Assessing Gridded Precipitation and Air Temperature Products in the Ayakkum Lake, Central Asia

Abstract

We evaluated the performance of gridded precipitation and air temperature datasets near the Ayakkum Lake at the southern margin of Xinjiang, arid central Asia. Statistical measures were applied to assess these climate products on a monthly basis from 2013 to 2018. For monthly precipitation amount, the European Centre for Medium-Range Weather Forecasts Reanalysis 5 shows a good performance among the five products based on most statistical measures, and the China Meteorological Forcing Dataset can also be used as an alternative, especially for estimating the long-term annual mean. For monthly air temperature, WorldClim historical weather data are recommended because of the low mean absolute error, root mean square error and distance between indices of simulation and observation. Better spatial and temporal coverages of in-situ observations are still needed to produce an optimal correction scheme for the mountainous regions of arid central Asia.

1. Introduction

As a typical mountain-basin system in arid central Asia, Xinjiang has experienced a warming and wetting trend during the past decades [1,2,3]. As observed at the surface meteorological stations from 1961 to 2018, the annual mean air temperature has increased by 0.30 °C per decade, and the precipitation amount has increased at a rate of 9.95 mm per decade [1]. In Xinjiang, there is a strong elevation dependency of precipitation amount and air temperature; that is, the mountains usually have lower air temperature and much more precipitation [4,5,6]. The surface water originating from rainfall and snowfall in the surrounding mountains is the critical water resource for the oasis, and plays an important role in regional sustainability in Xinjiang [7,8,9]. To understand the water resources available for the oasis in Xinjiang, it is necessary to accurately assess meteorological conditions in these high-elevation mountains.

The existing meteorological stations in Xinjiang are mostly distributed at low-lying oases, and the surrounding mountains are poorly gauged [10,11,12]. Among all the national meteorological stations (>100 stations) in Xinjiang, there are only three stations higher than 2500 m above sea level [1]. The gridded climate products derived from surface observations and remote sensing may provide a supplement in climate studies, especially for the sparsely gauged mountains, but the accuracy of these products should be carefully treated. As indicated in some assessments [6,13,14,15,16,17], most gridded precipitation and air temperature products can describe the spatial pattern constrained by the topography. It should be noted that the mountainous meteorological stations present are mainly located in the Tianshan Mountains in middle Xinjiang, and the Kunlun and the Altun mountains in the southern margin are not well gauged [1]. Although the Kunlun and the Altun mountains are generally arid and sparsely populated, the mountains generate valuable water resources in southern Xinjiang. Because of the absence of surface meteorological measurements across these mountains, the accuracy of the existing climate products in the southern margin of Xinjiang is not clear.

The Ayakkum Lake is a high-elevation salt lake at the southern margin of Xinjiang. Lying in an intermountain basin between the Kunlun and the Altun Mountains, the Ayakkum Lake is cold and arid. Against a warming background, the lake area has rapidly increased during the past several decades [18,19,20,21]. A lake area assessment from 2001 to 2016 showed an increasing trend of 13 km² per year, which is statistically significant at the 0.001 level [20], and another assessment from 1995 to 2015 presented a higher trend magnitude of approximately 20 km² per year [21]. The specific lake area and increasing trend in different publications vary, which is associated with the natural lake variability (e.g., seasonal and inter-annual variability in response to precipitation, evaporation and runoff) and technical aspects (e.g., spatial uncertainties of different satellite images) [22]. However, the remote sensing-derived area is around or slightly larger than 1000 km², indicating the Ayakkum Lake, instead of the Bosten Lake, may be the largest lake in Xinjiang in recent years. As a mountain lake sensitive to climate change, the Ayakkum Lake is a good study region to understand the changing hydrological processes in a cold and arid climate. There were only a few short-term meteorological records in 1984 and 2011–2012 described in previous publications [23,24], and measurements of less than one year are not sufficient to understand the intra-annual and inter-annual variations of climate in this high-elevation region.

