Examining long-term natural vegetation dynamics in the Aral Sea Basin applying the linear spectral mixture model

Linear spectral mixture model

The mixed pixels do not completely belong to a certain kind of feature. In order to achieve higher classification accuracy, it is necessary to decompose the mixed pixels into the percentage (abundance) of a kind of feature in the pixel, that is, the decomposition of mixed pixels (Small, 2004; Deng & Wu, 2016). LSMM has obvious physical meaning, simple operation, reliable accuracy, directly related to ground coverage, and solves the problem of mixed pixels (Deng et al., 2018). The resolution of MODIS data is 0.5 km, and there is a phenomenon of pixel mixing. LSMM addresses the issue by providing valuable sub-pixel information (Zhang et al., 2019b), so we used LSMM for classification. This study is a large-scale, long-term monitoring of natural vegetation coverage, while traditional platforms require a long time to process data. Therefore, we improved the suitability of the LSMM method for MODIS data and introduced it to the GEE platform. LSMM refers to the spectral reflectance of the pixel in a certain band, which is a linear combination of the reflectance of the components that make up the pixel with the proportion of their area as the weight coefficient. This is expressed as: (1) $$\{\begin{array}{c}{R}_{i\lambda }=\sum _{k=1}^n{f}_{ki}{r}_{k\lambda }+{\xi }_{i\lambda }\\ \sum _{k=1}^m{f}_{ki}=1\end{array}0\le f_{ki}\le 1$$where $R_{i\lambda }$ is the spectral reflectance of the i-th pixel in the λ band, $r_{k\lambda }$ is the spectral reflectance of the $k$-th basic component in the $\lambda$ band, $f_{\mathrm{k}\mathrm{i}}$ is the abundance of the $k$-th end element in the $i$-th pixel; $n$ is the number of end elements; and ${\xi }_{i\lambda }$ is the residual error value, which somewhat represents the multiple reflection and transmission of light between image tuples. It has a nonlinear mixing effect. $n$ is the number of end elements (Wu & Murray, 2003).

Endmember selecting

It is important to select the end element in the LSMM. The quality of the end element directly affects the overall accuracy of the experiment. Generally, three to four end elements are selected through tests, and the inversion effect is good (Zanotta et al., 2014). Because the image will be affected by the atmosphere, terrain, and sensors, the actual spectral curve differs from the spectral curve in the ground measurement and ground object spectral library, so it is more accurate to obtain the end element from the image (Zare & Ho, 2014). There are many methods to select image endmembers. But the disadvantage of the PPI algorithm is that the selected vector has a high degree of arbitrariness; the performance of the N-FINDR algorithm is largely related to the nature of the initially selected endmember (Zhong, Zhao & Zhang, 2014; Jafarzadeh & Hasanlou, 2019); and the nature of the AMEE algorithm also depends on the relationship between the spatial characteristics of the structure element and the spectral distribution in the scene (Li et al., 2011). Generally speaking, all kinds of endmember extraction algorithms are still in the exploratory stage, and each algorithm has its own advantages and disadvantages, which needs further improvement. Here, we tended to choose the final elements with a basis to improve the accuracy of the classification results, and visual interpretation is also widely used to verify the accuracy of machine classification (Yu et al., 2017), so visual interpretation was selected for endmember selection (Feng et al., 2019; Jiang, Van der Werff & Van der Meer, 2020). The specific experimental steps were (1) MOD13A1 was overlapped with Landsat 7 and Google Earth images; (2) the pure ground object pixel of Landsat 7 and Google Earth images were interpreted and recognized visually; and (3) the pixel values in MOD13A1 at the same position as the pure ground object pixel were taken as the end element values.

According to the distribution of ground features in the study area and the principle of selecting end elements (Micklin, 2016), four types of end elements were selected: cultivated land, vegetation, desert, and water body. The experiment selected 1,064 end elements from April to November in each year from 2000 to 2018. The water body of the MOD13A1 product was Nodata. Figure 3 shows the other end member values in 2018. The endmember values of each month in other years were selected using the same method.

Trend line analysis

Regression analysis is a mathematical model used to study the relationship between multiple variables and is widely employed in long-term studies and analysis of land cover change (Zhu et al., 2019; Parvin, 2019). The mathematical model implemented in this study was: (3) $$y=kx+a+\mathrm{\partial }$$where a and k are unknown constants, $\mathrm{\partial }$ is the error, and x is the time independent variable. y is the dependent variable for ground object cover areas, k can be calculated by the x, y observations: (4) $$k={\displaystyle \frac{{\sum }_{i=1}^n\left(x_i-{\displaystyle \frac{1}{n}{\sum }_{i=1}^n{x}_i}\right)\left(y_i-{\displaystyle \frac{1}{n}{\sum }_{i=1}^n{y}_i}\right)}{{\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}}$$where n is the year.

The slope of the equation k > 0 indicates that the ground cover increases with the year. The fitting accuracy is checked using the decision coefficient R². The closer x is to 1, the higher the goodness of fit.

