Advertisement

Remote sensing-based land surface change identification and prediction in the Aral Sea bed, Central Asia

  • H. Shen
  • J. AbuduwailiEmail author
  • L. Ma
  • A. Samat
Original Paper

Abstract

The human-induced desiccation of the Aral Sea has generated large amounts of salt dust and has been posing a great threat to local ecological environment and human health. Monitoring its land cover changes is essential to obtaining information about the desertification process and dynamics of potential salt/sand dust source. To this end, long-term Landsat imagery was applied for the land use/cover change analysis based on support vector machine approach. The land cover distribution of the study area for 1977, 1987, 1996, 2006 and 2015 was mapped. In addition, the Markov–cellular automata integrated approach was used to predict the land cover change in 2015 and project changes in 2025 by extrapolating current trends. The classification results revealed that water surface of the Aral Sea shrunk by more than 66%, leading to the dramatic expanding of the salt soil and bare area. Change detection analysis indicated a serious land degradation trend as well as a major land cover evolution mode in the Aral Kum that could predict shifts in dust composition. The Markov–cellular automata technique was successful in predicting land cover distribution in 2015, and the projected land cover for 2025 revealed more desertification of the landscape with potential expansion in the salt soils and bare area. It is worth noting that the vegetation cover of the region represented an obvious increase in recent years that may be a good signal of ecological recovery.

Keywords

Remote sensing Aral Sea Markov–cellular automata Land use and land cover Support vector machine 

Introduction

Land use/land cover (LULC) transformation is now identified as one of the key drivers of global change and one of the most important environmental issues of global concern for its increasing impacts on ecosystems, climate and social development (Bajocco et al. 2012; Guan et al. 2011). LULC changes and associated habitat loss or fragmentation could lead to loss of biodiversity and latent land degradation processes, especially in the arid zone where ecosystems are much more vulnerable, and desertification has been the most widespread and troublesome consequence. Landscape transformations represent the visible result of human interaction with land and can be analyzed and interpreted in terms of “trajectories” as typified. Rather than global climate change, human-induced changes including urbanization, shifting to intensive agriculture and overgrazing are the three major causes of LULC dynamics in the modern era (Halmy et al. 2015).

The Aral Sea, once the fourth largest lakes on the planet, rapidly shrunk in recent decades due to over-exploitation of water resources for agriculture and construction of water reservoirs and hydropower stations along the two main feed rivers (the Amu Darya and the Syr Darya) as well as the impact of climate change. Thus, the vast exposed lake bed turned into the special Aral Kum desert that comprises a large area of salt soils and loose sand dunes. These newly formed landscapes are highly unstable and considered to be the main source of salt and dust storms (Semenov 2012; Singer et al. 2003). Frequent salt/sand dust events and aerosols have been a posing great threat to the ecological environment and local inhabitants in the Circum-Aral region (Löw et al. 2013; O’Hara et al. 2000; Wiggs et al. 2003). The Aral Sea basin has experienced the most significant LULC dynamics which include primary vegetation succession during the past several decades (Wucherer and Breckle 2001). Nowadays, the vulnerability of the study area and the local humane society has increased to the crucial level. Mitigating this human-induced crisis needs timely monitoring information and efficient management measures.

The increasing availability of remotely sensed data and growing advances in their temporal, spatial and spectral resolutions provide an effective tool for detecting LULC changes in the dried Aral Sea bed, which could not only provide the exact location and the surface characteristics of the source regions for salt and dust events, but also help to investigate the water and other natural resources (Foody 2003; Yuan et al. 2005). Early in the 1970s, satellite imagery has been successfully used to monitor water regression and desertification process of the Aral Sea basin (Shi and Wang 2015; Shi et al. 2014). Several studies have mapped or assessed land cover patterns or soil types distribution in the Aral Sea basin based on remote sensing (Dukhovny et al. 2008; Kozhoridze et al. 2012; Micklin 2008; Stulina and Sektimenko 2004). Also, the major dust source regions and the intensity of dust removal were identified, which showed rapid shift over time (Kondratyev et al. 1985; Spivak et al. 2012). For instance, analysis of LULC changes indicated that the northeastern part of the Aral Kum desert created the most dust emission of the region, while from 2005 to 2008 the dust sources were detected on the eastern and southern shores and the Vozrojdenie Peninsula (Indoitu et al. 2015). However, these studies mainly focused on the southern part of the Aral Kum, no studies have been conducted beyond 2008 and the application of long-term satellite imagery for LULC transformation assessment is still missing. However, the ongoing activities in the area may have led to more changes in the landscape that can influence the distribution of dust source in the area. With this respect, the freely available continuous Landsat data that began from 1973 might be the optimal choice for long-term LULC change detection (Abd El-Kawy et al. 2011; Wulder et al. 2008).

In addition, predicting LULC change is also important for understanding and highlighting the potential modifications that might happen over landscapes in the near future (Guan et al. 2011). A better understanding of land trajectories associated with land degradation contributes to the assessment of past changes and to run short-term scenario analysis with reliable prediction rules. Many approaches and techniques have been proposed for modeling and predicting LULC change (Overmars et al. 2003). For instance, Markov chain analysis is a robust modeling approach that has been widely used to analyze the dynamics of LULC at different scales (Baker 1989; Muller and Middleton 1994). It is suitable for the spatially dependent land use data for not assuming statistical independence of data as required by logistic regression (Overmars et al. 2003). It is also able to predict all multidirectional LULC changes among all categories available. However, Markov chain analysis does not provide spatial distribution of the change, as well as it is not spatially explicit. This shortcoming can be overcome through the integration with other different dynamic and empirical models, like cellular automata (CA) models. The integrated approach, Markov–CA model, has the advantage of predicting two-way transitions among the available LULC classes and outperform regression-based models (Pontius and Malanson 2005); thus, it is commonly used in predicting LULC (Sang et al. 2011).

Focusing on the above missing work, mapping of land cover transformation and predicted future patterns was performed for the Aral Sea lake basin in the year of 2015 after the field survey carried out in the same year. The aims of this paper were to: (1) map LULC in the region between 1977 and 2015; (2) apply the Markov–CA technique to these LULC maps to predict potential changes in 2025 and (3) discuss the development of potential source areas for saline and dust storm. This paper contributes to the timely investigations of the land cover dynamics in the vulnerable dryland ecosystems characterized by increased intensive human pressure and prone to desertification, which would help to design better land use management plans and policies on water allocation.

