The effect of farmland on the surface water of the Aral Sea Region using Multi-source Satellite Data

Study area

The Aral Sea Region is located on the border between northwest Uzbekistan and southwest Kazakhstan. It covers 130,000 km², and includes North Aral, East Aral, West Aral and irrigated farmland of the Aral Sea Region (Fig. 1). This is a typical arid area with an average annual temperature of 11.62 °C and precipitation of 126.26 mm.

Data preparation

Detection algorithm

Detection algorithm of farmland change

LandTrendr (Landsat-based Detection of Trends in Disturbance and Recovery) algorithm is an approach to extract the spectral trajectories of land surface from Landsat time-series stacks (Watts & Laffan, 2014). This algorithm can reflect farmland variations by detecting the temporal-spectral trajectory of each pixel. The input of every pixel for the algorithm is a time series with one spectral value or index. This algorithm also needs to set parameters, such as noise-induced spikes (outliers), potential vertices (breakpoints), fitting trajectories, and the optimal number of segments (Watts & Laffan, 2014).

LandTrendr algorithm on the GEE platform was used to detect the farmland change in the Aral Sea Region (Fig. 2A). Before running this algorithm, the features of farmland variations were analyzed, tested different combinations of parameter values. The NDVI was determined as the spectral index, and other parameters were calculated (Table S1). NDVI of farmland ranged from 0.45 to 0.69 in the Aral Sea Region and plummeted after abandonment to range from 0.2 to 0.3. Thus, NDVI of abandoned farmland was set between 0.2 and 0.3. The mentioned thresholds were adopted to detect the change patterns of farmland, the change time of occurrence and end, and duration.

Detection algorithm of waters change

Via the GEE platform, the random forest algorithm was adopted to detect the variation of water area (Fig. 2B). Random forest is a machine learning algorithm that is complies with the decision tree. This algorithm uses CART clustering for prediction and classification (Breiman, 2001). Two primary parameters should be set before implementing the random forest algorithm. To be specific, the first refers to the number of splitting times (mtry) of the branch tree in each model. This parameter should be set as the square root value of variables and 2 times or 1/2 of the prediction’s open value. In this paper, the variables specifically refer to the bands of remote sensing images (Table S2). The second refers to the number of generated trees (ntree) in model operation, and the amount of model calculation is proportional to the value of ntree. The ntree was set as 500 because it will be pretty obvious with a bit of tinkering that the predictions from the model will not change much after this value.

The accuracy of model depends more on quality of training samples. Therefore, we selected 200 pixels representing water and another 200 pixels representing non-water based on Landsat imagery. In order to access accuracy of results, we visually examined all selected pixels (from 1987 to 2019) and used high-spatial-resolution Google Earth imagery for further manual interpretation of land cover (Fig. 3). Therefore, we determined and recorded whether and when water pixels change occurred.

The ICESat data was adopted to analyze the alteration in water level in the Aral Sea. At first, the data format of GLA14 was converted. Second, GLA14 was adopted to extract the water level of the Aral Sea (Eq. (2)). Lastly, the water volume was calculated according to the alteration in water level and the variation of water area of the Aral Sea (Eq. (3)). Meanwhile, researchers compared ICESat data with airborne LiDAR data and DEM from GPS data. The error of ICESat data is less than 3 cm (Kurtz et al., 2008; Fricker, 2005).

(2) $$W{L}_T={\displaystyle \frac{{\sum }_{i=1}^{n}wl}{n}}$$

(3) $$WV=W{L}_T\ast WA$$where WL_T is mean water level. wl is the land elevation of every ICESat point. WV is water volume. WA is water area.