Sen trend analysis and Mann–Kendall trend test

Sen trend analysis can calculate the trend degree, and the Mann–Kendall trend test can test the significance of the change trend (Drápela & Drápelová, 2011). The combination of the two methods has formed a new long time series trend analysis method, which can reduce the impact of data noise (Bi et al., 2020). The Sen trend analysis combined with the Mann–Kendall test can be utilized to determine the variation trend of long-term series data (Jiang et al., 2015). The calculation formula is: (5) $$\beta =\mathrm{M}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\left({\displaystyle \frac{x_j-x_i}{j-i}}\right),j>i,i=1,2,3\dots ,N$$Among, $x_j$ and $x_i$ are elements of the time series of the trend to be analyzed. When β > 0, it indicates an upward trend; when β < 0, it indicates a downward trend.

Mann–Kendall trend test does not require the test data to follow certain distribution rules, and can avoid the influence of a few outliers. The test statistic S is calculated as follows: (6) $$S=\sum _{K=1}^{n-1}\sum _{j=k+1}^n\mathrm{S}\mathrm{g}\mathrm{n}(X_J-X_K)$$Inside, (7) $$\mathrm{S}\mathrm{g}\mathrm{n}(X_J-X_K)=\{\begin{array}{c}+1X_J-X_K>0\\ 0X_J-X_K=0\\ -1X_J-X_K<0\end{array}$$S is a normal distribution and the mean value of 0, variance $\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)={\displaystyle \frac{n\left(n-1\right)\left(2n+5\right)}{18}}$, When n > 10, the standard normal variable is calculated as follows: (8) $$Z=\{\begin{array}{c}{\displaystyle \frac{s-1}{\sqrt{V_{ar\left(S\right)}}}S>0}\\ 0S=0\\ {\displaystyle \frac{S+1}{\sqrt{V_{ar\left(S\right)}}}S<0}\end{array}$$

In this study, the confidence level α = 0.01 or α = 0.05, Z > 0 was the rising trend; Z < 0 is a downward trend, and when |Z| ≥ 1.96 and |Z| ≥ 2.576, the reliability was 95% and 99%, respectively (Gocic & Trajkovic, 2013).

Principal component analysis

Principal Component Analysis (PCA) is a statistical analysis method that converts multiple variables into unrelated comprehensive variables (Zhang et al., 2020). The basic principle of PCA is to ensure the minimum loss of information, reduce the dimensionality of the original data, and improve research efficiency while ensuring research accuracy (Bunnell et al., 2020). PCA as a descriptive tool needs no distributional assumptions and, as such, is very much an adaptive exploratory method which can be used on numerical data of various types (Jolliffe & Cadima, 2016). Its mathematical model is as follows: (9) $$X=\left(X_{ij}\right)A\ast B,i=1,2,\cdots ,B$$

A is number of study areas, original sample matrix X of B selection indicators.

Calculate the correlation coefficient matrix between each indicator $R_{b\ast b}$, eigenvalues and normal eigenvectors $e_j$, this gives the principal component $T_i$, as follows: $T_i=X_{ej}$.

The contribution rate formula is as follows: $a=\sum _i^q{a}_j$

When the j-th principal component variance contribution rate is 85–95%, the original B index information can be reflected. At the same time, the comprehensive score of ecological security in the study area can be obtained, as follows: $W=a{X}_1+b{X}_2+\cdots +x{X}_x$, where X is eigenvectors; a, b, … , x are standardized data of the original data (Licciardi et al., 2012).

Spatio–temporal variation characteristics of ground cover

In order to better explore the driving forces affecting natural vegetation coverage, this study quantitatively analyzes the dynamic changes in the abundance maps of other features of the Aral Sea Basin (Table 3). The average area of cultivated land in the Aral Sea Basin was 136,600 km², accounting for 10.54% of the total basin area. The area of cultivated land increased from 96,600 km² in 2000 to 135,000 in 2018, an increase of 3.10%. The growth rate was large in 2000–2005 and 2011–2014, and the total increase in area was 166,000 km², mainly distributed on both sides of the Amu Darya, upstream of the Syr Darya, and the Zeravshan River. The reduction was large in 2005–2008 and 2014–2017, with an area reduction of 127,600 km², mainly distributed in Kokand and northern Tashkent. Overall, the distribution decreased inside the city and increased around the city periphery and the river (Figs. 5B and 6A).

The average desert area of the Aral Sea Basin is 778,300 km², accounting for 62.85% of the basin area, nearly 3/5. From 2000 to 2018, the average annual desert area decreased from 858,500 km² to 778,300 km², a decrease of 6.47%. The reduced area is striped along the Aral Sea basin boundary. From 2005–2008 and 2009–2011 there was an increasing trend, with a total growth area of 300,600 km². The more obvious increase was observed around the South Aral Sea. As the lake subsided, new deserts formed around the edge. The distribution of deserts decreased around the edge and increased in the center (Figs. 5C and 6B).