Since 2013, an automatic weather station has been operated near the Ayakkum Lake [25], which provides a platform to assess the accuracy of gridded climate products. In this paper, based on the observed precipitation amount and air temperature data from 2013 to 2018, we assessed several gridded climate products on a monthly basis for this region, which may be useful to understand the changing climate in the mountains of southern Xinjiang.

2. Data and Methods

2.1. Study Area

The Ayakkum Lake is located in the Qimantag Township, Ruoqiang (Qarkilik) County, Bayingolin Mongol Autonomous Prefecture, Xinjiang Uygur Autonomous Region, China (Figure 1). The lake is in the Kumkol Basin surrounded by the Altun Mountains to the north and Kunlun Mountains to the south. The lake area is approximately 1000 km² [18,19,21], and the drainage area is approximately 2.5 × 10⁴ km² [26]. The lake water level elevation is approximately 3880 m [21], and the water volume is approximately 5.5 × 10⁹ m³ [26]. The annual mean air temperature is approximately 0 °C, and the annual precipitation amount ranges from 100 mm to 200 mm; the annual evaporation amount is between 2300 mm and 2500 mm, and the annual sunshine duration is approximately 2900 h [27].

2.2. Meteorological Observations

The automatic weather station (37.54° N, 88.80° E, 4300 m above sea level) is located on the western side of the Ayakkum Lake (Figure 1b). The vegetation type is alpine desert, and the soil type is inceptisols [25]. In this study, the measurements from May 2013 to July 2018 were used to assess the gridded precipitation and air temperature products. The daily observations were compiled into a monthly time series. Two months (August and September 2015) with missing values > 4 days per month were discarded. To acquire an annual mean, we calculated the arithmetic mean for each month from January to December using the data available, and then summed (and averaged) the twelve values into an annual precipitation amount (and air temperature).

2.3. Gridded Climate Datasets

In this study, we collected five gridded precipitation products and five gridded air temperature products (

Table 1. Gridded precipitation and air temperature products used in this study.

Product	Parameter	Period	Spatial Resolution	Spatial Coverage
CMFD	Precipitation and temperature	1979–2018	6′ × 6′	China
CRU	Precipitation and temperature	1901–2021	30′ × 30′	Globe
ERA5	Precipitation and temperature	1950–2022	6′ × 6′	Globe
WorldClim	Precipitation and temperature	1960–2018	2.5′ × 2.5′	Globe
CPC	Precipitation	1979–2022	30′ × 30′	Globe
GHCN	Temperature	1948–2022	30′ × 30′	Globe

), and most products were produced on a global scale. Although a lot of gridded datasets have been released in the past decade, here, we only focused on the products with: (1) an output spatial resolution finer than 30′ (latitude) by 30′ (longitude), (2) the series updated to at least December 2018, and (3) the available time series longer than 30 years. The nearest grid box was applied to assess the accuracy of the climate products.

The China Meteorological Forcing Dataset (CMFD) is a gridded climate product covering the terrestrial area of China, and was generated using multiple data sources including remote sensing, reanalysis and surface observations [28,29]. The spatial resolution is 6′ (latitude) by 6′ (longitude), and the time series is available from January 1979 to December 2018.

The Climatic Research Unit (CRU) gridded time series (TS v. 4.06) is a gridded climate product using an improved interpolation function [30]. The spatial resolution is 30′ (latitude) by 30′ (longitude), and the period available is from January 1901 to December 2021. This dataset is based on the interpolation of monthly meteorological anomalies from extensive in-situ surface observations.

The European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis 5 (ERA5) is a global gridded climate product based on the Integrated Forecasting System, and benefits from recent techniques of model physics, core dynamics as well as data assimilation [31]. The dataset is the fifth generation of ECMWF atmospheric reanalysis, and covers the period from January 1950 to the present. Here, the spatial resolution of 6′ (latitude) by 6′ (longitude) was applied.

