Dynamic monitoring of aeolian desertification based on multiple indicators in Horqin Sandy Land, China

Science of the Total Environment 650 (2019) 2374–2388

Contents lists available at ScienceDirect

Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv

Dynamic monitoring of aeolian desertification based on multiple indicators in Horqin Sandy Land, China Hanchen Duan ⁎, Tao Wang, Xian Xue, Changzhen Yan Key Laboratory of Desert and Desertification, Northwest Institute of Eco-Environment and Resources, Chinese Academy of Sciences, Lanzhou 730000, China

H I G H L I G H T S

G R A P H I C A L

A B S T R A C T

• Auto-extraction method of aeolian desertified land based on multiple indicators. • The time series dataset of aeolian desertified land was extracted from 2000 to 2015. • Aeolian desertified land decreased significantly at a rate of 2388.60 km2 y−1. • Climate and human factors both responsible for aeolian desertification reversal.

a r t i c l e

i n f o

Article history: Received 19 July 2018 Received in revised form 17 September 2018 Accepted 30 September 2018 Available online 01 October 2018 Editor: Ralf Ludwig Keywords: Aeolian desertified land (ADL) Time series Spatial-temporal dynamics Driving factors Horqin Sandy Land

⁎ Corresponding author. E-mail address: [email protected] (H. Duan).

https://doi.org/10.1016/j.scitotenv.2018.09.374 0048-9697/© 2018 Published by Elsevier B.V.

a b s t r a c t Aeolian desertification has become one of the most serious environmental and socioeconomic problems facing the world today. Quantitative remote sensing technology is an important means to achieve the development trends of aeolian desertified land (ADL). To compensate for the shortcomings in the time scale of Landsat Thematic Mapper and other high-spatial-resolution remote sensing data, this study introduces Moderate Resolution Imaging Spectroradiometer time series data and products to invert the monitoring indicators of ADL. The QUEST (quick, unbiased, and efficient statistical tree) classification method was used to establish the extraction model of ADL based on multiple indicators. The ADL time series dataset was extracted from 2000 to 2015, and the characteristics of ADL and its spatial-temporal dynamics were analyzed. These results were combined with meteorological data and socioeconomic statistics to discuss the main factors influencing ADL. The results showed that, by the end of 2015, the total area of ADL was 32,633 km2, accounting for 26.02% of the study area. The slight, moderate, severe, and extremely severe ADL accounted for 51.39%, 34.11%, 10.31%, and 4.20%, respectively. The total area of ADL decreased significantly at a rate of 2388.60 km2 y−1 from 2000 to 2015. The decreasing area was dominated by the slight and moderate ADL. The reversal of ADL exhibited significant correlations with an increase of annual precipitation and a decrease of annual maximum wind velocity (p b 0.01). The impact of annual maximum wind velocity on ADL is more important than annual precipitation. Increases in population density and the number of livestock did not promote the development of ADL. A series of ecological protection projects and policies created advantageous conditions for the reversal of ADL. This research provides a new method for monitoring ADL and useful information for controlling and managing aeolian desertification in this region. © 2018 Published by Elsevier B.V.

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

1. Introduction Desertification can be defined as a land degradation process that occurs in arid, semi-arid, and dry subhumid regions as a result of various factors, including climate change and unsustainable human activities (UNCCD, 2004; UNEP, 1994). It is composed of aeolian desertification, water erosion, salinization, and freeze or thawing according to the chemical and physical processes of soil degradation. Aeolian desertification is one of the main types of desertification that occurs in northern China (Wang and Zhu, 2003), which is one of the most serious environmental problems, especially in arid, semi-arid, and dry subhumid zones (Helldén, 2008; Kassas, 1995; UNCED, 1992; Wang et al., 2015). The occurrence and development of aeolian desertification cause serious harms to the environment, natural resources, social economy, and people's lives (Oswald and Harris, 2016; Song et al., 2015; Zhao et al., 2014). The expansion of aeolian desertification has converted land areas from non-desertified land as large as a midsize county each year, and the economic loss directly attributed to aeolian desertification was approximately 54 billion RMB per year in China (Ge et al., 2016; Wang and Zhu, 2001). These losses seriously restricted the regional socioeconomic sustainable development and affected nearly 300 million people in northern China (Li et al., 2007; Wang et al., 2004; Zhu and Wang, 1993). Therefore, the Chinese government has supported many studies in the field of aeolian desertification and conducted a series of projects over the past 40 years. However, after many years of effort, the aeolian desertification situation remains severe with the possibility that restoration gains could be thwarted. This makes it necessary to continue to strengthen comprehensive research on aeolian desertification. Historically, the Horqin Sandy Land was a lush forest and grassland landscape as well as a traditional grazing area. The ecology of the area was seriously damaged, however, because of the influence of climate factors, such as drought and wind; soil factors, such as a sandy substrate with low organic content; and excessive reclamation, overgrazing, and unsustainable human economic activities. This damage has been especially pronounced over the past century. The sparse forest sandy grassland gradually degenerates into a sandy landscape with frequent sandstorms, scattered sand dunes, and semifixed dunes. This has resulted in this region becoming an agropastoral area (Wu et al., 2003). The Horqin Sandy Land is one of the most ecologically fragile areas in China. It is close to the densely populated and economically developed regions in northeast and North China. Because of its unique geographical location, it affects the economic and social development of the region as well as the environmental quality of the Beijing-Tianjin-Tangshan region. The problem of aeolian desertification in the Horqin Sandy Land has been a widespread concern among the government, public, and academia since the 1950s. For nearly 50 years, scientific researchers have conducted work focused on the cause, process, and comprehensive improvement of aeolian desertification and they have published hundreds of academic papers and monographs (Ge et al., 2016; Li et al., 2017; Wang et al., 2016, 2017). Some fundamental and vital components of understanding, including the time series of aeolian desertification in the Horqin Sandy Land, still have not been fully elucidated. The combination of remote sensing (RS) technology and geographic information systems (GIS) provides an effective method for monitoring aeolian desertification (Lambin, 2001). At present, most of the published work has focused on the analysis of dynamic changes in aeolian desertification via comparison of RS data in different periods that have a certain time interval. Taking Landsat Multi-spectral Scanner (MSS), Thematic Mapper (TM), and Enhanced Thematic Mapper Plus (ETM +) images as the data source, a human-computer interactive visual interpretation method was used to extract the information of aeolian desertification. Hu et al. (2015) monitored the spatiotemporal changes of aeolian desertification of the Zoige Basin five times over the course of

2375

35 years (1975, 1990, 2000, 2005 and 2010). Xue et al. (2013) evaluated the evolution and status of aeolian desertification between 1975 and 2010 in the northwestern Shanxi Plateau by using MSS and TM images (acquired in 1975, 1991, 2000, 2006, and 2010). On the other hand, based on the same data source, computer automatic classification method also is used widely to extract and monitor the aeolian desertification. Qi et al. (2012) monitored the temporal and spatial dynamics of desertification by using the supervised classification method from 1986 to 2003 in the agropastoral transitional zone of northern Shaanxi Province, China. Salih et al. (2017) adopted spectral mixture analysis, objectbased oriented classification, and change vector analysis methods to describe the status and rate of land degradation and desertification processes by using multi-temporal imagery of TM and ETM+ respectively. Although these methods have high precision, they are difficult to monitor the dynamics of aeolian desertification at a large scale and over long time series because of the large time scale of the process and the influence of artificial factors. With recent improvements in RS satellite technology, the application of low and moderate spatial resolution RS data can meet the requirements necessary to produce a time series of aeolian desertification. For example, the Advanced Very High Resolution Radiometer (AVHRR), Moderate Resolution Imaging Spectroradiometer (MODIS), and Systeme Probatoire d'Observation de la Terre (SPOT) can be used widely to monitor the aeolian desertification. Many researchers have begun to employ special indices to extract the information of aeolian desertification (Badreldin et al., 2014; Li et al., 2016). Eckert et al. (2015) detected land degradation and regeneration in Mongolia by using MODIS Normalized Difference Vegetation Index (NDVI) time series and found that NDVI time series trend analysis was suitable for identifying land degradation and regeneration. Feng et al. (2016) developed the Normalized Difference Desertification Index (NDDI) to analyze the spatial-temporal change of aeolian desertification and its possible influencing factors in northern China. Ma et al. (2011) constructed the desertification-monitoring model based on the Albedo-NDVI feature space, which can effectively achieve the automatic identification of desertified land. Related indicators or indices used in these studies can well represent the spatial and temporal changes of land desertification. However, most of these studies used a single indicator or constructed the feature space through the relationship among several indicators to describe the process of aeolian desertification. Because the causes and existing conditions of aeolian desertification are complicated, it is difficult to extract and evaluate the aeolian desertification comprehensively and accurately by using a single indicator, which cannot fully reflect the comprehensive situation of aeolian desertification due to the inherent uncertainty. To fill the gaps produced by utilizing a single spectral fingerprint, some researchers proposed methods suitable for large-scale monitoring of aeolian desertification based on a combination of multiple indices or indicators (Han et al., 2013; Lamchin et al., 2016; Liu et al., 2007; Ma et al., 2007; Mamatsawut et al., 2008). These studies have shown that this method not only can guarantee the monitoring accuracy but also can meet the requirements of time series. The combination of multiple indicators has become the latest development trend for aeolian desertification long-term monitoring and assessment (Jiang and Lin, 2018; Zucca et al., 2012). In addition, the analysis of aeolian desertified land (ADL) time series can reflect changes of ADL in drought or wet years and can also distinguish the impact of human activities and climate change more precisely. The objective of this research is to use MODIS data and its products as the monitoring indicators to extract the time series data of aeolian desertification. These indicators include Fractional Vegetation Cover (FVC), Modified Soil Adjusted Vegetation Index (MSAVI), Albedo, Land Surface Temperature (LST), and Modified Temperature Vegetation Drought Index (MTVDI). FVC and MSAVI reflect the ecological status of ADL, and the other three indicators reflect the physical properties of ADL. All of these indicators can reflect the characteristics of different types of ADL and also can be obtained by quantitative RS inversion. By constructing a monitoring system of

