Aerosol optical depth (AOD): spatial and temporal variations and association with meteorological covariates in Taklimakan desert, China

Overview of the study area

The Taklimakan Desert is located in the Tarim Basin of Xinjiang, China. It is the largest mobile desert in China and the second largest in the world. The Taklimakan Desert has an extremely dry climate, sparse vegetation, and scarce rainfall. It has rich sand sources and maximal bare ground and only scattered tamarix shrubs and reeds are distributed between the ridges. It mainly comprises flowing sand, dry soil, loose structures, and exposed surfaces with flowing shapes that are prone to wind erosion and dusty weather. Easterly winds prevail throughout the year with an annual average of more than 157 days of dust weather and approximately 16 days of sandstorms. The dust and aerosol concentrations in the atmosphere are considerably higher in this region than in cities.

Satellite RS data

Meteorological data

Meteorological data are provided by the Tazhong station. The data include average wind speed, air temperature, relative humidity, and evaporation. According to the classification standard of the general meteorological classification method, the months of March–May are classified as spring, June–August summer, September–November autumn, and December–February winter. These divisions correspond to the time scale of downloaded MODIS AOD data. Weather data for each of the four intervals of 2000, 2005, 2010, and 2015 were selected (for a total of 48 months), and averages from March to May, June to August, September to November, and December to February calculated. The average values calculated were seasonal and annual average temperature, average relative humidity, average wind speed (the data is recorded every 2 s), maximum wind speed, and evaporation.

Ground-based AOD data

A CE-318 sun sky lunar multispectral photometer is located in the Tazhong Desert Atmospheric Environment Observation Experimental Station of the China Meteorological Administration (38°58′N, 83°39′E, altitude 1,090 m) as shown in Fig. 1. This meteorological station is nearly 200 km deep in the Taklimakan Desert and monitors desert weather and climate.

An 8-band CE-318 sun sky lunar multispectral photometer (Cimel Electronique, Paris, France) was used for AOD observations in this study. This instrument can measure field solar and sky radiations with high accuracy, including 440-, 670-, 870-, 1,020-nm aerosol bands, 936-nm water vapor band, and three 870-nm polarimetric bands. The bandwidth of each channel is 10 nm. The observation time was from March to December 2015 and observations were not conducted in November. The ground-based AOD data correspond to the average data 30 min before and after the satellite transit time. The daily sun photometer data were integrated to obtain monthly ground-based AOD data. These data were then used to obtain the AOD value in the 550-nm band using the Angstrom exponent calculation, compared with the AOD values of the MODIS products, and used to verify AOD data of the MODIS products. The AOD value is calculated as follows: $$\tau =\beta \epsilon -\alpha$$where τ(ε) is the AOD, α is the Angstrom exponent that represents the proportion of large or small particles in the aerosol component, and β is the Angstrom turbidity coefficient that measures the aerosol concentration in the atmosphere.

The Angstrom exponent calculation formula is as follows: $$\alpha ={\displaystyle \frac{\ln [\tau (\lambda 1)/\tau (\lambda 2)]}{\ln (\lambda 1/\lambda 2)}}$$where τ(λ) is the AOD and λ₁ and λ₂ are the wavelengths (nm).

Random forest model

The random forest (RF) algorithm is a machine learning algorithm for classification and regression. It integrates multiple decision trees using the concept of holistic learning. The RF algorithm can process high-dimensional data and can be used in collections. It uses a large number of trees (Sun et al., 2018). Partition variables between tree correlations in the minimized set are randomly selected using RF for regression, and the output is the average of the outputs from all decision trees (Chen et al., 2019; Wang et al., 2018). The RF method utilizes many variables and few samples. It can rank the control factors of variables and provide relative importance of each variable (Hong et al., 2019). Moreover, for unbalanced samples, RF can balance errors and has its own advantages of avoiding overfitting and reducing generalization errors. In this study, the AOD was the dependent variable and six environmental covariates, such as average temperature, average relative humidity, average wind speed, maximum wind speed, evaporation, and NDVI, were the independent variables. We used environmental variables to estimate the AOD value and obtained the importance of environmental covariate under the framework of the RF model.

MODIS product verification

The results of the MODIS aerosol products (MCD19A2) were verified by comparing them with measured data obtained using a CE-318 solar photometer (Fig. 2), obtaining R = 0.954 and R² = 0.910. The fitting curve is nearly 1:1. The line height is consistent with high precision. Based on these results, the quality of AOD data calculated using the MODIS AOD product is considered reliable.

Spatiotemporal distributions of AOD

The average AOD spatial distribution in the Taklimakan Desert of each month in 2000, 2005, 2010, and 2015 is shown in Fig. 3. The low-value area of AOD in the southern part of the study area and its margins is relatively stable over time. High-value areas occupy the center of the Desert and its northern margins. The spatial distribution of AOD is mainly affected by topographical factors and shows an obvious step-up from south to north.

The white areas in the picture are null values. The inversion algorithm of MODIS AOD uses the dark pixel algorithm. Because snow covers the dust particles, the ground surface brings a bright background, this situation causes a change in ground surface reflectivity. Hence, the AOD value is missing in winter.