The WorldClim historical monthly weather data (WorldClim, for short) is a gridded climate product downscaled from the CRU TS v. 4.03 using the long-term climatology WorldClim v. 2.1 as bias correction [32]. The spatial resolution is 2.5′ (latitude) by 2.5′ (longitude). This product covers a period from January 1960 to December 2018, and has finer spatial resolution than the CRU time series.

The Climate Prediction Center Global Unified Gauge-based Analysis of Daily Precipitation dataset (CPC, for short) is a part of the products suite of the CPC Unified Precipitation Project [33]. The spatial resolution is 30′ (latitude) by 30′ (longitude). The product covers a period from January 1979 to the present.

The Global Historical Climatology Network and Climate Anomaly Monitoring System Gridded 2-m Temperature (Land) dataset (GHCN, for short) is a gridded dataset using the anomaly interpolation method with varying lapse rates of temperature for topography calibration [34]. The spatial resolution is 30′ (latitude) by 30′ (longitude). This monthly product covers a period from January 1948 to the present.

2.4. Methods

Several statistical measures were used in this study to assess the accuracy of gridded climate datasets, including the mean bias error (MBE), the mean absolute error (MAE), the root mean square error (RMSE) and the Pearson’s correlation coefficient (R):

$$\mathrm{MBE}=\frac{{\displaystyle \sum _{i=1}^n\left(M_i-O_i\right)}}{n}$$

(1)

$$\mathrm{MAE}=\frac{{\displaystyle \sum _{i=1}^n\left|M_i-O_i\right|}}{n}$$

(2)

$$\mathrm{RMSE}=\sqrt{\frac{{\displaystyle \sum _{i=1}^n{\left(M_i-O_i\right)}^2}}{n}}$$

(3)

$$R=\frac{{\displaystyle \sum _{i=1}^n\left(M_i-\overline{M}\right)\left(O_i-\overline{O}\right)}}{\sqrt{{\displaystyle \sum _{i=1}^n{\left(M_i-\overline{M}\right)}^2}}\sqrt{{\displaystyle \sum _{i=1}^n{\left(O_i-\overline{O}\right)}^2}}}$$

(4)

where M and O are predicted and observed meteorological parameters, respectively, and n is sample size.

We also used a new measure, the distance between indices of simulation and observation (DISO) [35], which is defined as

$$\mathrm{DISO}=\sqrt{{\left(R-1\right)}^2+{\mathrm{MBE}}_{\mathrm{nor}}^2+{\mathrm{RMSE}}_{\mathrm{nor}}^2}$$

(5)

where R is correlation coefficient, and MBE_nor and RMSE_nor are normalized MBE and RMSE (i.e., divided by the absolute observation mean). The DISO value close to zero indicates good performance. The advantage of DISO against the Taylor diagram is that the comprehensive performances of different models are still not quantified in the Taylor diagram [35,36]. This measure has been applied in many recent studies [37,38].

3. Results

3.1. Comparison of Precipitation Products

The observed annual mean precipitation in the automatic weather station is 159.8 mm, indicating a semi-arid climate in the study region. The product-derived precipitation amount is 158.6 mm (CMFD), 92.5 mm (CRU), 126.7 mm (ERA5), 125.8 mm (WorldClim) and 41.4 mm (CPC) per year. CMFD is quite close to the observation. The product-derived annual precipitation amount of CRU and CPC is much lower than the observation. As shown in Figure 2, the observed precipitation is concentrated in the months from May to September (94.5%), especially from June to August (77.8%). The proportions of precipitation from May to September for CMFD, CRU, REA5, WorldClim and CPC are 79.9%, 75.1%, 83.6%, 97.1% and 90.3%, respectively, and the proportions from June to August are 58.9%, 71.6%, 59.3%, 84.5% and 76.4%, respectively. The seasonal variation can be described by all the gridded products. It should also be noted that some months with large precipitation amounts cannot be well predicted in the grid products. For example, the observed precipitation amount in August 2016 is 89.9 mm, which is the largest month during the observation period; however, the product-derived precipitation amounts are all less than 35 mm.