2376

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

aeolian desertification based on a combination of multiple indicators, we quantitatively extracted the ADL information and discussed its changing trends and causes. 2. Materials and methods 2.1. Study area The Horqin Sandy Land (41° 41′–46° 05′ N, 117° 49′–123° 42′ E) is one of the four largest sandy lands in China, with a total area of 12.51 × 104 km2 and an elevation that ranges from 0 to 2024 m. It is located in the western part of the Northeast Plain, in the transition zone between the Inner Mongolian Plateau and the Northeast Plain of China. The east-west extent of the study area reaches from the Qilaotu Mountains to the Songliao Plain and the north-south extent reaches from the Daxinganling Mountains to the Nuluerhu Mountains. The area is a 400 km long sand belt from east to west. The elevation is high in the west, south, and north, but is low in the east and central region (Fig. 1). It mainly contains 15 counties of the Inner Mongolian Autonomous Region. The study area has a typical semi-arid continental monsoon climate. The winter and spring are windy and dry, summer is warm and comparatively wet, and autumn is short and cool. The annual mean temperature ranges from 3 to 7 °C. The average temperature is within −12 to −17 °C in the coldest month (January) and 20 to 24 °C in the hottest month (July). Extreme maximum and minimum temperatures are 39 °C and −29.30 °C, respectively. Spatial distribution of the annual precipitation is restricted by the atmospheric circulation, and precipitation is concentrated mainly in the summer and autumn. The mean annual precipitation is in the range of 350 to 500 mm, of which 70% is concentrated in summer. The maximum and minimum values of annual mean precipitation are 500 mm and 250 mm, respectively. The change rate of the annual precipitation is N20%. The mean annual potential evaporation ranges from 1500 to 2500 mm, which is more than five times that of

the mean annual precipitation (Li et al., 2012; Zhao et al., 2007). The aridity coefficient is within the range of 1.00 to 1.80. The study area is located in the midlatitude westerlies, where the western, northwestern, and northern cold air flows. In winter, western and northern wind prevails as a result of the cold anticyclone of Mongolia. During the summer, the southern and southwestern wind is dominant due to a continental infrabar and subtropical anticyclone. Spring and autumn are the transition seasons, so the wind direction changes frequently. The annual mean wind velocity ranges from 3.40 to 4.40 m s−1, and the maximum velocity is within the range of 19 to 31 m s−1. The threshold wind velocity for sand movement (≥5 m s−1) is exceeded in roughly 210 to 310 days y−1 (mainly in spring and winter). Sandstorms occur within the range of 10 to 15 days y−1, mostly in the spring (Li et al., 2009; Zhang et al., 2004). The soil is classified as Cambic Arenosols of sandy origin in the Food and Agriculture Organization of the United Nations (FAO) soil classification system (FAO, 2006), and it is characterized by coarse texture and loose structure. Because of these characteristics, the soil is particularly susceptible to wind erosion (Cao et al., 2011; FAO, 2006). The native vegetation is a tree-sparse grassland that contains mainly Ulmus pumila and Quercus mongolica (Wang, 2011). As a result of human destruction and aeolian desertification, the natural vegetation is dominated by semi-arid shrubs, subshrubs, and sparse forests. The Horqin Sandy Land is an agropastoral transitional zone that is typical in semi-arid and subhumid regions of China, and it is also one of the most serious desertified areas in China. The ecological environment is extremely fragile. Aeolian desertification of this region has been a longstanding hotspot for desertification research (Li et al., 2012; Liu et al., 2011). 2.2. Data sources and preprocessing 2.2.1. RS data MODIS standard land products provided by the Earth Observing System (EOS) of the National Aeronautics and Space Administration

Fig. 1. Sketch map of the Horqin Sandy Land based on elevation.

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

(NASA) were used in this research. These products include MOD 13A2 for the inversion and calculation of FVC, MSAVI, and MTVDI; MOD 11A2 for the extraction of LST; and MCD 43B3 for the extraction of albedo (Table 1). The track numbers of MODIS products in the study area were h26v04 and h27v04. The data acquisition time was the growing season (June to September) from 2000 to 2015. The original format of MODIS products is HDF-EOS, and the projection is sinusoidal. To facilitate the statistical analysis of the classified results and make the results correspond with field data, we performed format conversion, projection transformation, resampling, and image mosaic of MODIS products via the MODIS Reprojection Tool (MRT). The projection of products after transformation is the Albers Equal Area, the datum is WGS-84, and the output format is GEOTIF. Because the data set has thousands of images, the batch processing of the large amounts of data was accomplished by writing a batch script (*.bat) to call the MRT function. MODIS products were then cut by using the vector boundary of the study area. The maximum value composite (MVC) method was used to eliminate the influence of clouds and atmosphere. This process produced monthly products of each indicator in the growing season from 2000 to 2015. Annual MSAVI was synthesized based on the cumulative monthly data of the growing season; annual NDVI was generated from the monthly NDVI of the growing season via the MVC method; and annual albedo, LST, and MTVDI were synthesized from their monthly products in the growing season by averaging. Finally, we established the annual time series data set of each index and indicator from 2000 to 2015. This was used to extract the information on aeolian desertification. To ensure the classification accuracy of aeolian desertification, this study also utilized Landsat TM and Google Earth high-resolution RS data. Training samples were selected and the classification accuracy was verified based on the results of TM high-resolution classification, Google Earth images, and field investigation. The combination of high and medium resolution RS data improved the accuracy of aeolian desertification classification and increased the temporal continuity of classified results. 2.2.2. Field data Because ADL exhibits different features in different regions, significant differences exist between the monitoring indicators. To test the spatial difference and accuracy of each indicator obtained by RS, monitoring indicators were grouped and separately assessed according to surface features in the study area. On this basis, we divided ADL into grassland, shrub, and cropland desertification. We also divided ADL into different degrees, including slight, moderate, severe, and extremely severe desertification. The field work was carried out from 5th August to 3rd September in 2011 and from 5th August to 25th August in 2015. The method of random sampling was chosen to set the sampling sites, and the sampling area was 1 km × 1 km. Global Position System (GPS) units were used to locate the latitude and longitude of the sampling points. Starting from the center point, we sampled along four directions. Each directional transect contained five points, the interval between two sampling points was 100 m, and each sampling site contained 21 points (Fig. 2). The values for the 21 sampling points were averaged for each site to assess the accuracy of RS data. The actual albedo was measured with a radiometer (MR–32, EKO Instruments Ltd., Japan; spectral range: 0.30–2.80 μm, sensitivity: 7 mV/kW∙m−2), the actual

