Analysis and Prediction of Land Use/Land Cover Changes in Korgalzhyn District, Kazakhstan

Abstract

Changes occurring because of human activity in protected natural places require constant monitoring of land use (LU) structures. Therefore, Korgalzhyn District, which occupies part of the Korgalzhyn State Natural Reserve territory, is of considerable interest. The aim of these studies was to analyze changes in the composition of the land use/land cover (LULC) of Korgalzhyn District from 2010 to 2021 and predict LU transformation by 2030 and 2050. Landsat image classification was performed using Random Forest on the Google Earth Engine. The combined CA-ANN model was used to predict LULC changes by 2030 and 2050, and studies were carried out using the MOLUSCE plugin. The results of these studies showed that from 2010 to 2021, there was a steady increase in the share of ploughable land and an adequate reduction in grassland. It is established that, in 2030 and 2050, this trend will continue. At the same time, there will be no drastic changes in the composition of other land classes. The obtained results can be helpful for the development of land management plans and development policies for the Korgalzhyn District.

1. Introduction

Two strands dominate discussions about the leading causes of environmental crises [1]. The first is a concept that strictly separates culture from nature, which allows modern people to put themselves above nature and the animal world [2,3,4,5,6]. The second is the assertion that the driving forces of ecological catastrophe lie mainly in economic structures that contribute to the constant increase in the exploitation of the Earth’s ecological systems [7,8,9,10,11], hindering sustainable development. In particular, the dominant form of land management worldwide has become used for agricultural needs [12]. According to the FAO, this process tends to continue [13]. The main driving force behind this phenomenon is the constant population increase and the growing demand for food. For example, world food demand is projected to grow to 70% by 2050. The consequence of this process is the expansion of arable land, which is especially typical for developing countries with economies in transition [12]. For their part, researchers use a variety of modeling and optimization techniques to identify sound agricultural land management strategies to achieve sustainable development goals. Nevertheless, due to the susceptibility of the context and purpose to transformation, optimizing LU in agriculture is becoming an increasingly complex problem [14]. Therefore, to date, no generally accepted method has changed Agent-Based Models (ABMs) [15,16]. Researchers widely use this possibility of modeling and forecasting to assess the consequences of various natural phenomena and anthropogenic activities [17,18,19]. LULC monitoring and forecasting are necessary for developing a sustainable development scenario [20] and assessing the LU of an entire country [21], a region [22], and the whole planet [23]. As a result, modeling the future state of land structures has become one of the main directions for studying the processes of LU change in all its aspects, including agricultural land.

Various methods are used to model the future states of features, each unique to solving complex LULC problems. For example, in [24], these methods are divided into six types: Statistical Models, Cellular Automata (CA) Models, Economic Models, Agent-Based Models (ABMs), Hybrid Models, and Time Series Modeling for LULC Change. In addition, Available Modeling Software Packages for LULC Predictions are provided, including CLUE-S, DYNAMICA EGO 7.7.0, Land Change Modeler (LCM) in TerrSet 2020, and SLEUTH 3. Studies [25] also note the applicability of several models of the types mentioned above for predicting the future state of land use. Despite this, to generate future forecasts based on the detected transition patterns as part of LULC, researchers often prefer the use of Artificial Neural Networks in combination with Cellular Automata models (CA-ANN) [26,27,28] as one of the most promising approaches to deep learning. The CA paradigm arose due to the search for perfect tools to simulate complex natural phenomena [29]. The start of a deep curiosity in the concept of cellular automata was the statement that “everything is connected with everything else, but close objects are more connected than distant objects” [30] and the vision of space as a grid of cells [31]. With the development of CA’s ability to create dynamic spatial patterns across different time and space scales, it has become a powerful tool for understanding LU systems and their inherent dynamics, especially when combined with other devices such as ANNs. NNs are one of the fundamental elements of the theory of artificial intelligence, and ANNs are considered an “intelligent tool” capable of learning. Because CA-ANN is based on “what-if” scenarios, it is suitable for land change modeling studies [32]. It is believed that CA-ANN accurately represents the complex spatial inhomogeneities of LULC [33]; therefore, it is successfully used to evaluate future LULC transformations. At the same time, the Modules for Land-Use Change Simulation (MOLUSCE) plugin, as part of QGIS, is often used to model future LU generations based on CA-ANN [34,35,36,37,38,39].