The AOD value of each month in these four periods ranged from 0.43 to 2.41 and show that the change in AOD is a flow-cycle process. In this study, the AOD for the 12 months of 2000, 2005, 2010, and 2015 were processed in a unified manner and changes were analyzed to support attribution studies assessing the role of climate variability and change on the Taklimakan Desert.

After the MODIS AOD data was processed, the average AOD for each month in 2000, 2005, 2010, and 2015 was calculated, analyzed and compared. The spatial distribution of AOD data was similar in all four periods. In January, February, September, November, and December, Fig. 4 shows the hot spot colors did not change significantly for the medium- and low-values. However in the other moths AOD values varied widely. Based on changes in color distribution, the AOD values of the four periods show a clear seasonal cycle. The high and low AOD values were mainly observed in April and winter, respectively. As shown in Fig. 5, the AOD values for the 12 months of 2000, 2005, 2010, and 2015 followed a similar seasonal pattern of decreasing-increasing-decreasing trend, expressed as an inverted “W.” Except for June 2010 (0.38), the highest values of AOD occurred in April 2000, 2005, and 2015 (0.41, 0.39, and 0.43, respectively) and low values all occurred in November or December. The AOD value reached its maximum in April–June. After October, cold air entered from the north and moved east to south, causing the weather to become colder. Although the main effect of cold air did not fully enter the desert and its margins, the desert was still covered by snow dusty weather decreased because of the southward path of cold air; this further reduced the AOD values. The four sets of AOD RS data show that monthly AOD values are mainly affected by the weather.

Seasonal time series of AOD

Figure 6 shows the spatial AOD distributions for the four seasons of 2000, 2005, 2010, and 2015, with high values in the north and low in the south. The low-value area shows little spatial variation, while the high-value areas show much more variation but are relatively evenly distributed in the middle and upper parts of the desert and its margins. From Fig. 7, the annual variation of AOD in 2000, 2005, 2010, and 2015 are almost the same. AOD values in spring and summer increased between 2000 and 2015, with the largest increase in spring. In autumn and winter, AOD value ranges from 0.150 to 0.230, without showing significant change. The seasonal cycle is relatively stable. High AOD occurs in spring and is often associated with sandstorm events occurring in this season. Wind speeds accelerate the movement of mobile dunes; upward air accumulation and associated dust on natural surfaces lead to the highest AOD concentrations in the Taklimakan Desert in spring. Figure 8 shows the highest and lowest AOD values occurred in spring and winter, respectively. There is a gradual decrease from spring to winter in 2000, and the decrease from summer to autumn becomes more abrupt over time. Only spring and early summer are accompanied by large numbers of dusty days, while dusty weather lessens in late summer and autumn with AOD values decreasing. The AOD values for all four seasons from 2000, 2005, 2010 and 2015 were below 0.4 with those in spring being highest (0.28, 0.27, 0.29, and 0.35, respectively). Low AOD values in winter is attributed to reduced winter wind and dust activity, fewer sandstorms, and snowfall.

Annual averages with 5-year intervals and seasonal variations of AOD in the temporal and spatial changes in the Taklimakan Desert and its margins

Figure 9 shows spatial distributions of average AOD in the four seasons of the Taklimakan desert from 2000 to 2015. They are all at the bottom of the desert and its margins. The trend is particularly obvious in autumn; the range of high values in summer and winter is small and scattered and gradually collapses in the outer areas.

Figure 10 shows annual distribution of the average AOD from 2000 to 2015. The low-value areas from 2000 to 2015 always remained in the bottom of the Desert and its margins. Differences were observed in the range of high-value areas. The high-value areas in 2000 diffused from the central areas, and the high-value areas above the middle areas occupied half of the area. The range of high-value areas in 2005 was small and scattered, and the median range was concentrated. In 2010 and 2015, the high-value range significantly expanded. The maximum of AOD gradually rose from 2000, 2005, 2010 and 2015.

As shown in Fig. 11A, the average AOD concentrations are highest in spring. As early summer is also the main season for frequent winds, the amounts of dust gradually decrease, while the fall and winter seasons are significantly lower. The average seasonal AOD decline rates were 3.6%, 33.3%, and 5.6% from spring to summer, from summer to autumn, from autumn to winter. The average AOD value of the four quarters has respectively decreased and from spring to winter the decrease in the AOD value is most obvious, as high as 39.3%. The downward trend increases from summer to autumn. This is reflected from the perspective that atmospheric conditions are more severe in spring than in autumn and winter and the air quality is relatively poor. The Taklimakan Desert has experienced significant seasonal differences in mean AOD over the years owing to decreases in certain meteorological factors (such as humidity), the average AOD value shows a significant downward trend from spring to winter.