2377

LST was measured by using an Infrared Thermometer (Fluke62MAX, Fluke Corporation, USA; temperature range: −30 to 650 °C, accuracy: ±1.00 °C or ±1.00% of reading), the actual soil moisture (0–10 cm) was detected by using Time-Domain Reflectometry (TDR 100 Campbell Scientific, Inc., USA; range: 0–saturation, accuracy: ±3.00% volumetric water content), and the actual FVC was measured by using a digital camera (EOS 350D, Canon Inc., Japan; effective pixels: 8 million, maximum resolution: 3456 × 2304 pixels) (Fig. 2). In this study, we randomly selected 40 sampling sites in the study area. ArcGIS was used to extract the pixel values of the monitoring indicators that correspond to the measured sampling points, and their correlation was analyzed to ensure the usability of each indicator. 2.2.3. Meteorological and statistical data The meteorological data used in this research obtained from the China Meteorological Data Sharing Service System (http://cdc.cma.gov.cn/). Nine meteorological stations were included in the data set: Ongniud Qi, Baoguo Tu, Naiman Qi, Bairin Zuoqi, Kailu, Tongliao, Jarud Qi, Bayal Tuhu Shuo, and Tuquan. The selected data include the monthly temperature, precipitation, and wind velocity from 2000 to 2015, and the annual meteorological data were synthesized from the monthly data. On this basis, the impacts of climate change on aeolian desertification were analyzed. The statistical data came from the Inner Mongolia Autonomous Region Bureau of Statistics from 2000 to 2015, covering 13 counties of the study area (Tuquan County and Huolin Gole City were excluded because of the small amount of ADL in these two regions). These counties belong to Tongliao City, Chifeng City, and Xing'an League at the administrative division levels. The social data used in this study included the total number of population and livestock at the end of the year, which are two major human factors affecting land desertification. 2.3. Methods 2.3.1. Grading basis of aeolian desertification The classification of desertification degree has important practical significance for understanding the temporal and spatial variation of ADL and for selecting the corresponding control measures. Therefore, a systematic and feasible grading standard for aeolian desertification is required. On the basis of the grading standard proposed in previous studies (Han et al., 2010; Wang et al., 2004; Yan et al., 2007), the desertification intensities were defined as extremely severe, severe, moderate, and slight according to the vegetation coverage and the ratios of shifting sand dunes (sand sheets). To make the extraction of land desertification more objective and accurate, we selected additional monitoring indicators (i.e., MSAVI, Albedo, LST, and MTVDI) based on this classification system. We then divided the ADL into the same four grades. 2.3.2. Selection of monitoring indicators for aeolian desertification Selection of monitoring indicators mainly follows the principles of relevance, reversibility, reliability, applicability, irreplaceability, sensitivity, and zonality (Liu, 2004). Under the guidance of these selection principles, monitoring indicators were selected based on the existing studies of indicator systems for aeolian desertification monitoring (Han et al., 2015; Lamchin et al., 2017; Liu et al., 2015). The selected monitoring indicators reflected the characteristics of different desertification types. The indicators could be retrieved by RS images to rapidly

Table 1 Detailed information of MODIS products (https://lpdaac.usgs.gov/dataset_discovery/modis/modis_products_table). Name

Land-surface parameters

Temporal resolution

Spatial resolution

Level

MOD 13A2 MOD 11A2 MCD 43B3

Vegetation indices Temperature, emissivity BRDF and Albedo

16-day 8-day 16-day

1 km 1 km 1 km

L3 L3 L4

2378

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

Fig. 2. Schematic diagram of field validation of monitoring indicators (including the Albedo, LST, Soil moisture, and FVC).

update monitoring systems of macroscopic land desertification patterns. Albedo, LST, MTVDI were the indicators selected to reflect the physical properties of ADL. FVC and MSAVI were chosen to reflect the natural properties and ecological conditions of ADL. These indicators are quantitative indicators and have few artificial interventions. Therefore, we could monitor and evaluate the aeolian desertification objectively in the process of inversion. 2.3.3. Inversion of monitoring indicators for aeolian desertification Based on MODIS data and products, the five monitoring indicators of aeolian desertification were extracted and inversed to obtain the time series data sets of each indicator from 2000 to 2015. (1) Calculation of MSAVI

A vegetation index is a numeric value indicating vegetation growth potential and biomass, and it is composed of multispectral data analysis and numerical operations (Zhao, 2003). Vegetation indices provide an important reflection of vegetation productivity in the desertification process and are the most important indicator of desertification development and reversal (Pang et al., 2012). MSAVI is one of the more widely used vegetation indices. The results of a study by Qi et al. (1994) showed that MSAVI reduced the impact of the soil background and enhanced the sensitivity to vegetation compared with other vegetation indices that were included in the study. Because of the low vegetation coverage in desertified areas, it is necessary to adopt a vegetation index that can eliminate the impact of soil background. Therefore, this study selected MSAVI as one of the indicators of aeolian desertification monitoring, and its formula is as follows: MSAVI ¼

 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2NIR þ 1− ð2NIR þ 1Þ2 −8ðNIR−RÞ =2

ð1Þ

where, NIR is the reflectivity of the near-infrared band, and R is the reflectivity of the red band. The 16-day MSAVI synthetic product was calculated based on Formula (1) and the monthly MSAVI was generated via the MVC method. The MSAVI values of the growing season (June to September) were accumulated to represent the annual MSAVI of the study area. (2) Calculation of FVC

FVC is the percentage of the vertically projected area of vegetation (including leaves, stems, and branches) with reference to the total statistical area (Jia et al., 2016; Song et al., 2017). It is an important indicator of surface vegetation status and has been used as an input parameter in land desertification assessment, soil erosion monitoring, and distributed hydrological modeling (Dymond et al., 1992; Li et al., 2004). The pixel dichotomy model is one of the simplest methods for mixed pixel decomposition that has been widely applied. Thus, the pixel dichotomy model (Jia et al., 2016) was used to calculate the fractional vegetation cover of the study area, and its formula is as follows:

fc ¼

NDVI−NDVIsoil NDVIveg −NDVIsoil

ð2Þ

where, fc is the fraction vegetation cover, NDVIsoil is the NDVI value of the bare soil pixels, and NDVIveg is the NDVI value of the pure vegetation covered pixels. Because of the impact of related factors, such as surface conditions and vegetation types, the values of NDVIsoil and NDVIveg also change with time and space. According to NDVI frequency statistics and field test results, the NDVI value with the cumulative frequency of 0.50% is the NDVIsoil value, and the cumulative frequency of 99.50% is the NDVIveg value.

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

(3) Extraction of Albedo

Land surface albedo characterizes the ability of the Earth's surface to reflect solar radiation. It is the ratio of the solar radiation emitted by the surface to the solar radiation energy reaching the surface. It is an important parameter of surface energy balance (Dickson, 1995). An increase of albedo in semi-arid areas leads to a decrease in solar radiation absorbed by the surface and a reduction of atmospheric convection. This results in a reduction in precipitation (Ma et al., 2007). An increase in albedo enhances aeolian desertification, and aeolian desertification also increases albedo. Therefore, albedo can be used as one of the indicators of aeolian desertification monitoring. The albedo product of MODIS (MCD 43B3), based on the bidirectional reflective distribution function (BRDF), includes the black-sky albedo (BSA) and white-sky albedo (WSA) of seven narrow bands and three wide bands (0.30 to 0.70 μm in the visible band, 0.70 to 5.00 μm in the near-infrared band, and 0.30 to 5.00 μm in the shortwave band). Because the WSA is the integral of the incident angles, it is closer to the surface albedo in the general sense. Thus, we selected the WSA of the MCD 43B3 product as the surface albedo from 2000 to 2015. We then used the MVC method and mean method to obtain the monthly albedo and annual growing season albedo, respectively, for aeolian desertification extraction. (4) Extraction of LST

LST is an important parameter to describe the material exchange and energy balance between the land surface and atmosphere. It is also a key factor in physical processes with global effects. It has important significance in climate change, the hydrological cycle, and the ecological environment (Qin et al., 2001). In this study, the MOD 11A2 land surface temperature product provided by NASA was selected, including the daytime surface temperature, nighttime surface temperature, and emissivity at 31 and 32 bands. The LST was calculated by establishing the split-window algorithm based on the bright-temperature linear combination of the 31 and 32 channels. The channel bright-temperature was determined by the radiance and 0.10 K threshold bright-temperature table. The emissivity required during the LST calculation was determined by the MODIS land cover product. The synthetic method was used to calculate the average value of the sky surface temperature within eight days. On the basis of the eight-day LST product, the monthly and annual LST were synthesized, respectively, by using MVC and averaging.

then extract the desertification information by using MTVDI. The formula of MTVDI is as follows:

MTVDI ¼

T s −ða1 þ b1  MSAVIÞ ða2 þ b2  MSAVI Þ−ða1 þ b1  MSAVIÞ

ð3Þ

where, Ts is the LST of any pixel and a1, a2, b1, and b2 are the regression coefficients. The a and b coefficients correspond to the intercept and slope of the dry and wet side equations in the Ts-MSAVI feature space, respectively. MTVDI is calculated via Formula (3), and its value is between 0 and 1. Higher MTVDI values indicate lower soil moisture and a drier land surface, and vice versa. 2.3.4. Decision tree classification A decision tree is a mathematical method used to classify novel data into homogenous subgroups based on decision rules that are generated through training sample induction and learning (Akkas et al., 2015). It exhibits a flowchart-like structure that consists of a root tree, a series of internal nodes, and leaf nodes (Liu et al., 2005). Each internal node possesses only one parent node and two or more child nodes, with each of the leaf nodes representing a class label that signifies the classification result (Fig. 3). The decision tree starts from the root node and tests the data samples, and the data samples are divided into different subsets according to different results. Each subset continues to be divided into new subnodes until it reaches a termination criterion (Na et al., 2009). The decision tree method is divided into two processes that contain decision tree learning and classification. The learning process is a machine learning process that generates classified rules by analyzing training samples. Classification then can be performed on the unknown samples based on the rules generated in the learning process (Jiang et al., 2011). The QUEST (quick, unbiased, and efficient statistical tree) algorithm is a binary classification method proposed by Loh and Shih to establish a decision tree (Loh and Shih, 1997). The fundamental idea of QUEST is to select the branch variables and split points with different strategies. The QUEST algorithm is applicable to both continuous and discrete variables. In contrast to other general decision tree algorithms that are more likely to select predictive variables with more potential points, the construction of the QUEST decision tree is basically unbiased in the selection of variables. QUEST can distinguish the category members and noncategory members through the hyperplanes constituted by multiple variables in the feature space. Its operation speed and classification accuracy are superior to other decision trees. Therefore, a decision tree classification method based on the QUEST algorithm was used to extract the desertification information in the ENVI 5.3 (64-bit) software environment.