Changes resulting from intensive agricultural activities require constant monitoring of LU development and forecasting of future changes [40]. The urgent need to monitor and predict future generations of LU in the Korgalzhyn District in Akmola Oblast (Kazakhstan) is associated with the presence of the Korgalzhyn State Nature Reserve (KSNR) in its territory. KSNR is located in the central migratory flyways of the World Tengiz-Korgalzhyn Lakes System, home to rich flora and fauna [41]. Nature reserves are one of the most essential tools for protecting biodiversity and limiting species extinctions in the region. However, these functions can only be performed if environmental impacts, in this case, economic activities aimed at the uncontrolled growth of arable land, do not hurt protected habitats [42]. For example, in many developing countries with economies in transition, like Kazakhstan, the growing requirement for food has led to expanding arable land [12]. In the available studies [43,44,45,46], only the state and territory of KSNR are assessed. They do not cover the agro-industrial activities of Korgalzhyn District, which is intensively engaged in rainfed agriculture. Therefore, it is not surprising that we did not find studies devoted to studying the spatiotemporal dynamics of LULC and assessing the future LU scenario. Because understanding the patterns of LULC change locally is vital to developing land management strategies, the purposes of the present study are as follows:

(1)

To analyze changes in the spatiotemporal trends of LULC in Korgalzhyn District from 2010 to 2021 by applying Landsat images and RF on the Google Earth Engine (GEE) platform;

(2)

To forecast the spatiotemporal changes in LULC in Korgalzhyn District, which may occur by 2030 and 2050, by applying the integral CA-ANN method with the MOLUSCE plugin;

(3)

To assess the impact of LULC transformation on future LU planning.

2. Materials and Methods

In this section, we discuss the sources of satellite data and the stages of their pre-processing, as well as the methods of LULC classification used in the study area. Additionally, we present the methods for assessing classification accuracy and analyzing changes in land cover that we employ to ensure the reliability of the results.

2.1. Study Area

Korgalzhyn District (50°34′55″ N; 70°00′49″ E.) is located in the south of the Akmola Oblast, which belongs to northern Kazakhstan and borders on Karaganda Oblast (Figure 1). The Area of Interest (AOI) is 9300 km². Two ancient, vast hollows represent the relief of the site, with absolute heights of 305–380 m. The hollows are oriented from the southwest to the northeast and are separated by a gently sloping watershed ridge up to 435 m in total height [47]. The Nura and Kulanotpes rivers flow through the region. The second river, as a rule, dries up in the summer. The climate is continental. Winter is cold, stormy, and long. Snow lies for about 180 days, but the thickness of the snow cover does not exceed 20 cm. The average January temperatures range from −18 to −17 °C. Summer is moderately hot and dry, lasting 50 days. The average July temperature is 20–21 °C. The average annual rainfall is 250–300 mm. In addition to zonal soils, the territory is characterized by soils of the semihydromorphic and hydromorphic series: meadow-chestnut, meadow, meadow-marsh, bog, solonetzes, and solonchaks [43]. Korgalzhyn District mainly specializes in grain production. With a predominance of spring wheat, the entire infrastructure is available. In addition, oilseeds, potatoes, and vegetables are cultivated in the region, and livestock breeding and processing of agricultural products are an integral part of the region’s activities [48]. The AOI has eight administrative divisions (rural districts) [49] and occupies the KSNR position, as shown in Figure 1.

2.2. Collecting Satellite Data and Image Pre-Processing

In this study, USGS Landsat 7 and 8 Collection 2 Tier 1 TOA Reflectance, VIIRS Stray Light Corrected Nighttime, DMSP OLS: Nighttime Lights Time Series Version 4, and SRTM3 images were used as the initial data, which served to obtain long-term LULC series of Korgalzhyn District. Variations in the Urban Land (UL) class were determined using the DMSP OLS: Nighttime Lights Time Series Version 4 and VIIRS Stray Light Corrected Nighttime data [50]. Some characteristics of the datasets used are shown in

Table 1. Features and particularities of satellites and data sets for the classification of Korgalzhyn District.