Study area

The Aral Sea, a sea-lake, belongs to two independent countries, namely Kazakhstan in the north and Uzbekistan in the south. It is located between 43°24′–47°56′N and 58°12′–62°59′E that covers about 68,000 km2 area (Fig. 1). The Aral region is within typical temperate continental climatic zone with 60–140 mm mean annual precipitation and 800–1300 mm potential annual evaporation, which could increase the susceptibility to environmental degradation. It has hot summers and cold winters with the average annual temperature of 7.8 °C that varies from − 28 to 46 °C (Breckle and Wucherer 2012). The Aral Sea basin is geographically surrounded by the Ustyurt and Trans-Unguz Plateaus to the west and south, the Kyzylkum desert and Pamir Plateau to the east and southeast and hilly plains of Bet-Pak-Dala desert to the north. The remaining areas comprise various forms of alluvial valleys, mountain valleys and arid or semiarid steppes.
Fig. 1

Location map of the Aral Sea. The red line represents the water surface of the study area before 1960s, and the numbers 1, 2, 3, 4 marked in yellow represent the remaining separated water bodies at present

In 1961 the Aral Sea’s level was about 52.71 m and close to the normal value. Water covered more than 65,000 km2, water volume was estimated to 1051 km3, its mean depth was 16.16 m with a maximum of 68 m, and the salt content was 1% (Cretaux et al. 2013). Since 1960s, the increasing water withdrawal from Amu Darya and Syr Darya as well as the Karakum Canal has resulted in 88% decrease in surface area and 92% in volume of the Aral Sea (Micklin 2010). Consequently, the complete Aral Sea was gradually separated into four small independent water bodies, namely small Aral Sea in the north, Tshche-bas Gulf in the northwest, Western basin large Aral and Eastern basin large Aral (Fig. 1). The newly formed Aral Kum desert estimated to be more than 57,000 km2 in 2011 (Indoitu et al. 2015) has significantly increased the availability of particle mass for deflation, reaching 22.8 ton ha−1 year−1 (Breckle and Wucherer 2012; Groll et al. 2013; Indoitu et al. 2012; Singer et al. 2003). What was worse, the dust or aerosol usually has a high salt content (sometimes up to > 90% in winter) and the mean annual salt dust transported by wind was 0.5 × 106 t to 20 × 106 ~ 30 × 106 t (Orlovsky and Orlovsky 2002). The solonchaks (salt crust and salt soil) are the most widely distributed land cover in the study area, occupying approximately 1.5 × 105 km2. The dry Aral Sea bed has accumulated more than 1 billion tons of salt, providing a convenient condition for salt dust emission (Kozhoridze et al. 2012; Orlovsky et al. 2004). Thus, the impact of dust covers a region of about 500,000 km2 surrounding the former Aral Sea with 40–110 dust storms from the eastern Aral Kum per year (Issanova et al. 2015). Besides, other land cover types of the region include sand dunes, stone/sand deserts. The main vegetation covers are mostly open grass land and open shrub land (involving reeds), and wood is sparsely distributed at wet areas. Along with the catastrophic shrinking of the lake, its microclimate deteriorates seriously and a significant warming of more than 6 °C over the desiccated sea floor was reported (Roy et al. 2014).

Materials and methods

LULC classification

Satellite imagery and ancillary data

The data collection adopted in this research could be divided into satellite data and ancillary data. The input satellite data come from the Landsat 2 Multispectral Scanner (MSS), Landsat 5 Thematic Mapper (TM) and Landsat 8 Operational Land Imager (OLI) sensors for the years 1977, 1987, 1996, 2006 and 2015, respectively. They were downloaded from the United States Geological Survey (USGS) Web site (http://glovis.usgs.gov/). Note that for each year the imagery was selected from the growing season because of snow and cloud covers are usually so extensive in non-growing periods that would increase classification errors. Moreover, each imagery should be carefully selected with cloud cover over the region as low as possible. The 1977, 1987, 1996 and 2006 scenes were geo-registered to the 2015 scene. The Landsat digital numbers (DNs) were transformed to absolute units of at-sensor spectral radiance and then to top of atmosphere (TOA) reflectance following the standard Landsat equations and scaling factors. The study area is so vast that it must use 8–10 panels to mosaic the complete region (Table 1).
Table 1

Landsat satellite data used in the study

Years

Satellite and sensor

Scenes

Path/row

Date

Spatial resolution

1977

Landsat 2 MSS

8

172/28-30,

173/28-30,

174/28-29

1977-09-01

1977-09-02

1977-09-03

79 m

1987

Landsat 5 TM

9

160/28-30,

161/27-30,

162/28-29

1987-08-28

1987-08-20

1987-09-11

30 m

1996

Landsat 5 TM

9

160/28-30,

161/27-30,

162/28-29

1996-07-20

1996-07-11

1996-08-20

30 m

2006

Landsat 5 TM

10

160/27-30,

161/27-30,

162/28-29

2006-07-31

2006-08-07

2006-08-14

30 m

2015

Landsat 8 OLI

8

160/28-30,

161/28-30,

162/28-29

2015-08-03

2015-08-10

2015-08-17

30 m

Based on Land Cover Classification System (LCCS) developed by the United Nations Food and Agriculture Organization (UN/FAO) and the previous job (Löw et al. 2013), a special classification scheme was designed for the Aral Sea basin. The landscapes were distinguished as two types: vegetated area and non-vegetated area. The primary vegetated areas consist of “Shrub land” and “Grass land”. And the non-vegetated areas were identified as the following classes: (1) “Water,” (2) “Salt soil,” (3) “Salt crust” and (4) “Bare area.” It should be noted that “Salt soil” and “Salt crust” are two subclasses of “Bare area” in general, while in this study the salt-affected soil was divided into “Salt crust” and “Salt soil,” because of their significant difference not only in wind erodibility, but also in spectral reflection. Other kinds of barren land were classified as “Bare area.” Figure 2 shows the photographs of the six land cover types in the study area and represents the spectral characteristics and three indices or band ratio of these landscapes captured in the Landsat 8 imagery. There is no doubt that using different band combinations from different sensors would cause biasness; thus, the common of bands were used (Green, Red, NIR and MIR) for classification in all cases.
Fig. 2

Spectral curves and related indices of land cover in the Aral Kum. The photographs are: A1, A2, A3 Water; B1, B2, B3 Grass land; C1, C2, C3 Shrub land; D1, D2, D3 Salt crust; E1, E2, E3 Salt soils; F1, F2, F3 Bare area

The study area has low vegetation cover and a highly reflective background, suggesting challenges for categorizing LULC classes based on spectral data alone. Approaches for integrating ancillary data with spectral data have been found to improve LULC classification accuracy. Therefore, ancillary data were used to help in classifying LULC classes in the area. The ancillary data included land surface parameters derived from DEM data (SRTM, 90 m), namely slope length and steepness (LS) factor. Visual interpretation, band ratios (MIR/Red) and some remote sensing indices such as normalized difference vegetation index (NDVI) and normalized difference water index (NDWI) were also included to improve classification accuracy. In addition, the tasseled cap spectral indices of greenness, brightness and wetness were also applied in the analysis.

The decision tree (DT) approach was the major technique for land cover classification in those earlier studies about Aral Sea (Kozhoridze et al. 2012; Löw et al. 2013). However, it is sensitive to noise of observations and over-fitting (Ghimire et al. 2012). In contrast, support vector machine (SVM) has been confirmed to perform better in feature spaces with high dimensionality (Hansen 2012), which is a significant advantage when abundant remote sensing data are available while little ground truth reference could be obtained. The development of SVM started early in 1970s and did not become popular in pattern recognition and classification until late 1990s (Vapnik 1998). SVM was mainly applied in the classification of hyperspectral image and object detection in remote sensing (Melgani and Bruzzone 2004); recently, it has been also widely used in the classification of multispectral remotely sensed data and obtained great success (Pal and Mather 2005). The primary advantage of SVM was good generalization capability with limited training samples (Shao and Lunetta 2012). SVM is the heuristic one that utilizes a user-defined kernel function to map a set of nonlinear decision boundaries in the original dataset into linear boundaries of a higher-dimensional construct, optimally separating two classes; the final boundary is the optimal super hyperplane (OSH) (Han et al. 2007). The data points lying at the edge of each individual class hyperplane in feature space are the so-called support vectors. A suitable choice of kernel allows the data to become mostly separable in the feature space despite being non-separable in the original input space (Szuster et al. 2011). A classification probability threshold of zero was also applied, keeping all image pixels to be classified and get exactly one class label (Petropoulos et al. 2011). As mentioned above, the SVM was applied for the classification in this research. The SVM kernel used in this study is a radial basis function whose equation is represented below.
$$K(x_{i} ,x_{j} ) = \exp \left( { - \,\gamma \left\| {(x_{i} ,x_{j} )} \right\|^{2} } \right),\quad \gamma > 0,$$
where xi belongs to Rn which is an n-dimensional vector, and γ is the width of the kernel function.