As shown in Fig. 11B, the average AOD are relatively stable with a gradual increase to 2010 a large increase to 2015 (the AOD value is the annual averages with 5-year intervals). The AOD growth rates were 4.8%, 1.8%, and 16.7% from 2000 to 2005, from 2005 to 2010, from 2010 to 2015. Among them, the AOD values were highest in 2015 at 0.26. Overall, the annual increase rate was as high as 23.3%. When temperature increases, the local wind will produce thermal differences, due to uneven heating and temperature difference, leading to increased turbulence, Wind speed increases above a certain threshold, will project more sand and dust into the air, increasing of aerosol concentration. When the wind speed reduces, dust particles under the action of gravity settle back to land. Moreover, precipitation is extremely scarce. The groundwater content is very low and dust cannot be consolidated, so the accumulation of dust will cause the aerosol content to show an increasing trend. Therefore, wind speed is the key factor affecting the strength of dust weather and the change of dust aerosol concentration.

The importance of environmental factors

The predicted values from application of the RF model are shown in Fig. 12. They are highly consistent with the daily composite values of the MODIS AOD products with R2 values as high as 0.91. The impact of environmental covariates on AOD of the desert and its margins is ranked as follows: relative humidity > temperature > wind speed (max) > wind speed(mean)> evaporation> Normalized Difference Vegetation Index. RH has the greatest effect on AOD, followed by temperature, while NDVI has the smallest effect with weights of 0.798, 0.436, and 0.138, respectively. NDVI is not a major influencing factor for AOD, probably because the Taklimakan Desert is far from the ocean, cold and humid air currents and the Indian Ocean warm and humid air currents have difficulty reaching the desert, precipitation is extremely low, the sand blown by the wind is high, and vegetation has difficulty surviving. Conclusively, the dominant factor directly or indirectly affecting AOD in the Taklimakan Desert is RH. Temperatures, wind speeds, evaporation, and other factors work together to affect AOD. This study shows that the use of the RF model to estimate AOD from selected environmental covariates has good potential for application.

Relationship between temperature and humidity in sandy weather in the desert and its margins

Based on the data of average temperature and average RH on days with dust from 2000 to 2015 provided by atmospheric environment observations and the Tazhong Desert Atmospheric Environment Observation Experimental Station of the China Meteorological Administration, we mainly analyzed temperature and humidity and showed that AOD is directly or indirectly affected by RH. In Fig. 13 (Note: the X axis represents the number of days where sand and dust weather occurred in different years). Dust weather refers to when strong wind at ground level picks up a large amount of dust, air turbidity increases, horizontal visibility is significantly reduced, is a common disastrous weather phenomenon. According to the definition of sand dust by China Meteorological Administration, it can be divided into floating dust, sand blowing and sand storm. The horizontal visibility of floating dust is less than 10 km, the horizontal visibility of sand blowing is less than or equal to 1 km to less than 10 km, and the horizontal visibility of sand storm is less than 1 km. The distance of horizontal visibility is used to determine the occurrence of dust weather shows that when dust weather occurred from 2000 to 2015 (this article mainly analyzes the frequency of dust weather), the average temperature had a strong negative relationship with average RH. We speculate that when dust becomes mobile, there is rapid drying at the surface and RH decreases very quickly. because moisture binds the dust grains and wind allows heat access between those dust and sand grains and everything dries out faster. Whereas, our result show that humidity and AOD are negatively correlated, thus AOD will increase result from dust weather. Here, we can consider the dust weather as an essential part of the increase in dust aerosols.

In addition, we use the simple Aridity Index created by De Martonne (calculated by two climate factors, temperature and precipitation) to indirectly testify whether or not a change in RH has occurred, thus affecting AOD. By using the annual temperature and precipitation data of 2000, 2005, 2010 and 2015 (excluding the missing data), we obtained the Aridity index of 10.01, 10.01, 10.00 and 10.00 respectively, showing an overall decreasing trend, indicating that evaporation was more intense and precipitation was very rare, so RH would also decrease. According to the meteorological data of the Taklimakan Desert, there has been a upward trend of dusty weather in recent years. Moreover, wind speed is a key factor affecting the intensity of sand and dust weather. Thus, we can confidently assume that if in the future, the occurrence of dusty weather (dust, sand, etc.) increases, AOD may show an upward trend; however, the increase or decrease in AOD is still affected by various conditions.

Correlation analysis between AOD and meteorological variables or factors

Figure 14 presents the correlations between AOD and various meteorological elements (factors). NDVI exhibits the smallest correlations with other variables. Because there is little accumulation of water in the desert and its margins, the desert is highly mobile, preventing the establishment and survival of vegetation. Vegetation growth is severely challenged, and vegetation coverage hardly expands. Monthly NDVI in 2000, 2005, 2010, and 2015 showed little seasonal variation. Therefore, NDVI is not a major factor affecting dust aerosols. In terms of meteorological elements, there is a clear negative correlation between relative humidity (RH) and AOD; their correlation coefficient is the greatest at 0.7, followed by mean wind speed (WSmean) at 0.6. The correlation coefficients of atmospheric temperature, maximum wind speed, evaporation, and AOD are all 0.5. Overall, AOD are most affected by Meteorological factors (Relative humidity).