Data Sets	Date	Bands	Resolution
Landsat 7 Collection 2 Tier 1 TOA Reflectance	2010–2017, 2021	B2–B5	30 m
Landsat 8 Collection 2 Tier 1 TOA Reflectance	2018–2020	B2–B5	30 m
VIIRS Stray Light Corrected Nighttime	2014–2021	avg_rad	463.83 m
DMSP OLS: Nighttime Lights Time Series Version 4	2010–2013	avg_vis	927.67 m
SRTM 3	2010–2021	elevation	90 m

. Detailed metadata for individual images used in the analysis, including their acquisition dates, are presented in Appendix A.

All analytic actions with RS images were executed on the GEE platform [50]. The EVI, NDVI, NDWI [50], NDBI [51], NDTI [52], NDMI [53], and MNDWI [54] indices, and SRTM3 [45] were used to enhance the classification efficiency, along with the original RS data.

The GEE platform also involved preprocessing of standard data collections, including 2 Tier 1 TOA Reflectance data for Landsat 7 and 8, VIIRS Stray Light Corrected Nighttime data, DMSP OLS: Nighttime Lights Time Series Version 4 data, and SRTM3 data. Preprocessing also involved the selection of appropriate images (only those with less than 20% cloud coverage were used in the analysis) and cropping the images to the area of interest.

The VIIRS Stray Light Corrected Nighttime Lights Time Series Version has been incorporated into the classification algorithm to improve the accuracy of identifying areas with low building density and high vegetation levels.

SRTM3 data were utilized to determine the terrain, as it significantly affects the reflection properties, which are critically important for remote sensing classification methods [55].

Based on selected images, median mosaics were created, which were subsequently subjected to LULC classification.

These composite median images enabled data averaging, resulting in more stable and accurate outcomes. This approach allowed us to address and correct issues related to the changes in natural and atmospheric conditions, along with challenges associated with the Landsat 7 sensor’s SLC-off condition [56,57,58,59]. The decision to employ this method was mainly based on its universality for correcting radiometric issues from various sources. The study structure is shown in Figure 2.

2.3. Satellite Image Classification

To assess LULC changes from 2010 to 2021, satellite data were employed. Classification for each year was performed using the RF method on the GEE platform.

The Random Forest model was trained on a training dataset that included spectral characteristics for each land use and land cover class [60]. The dataset for each year was selected based on visual interpretation of vegetation indices and higher-quality satellite images from Google Earth Pro [61]. A training dataset for LULC classification was compiled based on pixel values within polygons that defined areas of interest for each class. Details of the training dataset are provided in Appendix B and

Table A1. The number of pixels in the training dataset over the years.

Year	Cropland	Pasture	Forest	Water Bodies	Urban Land
2010	1,085,821	1,283,513	62,855	822,395	81,026
2011	1,123,190	1,277,849	70,649	831,939	81,026
2012	1,014,292	1,015,332	70,649	831,939	81,026
2013	1,022,968	983,676	50,843	831,939	81,026
2014	990,490	792,385	28,071	851,719	81,026
2015	1,030,806	561,108	28,071	911,412	81,026
2016	1,111,232	610,795	30,760	914,232	81,026
2017	1,206,375	672,112	30,760	962,772	92,219
2018	1,204,947	795,583	53,716	988,529	92,219
2019	1,219,286	914,479	56,499	1,061,462	92,219
2020	1,213,400	1,094,677	56,499	1,061,462	92,219
2021	1,206,375	672,112	96,505	962,772	94,731

.

2.4. Classification Accuracy Assessment

In the context of this research, a test dataset consisting of 500 data points was employed to assess classification accuracy. Considering the spatial distribution characteristics of the LULC classes in the study area and the small area of some of them (Urban Land), the locations of these points were randomly selected, while accounting for stratification by classes to ensure sample representativeness.