Training data

Reference data mainly came from field survey in 2015, Google Earth imagery and historic maps, which were used to assess and improve the classification accuracy. The ground reference data consisted of 180 sampling points (See in Fig. 1) and were collected based on stratified random sampling (Boschetti et al. 2016). Each plot was assigned to an LULC class based on the field observations. For the 1987, 1996 and 2006 scenes, reference data were collected from the following historic maps:
  • “Land cover map of the desiccated Aral Sea bed,” as of 2006 (Dukhovny et al. 2008), in which the landscapes were divided into four major classes, namely water, solonchak, sands, plain delaic and of deposition with 17 subclasses;

  • “South Prearalie Landscape Map,” as of 1996 (digitized by SIC ICWC);

  • “Water and solonchak soil classification map,” as of 1987 (Kozhoridze et al. 2012);

Unfortunately, information material of a similar level of detail is not available for 1977. Therefore, it is attempted to compare the results of water surface of 1977 with the “Map of long term variations of Aral Sea shorelines in 1957–2008” (Kravtsova and Tarasenko 2010), while reference data of other classes were selected from the classification map of 1987 by visual interpretation according to the spectral features. In addition, for 2006 the Google Earth imagery provided reference samples for salt crust and grass land where high-resolution imagery in the Aral Kum existed.

In order to avoid spatial autocorrelation of validation sample, the mapping accuracies were estimated with a tenfold cross-validation. In each cross-validation fold the reference data were divided at random into two subsets: About 10% was training sets that guide the software what kinds of pixels are to be clustered for one land cover class indispensable, whereas the rest 90% was used for the accuracy assessment.

Accuracy assessment

The assessment of classification precision is performed by comparing classification results with 100 or more random sampled reference points based on ground truth data and visual interpretation. Then the confusion matrices calculating the user’s accuracy (UA), producer’s accuracy (PA), overall accuracy (OA) and Cohen’s kappa statistic (κ) are applied to evaluate the mapping accuracy and reduce classification errors as well.

Change detection analysis that describes the transformations in amount and location between imagery of the same scene at different times is essential to understanding the drivers of change and the interaction processes (Congalton and Green 2009). There are different approaches available to detect changes of land covers, and in this study a post-classification detection method was applied. To determine the quantity and quality of conversions from one category to other and their corresponding areas between two selected years, cross-pixel-based tabulation analysis compared classified image pairs of two different years to produce change information and thus interpret the changes more efficiently taking the advantage of “from–to” information. A two-way cross-matrix was obtained by the application of this procedure. Finally, statistical data of the overall land cover transformations as well as gains and losses in each class of the five intervals were compiled.

LULC change prediction

As mentioned above, Markov–CA is proved to be a robust approach for predicting LULC change. It is the combination of cellular automata/Markov chain/multicriteria/multiobjective land allocation and adds an element of spatial contiguity as well as knowledge of the likely spatial distribution of transitions to Markov chain analysis (Sang et al. 2011).

In this combination, the Markov model focuses on the quantity in predictions for land cover changes. However, its spatial parameters are weak and the various types of land cover changes in the spatial extents are unknown (Wickramasuriya et al. 2009). Markov chain is a prediction and optimal control theory method that is based on the process of the formation of Markov random process systems (Sang et al. 2011). It not only provides the quantification of conversion states between the land covers, but also reveals the transfer rate among different types. The prediction of land cover changes is calculated on the basis of the Bayes conditional probability formula as follows (Pijanowski et al. 2002):
$$S_{t + 1} = S_{t} \times P_{ij} ,$$
in which St, St+1 are the system status at the time of t or t + 1; Pij is the transition probability matrix in a state which is expressed as follows (Pijanowski et al. 2002):
$$P_{ij} = \left[ {\begin{array}{*{20}l} {P_{11} } \hfill & {P_{12} } \hfill & \cdots \hfill & {P_{1n} } \hfill \\ {P_{21} } \hfill & {P_{22} } \hfill & \cdots \hfill & {P_{2n} } \hfill \\ \vdots \hfill & \vdots \hfill & \vdots \hfill & \vdots \hfill \\ {P_{m1} } \hfill & {P_{m2} } \hfill & \cdots \hfill & {P_{mn} } \hfill \\ \end{array} } \right].$$
In contrast, the CA model has a strong capability of space–time dynamic evolution with complex space systems. Some uncertainties like the interaction between model elements, structures and the quality of input data sources of the CA model would have an impact on the model’s behavior (Peterson et al. 2009). It is mainly concerned the local interaction and the powerful computing power of the cells with obvious spatiotemporal coupling characteristics, especially for dynamic simulation and display of self-organizing feature system. The application of geographic cellular automata to land cover change simulations takes the historical trends of land cover into account. The CA model can be expressed as follows (Pijanowski et al. 2002):
$$S_{t + 1} = f[S_{t} ,\,\,N],$$

In the equation S represents the set of limited and discrete cellular states, N is the cellular field, t and t + 1 indicate the different times, and f is the transformation rule of cellular states in local space.

Thus, the Markov–CA model is about the time series and space for the advantages of forecasting and can obtain better simulation for temporal and spatial patterns of land cover changes both in quantity and in space (Han et al. 2009). In this paper the Markov–CA model has been applied to simulate land cover changes in the following three steps:
  1. 1.

    applying the Markov chain analysis to the 1996, 2006 and 2015 LULC maps that were resulting from the classification process for calculating class transition probability matrices between 1996 and 2006, 2006 and 2015, and 1996 and 2015 using a first-order Markov model. These transition matrices were then used in the successive step.

     
  2. 2.

    calculating transition suitability maps of LULC; these maps define the suitability of each pixel for transition to any LULC class created using both Boolean logical operations and the weighted linear combination methods. The value of each pixel ranges from 0 to 255, 0 representing unsuitable and 255 representing highly suitable. Field surveys were used to define transition rules. The physical closeness to an existing class is likely to be a driver of change to this class in the future. In this study, the transition suitability maps were used to predict LULC distribution in 2015 and simulate the distribution in 2025.