(5) Calculation of MTVDI

Soil moisture is an important environmental factor used to study soil moisture conditions and its relation to vegetation. It plays an important role in material and energy exchange between the land surface and atmosphere, and it is of great importance for global change and desertification research (Jiao et al., 2016). TVDI is established based on the relationship between LST and the vegetation index, and it can provide a good estimate of the moisture status of the soil surface. TVDI is the ratio obtained directly from the T s-NDVI feature space, which considers both changes in the vegetation index and changes in the LST corresponding to the same vegetation index. This index shows only the relative state of soil moisture. Compared with NDVI, MSAVI can better describe the vegetation cover and soil background. Therefore, this study used MSAVI instead of NDVI to construct the Ts -MSAVI feature space to reflect the soil moisture status in the study area. We

2379

Fig. 3. Typical structure of a simple decision tree.

2380

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

To ensure the accuracy and irreplaceability of the monitoring indicators, field data was used to verify their accuracy. All the indicators were then overlaid and combined, and the QUEST decision tree was applied to extract the ADL information via the same training samples.

Table 2 Accuracy assessment of aeolian desertified extraction based on the QUEST classified method. Desertification degree

Extremely severe

Severe Moderate Slight Non

3. Results 3.1. Accuracy verification of ADL monitoring indicators Satellite products and their associated inversion indicators need to be verified with ground observations. This guarantees data quality and also can improve the inversion results (Yu et al., 2010). By analyzing the correlation between desertification monitoring indicators and their corresponding ground-truth values, we obtained a scatterplot of the measured value and the inversion value of each indicator. As shown in the scatterplots (Fig. 4), a significant correlation exists between the estimated values and measured values of each indicator. Their R2 values were 0.77, 0.68, 0.66, and 0.39 dividedly, and correlations were significant at 0.01 levels (p b 0.01). In particular, 97.89% of the absolute deviation for albedo between estimated values and measured values were within the range of ±0.05. Therefore, all of the indicators obtained from related models exhibited high accuracy and matched the research requirements for aeolian desertification extraction. On the basis of these monitoring indicators, we extracted the ADL by using the QUEST decision tree classification method. To test the reliability of the QUEST classified accuracy, this paper utilized field data and desertification vector data collected during 2010. This data set utilized Landsat TM images as the reference data set. Sixty-six ROIs (regions of interest) of different desertification types were selected to construct the confusion matrix for the evaluation of the desertification classified accuracy. This enabled us to use high-resolution classified results to verify the low-resolution classification results. As shown in Table 2, the overall classification accuracy of aeolian desertification was 96.50%, and the Kappa coefficient was 0.90. The classification accuracy of extremely severe desertification, severe desertification, moderate

Extremely 215 18 severe Severe 1 205 Moderate 0 42 Slight 0 0 Non 0 0 Total 216 265 Prod. acc. (%) 99.54 77.36 Overall acc. = 96.50%; Kappa = 0.90

Total User acc. (%)

0

0

0

233

92.27

0 211 25 3 239 88.28

0 0 72 45 117 61.54

0 0 1 3026 3027 99.97

206 253 98 3074 3864 –

99.51 83.40 73.47 98.44 – –

desertification, and slight desertification were 92.27%, 99.51%, 83.40%, and 73.47%, respectively. In summary, the use of the QUEST classification method exhibited a high accuracy in effectively classifying the aeolian desertification, and it able achieved the research requirements. 3.2. Change trends of ADL monitoring indicators The annual change trends for each of the ADL monitoring indicators during 2000–2015 are shown in Fig. 5. According to the figure, MSAVI and FVC showed an overall increasing trend in volatility from 2000 to 2015, but the trend was not significant (p N 0.05). LST, albedo, and MTVDI all showed a decreasing trend. The decreasing trend of albedo and MTVDI was significant (p b 0.01 and p b 0.05). The decreasing trend for LST, however, was not significant. MSAVI and FVC both exhibited local maxima in 2005, 2008, and 2012, whereas LST and albedo presented local minima in the corresponding years. This indicated that the LST and albedo decreased with increasing vegetation coverage. MTVDI showed local minima in 2005 and 2008, which was similar to the albedo trend, but MTVDI exhibited a peak in 2012, which was in contrast to the previous two years. The change of MTVDI, reflecting land surface

Fig. 4. Relationship between estimated value and measured value of (a) FVC, (b) Albedo, (c) LST, and (d) soil moisture.

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

2381

Fig. 5. The annual trends of ADL monitoring indicators from 2000 to 2015: (a) LST and Albedo; (b) MSAVI and FVC; (c) MTVDI.

dryness, showed a downward trend with increasing vegetation in 2005 and 2008. However, the vegetation coverage increased after the surface became drier in 2012. This indicated that the impact of soil moisture on

vegetation coverage weakened during this period. Therefore, we indirectly inferred that this increase in vegetation coverage may have been caused by human factors. The same result was found in 2007

Fig. 6. Spatial distribution of ADL in the Horqin Sandy Land in 2015.

2382

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

and 2009, that is, MSAVI and FVC both exhibited local minima, whereas albedo and MTVDI showed peaks. This meant that albedo increased with a reduction in vegetation coverage, and the amount of surface drought intensified. Thus, we identified a corresponding relationship of vegetation coverage with LST, albedo, and MTVDI from this analysis. 3.3. Current status and spatial distribution of ADL in 2015 Based on the accuracy assessment of ADL extraction, the decision tree classification method was used to extract the aeolian desertification in the Horqin Sandy Land in 2015, and the current status and spatial distribution characteristics of ADL were obtained (Fig. 6). As shown in Fig. 6, the ADL of the Horqin Sandy Land is mainly distributed in the administrative regions of Barin Youqi, Ar Horqin Qi, Ongniud Qi, Naiman Qi, Hure Qi, and Horqin Zuoyi Houqi. The desertification degree of Ongniud Qi, Naiman Qi, and Hure Qi are more serious. Extremely severe desertified land is distributed mainly in the mid-eastern part of Ongniud Qi and the northwest part of Naiman Qi. These two regions also are located along both sides of the Laoha River, which exhibits an area of extremely severe desertified land that accounts for 86.51% of the total area of desertified land in the study area. An obvious belt of extremely severe desertified land also is found in the northern part of Hure Qi. Therefore, these counties are key regions for aeolian desertification control. Severe desertified land is distributed mainly along the periphery of extremely severe desertified lands, and these areas typically represent the transitional zones between severe desertified land and moderate desertified land. A large area of moderate and slight desertified land is distributed mainly along both sides of the Xiliao River. The ADL on the south bank of Xiliao River shows a regular east-to-west distribution and a southto-north arrangement. The distribution of ADL in Horqin Zuoyi Houqi is especially typical of this pattern. The ADL on the north bank of the Xiliao River is distributed mainly in Horqin Zuoyi Zhongqi, Horqin Youyi Zhongqi, and Ar Horqin Qi. Over time, the transformation between slight desertified land and moderate desertified land is the most obvious pattern, and the two categories of ADL are scattered in each county of the study area. The total area of ADL in the Horqin Sandy Land was 32,633 km2 at the end of 2015, accounting for 26.02% of the total area (Table 3). The areas of slight, moderate, severe, and extremely severe desertified land are 16,769 km 2 , 11,130 km2, 3363 km2 , and 1371 km 2 , respectively. These areas account for 51.39%, 34.11%, 10.31%, and 4.20% of the total area of ADL. Slightly desertified land is the most widely distributed. The area of ADL of Ongniud Qi is 5878 km2 which accounts for 18.01% of the total area of ADL in the study area. This is the largest proportion of the total study area.