We also compared the proportion distribution of the observed and product-derived precipitation amounts in the Ayakkum Lake in Figure 3a. The months with weak precipitation of less than 5 mm per month play a dominant role in the study region. For the observed data, the proportion with precipitation < 5 mm per month is approximately 57.4%. This proportion is overestimated by CPC (78.7%), WorldClim (70.5%) and CRU (67.2%), and is underestimated by ERA5 (45.9%) and CMFD (39.3%). As the precipitation amount increases, the proportion shows a decreasing trend, and the contribution of months with precipitation larger than 50 mm per month is usually less than 10%. The product-derived proportions of ≥50 mm/month range from 0.0% (ERA5 and CPC) to 4.9% (WorldClim), but the observed proportion (6.6%) is slightly higher than these product-derived proportions. In Figure 3b, the median values of monthly precipitation amount derived from CRU (0.5 mm), WorldClim (1.0 mm) and CPC (0.0 mm) are less than the observed median (3.4 mm), and the medians derived from CMFD (7.4 mm) and ERA5 (7.1 mm) are larger than the observation.

Table 2. Comparison of the accuracy of five gridded precipitation products.

Product	MBE (mm)	MAE (mm)	RMSE (mm)	R2	DISO	n
CMFD	0.12	9.06	15.09	0.42	1.18	61
CRU	–5.14	9.90	16.40	0.39	1.33	61
ERA5	–2.65	6.23	11.59	0.80	0.89	61
WorldClim	–1.97	8.73	18.60	0.43	1.43	61
CPC	–9.63	10.76	19.27	0.30	1.66	61

shows some statistical measures (i.e., MBE, MAE, RMSE, R² and DISO) calculated for monthly amounts. Regarding the MBE value, CMFD is very close to zero (MBE = –0.12 mm), indicating a good approximate of the total amount; however, CMFD is not always the best when using other parameters. ERA5 shows the lowest MAE and RMSE values, and the highest R² value; CMFD, CRU and WorldClim generally show a similar performance according to MAE, RMSE and R². CPC usually exhibits a weak performance among the five precipitation products, and has the largest MBE, MAE and RMSE, as well as the lowest R² value. The DISO value is a statistical measure merging different metrics, including MBE, RMSE and R², in a three-dimensional coordinate system [35,36]. The calculated DISO values are 1.18 (CMFD), 1.33 (CRU), 0.89 (ERA5), 1.43 (WorldClim) and 1.66 (CPC), indicating that ERA5 is the recommended one among the five gridded products, and the alternative is CMFD. The DISO-based optimal product is consistent with the results using MAE, RMSE and R².

Here, we present the relationship between observed and product-derived precipitation amounts on a monthly basis (Figure 4). A lot of points are concentrated close to 0 mm, indicating that low precipitation is common near the Ayakkum Lake, which is consistent with the monthly variation shown in Figure 2. During the winter months, the observed precipitation is very limited, and the dry conditions can be predicted in different gridded precipitation products. As the precipitation amount increases, the predicted amount generally departs from the diagonal line (y = x). The best-fitting lines show a lower slope than the diagonal line, which indicates that the higher precipitation amounts are usually underestimated.

3.2. Comparison of Air Temperature Products

In this study, the observed annual mean air temperature in the Ayakkum Lake is –4.8 °C, and the product-derived air temperatures are –6.4 °C (CMFD), –3.2 °C (CRU), –5.9 °C (ERA5), –3.8 °C (WorldClim) and –5.1 °C (GHCN), respectively. As shown in Figure 5, the surface air temperature is higher in the summer months and lower in the winter months, and the air temperature derived from different gridded products presents similar intra-annual variations. The observed highest monthly air temperature usually occurs in July or August, and ranges between 6.8 °C and 7.9 °C each year. The monthly air temperature maximum is usually underestimated by CMFD and ERA5, and overestimated by CRU, WorldClim and GHCN. The lowest monthly air temperature for each winter, ranging from –18.9 °C to –15.7 °C, is usually found in January, except in one case in December 2015. CMFD and GHCN always underestimate the minimum, while CRU usually overestimates the minimum.