     
  3. 3.

    application of the CA model to the transition matrices and the transition potential maps to predict the spatial distribution of LULC. The transition probabilities for the period 1996–2006 with the 2006 LULC base map were used to model LULC in 2015 based on the Markov–CA approach. These areas divided by the number of timesteps involved in the simulation provided the areas to be converted per iteration. The future assignment for each pixel to specific LULC class was produced according to how much suitable the pixel to this LULC class is and how near the pixel is to other pixels of the same class. The simulation continues until the area to be converted of each LULC per iteration is met. Then using the kappa statistic to assess the agreement between the predicted 2015 LULC map with the classified 2015 LULC map. To consider the agreement in the spatial pattern in assessing the validity of the model the agreement and disagreement components between the two maps were calculated. Following the same process, the Markov–CA model was applied to project LULC in the study area by 2025. The transition probabilities for the period 2006–2015 and the LULC base map of 2015 were used.

     

Results and discussion

Land cover classification map of 1977–2015

The use of SVM to classify the Landsat imagery for the year 1977, 1987, 1996, 2006 and 2015 produced LULC maps (Fig. 3). Table 2 is cross-validation result for the five selected year. User’s and producer’s accuracies for individual categories were varying from 74.43 to 96.04%. The root-mean-square error (RMSE) can be estimated as an accuracy measurement when training the SVM with all samples. The RMSE estimated is, respectively, 0.4632, 0.4874, 0.5121, 0.4924 and 0.4212. The obtained overall accuracies were all higher than 79%, with κ indices of agreement of more than 0.75, which indicates that the land cover maps show a strong agreement with the reference maps and are accurate enough for further analysis.
Fig. 3

Land cover classification maps of the study area for the five stages: a 1977; b 1987; c 1996; d 2006; e 2015, and Markov–CA model predictions: f 2015; g 2025

Table 2

Tenfold cross-validation estimation for the classification results for the Landsat data using SVM

Land cover

1977

1987

1996

2006

2015

UA

PA

UA

PA

UA

PA

UA

PA

UA

PA

Water

91.28

92.13

91.53

90.38

92.15

93.02

92.32

92.08

95.18

96.04

Grass land

77.05

78.43

81.32

82.55

83.76

82.54

82.15

84.47

86.75

84.07

Shrub land

78.13

76.97

80.15

81.23

81.46

83.23

80.53

81.25

81.25

80.16

Salt crust

76.55

75.12

81.33

83.02

82.14

83.27

81.21

80.57

84.37

85.02

Salt soil

75.38

74.43

80.84

79.19

80.53

83.09

85.11

84.34

84.33

83.08

Bare area

80.22

81.03

81.32

83.77

84.76

83.54

86.07

87.41

87.24

88.21

OA

79.54

80.23

81.36

 

82.72

84.25

κ

0.75

0.77

0.80

0.81

0.83

RMSE

0.4632

0.4874

0.5121

0.4924

0.4212

*RMSE root-mean-square error

The dramatic changes occurred on “Water,” “Bare area” and “Salt soil” from 1977 to 2015. In comparison with the water area in 1977 as presented in Fig. 3a, the other five classes only covered a small part of the study area and mainly distributed in the Eastern Aral Sea bed. Subsequently, the shoreline of the lake receded westward and northward at an astonishing rate in the followed 4 decades. Thus, the islands among the lake grew “bigger” and eventually connected with the early exposed sea floor, dividing the formal complete Aral Sea into several separated water bodies. Meanwhile, most of the desiccated sea bed turned to “Salt soil,” “Bare area” and small part of “Salt crust.” The formation of “Salt crust” requires unique conditions that the marshy solonchaks soils must be periodically flooded. Therefore, it mainly distributes in a narrow stripe around the receding shoreline of the Aral Sea. Only in the estuary and other moist areas some vegetation started to develop, such as Amu Darya delta in the south and Syr Darya outlet in the northeast where the soil and water are suitable for plant growth.

The land cover maps for the five dates offer direct insights into trends and rates of the land cover transformation. And the statistics uncover change process and dynamics of each individual category in detail. In 1977 water surface was absolutely the dominant land cover in the Aral Sea basin, occupying 81.73% of the study area, and then rapidly decreased to 61.12% in 1987, to 14.34% at present. The significant declining trend is highly consistent with the results of other associated studies (Cretaux et al. 2013; Micklin 2010). Accompanied with the lake shrinkage is the drastic expansion of unstable barren areas, including “Bare area,” “Salt soil” and “Salt crust,” which means the intensified desertification and wind erosion potential in the Aral Sea basin. At the same time, the “Salt soil” and “Bare area” expanded rapidly over the study period, respectively, increasing from 5.24 and 11.83% at the beginning to 28.86 and 38.99% in 2015, which offered large amounts of saline dust and sand for deflation process.

In addition, vegetation covers (“Shrub land” and “Grass land”) have occurred on the exposed lake bed since its shrinking, but represented a fluctuated trend. Especially in recent years, they occupied a relatively high proportion of the Aral Sea bed, adding up to 15.38%. As is known to all, vegetation plays an important role in terrestrial ecosystem, especially “precious” in arid land for its scarcity. The increase in vegetation cover could be taken as a good signal for environmental improvement. To great extent, this should be owing to the ecological recovery measures made by the two associated countries in the Aral region. However, this improvement is far from enough to relieve the current catastrophic desertification of the region (Table 3).
Table 3

Estimation area and percent (%) of individual land covers (km2) derived from the classification process

Land cover

1977

1987

1996

2006

2015

Area

(%)

Area

(%)

Area

(%)

Area

(%)

Area

(%)

Water

55,317.88

81.73

41,369.31

61.12

33,101.17

48.90

15,271.45

22.56

9707.50

14.34

Grass land

526.98

0. 78

2508.57

3.70

679.84

1.01

652.48

0.96

9025.65

13.34

Shrub land

199.02

0.29

2721.31

4.02

913.37

1.35

1742.38

2.57

1380.74

2.04

Salt crust

89.99

0. 13

296.15

0.44

169.48

0.25

3146.78

4.65

1641.71

2.43

Salt soil

3547.68

5.24

12,385.60

18.30

18,924.40

27.96

19,053.31

28.15

19,536.52

28.86

Bare area

8003.37

11.83

8403.98

12.42

13,896.66

20.53

27,818.52

41.11

26,392.80

38.99

Total

67,684.92

100.0

67,684.92

100.0

67,684.92

100.0

67,684.92

100.0

67,684.92

100.0

Change detection analysis

Change matrices for the period 1977–1987, 1987–1996, 1996–2006, 2006–2015 and 1977–2015 were produced (Table 4), and the change detection maps are displayed in Fig. 4. The most significant land cover transformation was seen in the period of 2006–2015, during which the no change percent was about 44.42%, smaller than those of the other three intervals, 1977–1987 (69.39%), 1987–1996 (70.78%) and 1996–2006 (55.31%). Overall, the proportion with land cover conversions was of the order of 76.24%, which means almost the whole basin was completely different from its original situation except for the remained water surface.
Table 4

Matrices of land cover changes (km2) during 1977–2015

1977

1987

Water

Grass land

Shrub land

Salt crust

Salt soil

Bare area

Total (1977)

a 1977–1987

 Water

40,955.35

1620.97

2316.42

205.43

6319.35

3900.36

55,317.88

 Grass land

17.1

89.54

50.95

5.02

120.43

243.94

526.98

 Shrub land

5.89

19.16

138.14

0.41

5.92

29.5

199.02

 Salt crust

2.39

11.76

1.18

2.11

37.78

34.77

89.99

 Salt soil

142.09

214.06

15.26

30.48

2365.21

780.58

3547.68

 Bare area

246.49

553.08

199.36

52.7

3536.91

3414.83

8003.37

 Total (1987)