The GRACE dataset was used to extract the equivalent water height (EWH). The data contained in this dataset are units of “Equivalent Water Thickness” which represent the deviations of mass in terms of the vertical extent of water in centimeters (Watkins et al., 2015). The EWH was processed at NASA Jet Propulsion Laboratory. Actually, this dataset was produced by three institutions: CSR (U. Texas/Center for Space Research), GFZ (German Research Center for Geosciences), and JPL (NASA Jet Propulsion Laboratory). However, since each institution independently produces the coefficients, the results may be slightly different. And the mass change maps of the three institutions have typically suffered from poor observability of east-west gradients, resulting in “N-S stripes” that are conventionally removed via empirical smoothing and/or “destriping” algorithms. In order to overcome these limitations, Watkins et al., developed GRACE(-FO) mass concentration (mascon)solutions (Watkins et al., 2015; Wiese, Landerer & Watkins, 2016; Landerer et al., 2020). The solutions allow for convenient application of a priori information derived from near-global geophysical models to prevent striping. Significantly, over the land, mascon solutions have more excellent resolution for smaller spatial regions, in particular when studying secular signals. Meanwhile, for different application scenarios, the dataset has its advantages. For instance, it is more suitable for water, ice mass, and ocean bottom pressure mass change application to use JPL mascon (Wiese, Landerer & Watkins, 2016; Landerer et al., 2020). Therefore, JPL data, which have been processed by the GRACE(-FO) mascon solutions, were used in this paper. And the measurement accuracy of GRACE data is credible (5,000 km wavelength: 0.001 mm; 500 km wavelength: 0.01 mm) (Zheng et al., 2009). Water storage variation in the Aral Sea Region was analyzed (Eq. (4)).

(4) $$WS=EWHT\ast A$$where EWHT is the equivalent water height from GRACE dataset; WS is water storage; A is the area of study region.

Accuracy assessment

Two hundred pixels were selected to build a confusion matrix (CM) to assess the accuracy of the five surface features in 1987 (i.e., water, farmland, desert, shoal and irrigation land). Meanwhile, we visually examined all surface features by high-spatial-resolution Google Earth imagery. The overall accuracy, user accuracy, and producer accuracy for each feature were concluded following the indexes defined by Foody and Green (Congalton & Green, 1993; Foody, 2002). Two hundred validating pixels were selected to represent the stable area of farmland. In addition, another 200 validating pixels represent the various area of farmland for each change mode. The stable area refers to the pixels that remain unchanged as farmland for years. While the various area represents the changed farmland pixels during the whole study period. The 20% pixels of training samples in water extraction were adopted to produce a CM to access the accuracy of water extraction results.

Water-farmland coupling degree model

The coupling degree model originated from the coupling coefficient model in physics. Farmland and surface water were combined to build the water-farmland coupling degree model to explore the impact of farmland on surface water. In such a model, four indicators were selected related to surface water and farmland (Fig. 2C). At first, every indicator was normalized, and then the entropy method was used to determine their weight. The index matrix is $\mathrm{X}={\left({\mathrm{x}}_{\mathrm{a}\mathrm{b}}\right)}_{\mathrm{i}\times \mathrm{j}}$ (where i is the evaluation index; j is the index, and $x_{ab}$ is the b-th index of the a-th sample,). The coupling degree model (C) is as follows (Chang, 2018).

(6) $$C={\left\{{\displaystyle \frac{f_1\left(x\right)\ast \cdots \ast f_i\left(x\right)}{{\left|{\displaystyle \frac{f_1\left(x\right)+\cdots +f_i\left(x\right)}{i}}\right|}^i}}\right\}}^{1/1ii}$$

(7) $$f_i\left(x\right)=\sum _b^j{w}_b{x}_b$$where $f_i\left(x\right)$ is the comprehensive evaluation function of subsystem; $w_b$ is the weight of each index; $x_b$ is each index; $i$ is the number of subsystems to be access. There are two subsystems which need to be evaluated, so $i=2$. ( $C\in \left[0,1\right]$). If the C value is closer to 1, it will mean that the subsystems are more related and coordinated (Congalton & Green, 1993).

Accuracy assessment

The CM for the classification of 1987 imagery (Table S3) revealed that the overall accuracy (OA) reached over 90%. The user accuracy (UA), and producer accuracy (PA) for each class mainly were more than 86%. The accuracy of farmland is less than 90%. The reasons are elucidated below. There were considerable irregular water bodies and farmland in irrigated land of the Aral Sea Region; thus, the farmland often tended to mix with water and Irrigation land. Moreover, some sparse vegetation was planted in the irrigation land. Thus, it was difficult to distinguish by employing Landsat images with a spatial resolution of 30 m. However, the overall accuracy reached 92.1%, suggesting the random forest model’s capability. A similar method was used to assess the accuracy of extraction of water area in the Aral Sea from 1987 to 2019. The accuracy of extraction of water area reached over 90%, demonstrating the credibility of the results.