Ongniud is followed by Ar Horqin Qi and Barin Youqi, which have 4257 km2 and 4016 km2 , respectively, accounting for 13.05% and 12.31% of the total area of ADL in the study area. The proportion of ADL to the total area of the county is highest in Ongniud Qi at 49.53%. This proportion is closely followed by Hure Qi and Naiman Qi with desertified areas that account for 47.95% and 45.71% of the total area of their counties, respectively. The proportion of extremely severe and severe desertified land in these three counties was higher than that in other counties. Although the degree of ADL in Ar Horqin Qi is comparatively lower, the area is extensive and the situation remains serious. As a typical sandy grassland and national nature reserve, the local government should give this issue greater attention. 3.4. Construction of time series dataset and spatial-temporal dynamic changes of ADL Based on the time series dataset of aeolian desertification monitoring indicators, the decision tree classification method was used to quantitatively extract the desertification information from 2000 to 2015. Then the time series dataset of ADL was constructed in the Horqin Sandy Land for the 16 year period. We identified the spatial distribution of ADL in different years (Fig. 7), and its range was similar in each year. These ADL results based on the multi-indicator quantitative inversion were compared with high-resolution ADL classification obtained by human–computer interactive visual interpretation (Duan et al., 2014). The comparison showed that our results exhibited high consistency in the overall spatial distribution and different degrees of desertified land distribution. This finding suggested that the multi-indicator inversion technique based on the QUEST classified method was effective in accurately extracting and monitoring the dynamics of aeolian desertification. We acquired the areas and trends for different degrees of ADL in each year (Table 4, Fig. 7). As shown in Table 4, the total area of ADL in the study area was the largest in 2002, followed by 2001. The year with the smallest ADL area was 2012. The years with higher extremely severe desertified land were 2000 and 2007 (i.e., N2000 km2 in both years). The year with the smallest extremely severe desertified land was 2011, with an area of 664 km2. The year with the highest severe desertified land was 2000, followed by 2007. The year with the highest moderate desertified land was 2002, with an area of 41,184 km2, accounting for 58.59% of the total area of ADL in 2002. This exceeded the total desertification area of some years. The year with the smallest moderate desertified land was 2012, accounting for b30% of the total area of ADL in 2012. The year with the highest slightly desertified land was 2008, which accounts for 59.37% of the total area of ADL in that year.

Table 3 Distribution of ADL of each county in 2015. Counties

Total area/km2

Extremely severe/km2

Severe/km2

Moderate/km2

Slight/km2

Total area of desertified land/km2

Percentage of desertified land/%

Aohan Qi Ar Horqin Qi Barin Youqi Barin Zuoqi Kailu Hure Qi Horqin Youyi Zhongqi Horqin Zuoyi Houqi Horqin Zuoyi Zhongqi Naiman Qi Tongliao Ongniud Qi Jarud Qi Huolinguole Tuquan Total

8288 12,868 9827 6417 4264 4649 12,528 11,608 9525 8204 3422 11,868 16,545 702 4700 125,415

14 3 32 0 0 113 1 12 5 174 0 1012 5 0 0 1371

38 141 227 0 25 327 189 116 125 373 0 1701 101 0 0 3363

641 2047 1838 74 166 547 352 1035 300 1998 3 1811 315 0 3 11,130

1396 2066 1919 405 503 1242 1419 2260 1586 1205 186 1354 1201 1 26 16,769

2089 4257 4016 479 694 2229 1961 3423 2016 3750 189 5878 1622 1 29 32,633

25.21 33.08 40.87 7.46 16.28 47.95 15.65 29.49 21.17 45.71 5.52 49.53 9.80 0.14 0.62 26.02

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

2383

Fig. 7. Spatial distribution of ADL in the Horqin Sandy Land during 2000–2015.

The year with the minimum slightly desertified land was 2000, accounting for only 24.13% of the total area of ADL in that year. Viewed from the degree of ADL, the desertified degree was relatively serious in 2000 and 2007. During this period, the proportion of severe and extremely severe desertified land was larger than that of other desertified degrees.

In the study area, we observed that the total area of ADL showed a decreasing trend from 2000 to 2015. The rate of decline was 2388.60 km2 y−1 (Fig. 8). The decline of ADL mainly was dominated by moderate and slight desertified land, in which the moderate desertified land exhibited a more significant decrease. The overall

2384

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

Table 4 Area of different ADL in Horqin Sandy Land during 2000–2015. Year

Extremely severe/km2

Severe/km2

Moderate/km2

Slight/km2

Total area of ADL/km2

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015

2270 1950 1846 1452 1618 1353 1529 2229 1426 1632 1262 664 755 1134 1139 1371

6041 4329 4268 3206 2973 2958 3836 4992 3790 4405 4364 3737 1714 1993 2245 3363

31,822 40,504 41,184 26,134 26,446 19,146 21,597 37,115 15,862 29,816 17,692 24,212 6310 7078 11,549 11,130

12,761 20,166 22,996 24,512 24,929 23,316 25,718 15,125 30,803 15,962 18,551 22,732 13,791 16,173 15,433 16,769

52,894 66,949 70,294 55,304 55,966 46,773 52,680 59,461 51,881 51,815 41,869 51,345 22,570 26,378 30,366 32,633

Fig. 9. Change trend of ADL and its three-year moving average from 2000 to 2015 (T1: 2000–2005, T2: 2005–2011, T3: 2011–2015).

to this period had not changed significantly. In T3, the ADL again showed a significant downward trend, and the precipitation during this period also increased. Therefore, we inferred that although the ecological protection project has controlled the expansion of desertification, sensitive climatic factors, especially precipitation, are affecting the frequent fluctuation of aeolian desertification.

trends of severe and extremely severe desertified land were not obvious. As shown in Fig. 8, the trends in slight and moderate desertified land basically were opposed in many years. The area change trend of moderate desertified land was most similar to the trend of total ADL area, followed by the slight desertified land. The dynamic of the total area of ADL was affected obviously by slight and moderate desertified land. Viewed from the yearly trend of ADL, all types of ADL from 2000 to 2012 showed a decreasing trend. The largest decline was from 2011 to 2012, reaching the lowest value in 2012.The trends of all types of ADL rebounded after 2012, showing gradual upward trends. To avoid the impact of extreme weather and climatic events, we adopted the three-year moving average and time segmentation methods for further analysis of aeolian desertification. The variation of ADL was discussed in three different periods (Fig. 9). The first period was T1 (2000–2005), the second period was T2 (2005–2011), and the third period was T3 (2011–2015). By comparing the change trend of original ADL and the trend of ADL after moving average, we found that the overall trend of both was consistent. The fluctuation range after the moving average was significantly lower than the original fluctuation range, which had a good smoothing effect on the year of extreme climate. Combined with climatic factors, the trend analysis revealed that during T1, the ADL showed a significant downward trend in the study area and precipitation increased significantly during the same period, indicating that the increase of precipitation played a positive role in the reversal of ADL during this period. In T2, the change trend of ADL was relatively gentle, and the precipitation corresponding

3.5.1. Climatic factors The results showed that the annual mean temperature increased slowly from 2000 to 2007. The rate of increase was 0.13 °C y−1, and it reached its maximum value of 8.32 °C in 2007. After 2007 it declined significantly, reaching its minimum value of 5.90 °C in 2012 with a decreasing rate of 0.43 °C y−1. After 2012, the temperature again increased significantly (Fig. 10). Therefore, the annual mean temperature in the study area has experienced a process of slow increase, significant decline, and then rapid increase. However, the overall trend in temperature was not obvious. The precipitation in the study area fluctuated upward overall for the 16 study years (Fig. 10), and its growth rate was 8.06 mm y−1. The annual precipitation in the study area was 327.68 mm. The annual potential evaporation was within the range of 1500 to 2500 mm, which was nearly five to eight times the precipitation. The growth of precipitation from 2000 to 2011 was relatively stable. An obvious peak occurred in 2012 when the precipitation reached 489.85 mm, which was more

Fig. 8. The change trends of different degrees of ADL in the Horqin Sandy Land from 2000 to 2015.

Fig. 10. Change trends of annual precipitation and annual mean temperature from 2000 to 2015 in the Horqin Sandy Land.

3.5. Causes of aeolian desertification

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

than two times the level in the year with the least precipitation (2009). The large annual precipitation change rate easily could cause frequent drought and flood disasters. The increase of precipitation in the study area created favorable natural conditions for the reversal of land desertification. The annual mean and annual maximum wind velocity in the study area exhibited significant overall declines, with decreasing rates of 0.56 m s−1 y−1 and 0.03 m s−1 y−1, respectively (Fig. 11). From 2000 to 2004, the annual mean wind velocity showed a slight increasing trend. The annual maximum wind velocity decreased slowly during the same period. However, their overall trends were not obvious. Therefore, the variation of wind velocity during this period did not play a significant role in promoting the reversal of land desertification. After 2004, the annual mean and maximum wind velocity showed a decreasing trend. The decrease in the wind velocity weakened the wind erosion, which was beneficial to the control and reversal of aeolian desertification. We found that the change of ADL has significant correlations with annual precipitation, annual mean wind velocity, and annual maximum wind velocity (p b 0.01; Table 5). However, ADL had no obvious correlation with annual mean temperature (p N 0.01; Table 5). To avoid including variables with low explanatory power, we used stepwise multiple linear regression to determine the regression equation with the driving factors that significantly affected ADL. Taking the area of ADL as the dependent variable and the four climatic factors as the independent variable, we applied a stepwise method with an F-test criterion (Sig. b 0.05) in SPSS Statistics 19. The related parameters are given in Table 6. The final optimization equation is as follows: Y ¼ 3:60 þ 0:31X3 −0:01X1