41,369.31

2508.57

2721.31

296.15

12,385.6

8403.98

67,684.92

1987

1996

Water

Grass land

Shrub land

Salt crust

Salt soil

Bare area

Total (1987)

b 1987–1996

 Water

32,744.72

134.88

209.12

77.54

4568.67

3634.38

41,369.31

 Grass land

46.36

83.51

56.98

4.29

1036.12

1281.31

2508.57

 Shrub land

131.95

184.48

515.24

1.52

604.81

1283.31

2721.31

 Salt crust

23.81

4.04

3.42

22.74

145.67

96.47

296.15

 Salt soil

81.66

89.95

26.24

40.03

9542.2

2605.52

12,385.6

 Bare area

72.67

182.98

102.37

23.36

3026.93

4995.67

8403.98

 Total (1996)

33,101.17

679.84

913.37

169.48

18,924.4

13,896.66

67,684.92

1996

2006

Water

Grass land

Shrub land

Salt crust

Salt soil

Bare area

Total (1996)

c 1996–2006

 Water

14,840.16

107.36

525.91

3084.5

5280.86

9262.38

33,101.17

 Grass land

12.97

58.79

153.21

0.23

53.13

401.51

679.84

 Shrub land

44.46

51.51

633.58

0

32.77

151.05

913.37

 Salt crust

14.33

0.13

0.09

45.46

53.36

56.11

169.48

 Salt soil

192.56

66.42

45.94

10.76

11,260.46

7348.26

18,924.4

 Bare area

166.97

368.27

383.65

5.83

2372.73

10,599.21

13,896.66

 Total (2006)

15,271.45

652.48

1742.38

3146.78

19,053.31

27,818.52

67,684.92

2006

2015

Water

Grass land

Shrub land

Salt crust

Salt soil

Bare area

Total (2006)

d 2006–2015

 Water

9146.08

49.03

163.24

1020.97

4820.06

72.07

15,271.45

 Grass land

1.5

493.09

92.88

17.92

10.18

36.91

652.48

 Shrub land

63.15

912.29

587.51

62.58

85.34

31.51

1742.38

 Salt crust

201.86

4.32

43.99

122.46

2690.51

83.64

3146.78

Salt soil

39.11

925.35

21.91

154.23

5482.29

12,430.42

19,053.31

 Bare area

255.8

6641.57

471.21

263.55

6448.14

13,738.25

27,818.52

 Total (2015)

9707.5

9025.65

1380.74

1641.71

19,536.52

26,392.8

67,684.92

1977

2015

Water

Grass land

Shrub land

Salt crust

Salt soil

Bare area

Total (1977)

e 1977–2015

 Water

9661.9

6676.10

1086.56

1594.05

19,238.59

17,060.68

55,317.88

 Grass land

5.23

237.07

46.27

0.48

8.76

229.17

526.98

 Shrub land

1.02

133.89

50.36

0.01

0.26

13.48

199.02

 Salt crust

0.03

15.94

0.01

0.73

15.79

57.49

89.99

 Salt soil

2.65

385.39

13.32

13.16

118.19

3014.97

3547.68

 Bare area

36.67

1577.26

184.22

33.28

154.93

6017.01

8003.37

 Total (2015)

9707.5

9025.65

1380.74

1641.71

19,536.52

26,392.8

67,684.92

Fig. 4

Change detection maps of land covers in the Aral Sea basin for the five phases: a 1977–1987; b 1987–1996; c 1996–2006; d 2006–2015; e 1977–2015

The comparison of each category of 1977 and 2015 revealed that the Aral Sea bed has experienced drastic LULC transformation. The “Water” surface diminished by 82.13%, most of which converted into “Salt soil” (34.78%) and “Bare area” (30.84%), followed by “Grass land” (12.07%), “Salt crust” (2.88%) and “Shrub land” (1.96%). Subsequently, the rest five land cover classes all expanded at an astonishing rate. Among these, the “Salt crust” grew up from 89.99 to 1641.71 km2, increasing by almost 18 times, and the followed are “Grass land,” “Shrub land,” “Salt soil” and “Bare area.” Other obvious land cover transformations are that 60.86% of “Grass land,” 51.44% of “Salt crust” and 86.62% of “Salt soil” as well as 20.56% of “Shrub land” converted into “Bare area” and 55.43% of “Shrub land” became “Grass land,” which revealed a strong desertification and vegetation degradation mainly resulting from deterioration of microclimate in the region.

Meanwhile, an important shift that must be noted is that about 18.84% of “Salt crust” converted into “Salt soil,” which was considered to be one of the major landscape evolution modes in the desiccated Aral Sea bed. In other words, the recession firstly resulted in extensive “Salt crust” immediately close to the sea, then “Salt crust” converted into “Salt soil” and afterward into “Bare area.” Additionally, the significant increase in the “Grass land” in 2015 was highly contributed by the “Bare area,” about 1891.84 km2 “Grass land” grew up on the formal “Bare area,” occupying 20.96% of the total area of “Grass land.” This transition direction suggested that the ecological environment of the Aral Sea bed improved a little in recent years, which mainly took place on parts of the northeastern and southern Aral Kum. But this cannot compensate for the large ecological loss of the water surface. Still overall the desertification of the region is at catastrophic level.

Prediction of LULC

The characteristics of LULC change over the three stages (1996–2006, 2006–2015 and 1996–2015) can be identified from the transition trend as revealed from the Markov transition matrices over the period between 1996 and 2015 (Table 4). From 1996 to 2015, the highest probability of undergoing no change is attained by water, bare area and shrub land with transition probabilities exceeding 0.85. The lowest probability of persistence was attained by salt crust and salt soil. During the period 2006–2015 the most stable LULC classes included water and bare area. The most dynamic LULC classes with the lowest transition persistence were the salt crust.

Comparison of the 2015 LULC map with the predicted map of Markov–CA method (Table 5) revealed a good agreement between the two with an overall accuracy above 88% and overall κ exceeding 0.87. All the classes attained PA and UA higher than 84%. The best agreement was attained by water, with κ values exceeding 0.95, whereas the least agreement was attained by shrub land and salt soil with κ of 0.83 and 0.84, respectively. The results indicate that the Markov–CA model was successful in predicting the LULC in 2015 and can be reliably used to predict future LULC change in the area given the assumption of stable rates of change. The prediction of potential distribution of the LULC classes in 2025 (Fig. 3g) shows expansion in salt soils. It also shows an increase in bare area and growth in the grass land.
Table 5

Markov chain matrix of LULC transition probabilities for the periods 1996–2015 and 2006–2015