Table S4 show the farmland detection accuracy of abandon and replanting by LandTrendr. The overall accuracy of 85.3% was low for abandoned farmland for the mixture of abandoned farmland and sparse planting. For the replanting, the PA, UA, and OA are above 80%. Thus, it is sufficient for farmland to detect the change model. In summary, the OA of abandoned farmland change detection was more than 85%, demonstrating the detection method’s reliability based on the LandTrendr.

Temporal and spatial variations of farmland

The distribution of abandoned farmland, in the Aral Sea Region during 1987–2019, was rendered the disturbance map with a gradient (Fig. 4) according to the occurrent years. The abandoned region was mainly diffused around the Syr Darya and the Amu Darya, especially around the Amu Darya (Table S5). Within 9 km of these Draya rivers, the percentage of abandoned farmland reached 59.96% from 1987 to 2019. Overall, 3,129 km² of farmland was abandoned. The annual change peaked between 1987 and 1993 (Table S6), after which percentage of abandoned farmland decreased under 3% from 2002 to 2016 (Table S6). However, the percentage of abandoned farmland was elevated to 5.7% in 2017 (Table S6), which resulted in the plummet of water volume in the Aral Sea.

Waters variations of the aral sea region

The Aral Sea has been shrinking at 1,606.36 km² per year from 1987–2019. In 2019, the area of the Aral Sea took up only 15.6% of its size in 1987 (Fig. 5). With an annual decrease of 1,306.18 km², the East Aral (820.22 km²) had almost dried up by 2019. The area of the East Aral, in 2019, only accounted for 1.92% in 1987. On the contrary, as impacted by the construction of the Dike Kokaral dam in 2005, the North Aral showed a weak expansion trend with an average annual growth of 10.05 km². The area of North Aral rose to 4,658.773 km² in 2019. Kazakhstan Government built the Kokaral dam at the junction of the North Aral and East Aral (Fig. 5). In order to keep the water level of North Aral and environment stable, the water flow from North Aral to East Aral was artificially reduced. Though this policy effectively prevented the loss of North Aral, it also reduced the water flowing into East Aral. Such a dam aggravated the dry-up of East Aral. The West Aral is small and far from the irrigated area. Accordingly, the West Aral had shrunk more minor than the East Aral at 314.83 km² each year.

Likewise, the water level of the Aral Sea tended to decrease. The curve of the water level of the North Aral significantly ascended as impacted by the construction of the dam (Fig. 6). The water level of the North Aral was elevated near 0.32 m per year from 2003 to 2009. In the East Aral, the water level drastically declined at 0.5 m per year. Moreover, the East Aral was only 0.05 m in 2019. Compared with the East Aral, declining West Aral was lower as impacted by its smaller range and farther from the irrigated area. To be specific, the water level of the West Aral decreased at 0.4 m per year and dropped to 2.1 m in 2019.

Figure 7 illustrates the water mass change of the GRACE analysis in the study area. The curve of equivalent water height shows a downward trend from 2002 to 2016. The change of equivalent water height is same as the water level. The water mass was overall declining, marking a decrease by 0.97 km³ each year. The decrease in water mass represented the descending of water storage, such as the Aral Sea, rivers, reservoirs and groundwater.

Construction of water-farmland coupling degree model

According to the results of this study regarding water and farmland in the Aral Sea Region, three indexes were selected to build a comprehensive evaluation index system (Table S7). This system included two subsystems (surface water system, farmland system). Water subsystem had two evaluation indexes, and farmland subsystem had one evaluation index (Table S7).

The entropy method was adopted to calculate the weight of each index. According to the mentioned weights, the coupling degree was obtained between the two subsystems from 2003 to 2009. In addition, the results indicated that the annual coupling degree value was above 70%, and the average coupling degree reached 0.903 (Table S8). By the criterion of coupling degree (Table S9) (Chang, 2018), only 2003 and 2005 pertain to intermediate level coupling coordination. Other years belong to high-quality coupling coordination. In addition, the average coupling degree belongs to the high-quality coupling coordination. The mentioned coupling degree results suggested that the greater coupling degree between the two subsystems, and farmland changes affected the variations of surface water in the Aral Sea Region.