The average area of the water body in the Aral Sea Basin was 63,800 km², accounting for 5.15% of the entire basin area, and its coverage was low. It decreased from 72,500 km² in 2000 to 46,400 km² in 2018, a decrease of 2.11%. The reduction was largely in 2004–2008, with a total area reduction of 88,400 km². The primary reason is that the amount of water in the South Aral Sea has decreased dramatically. The increased water areas are mainly distributed in the North Aral Sea and the Pamirs. The reason for the reversal of the South Aral Sea and the North Aral Sea is that Kazakhstan built a dam in 2005 to introduce the Syr Darya from the South Aral Sea to the North Aral Sea. Over the past nineteen years, water bodies reduced by a total area of 22,300 km² more than the increase in area, and the water body is decreasing (Figs. 5D and 6C).

Correlation analysis of desert and cultivated land cover change

By analyzing the inter-annual spatial distribution characteristics of the cover area of cultivated land and deserts, we obtained a slope of k < 0 and a determination coefficient of 0.44 (Fig. 7). This shows that the change in cultivated land coverage has a significant relationship with the desert area. Areas with significantly increased cultivated land coverage showed an expanding trend from the original farmland to the surrounding areas (the red part in Fig. 6A). The desert shows a degrading trend (black part in Fig. 6B). The spatial distribution of desert degradation areas and farmland growth areas is similar, mainly distributed along the banks of the Amu Darya, deltas, and around cities. In desert areas, the temperature is relatively high, and the surface water evaporates quickly which limits the growth of crops. The development of irrigated farmland in this area requires a large amount of river water to be directed to the cultivated land, and the soil in Central Asia has a high salt content. In order to reduce the salinity, cultivated land must be ‘washed’ regularly. According to statistics, more than 1 billion km³ of water is used to clean farmland every year. A large amount of surface runoff is used for irrigation, resulting in a decrease in the amount of water in the lower Amu Darya and Syr Darya Basin, which exacerbated the shrinkage of the Aral Sea (Fig. 6C).

Impact of climate variability on natural vegetation cover

Climate change profoundly affects the growth and distribution of plants, temperature affects the growth cycle of vegetation, and precipitation affects the composition and physiological changes of vegetation. The Aral Sea is in a semi-arid area, and the climate has a greater impact on vegetation cover changes. As shown in Fig. 8A, the temperature growth rate from 2000 to 2018 is 0.56 °C/a, which indicates climate warming in this region in the past 19 years. However, the temperature in winter showed the opposite trend, and the temperature in the Pamirs dropped significantly. The region has a climate with hotter summers and colder winters. The precipitation in this area also has a slight increase trend, and there is a clear increasing trend of precipitation in Kazakhstan, where Rain Fed Crops account for 10.12% of the total cultivated land area in Central Asia (Lioubimtseva, 2014).

Impact of human activities on natural vegetation cover

The Aral Sea Basin is located in five Central Asian countries. Since the disintegration of the Soviet Union in 1991, the economic and social patterns of the region have varied. Among them, agriculture has changed from state-owned to private and free operations, and human activities inevitably affect the growth of regional vegetation. Population quantity, distribution, migration and human production activities can destroy the surface and land cover, and the physical and chemical properties of soil and water. They are closely related to changes in regional vegetation cover. The state of economic development reflects the degree of urbanization and land use in the region. In the past nineteen years, the GDP growth rate was 6.7% (Fig. 8B).

Driving forces of natural vegetation cover changes

We selected five factors to determine natural vegetation cover changes: annual average temperature, total precipitation, total population, GDP, and cultivated land area. PCA was used to quantitatively analyze the impact of various factors on vegetation coverage. The total population, GDP and arable land area contributed 0.82 to the first principal component, which can somewhat reflect the impact of human activities on vegetation coverage. The annual average temperature and annual total precipitation contributed 0.87 to the second principal component, which can somewhat represent the impact of climate factors on regional vegetation (Table 4). In the past nineteen years, the impact of climatic factors on the Aral Sea Basin generally initially increased and then decreased. The impact of human factors on the vegetation coverage of the Aral Sea Basin initially decreased and then increased (Fig. 8C). Overall, human activities were the major driving factor that affected the changes in vegetation cover in the Aral Sea Basin, of which the change in cultivated land area had a higher impact. Among the natural factors, precipitation was the main factor affecting the change in vegetation coverage in the Aral Sea Basin, followed by temperature.

Compared with natural factors, the impact of human activities on vegetation cover changes is more rapid and direct. According to statistics, the petrochemical, non-ferrous metallurgical, and power industries in the region show a growth trend, while the output value of the light and wood industries shows a downward trend. The rapid industrialization has caused substantial damage to vegetation coverage. Regional economic development has a profound impact on regional population growth and migration. Combining the results of the Mann–Kendall test and Sen median trend analysis can effectively characterize the trend changes of ground features. By analyzing the change curve of natural vegetation and arable land, we found that the area of arable land increased in 2004–2005, 2007–2013 and 2015–2018; the vegetation area decreased. Jiang et al. (2017) also pointed out that the expansion of irrigated farmland led to the degradation of some natural vegetation. The Aral Sea Basin is also a hot spot for abandoned farmland. The residents of the former prosperous ports migrated to Kazakhstan and Uzbekistan due to the dryness of the Aral Sea, and a large amount of arable land was abandoned. Abandoned farmland was gradually restored to other vegetation types.