Land cover

Period

To

Water

Grass land

Shrub land

Salt crust

Salt soil

Bare area

From

 Water

1996–2015

0.97

0.04

0.12

0.09

0.24

0.23

2006–2015

0.96

0.11

0.14

0.07

0.31

0.31

 Grass land

1996–2015

0.00

0.51

0.01

0.00

0.01

0.13

2006–2015

0.00

0.44

0.00

0.00

0.03

0.11

 Shrub land

1996–2015

0.01

0.32

0.86

0.00

0.01

0.08

2006–2015

0.03

0.25

0.71

0.00

0.02

0.12

 Salt crust

1996–2015

0.00

0.00

0.00

0.24

0.22

0.07

2006–2015

0.02

0.00

0.00

0.37

0.19

0.09

 Salt soil

1996–2015

0.00

0.01

0.00

0.00

0.35

0.11

2006–2015

0.00

0.02

0.00

0.00

0.24

0.24

 Bare area

1996–2015

0.00

0.11

0.01

0.00

0.00

0.85

2006–2015

0.00

0.13

0.02

0.00

0.00

0.77

Drivers of land cover dynamics

Land cover evolution of the Aral Sea is controlled by several factors, which could be simply divided into two kinds, namely natural and anthropogenic. Above all, problems that occurred on water resource must be the core of the Aral Crisis and the associated land cover dynamics. Hence, the drivers affecting water resources of the Aral Sea basin must be responsible for the land cover dynamics. In this case, the natural factor mainly refers to climate variations and the anthropogenic factor is human activities. More specially, it is the agricultural activities in the drainage basin of the Aral Sea that broke the balance between local society and its natural ecosystem (Table 6).
Table 6

Assessment of the agreement between predicted and actual LULC in 2015

 

Water

Grass land

Shrub land

Salt crust

Salt soil

Bare area

Producer’s accuracy

94.25

88.12

83.09

85.65

83.11

90.12

User’s accuracy

95.11

84.36

84.87

87.01

84.95

88.36

Kappa/class

1.00

0.85

0.83

0.88

0.84

0.92

Overall kappa

0.87

Overall accuracy

88.65

Mean omission/class

7.27 ± 8.02

Mean commission/class

7..35 ± 7.12

Climate change

In this region, climate change is considered to alter the water availability with declining summer precipitation and potentially increasing winter rainfall, which could significantly affect the precipitation-sensitive vegetation (Lioubimtseva and Henebry 2009). The Aral Sea is fed by Central Asia’s two important rivers, the Amu Darya and the Syr Darya, with a flow, respectively, of about 70 and 35 km3 per year on average before 1960 (Glantz 2005). It was the huge water volume of the two rivers that kept the Aral Sea the global fourth largest lake for a long period. However, the total annual inflow runoffs to the sea decrease to 1–2 km3 in recent decades (See Fig. 5d). The regional climate change seems to play a negligible role in driving the Aral Sea’s desiccation and associated land surface dynamics, as the obvious rising air temperatures augment evaporation from the sea surface. Figure 5a, b, c gives the trends of temperature and precipitation recorded by 3 meteorological stations in the Aral region since 1960, and all of them showed obvious rise in temperature, while the precipitation represented strong inter-annual variations. In accordance with the fourth report of IPCC, the average annual air temperature of the Aral region in 2005 was about 1.1–1.7 °C higher than that of 1901. While a slight rise in atmospheric precipitation increases runoff of the Syr Darya and Amu Darya Rivers within the Aral drainage basin, the average monthly precipitation over the catchment areas of the Amu Darya River for the period 1979–2001 revealed an obvious decreasing trend from 7~8 to 4~5 km3 (Nezlin et al. 2004). There is no doubt that these climatic variations have made contribution to the associated land cover changes. However, the simulated precipitation in climate projections for this region remains highly uncertain in the recent rounds of climate model inter-comparisons (de Beurs et al. 2015).
Fig. 5

Climate changes in the Aral Sea region and hydrological variations of the Aral Sea during 1960–2015. a, b, c displays the temperature and precipitation dynamics according to the three meteorological stations located around the Aral Sea; d shows the changes of water level, area and inflow runoffs of the sea)

Agricultural activities

Arid lands throughout the world have been confirmed to be highly susceptible to the impacts of human intervention. To some extent climate change did increase the water incomings as well out-comings of the Aral Sea. But it has been generally accepted that the most dramatic land use changes were driven by the rapid and massive expansion of irrigation, water diversion and conversion of desert rangelands into irrigated croplands. For example, in the Aral Sea basin most of the fresh water in the rivers is diverted through one of the largest irrigation systems in the world to ~ 8.5 million ha of cropland (Bekchanov et al. 2016). And its upstream irrigation schemes for growing rice and cotton consumed 90% of the natural flow of water from the Tian Shan Mountains during the 1960s (Micklin 2007). During the summer months, when demand for irrigation is at its highest, little water reaches the Sea, while in winter diversions for irrigation and relatively large amounts of water used for leaching and to upstream reservoirs to produce electricity mean almost no inflow to the sea in the whole year. In the past 50 years, irrigation water withdrawal from both the Syr Darya and the Amu Darya continuously was carried on and the volume of the remaining runoff in the rivers and inflow into the Aral Sea is critically less than those before 1961 (Fig. 5d).

Land cover and ecological environment changes of the Aral Sea have become so serious that they are at the flashpoint stage for some locations. The deep causes of the Aral Crisis are the improper managements of a complex natural economic system that roots in the Aral Sea catchment. Hence, there are several pieces of suggestions for the competent management and conservation of ecological environment in the Aral drainage basin. Appropriate land use management, including urban planning and agricultural transformation, could reduce some of the pressures on land degradation. At the same time, more effective water resource management practices should be carried out immediately based on degradation level to minimize the human-induced impacts. It has been confirmed that the generalized productivity of irrigated lands in Central Asia does not depend on the irrigation water volume (Esenov et al. 1992). Thus, the temporary removal of these lands from crop rotation will not only free the necessary river runoff fed for the Aral Sea, but also preserve the arable land in the region. Secondly, appropriate measures to mitigate current land desertification process as well as to prevent their further degradation should be taken immediately. To this end, one proved effective means is to carry out large-scale afforestation based on identifying highly unstable regions that are suitable for tree or grass planting in the study area, especially those salt- and arid-tolerant species that could successfully survive these hard environmental conditions and fulfill their ecological function.

Conclusion

SVM technique was successful in classifying the LULC in the Aral Kum desert region using Landsat imagery integrated with environmental variables, suggesting that this technique can be reliably used for mapping LULC in similar arid and semiarid ecosystems. The results of nearly 40 years of observations demonstrate that Landsat classifications can provide accurate landscape change maps and statistics of the study area. At present, the bare area and salt soil, respectively, cover 38.99 and 31.29% of the old Aral Sea and are prone to wind deflation and developing active sources of salt dust. In contrast, the occurrence of vegetation is relatively scarce, only occupying 15.38%. Change detection analysis confirmed that the major land cover evolution trend in the desiccated bed is as follows: “Water” disappeared and the exposed salt-abundant sea floor immediately formed “Salt crust” along the shoreline, and then it converted into “Salt soil” and finally into “Bare area.” This could be used to predict shifts in dust composition that will lead to decreasing salt loads in eolian dust and subsequently reduced deposition of salt in downwind areas. Still, the ongoing shrinking of the Aral Sea means the expansion of salt-affected surfaces will continue, as well as the generation of huge amounts of deflated salt dust.

The performance of Markov–CA model in predicting LULC distribution for 2015 reveals the potential and merit of applying this approach for projecting future LULC change in similar arid regions. The simulated potential distribution of the LULC classes in 2025 indicated that the changes the landscape has experienced in the recent past are likely to go on. However, it must be pointed out that the Markov–CA model only takes into account the surrounding natural environment, and the cellular states are not moving. It is not considered to play a decisive role in the social environment and the interaction of the dynamic changes in land cover. Therefore, the simulated human decision-making model is still a weak link that needs further research. The LULC changes occurring in the region are the results of human activities as well as climate variations of the catchment. The simulation results could be used as a guide to ecological recovery planning in the area, helping policy-makers improve land use management plans to balance economic development and environmental conservation.