ð4Þ

where, Y is the area of ADL, X3 is the annual maximum wind velocity, and X1 is the annual precipitation. The regression equation has a coefficient of determination (R2) of 0.89, and it passed the 1% significance test (Sig. b 0.01). The result indicated that the impacts of annual maximum wind velocity and annual precipitation on ADL were greater than other climatic factors, making them the two most obvious climatic factors affecting land desertification. The regression coefficient of annual maximum wind velocity was 0.31, which was greater than that of annual precipitation (−0.01, Table 6). This indicated that the impact of annual maximum wind velocity on ADL was more important than annual precipitation. 3.5.2. Human factors Based on the analysis of existing research on aeolian desertification and its driving factors in conjunction with the actual status of the

2385

Table 5 Correlation between the areas of ADL and climatic factors. Desertification types

Annual mean temperature

Annual precipitation

Annual mean wind velocity

Annual maximum wind velocity

Extremely severe Severe Moderate Slight Total area

0.37 0.13 0.26 0.25 0.32

−0.67⁎⁎ −0.83⁎⁎ −0.84⁎⁎ −0.09 −0.80⁎⁎

0.57⁎ 0.43 0.68⁎⁎ 0.33 0.72⁎⁎

0.65⁎⁎ 0.54⁎ 0.83⁎⁎ 0.39 0.88⁎⁎

⁎ indicates a significant test by 0.05, ⁎⁎ indicates a significant test by 0.01.

study area and the availability of statistical data, we selected population density and the number of livestock to evaluate the effects of human factors on aeolian desertification. Population density was calculated, and its changing trends were analyzed from 2000 to 2015 (Fig. 12). The results show that the population density increased significantly during the study period (p b 0.01). It increased from 42.50 persons/km2 to 44.20 persons/km2. The population density reached a maximum of 44.70 persons/km2 in 2011. The average population density of the whole area during the study period was 43.80 persons/km2, which far exceeded the standard of 20 persons/km2 for semi-arid areas proposed by the United Nations. The increase in population density leads to increases in population pressure and demand for all kinds of resources. In principle, these trends are adverse to the reversal of ADL. However, the change trend of ADL showed that the ADL did not expand and instead showed a reversal trend from 2000 to 2015. This trend has two likely reasons. First, the local government has strengthened its efforts to combat aeolian desertification since 2000. The implementation of a series of ecological projects has played a positive role in promoting the reversal of aeolian desertification. Second, the acceleration of urbanization in China has resulted in reduced population pressures in the relatively remote desertification areas and has created favorable conditions for the reversal of aeolian desertification. Overgrazing is another human factor that cannot be ignored in the process of aeolian desertification (Bo et al., 2013). Increases in the number of livestock lead to the excessive use of grassland, which results in the decline of vegetation cover and grassland degradation. Highintensity grazing and trampling of livestock destroys the structure of the land surface and facilitates aeolian desertification. The number of livestock in the study area showed an increasing trend from 2000 to 2015, with an annual mean growth rate of 74.85 × 104 heads y−1 (Fig. 12). The most rapid growth period for the number of livestock was from 2000 to 2007. The growth rate in this period reached up to 142.70 × 104 heads y−1, which was nearly two times that of the overall growth rate of livestock in the study area. The upward trend has been slowing since 2007, with a growth rate of 16.58 × 104 heads y−1. However, the ADL did not expand with the increase of the number of livestock. Field investigation indicated that this was mainly due to a limitation imposed by the local government on the capacity of pasture grazing. The government has implemented a regime of grassland

Table 6 Result of stepwise multiple linear regression analysis. Model

Fig. 11. Change trends of annual mean wind velocity and annual maximum wind velocity from 2000 to 2015 in the Horqin Sandy Land.

(Constant) Annual maximum wind velocity (X3) Annual precipitation (X1)

Unstandardized coefficients B

Std. error

3.60 0.31

1.41 0.06

−0.01 0.00

Standardized coefficients

t

Sig.

0.62

2.56 5.44

0.02 0.00

−0.43

−3.72 0.00

2386

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

Fig. 12. Changes in population density and the number of livestock from 2000 to 2015 in the Horqin Sandy Land.

enclosure and confinement feeding, and this has resulted in human produced material replacing natural pastures as forage. These measures have restored and improved the ecological functions of the pasture, thereby curbing the further expansion of aeolian desertification. The effect of relevant ecological projects carried out by the government in the study area was also evaluated. This data was obtained from the Three North Shelter Forest system construction network of China (http://tnsf.forestry.gov.cn/site/1/html/zxfw//list_1075.htm). The larger ecological projects mainly include the Beijing-Tianjin sandstorm source control project, a project returning farmland to forestland or grassland, and the Three North Shelterbelt Project (TNSP). The Horqin Sandy Land is the key project area of TNSP, and therefore the effectiveness of the project is relevant to our study. According to the fifth national monitoring results of desertification, the expanded trend of aeolian desertification in the Horqin Sandy Land has been reversed completely. The area of ADL was reduced and the severity degree was alleviated. The six counties on which the project focused were Tuquan, Horqin Youyi Zhongqi, Jarud Qi, Naiman Qi, Hure Qi, and Kailu. Since 2001, 2.85 × 104 ha of the project afforestation was completed in Tuquan County and 3.64 × 104 ha was implemented in Horqin Youyi Zhongqi. By the end of 2014, a total of 7.65 × 104 ha of shelter forest was established in Jarud Qi and 3.82 × 104 ha of ADL was controlled. From 2001 to 2015, Naiman Qi completed 4.49 × 104 ha of the reforestation task and 0.80 × 104 ha of trees had been planted to fix the sand. By the end of 2014, 7.90 × 104 ha of trees had been planted to prevent the wind in Hure Qi. The effect of sand fixation was significant, and the land desertification was curbed effectively. Since 2001, a total of 2.85 × 104 ha of the afforestation project was completed in Kailu County, and 0.55 × 104 ha of trees (or grass) had been planted to fix the sand. Different counties in the TNSP have different emphases. Tuquan, Horqin Youyi Zhongqi, and Jarud Qi are the key areas of grassland protection. Naiman and Hure Qi are the main areas of aeolian desertification control. Kailu is the main area of farmland protection. Statistical results indicated that the implementation of ecological protection projects has had remarkable effects, and these projects have played an important role in promoting the improvement of the ecological environment. Thus, it is evident that human activities can have a positive effect on aeolian desertification through ecological protection projects. 4. Discussion Land aeolian desertification is the comprehensive outcome of energy dynamics and landscape factors under the effects of natural and human influences (Guo et al., 2005). Many indicators can be used for aeolian desertification monitoring and evaluation, but studies traditionally

have deployed a method that utilizes a single indicator or qualitative description to evaluate the aeolian desertification. This practice severely limits the accuracy and practicality of research. The extraction of aeolian desertification based on multiple indicators has the ability to overcome the shortcomings and deficiencies of single indicator studies. Multiple indicators, which reflect the natural attributes and the ecological status of ADL, produce a more objective and accurate quantitative extraction of ADL. Based on multi-indicator extraction technology, the classification efficiency of ADL improved significantly. Furthermore, the current status and change information of ADL can be updated at any time. However, because the monitoring indicators are all dynamic, they are extremely sensitive to climate change and human activities. Under the influence of extreme climatic events, extreme weather will have a significant influence on changes in the monitoring indicators for the year, and then will make the ADL in that year fluctuate greatly, thus affecting the change trend of aeolian desertification throughout the entire monitoring period. Result based on multi-indicator provides an objective reflection of the reality of aeolian desertification. Some researchers believe that aeolian desertification is a relatively stable and long-term process, that large fluctuation in a short period is unreasonable, and that the general monitoring period should be three to five years. Because of the limitation of the availability of early RS images, annual images cannot be obtained for accurate monitoring of ADL. Therefore, the monitoring period generally is defined as three to five years. We determined the deterioration and reversal of ADL by analyzing dynamic changes in different periods. The disadvantage of this approach is that if the selected image is an extreme weather year or month, the trend of aeolian desertification may be completely opposite, which is also the main reason for the inconsistency of ADL monitoring results in the same area among many studies. At the same time, the interval of three to five years also makes it impossible to know the specific process and status of ADL during the monitoring period, and it is impossible to establish a good corresponding relationship with influencing factors when exploring its driving mechanism. Therefore, to avoid the misunderstandings caused by sharp fluctuations in the annual data of ADL on a short time scale, and to satisfy the stable status of ADL in a certain period, we adopted a combination of three-year moving average and time segmentation methods to discuss the dynamic process of ADL. The results showed that the change trend of ADL after the moving average was consistent with the original trend while also reflecting abundant ADL information on the annual scale. Its fluctuation range also reduced obviously, and the status was relatively stable. It was evident that this combination of time series monitoring and time segmentation monitoring better reflected the dynamic process of aeolian desertification, and thus exploration in this field should be strengthened in future work. Based on MODIS data, monitoring indicators of ADL were obtained by RS inversion. The time series data sets of ADL based on multiple indicators were then extracted. This method did not suffer from the short time scales found in land desertification monitoring based on highresolution RS data, such as Landsat TM/ETM/OLI. It also avoided the influence of extreme weather (e.g., extremes in drought and flood years) on the monitoring results. By analyzing previous monitoring results of ADL based on human-computer interactive visual interpretation and the dynamic transfer matrix of different degrees of ADL in different periods, we found that the classification results of ADL based on multiple indicators and human–computer visual interpretation exhibited good consistency in the overall spatial distribution. Furthermore, a close mutual transformation relationship existed between slight and moderate desertified land. The shift between these two lands caused the complementarity of slight and moderate desertified land. The time series monitoring results helped elucidate the process of land desertification change. Climate change is one of the main natural factors affecting the development and reversal of land desertification (Lam et al., 2011; Ren et al., 2016). Dry weather and high winds can lead to the occurrence and

