmip-pcmdi.llnl.gov/index.html.