Notes

Acknowledgements

The authors thank the USGS for providing the Landsat data. And special thanks go to the editor’s and three reviewers’ precious comments and suggestions for the manuscript. This research was conducted under the support of the Strategic Priority Research Program of Chinese Academy of Sciences (Grant No. XDA2006030102) and the National Natural Science Foundation of China (Grant Nos. U1603242, 41471098, 41601440).

References

  1. Abd El-Kawy OR, Rød JK, Ismail HA, Suliman AS (2011) Land use and land cover change detection in the western Nile delta of Egypt using remote sensing data. Appl Geogr 31:483–494CrossRefGoogle Scholar
  2. Bajocco S, De Angelis A, Salvati L (2012) A satellite-based green index as a proxy for vegetation cover quality in a Mediterranean region. Ecol Ind 23:578–587CrossRefGoogle Scholar
  3. Baker WL (1989) A review of models of landscape change. Landsc Ecol 2:111–133CrossRefGoogle Scholar
  4. Bekchanov M, Ringler C, Bhaduri A, Jeuland M (2016) Optimizing irrigation efficiency improvements in the Aral Sea Basin. Water Resour Econ 13:30–45CrossRefGoogle Scholar
  5. Boschetti L, Stehman SV, Roy DP (2016) A stratified random sampling design in space and time for regional to global scale burned area product validation. Remote Sens Environ 186:465–478CrossRefGoogle Scholar
  6. Breckle SW, Wucherer W (2012) Climatic conditions in the Aralkum. In: Breckle S-W, Wucherer W, Dimeyeva LA, Ogar NP (eds) Aralkum—a man-made desert: the desiccated floor of the Aral Sea (Central Asia). Springer, Berlin, pp 49–72CrossRefGoogle Scholar
  7. Congalton RG, Green K (2009) Assessing the accuracy of remotely sensed data: principles and practices. CRC Press, Taylor and Francis, Milton ParkGoogle Scholar
  8. Cretaux J-F, Letolle R, Bergé-Nguyen M (2013) History of Aral Sea level variability and current scientific debates. Glob Planet Change Part A 110:99–113CrossRefGoogle Scholar
  9. de Beurs KM, Henebry GM, Owsley BC, Sokolik I (2015) Using multiple remote sensing perspectives to identify and attribute land surface dynamics in Central Asia 2001–2013. Remote Sens Environ 170:48–61CrossRefGoogle Scholar
  10. Dukhovny V, Navratil P, Rusiev I, Stulina G, Roshenko Y (2008) Comprehensive remote sensing and ground based studies of the dried Aral Sea bed. SIC ICWC, Tashkent, p 173Google Scholar
  11. Esenov SE, Sydykov ZS, Altunin VS, Tursunov AA, Telemtaev MM (1992) The Aral Sea problem must be solved. Hydrotech Constr 26:131–133CrossRefGoogle Scholar
  12. Foody GM (2003) Remote sensing of tropical forest environments: towards the monitoring of environmental resources for sustainable development. Int J Remote Sens 24:4035–4046CrossRefGoogle Scholar
  13. Ghimire B, Rogan J, Galiano VR, Panday P, Neeti N (2012) An evaluation of bagging, boosting, and random forests for land-cover classification in cape cod, Massachusetts, USA. GISci Remote Sens 49:623–643CrossRefGoogle Scholar
  14. Glantz MH (2005) Water, climate, and development issues in the Amu Darya basin. Mitig Adapt Strat Glob Change 10:23–50CrossRefGoogle Scholar
  15. Groll M, Opp C, Aslanov I (2013) Spatial and temporal distribution of the dust deposition in Central Asia: results from a long term monitoring program. Aeolian Res 9:49–62CrossRefGoogle Scholar
  16. Guan D, Li H, Inohae T, Su W, Nagaie T, Hokao K (2011) Modeling urban land use change by the integration of cellular automaton and Markov model. Ecol Model 222:3761–3772CrossRefGoogle Scholar
  17. Halmy MWA, Gessler PE, Hicke JA, Salem BB (2015) Land use/land cover change detection and prediction in the north-western coastal desert of Egypt using Markov-CA. Appl Geogr 63:101–112CrossRefGoogle Scholar
  18. Han D, Chan L, Zhu N (2007) Flood forecasting using support vector machines. J Hydroinf 9:267–276CrossRefGoogle Scholar
  19. Han J, Hayashi Y, Cao X, Imura H (2009) Application of an integrated system dynamics and cellular automata model for urban growth assessment: a case study of Shanghai, China. Landsc Urb Plan 91:133–141CrossRefGoogle Scholar
  20. Hansen MC (2012) Classification trees and mixed pixel training data. In: Giri CP (ed) Remote sensing of land use and land cover, principles and applications. CRC Press, Taylor and Francis Group, Boca Raton, pp 127–136CrossRefGoogle Scholar
  21. Indoitu R, Orlovsky L, Orlovsky N (2012) Dust storms in Central Asia: spatial and temporal variations. J Arid Environ 85:62–70CrossRefGoogle Scholar
  22. Indoitu R, Kozhoridze G, Batyrbaeva M, Vitkovskaya I, Orlovsky N, Blumberg D, Orlovsky L (2015) Dust emission and environmental changes in the dried bottom of the Aral Sea. Aeolian Res 17:101–115CrossRefGoogle Scholar
  23. Issanova G, Abuduwaili J, Kaldybayev A, Semenov O, Dedova T (2015) Dust storms in Kazakhstan: frequency and division. J Geol Soc India 85:348–358CrossRefGoogle Scholar
  24. Kondratyev K, Grigoryev A, Zhvalev V, Melentyev V (1985) An integrated study of dust storms in the Aral region. Meteorol Hydrol 4:32–38Google Scholar
  25. Kozhoridze G, Orlovsky L, Orlovsky N (2012) Monitoring land cover dynamics in the Aral Sea region by remote sensing. vol. 8538. pp 85381V–85381VGoogle Scholar
  26. Kravtsova VI, Tarasenko TV (2010) Space monitoring of Aral Sea Degradation. Water Resour 37:285–296CrossRefGoogle Scholar
  27. Lioubimtseva E, Henebry GM (2009) Climate and environmental change in arid Central Asia: impacts, vulnerability, and adaptations. J Arid Environ 73:963–977CrossRefGoogle Scholar
  28. Löw F, Navratil P, Kotte K, Schöler HF, Bubenzer O (2013) Remote-sensing-based analysis of landscape change in the desiccated seabed of the Aral Sea—a potential tool for assessing the hazard degree of dust and salt storms. Environ Monit Assess 185:8303–8319CrossRefGoogle Scholar
  29. Melgani F, Bruzzone L (2004) Classification of hyperspectral remote sensing images with support vector machines. IEEE Trans Geosci Remote Sens 42:1778–1790CrossRefGoogle Scholar
  30. Micklin P (2007) The Aral Sea disaster. Ann Rev Earth Planet Sci 35:47–72CrossRefGoogle Scholar