Benefits of methods

The results shown that, the loss of water by abandoned farmland will increase the burden of normal water use. During the past years, many scholars were committed to analyzing the reasons for the decrease of surface water in the Aral Sea Region. Compared with previous researches, our work has similarities and differences. For similarities, the surface water loss was caused by improper irrigation management. On the contrary, our research was from the perspective of farmland abandonment. Few studies have analyzed the loss of water caused by abandoned farmland. They were analyzed effect of land water storage on surface water (Jin et al., 2017; Gaybullaev, Chen & Gaybullaev, 2012; Parajuli & Yang, 2017).

The coupling degree model was used for the first time to monitor the effect of farmland on the surface water of the Aral Sea Region. Most previous researches regarded farmland and surface water as independent objects, which could not acquire the natural connection between them. Therefore, changes in farmland and surface water were put into a quantifiable system that can be used to explore the substantial impact of farmland changes on the surface water. According to the result, the coupling degree model showed good applicability in this field. In addition, it could provide a new method for the internal connection between farmland, surface water, and ecology in the Aral Sea Region and even the Aral Sea Basin.

Effect of abandoned farmland on surface water in aral sea region

The abandoned farmland will cause a lot of surface water loss. The main reason is that irrigation canals of abandoned farmland are not blocked because of the negligence of management. It directly led to majority of the water was being soaked up by the desert and blatantly loss (Ibragimov et al., 2007; Maksud et al., 2016; Jabbarov et al., 2013). According to the results of section 3.2, we extracted the area of abandoned farmland (Table S6). The area of abandoned farmland reached per year 97.843 km². And the abandoned farmland will cause a lot of water waste. Farmland in Aral Sea Region is concentrated on planting cotton, wheat and rice (Chapagain, Gautam & Hoekstra, 2005). Table S10 shows that wheat is the thirstiest crop. And the water use efficiency of cotton is the lowest (0.451 * 108 kg/km³). Table S11 shows the crop production in precious decades.

Based on the above results, the loss volume of water loss by abandoned farmland can be extracted. If all the abandoned farmland were used to plant cotton in the past few years, the water loss would reach 0.028 km³ every year (Table S12). According to the current distribution of available water per capita from the World Bank (Uzbekistan, 2,596 m³/per person/annually; Kazakhstan, 1,943 m³/per person/annually), the water loss can provide annual domestic water for tens of thousands of people in the Aral Sea Region. The specific numbers are as follows: Kazakhstan are 14,410 persons, and Uzbekistan are 10,785 persons. Therefore, abandoned farmland caused additional loss of normal domestic water for people in the Aral Sea Region.

The previous researches gave us much inspiration. Micklin macroscopically analyzed the reason and effect of water changes in the Aral Sea (Micklin, 2010). Shibuo, Jarsjö & Destouni (2007) found nearly 100% return flows of the applied irrigation water in the Aral Sea Drainage Basin. Cretaux, Letolle & Bergé-Nguyen (2013) analyzed effect of groundwater on the hydrological water budget of the Aral Sea by remote sensing data. Compared with above researches, our contributions are as follows:

1. Based on the coupling degree model, we quantitatively characterized the effect of farmland changes on surface water.

2. The coupling degree model was used for the first time to study the effect of farmland changes on surface water. We believe that this attempt will help broaden the thinking of researchers in the fields of remote sensing, environment and more. They can connect independent systems through the coupling degree model.

3. The effect mechanism was analyzed on surface water changes by calculating the amount of water loss in abandoned farmland. Moreover, the normal domestic water for people caused by water loss was analyzed.

Limitations of research

At first, vegetation indices can reduce other factors’ impacts in monitoring land cover change (Zhang et al., 2013). The NDVI can reduce the influence of external environmental factors on the surface reflectance to some extent. Previous studies indicated that the accuracy of NDVI in detecting farmland is better than others (Tong et al., 2017). However, due to the disturbance by other vegetation in the arid region, using NDVI to the LandTrendr algorithm has some limitations (Watts & Laffan, 2014). A typical example is that the sparse vegetation of the Aral Sea Region generated the same NDVI signal with farmland and affected the detection results. Therefore, it is necessary to add additional indices to improve accuracy.

Secondly, the data on water level and water storage cannot match the water area’s time series. However, this fusion of data yet offers valuable and unique information on the surface water. Simultaneously, in the next step of our research, we need to improve data integrity of the surface water.