Most months show low air temperatures below freezing point. In Figure 6a, the air temperature generally conforms to the normal distribution. The air temperature interval between –5 °C and 5 °C has the largest proportion, which is true for observed and product-derived data. The observed proportion of air temperature between –5 °C and 5 °C is 37.7%, and the product-derived proportion ranges from 32.8% (GHCN) to 39.3% (CMFD). In Figure 6b, the observed median air temperature is –4.4 °C. The medians derived from CMFD (–5.6 °C) and ERA5 (–5.4 °C) are lower than the observation, and the medians derived from CRU (–2.8 °C), WorldClim (–3.5 °C) and GHCN (–3.5 °C) are higher than the observation.

According to the R² values (>0.98) for each product in

Table 3. Comparison of the accuracy of five gridded air temperature products.

Product	MBE (°C)	MAE (°C)	RMSE (°C)	R2	DISO	n
CMFD	–1.64	1.64	1.94	0.99	0.55	61
CRU	1.62	1.70	1.91	0.99	0.54	61
ERA5	–1.17	1.39	1.58	0.99	0.42	61
WorldClim	0.98	1.28	1.52	0.99	0.39	61
GHCN	–0.28	1.54	1.85	0.98	0.40	61

, the performances of gridded air temperature products are better than those of precipitation, as shown in Table 2. Judged by MAE and RMSE, WorldClim shows the best performance (MAE = 1.28 °C and RMSE = 1.52 °C). GHCN shows an MBE value close to zero (MBE = –0.28 °C), but does not have a very low MAE or RMSE value. According to the DISO value, WorldClim shows the best performance (0.39), although GHCN (0.40) and ERA5 (0.42) also have relatively low values. The optimal product based on DISO is consistent with that based on MAE and RMSE.

The relationships between observed and product-derived monthly air temperature are also shown in Figure 7. The points are generally close to the diagonal line (y = x), indicating generally good performance. The slopes of the best-fitting lines range from 0.913 (ERA5) to 1.152 (GHCN). For CMFD, the best-fitting line is always at the bottom-right side of the diagonal line; in contrast, the CRU-based best-fitting line is at the top-left side, indicating an overestimation of air temperature. For the three remaining air temperature products, the best-fitting line and the y = x line are cut across.

4. Discussions

4.1. Accuracy of Gridded Climate Products

In this study, several precipitation and air temperature products were evaluated in the Ayakkum Lake. Regarding the monthly precipitation amount, ERA5 shows a good performance based on most statistical measures, including MAE, RMSE, R² and DISO, and CMFD can also be an alternative, especially for estimating the long-term annual mean. Regarding the monthly air temperature, WorldClim is recommended because of the low MAE, MBE and DISO, as well as high R², and ERA5 and GHCN can be alternatives. It should be noted that the new statistical measure DISO [35,36] provides an effective way to evaluate the prediction of gridded climate products. Although previous studies examined gridded climate products in arid central Asia, there is no meteorological station available across the Ayakkum Lake basin in these works. The result of this study is useful to understand the performance of gridded climate products in the Ayakkum Lake basin, as well as in other central Asian mountains.

As a global reanalysis dataset, ERA5 has been widely used across arid central Asia [39,40,41]. Although some assessments have indicated ERA5 does not always show optimal performance for the mountains [39,42], ERA5 is generally recommended in the Ayakkum Lake according to the statistical measures in this study. CMFD is another alternative for precipitation estimation in this study. Because satellite-based precipitation information is merged into CMFD, the performance of CMFD is also confirmed in many studies in western China [43,44,45]. Interpolation-based WorldClim shows a good performance for air temperature, but not for precipitation amount, indicating that the lapse rate of air temperature is more stable than that of precipitation amount. The precipitation usually has obvious spatial heterogeneity, and is constrained by local and regional topography.