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

development of aeolian desertification, but favorable water and thermal conditions in conjunction with reduced wind are conducive to the reversal of aeolian desertification. The relationships between the ADL trend and climatic factors were analyzed via a stepwise multiple linear regression method. The analysis showed that annual maximum wind velocity and annual precipitation were the two most obvious climatic factors affecting land desertification. During the monitoring period, the significant decrease in wind velocity and the obvious increase in precipitation provided favorable conditions for the reversal of aeolian desertification. In addition to climate factors, the impacts of human activities on aeolian desertification included two main aspects. On the one hand, unsustainable human activities destroyed the balance of natural ecosystems and triggered a series of environmental problems (Su et al., 2006). The ecological environment is inherently fragile in arid and semi-arid agropastoral transitional zones, and even a slight disturbance by human activities can destroy the original balance of the ecosystem and induce aeolian desertification. On the other hand, humans have directed a series of ecological protection projects and policies directed at land desertification control. These policies and projects have restrained the expansion of land desertification. However, the impact of human factors (e.g., policy) on aeolian desertification cannot be quantified due to the limited availability and accuracy of data. As a result of these limitations, it is difficult to quantitatively judge the contribution of human factors on land desertification. Therefore, we simply analyzed the changing trends in human factors based on the stepwise multiple linear regression analysis of climatic factors. We indirectly inferred the effect of human factors on aeolian desertification by combining the influence of climatic factors and the management effects of ecological protection projects in the study area. The quantitative impact of human factors on aeolian desertification should be the topic of future studies. 5. Conclusion The QUEST decision tree classification method was used to construct the ADL auto-extraction model based on a combination of multiple indicators. The time series data set of ADL was then extracted for the 2000 to 2015 period. The driving forces of aeolian desertification were discussed by combining the natural factors and human factors. The results showed that all indicators have high accuracy. These indicators can be incorporated into a monitoring system that can be updated rapidly to reflect extent and characteristics of aeolian desertification in real-time, and they can be used for the quantitative extraction of aeolian desertification. The overall classification accuracy based on the QUEST classifier was 96.50%, and the Kappa coefficient was 0.90. Compared with the human-computer interactive visual interpretation, this method has obvious advantages for ADL monitoring at large scales and over long time series. Time series monitoring avoids the influence of extreme weather years on monitoring results. The monitoring results indicate that the total area of ADL in the study area showed an obvious decreasing trend with a declining rate of 2388.60 km2 y−1. This decrease was mainly dominated by a decline in slight and moderate desertified land. The trends in severe and extremely severe desertified land were not obvious. This trend indicated that the moderate and slight desertified land had the greatest impact on the overall trend in total ADL. All degrees of ADL declined from 2000 to 2012. The trends then rebounded and increased gradually. Analysis of climatic and human factors showed that the effect of human activities on aeolian desertification was positive during this period. Therefore, we speculated that a warm and dry climate may be a primary reason for the rebound in ADL. The analysis of factors driving ADL showed that annual maximum wind velocity and annual precipitation are two most obvious climatic factors affecting land desertification. Furthermore, the impact of annual maximum wind velocity on ADL is more important than annual precipitation. Temperature had less of an effect on the aeolian desertification over the 16 year period. The overall trend of climate change is beneficial

2387

to the control and reversal of aeolian desertification. The area of ADL did not expand with the increase of population density and number of livestock. Rather, ADL showed a reversal trend from 2000 to 2015. Combined with this analysis of climatic factors, we preliminarily conclude that human activities played a positive role in the reversal of ADL. Our data suggested that the implementation of a series of ecological protection projects and optimization of related regimes carried out by local governments have curbed the expansion of ADL. As a result, the ecological environment has been effectively protected and improved.

Acknowledgments This work was supported by the National Natural Science Foundation of China (41401109), the Project of National Key Research and Development Program of China (2016YFC0500902), and the China Scholarship Council. We express our great thanks to the anonymous reviewers for their constructive comments and suggestions. References Akkas, E., et al., 2015. Application of Decision Tree Algorithm for classification and identification of natural minerals using SEM-EDS. Comput. Geosci. 80, 38–48. Badreldin, N., et al., 2014. Assessing the spatiotemporal dynamics of vegetation cover as an indicator of desertification in Egypt using multi-temporal MODIS satellite images. Arab. J. Geosci. 7, 4461–4475. Bo, T.L., et al., 2013. Modeling the impact of overgrazing on evolution process of grassland desertification. Aeolian Res. 9, 183–189. Cao, C.Y., et al., 2011. Spatial variability of soil nutrients and microbiological properties after the establishment of leguminous shrub Caragana microphylla Lam. plantation on sand dune in the Horqin Sandy Land of Northeast China. Ecol. Eng. 37, 1467–1475. Dickson, R.E., 1995. Land processes in climate models. Remote Sens. Environ. 51, 27–38. Duan, H.C., et al., 2014. Dynamics of aeolian desertification and its driving forces in the Horqin Sandy Land, Northern China. Environ. Monit. Assess. 186, 6083–6096. Dymond, J.R., et al., 1992. Percentage vegetation cover of a degrading rangeland from spot. Int. J. Remote Sens. 13, 1999–2007. Eckert, S., et al., 2015. Trend analysis of MODIS NDVI time series for detecting land degradation and regeneration in Mongolia. J. Arid Environ. 113, 16–28. FAO, 2006. FAO/IUSS Working Group WRB, World reference base for soil resources 2006. World Soil Resources Reports. 103. FAO, Rome. Feng, L.L., et al., 2016. The dynamic monitoring of aeolian desertification land distribution and its response to climate change in northern China. Sci. Rep. 6. https://doi.org/ 10.1038/srep39563. Ge, X.D., et al., 2016. Impact of land use intensity on sandy desertification: an evidence from Horqin Sandy Land, China. Ecol. Indic. 61, 346–358. Guo, J.Y., et al., 2005. Ecological research on shrub vegetation in Hunshandake Sandy Land. Inn. Mong. For. Sci. Technol. 1–4. Han, Z.W., et al., 2010. Change trends for desertified lands in the Horqin Sandy Land at the beginning of the twenty-first century. Environ. Earth Sci. 59, 1749–1757. Han, L.Y., et al., 2013. Desertification assessments of Hexi regions in Gansu province by remote sensing. Arid Land Geogr. 36, 131–138. Han, L.Y., et al., 2015. Desertification assessments in the Hexi corridor of northern China's Gansu Province by remote sensing. Nat. Hazards 75, 2715–2731. Helldén, U., 2008. A coupled human–environment model for desertification simulation and impact studies. Glob. Planet. Chang. 64, 158–168. http://cdc.cma.gov.cn/ http:// tnsf.forestry.gov.cn/site/1/html/zxfw//list_1075.htm https://lpdaac.usgs.gov/dataset_ discovery/modis/modis_products_table. Hu, G.Y., et al., 2015. The developmental trend and influencing factors of aeolian desertification in the Zoige Basin, eastern Qinghai-Tibet Plateau. Aeolian Res. 19, 275–281. Jia, K., et al., 2016. Fractional vegetation cover estimation algorithm for Chinese GF-1 wide field view data. Remote Sens. Environ. 177, 184–191. Jiang, M., Lin, Y., 2018. Desertification in the south Junggar Basin, 2000–2009: part I. Spatial analysis and indicator retrieval. Adv. Space Res. 62, 1–15. Jiang, L.H., et al., 2011. Classification methods of remote sensing image based on decision tree technologies. In: Li, D., Liu, Y., Chen, Y. (Eds.), Computer and Computing Technologies in Agriculture IV. CCTA 2010. IFIP Advances in Information and Communication Technology. vol. 344. Springer, Berlin, Heidelberg. Jiao, Q., et al., 2016. Impacts of re-vegetation on surface soil moisture over the Chinese Loess Plateau based on remote sensing datasets. Remote Sens. 8. https://doi.org/ 10.3390/rs8020156. Kassas, M., 1995. Desertification - a general-review. J. Arid Environ. 30, 115–128. Lam, D.K., et al., 2011. Tracking desertification in California using remote sensing: a sand dune encroachment approach. Remote Sens. 3, 1–13. Lambin, E.F., 2001. Remote sensing and geographic information systems analysis. Int. Encycl. Soc. Behav. Sci. 13150–13155. Lamchin, M., et al., 2016. Assessment of land cover change and desertification using remote sensing technology in a local region of Mongolia. Adv. Space Res. 57, 64–77. Lamchin, M., et al., 2017. Correlation between desertification and environmental variables using remote sensing techniques in Hogno Khaan, Mongolia. Sustainability 9. https:// doi.org/10.3390/su9040581.