  31. Micklin P (2008) Using satellite remote sensing to study and monitor The Aral Sea and adjacent zone. In: Qi J, Evered KT (eds) Environmental problems of Central Asia and their economic, social and security impacts. Springer, The Netherlands, pp 31–58CrossRefGoogle Scholar
  32. Micklin P (2010) The past, present, and future Aral Sea. Lakes Reserv Res Manag 15:193–213CrossRefGoogle Scholar
  33. Muller MR, Middleton J (1994) A Markov model of land-use change dynamics in the Niagara region, Ontario, Canada. Landsc Ecol 9:151–157Google Scholar
  34. Nezlin NP, Kostianoy AG, Lebedev SA (2004) Interannual variations of the discharge of Amu Darya and Syr Darya estimated from global atmospheric precipitation. J Mar Syst 47:67–75CrossRefGoogle Scholar
  35. O’Hara SL, Wiggs GFS, Mamedov B, Davidson G, Hubbard RB (2000) Exposure to airborne dust contaminated with pesticide in the Aral Sea region. The Lancet 355:627–628CrossRefGoogle Scholar
  36. Orlovsky L, Orlovsky L (2002) White sand storms in Central Asia, global alarm: dust and sand storms from the World’s Drylands. United Nations Convention to Combat Desertification, Bangkok, pp 169–201Google Scholar
  37. Orlovsky L, Tolkacheva G, Orlovsky N, Mamedov B (2004) Dust storms as a factor of atmospheric air pollution in the Aral Sea basin. Adv Air Pollut Ser 14:353–362Google Scholar
  38. Overmars KP, de Koning GHJ, Veldkamp A (2003) Spatial autocorrelation in multi-scale land use models. Ecol Model 164:257–270CrossRefGoogle Scholar
  39. Pal M, Mather PM (2005) Support vector machines for classification in remote sensing. Int J Remote Sens 26:1007–1011CrossRefGoogle Scholar
  40. Peterson LK, Bergen KM, Brown DG, Vashchuk L, Blam Y (2009) Forested land-cover patterns and trends over changing forest management eras in the Siberian Baikal region. For Ecol Manag 257:911–922CrossRefGoogle Scholar
  41. Petropoulos GP, Kontoes C, Keramitsoglou I (2011) Burnt area delineation from a uni-temporal perspective based on Landsat TM imagery classification using Support Vector Machines. Int J Appl Earth Obs Geoinf 13:70–80CrossRefGoogle Scholar
  42. Pijanowski BC, Brown DG, Shellito BA, Manik GA (2002) Using neural networks and GIS to forecast land use changes: a land transformation model. Comput Environ Urb Syst 26:553–575CrossRefGoogle Scholar
  43. Pontius GR, Malanson J (2005) Comparison of the structure and accuracy of two land change models. Int J Geogr Inf Sci 19:243–265CrossRefGoogle Scholar
  44. Roy SB, Smith M, Morris L, Orlovsky N, Khalilov A (2014) Impact of the desiccation of the Aral Sea on summertime surface air temperatures. J Arid Environ 110:79–85CrossRefGoogle Scholar
  45. Sang L, Zhang C, Yang J, Zhu D, Yun W (2011) Simulation of land use spatial pattern of towns and villages based on CA–Markov model. Math Comput Model 54:938–943CrossRefGoogle Scholar
  46. Semenov OE (2012) Dust storms and sandstorms and aerosol long-distance transport. In: Breckle S-W, Wucherer W, Dimeyeva LA, Ogar NP (eds) Aralkum—a man-made desert: the desiccated floor of the Aral Sea (Central Asia). Springer, Berlin, pp 73–82CrossRefGoogle Scholar
  47. Shao Y, Lunetta RS (2012) Comparison of support vector machine, neural network, and CART algorithms for the land-cover classification using limited training data points. ISPRS J Photogramm Remote Sens 70:78–87CrossRefGoogle Scholar
  48. Shi W, Wang M (2015) Decadal changes of water properties in the Aral Sea observed by MODIS-Aqua. J Geophys Res Oceans 120:4687–4708CrossRefGoogle Scholar
  49. Shi W, Wang M, Guo W (2014) Long-term hydrological changes of the Aral Sea observed by satellites. J Geophys Res Oceans 119:3313–3326CrossRefGoogle Scholar
  50. Singer A, Zobeck T, Poberezsky L, Argaman E (2003) The PM10and PM2·5 dust generation potential of soils/sediments in the Southern Aral Sea Basin, Uzbekistan. J Arid Environ 54:705–728CrossRefGoogle Scholar
  51. Spivak L, Terechov A, Vitkovskaya I, Batyrbayeva M, Orlovsky L (2012) Dynamics of dust transfer from the desiccated Aral Sea bottom analysed by remote sensing. In: Breckle S-W, Wucherer W, Dimeyeva LA, Ogar NP (eds) Aralkum—a man-made desert: the desiccated floor of the Aral Sea (Central Asia). Springer, Berlin, pp 97–106CrossRefGoogle Scholar
  52. Stulina G, Sektimenko V (2004) The change in soil cover on the exposed bed of the Aral Sea. J Mar Syst 47:121–125CrossRefGoogle Scholar
  53. Szuster BW, Chen Q, Borger M (2011) A comparison of classification techniques to support land cover and land use analysis in tropical coastal zones. Appl Geogr 31:525–532CrossRefGoogle Scholar
  54. Vapnik V (1998) Statistical learning theory. Wiley, New YorkGoogle Scholar
  55. Wickramasuriya RC, Bregt AK, van Delden H, Hagen-Zanker A (2009) The dynamics of shifting cultivation captured in an extended Constrained Cellular Automata land use model. Ecol Model 220:2302–2309CrossRefGoogle Scholar
  56. Wiggs GFS, O’Hara SL, Wegerdt J, Van Der Meer J, Small I, Hubbard R (2003) The dynamics and characteristics of aeolian dust in dryland Central Asia: possible impacts on human exposure and respiratory health in the Aral Sea basin. Geogr J 169:142–157CrossRefGoogle Scholar
  57. Wucherer W, Breckle S-W (2001) Vegetation dynamics on the dry sea floor of the Aral Sea. In: Breckle S-W, Veste M, Wucherer W (eds) Sustainable land use in deserts. Springer, Berlin, pp 52–68CrossRefGoogle Scholar
  58. Wulder MA, White JC, Goward SN, Masek JG, Irons JR, Herold M, Cohen WB, Loveland TR, Woodcock CE (2008) Landsat continuity: issues and opportunities for land cover monitoring. Remote Sens Environ 112:955–969CrossRefGoogle Scholar
  59. Yuan F, Sawaya KE, Loeffelholz BC, Bauer ME (2005) Land cover classification and change analysis of the Twin Cities (Minnesota) metropolitan area by multitemporal landsat remote sensing. Remote Sens Environ 98:317–328CrossRefGoogle Scholar

Copyright information

© Islamic Azad University (IAU) 2018

Authors and Affiliations

  1. 1.State Key Laboratory of Desert and Oasis Ecology, Xinjiang Institute of Ecology and GeographyChinese Academy of SciencesUrumqiChina
  2. 2.Chinese Academy of Sciences Research Center for Ecology and Environment of Central AsiaUrumqiChina
  3. 3.University of Chinese Academy of SciencesBeijingChina

Personalised recommendations