In the Ayakkum Lake basin, there is no national meteorological station, and some neighboring stations are located in the low-lying Tarim Basin, which cannot describe the plateau or mountain climate. The absence of meteorological stations logically increases the uncertainty in mapping precipitation and air temperature. In recent years, the local meteorological administrations, nature reserve administrations and some research institutes have established new automatic weather stations in the mountains at the southern margin of Xinjiang [23,25]; the Third Xinjiang Scientific Expedition Program and the Second Tibetan Plateau Scientific Expedition and Research Program may also promote field meteorological observations. However, longer in-situ observations are still needed.

In arid central Asia, heavy precipitation events do occur, especially in recent years [46,47,48], which are usually associated with the large-scale circulation as well as the local topography [49,50,51]. As shown in this study, the precipitation amount may be close to 90 mm per month, which accounts for more than half of the long-term annual precipitation amount (Figure 2). However, the high-precipitation months are usually difficult to predict in most gridded climate products.

4.2. Implications for Lake Water Budget Studies

The Ayakkum Lake is a representative mountainous lake in southern Xinjiang of arid central Asia. In this sparsely populated region, the meteorological and hydrological observations are very limited. The entire lake belongs to a national nature reserve, and tourists are not allowed to visit without permission. During the past few decades, the water budget of the Ayakkum Lake has been sensitive to global warming. Many researchers have examined the growing lake area and rising water level using remote sensing techniques [18,19,20,21].

It is an interesting question why the lake area shows an increasing trend. Due to the limited in-situ observations, most studies of the lake budget are just qualitative speculation. For example, Chen et al. [21] attributed the Ayakkum Lake’s area variation to the increasing air temperature and precipitation amount, as well as the decreasing evaporation amount; however, the linearly interpolated meteorological conditions (summer air temperature ranging between 15 °C and 17 °C, and annual precipitation ranging between 0 mm and 70 mm) are quite different from those observed in this study and previous measurements. Li et al. [52] also noticed the rapid warming and wetting trend in the Ayakkum Lake, and applied an interpolation-based gridded climate product at the spatial resolution of 0.5° by 0.5°; however, the annual precipitation amount has increased from 120 mm to 225 mm in the past five decades, and the trend magnitude is, abnormally, higher than in the known observations in western China. Wang et al. [53] linked the Ayakkum Lake’s area to the glacier area against a warming background, but only one meteorological station (annual precipitation ranging between 12 mm and 96 mm, and air temperature ranging between 2.7 °C and 4.8 °C) outside the lake basin was used; the meteorological station used suggested much drier conditions than the observations. With an improved understanding of precipitation and air temperature, the water budget can be estimated more accurately.

5. Conclusions

We evaluated the performance of several gridded precipitation and air temperature datasets near the Ayakkum Lake at the southern margin of Xinjiang, arid central Asia. For precipitation, ERA5 shows a good performance among the five products based on most statistical measures (i.e., MAE, RMSE, R² and DISO), and CMFD can also be an alternative, especially for estimating the long-term annual mean. For air temperature, WorldClim is recommended because of the low MAE, MBE and DISO, as well as high R², and ERA5 and GHCN can be alternatives. Some gridded products may have large uncertainties, and are not suitable to be directly applied before careful calibration or downscaling. This study provides a basis to extend the temporal coverage in order to understand the water budget of the lake, and is helpful to identify the optimal prediction of common meteorological parameters. Longer temporal coverage and enhanced spatial coverage of in-situ observations are still needed to validate the gridded climate products and produce an optimal correction scheme for the central Asian mountains.