2388

H. Duan et al. / Science of the Total Environment 650 (2019) 2374–2388

Li, M.M., et al., 2004. Estimation of vegetation fraction in the Upper Basin of Miyun reservoir by remote sensing. Resour. Sci. 26, 153–159. Li, A.M., et al., 2007. Remote sensing monitoring on dynamic of sandy desertification degree in Horqin Sandy Land at the beginning of 21st century. J. Desert Res. 27, 546–551. Li, Y.L., et al., 2009. Effectiveness of sand-fixing measures on desert land restoration in Kerqin Sandy Land, northern China. Ecol. Eng. 35, 118–127. Li, Y.Q., et al., 2012. Mongolian pine plantations enhance soil physico-chemical properties and carbon and nitrogen capacities in semi-arid degraded sandy land in China. Appl. Soil Ecol. 56, 1–9. Li, Q., et al., 2016. Quantitative assessment of the relative roles of climate change and human activities in desertification processes on the Qinghai-Tibet Plateau based on net primary productivity. Catena 147, 789–796. Li, J.Y., et al., 2017. Historical grassland desertification changes in the Horqin Sandy Land, Northern China (1985–2013). Sci. Rep. 7. https://doi.org/10.1038/s41598-01703267-x. Liu, A.X., 2004. Remote Sensing Monitoring of Desertification in China and Central Asia. Chinese Academy of Sciences. Liu, Y.H., et al., 2005. Research and application of the decision tree classification using MODIS data. J. Remote. Sens. 9, 405–412. Liu, A.X., et al., 2007. Method for remote sensing monitoring of desertification based on MODIS and NOAA/AVHRR data. Trans. CASE 23, 145–150. Liu, R.T., et al., 2011. Facilitative effects of shrubs in shifting sand on soil macro-faunal community in Horqin Sand Land of Inner Mongolia, Northern China. Eur. J. Soil Biol. 47, 316–321. Liu, S.L., et al., 2015. Several challenges in monitoring and assessing desertification. Environ. Earth Sci. 73, 7561–7570. Loh, W.Y., Shih, Y.S., 1997. Split selection methods for classification trees. Stat. Sin. 7, 815–840. Ma, J.H., et al., 2007. Qualified evaluating on the remote sensing of desertification - a case study of the Erdos region. J. Lanzhou Univ. (Nat. Sci.) 43, 1–6. Ma, Z.Y., et al., 2011. The construction and application of an Aledo-NDVI based desertification monitoring model. Procedia Environ Sci 10, 2029–2035. Mamatsawut, et al., 2008. Decision tree classification for extracting information on sandy desertification land in the southern Taklamakan. Res. Environ. Sci. 21, 109–114. Na, X.D., et al., 2009. Freshwater marsh wetland information extraction based on QUEST decision tree integrating with multi-source data. Chin. J. Ecol. 28, 357–365. Oswald, J., Harris, S., 2016. Chapter 11.1 - desertification A2 - Shroder, John F. In: Sivanpillai, R. (Ed.), Biological and Environmental Hazards, Risks, and Disasters. Academic Press, Boston, pp. 229–256. Pang, J.L., et al., 2012. RS-based study on dynamics of the vegetation coverage in recent 10 years in Yanchi County. Res. Soil Water Conserv. 19, 112–121. Qi, J., et al., 1994. A modified soil adjusted vegetation index. Remote Sens. Environ. 48, 119–126. Qi, Y.B., et al., 2012. Temporal-spatial variability of desertification in an agro-pastoral transitional zone of northern Shaanxi Province, China. Catena 88, 37–45. Qin, Z., et al., 2001. A mono-window algorithm for retrieving land surface temperature from Landsat TM data and its application to the Israel-Egypt border region. Int. J. Remote Sens. 22, 3719–3746. Ren, X.B., et al., 2016. A GIS-based assessment of vulnerability to aeolian desertification in the source areas of the Yangtze and Yellow Rivers. Remote Sens. 8. https://doi.org/ 10.3390/rs8080626. Salih, A.A.M., et al., 2017. Spectral mixture analysis (SMA) and change vector analysis (CVA) methods for monitoring and mapping land degradation/desertification in arid and semiarid areas (Sudan), using Landsat imagery. Egypt. J. Remote. Sens. Space Sci. 20, S21–S29.

Song, X., et al., 2015. Monitoring and analysis of aeolian desertification dynamics from 1975 to 2010 in the Heihe River Basin, northwestern China. Environ. Earth Sci. 74, 3123–3133. Song, W., et al., 2017. Estimating fractional vegetation cover and the vegetation index of bare soil and highly dense vegetation with a physically based method. Int. J. Appl. Earth Obs. Geoinf. 58, 168–176. Su, Z.Z., et al., 2006. Potential impact of climatic change and human activities on desertification in China. J. Desert Res. 26, 329–335. UNCCD, 2004. Preserving our common ground. UNCCD 10 years on. United Nations Convention to Combat Desertification, Bonn, Germany. UNCED, 1992. Managing Fragile Ecosystems: Combating Desertification and Drought (United Nations Conference on Environment and Development). UNEP, 1994. Development of guidelines for assessment and mapping of desertification and degradation in Asia/Pacific. Proceedings of Draft Report of the Expert Panel Meeting. United Nations Environment Programme. Wang, T., 2011. Desert and Aeolian Desertification in China. ELSEVIER and Science Press, Beijing. Wang, T., Zhu, Z.D., 2001. Some problems of desertification in northern China. Quat. Sci. 21, 56–65. Wang, T., Zhu, Z.D., 2003. Study on sandy desertification in China-1. Definition of sandy desertification and its connotation. J. Desert Res. 23, 209–214. Wang, T., et al., 2004. Spatial-temporal changes of sandy desertified land during last 5 decades in northern China. Acta Geograph. Sin. 59, 203–212. Wang, H.B., et al., 2015. Monitoring the recent trend of aeolian desertification using Landsat TM and Landsat 8 imagery on the north-east Qinghai-Tibet Plateau in the Qinghai Lake basin. Nat. Hazards 79, 1753–1772. Wang, Y.F., et al., 2016. Evolvement characters of Aeolian desertification of Horqin sandy land in the past 34 years-a case study of Naiman Banner. Proceedings of the 7th Annual Meeting of Risk Analysis Council of China Association for Disaster Prevention. 128, pp. 787–792. Wang, Y.F., et al., 2017. Monitoring the trends of aeolian desertified lands based on timeseries remote sensing data in the Horqin Sandy Land, China. Catena 157, 286–298. Wu, W., et al., 2003. Dynamic assessment for desertification based on the changes in landuse structure: an example from Naiman Qi of the Inner Mongolia, China. Acta Sci. Nat. Univ. Pekin. 39, 481–488. Xue, Z.J., et al., 2013. Evaluation of aeolian desertification from 1975 to 2010 and its causes in northwest Shanxi Province, China. Glob. Planet. Chang. 107, 102–108. Yan, C.Z., et al., 2007. Surveying sandy deserts and desertified lands in north-western China by remote sensing. Int. J. Remote Sens. 28, 3603–3618. Yu, Y., et al., 2010. Comparison of surface albedo measurement with MODIS product at Namco Station of Tibetan plateau. Plateau Meteorol. 29, 260–267. Zhang, T.H., et al., 2004. A comparison of different measures for stabilizing moving sand dunes in the Horqin Sandy Land of Inner Mongolia, China. J. Arid Environ. 58, 203–214. Zhao, Y.S., 2003. The Principle and Method of Analysis of Remote Sensing Application. Science Press, Beijing. Zhao, H.L., et al., 2007. Shrub facilitation of desert land restoration in the Horqin Sand Land of Inner Mongolia. Ecol. Eng. 31, 1–8. Zhao, H.L., et al., 2014. Effects of desertification on temporal and spatial distribution of soil macro-arthropods in Horqin sandy grassland, Inner Mongolia. Geoderma 223, 62–67. Zhu, Z., Wang, T., 1993. Trends of desertification and its rehabilitation in China. Desertification Control Bull. 27–30. Zucca, C., et al., 2012. Towards a World Desertification Atlas. Relating and selecting indicators and data sets to represent complex issues. Ecol. Indic. 15, 157–170.