Detailed location designations of control points used for each category are presented in Appendix C. The study focuses on five primary classes that are characteristic of the study area. These classes are described in detail in

Table 2. The classes of LULC in the AOI.

Code	Level One	Level Two
CL	Cropland	Regular cultivated lands, different fallows
PE	Pasture	Grasslands and hays
FT	Forest	Trees and shrubs
WB	Water Bodies	Lakes and rivers ponds and various reservoirs
UL	Urban Land	Impenetrable surfaces

.

To estimate the classification exactness using the RF method, a confusion matrix was also created, and various metrics were used, such as overall accuracy (OA), user’s accuracy (UA), and producer’s accuracy (PA) [62]. Due to the ambiguity of the Kappa coefficient [63], the following were used instead: F1-score [64], Quantity Disagreement (QD), and Allocation Disagreement (AD) [65]. QD is an error related to the misrepresentation of pixels in different categories. It evaluates the difference in the number of pixels assigned to other classes compared with the actual values. AD is an error related to the incorrect placement (arrangement) of pixels in various categories. It measures the proportion of pixels incorrectly assigned to classes that do not match their locations.

2.5. Change Detection

LULC changes were analyzed in a continuous time series (“time-series stability” principle). Using a constant time series to analyze changes in LULC has several advantages over analyzing individual periods. A continuous time series allows one to capture the dynamics of LULC changes over an entire period, which helps to identify key trends that are especially important when generating long-term forecasts [66]. A Sankey chart was built using the ChartExpo plugin in Microsoft Excel LTSC MSO (16.0.14332) [67].

2.6. Prediction of Future LULC

To calculate the forecast models, the period from 2010 to 2016 was chosen since it has a high linear relationship throughout the entire time interval under consideration, as confirmed by Pearson’s correlation value [68], which was 0.94. To create a forecast model and calculate LULC changes for 2030 and 2050, the transition matrix for 2010–2016 was first created, and then the combined CA-ANN method was used to model future changes. Research on creating a forecast model and calculating LULC changes for 2030 and 2050 was conducted using the MOLUSCE plugin of QGIS software version 2.18.10. This plugin is open source for QGIS 2.0 and above [25] and also combines well-known methods of tooling for land-use change simulation [39]. Therefore, MOLUSCE is available to all interested researchers with two sets of data [36]. The first data set is LULC classes, usually obtained by classifying raster images using different algorithms over a long period or in time intervals. The second set of data is drivers of LULC change, or explanatory variables. The choice of the latter depends on the spatial structure of the studied territories as well as the goals and objectives of the study. Geographical variables that help pinpoint where changes might occur in the future typically include DEM, slope, and distances from roads, rivers, and settlements. But they may not be limited. Therefore, the integrated CA-ANN model using MOLUSCE is one of the trustworthy methods for modeling the forecast of complex spatial systems such as LULC, which provides the basis for its application in our study.

To calculate the ANN, we took into account several predictor variables (Figure 3) that play a role in changing LULC; they included maps of buffer zones of roads (Figure 3a), settlements (Figure 3b), and rivers (Figure 3c) using the Euclidean distance method for the study area3dslope map (Figure 3e) and aspect ratio (Figure 3f). To create the DEM, the SRTM3 dataset was used, and from it, the slope map and aspect ratio were calculated. Data for buffer zones were obtained from vector data in the GIS of the Korgalzhyn district. All maps were created in ArcGIS Pro 3.1.2.

The ANN was calculated based on the Transition Potential Modeling module, after which the Cellular Automata Simulation module was used [69]. The values of priorities 4a–4d were assessed using the Cramer coefficient [70] in the MOLUSCE plugin. To consider spatial interactions between adjacent cells, we chose a 3 × 3 pixel neighborhood. During the training phase of our model, we ran 1000 iterations with a learning rate of 0.1. In each iteration, the neural network weights and biases were adjusted to reduce the error between the predicted and observed LULC categories.

Table 3. Cramer’s V value of spatial variables.

Spatial Variables	Cramer’s V
distances from roads	0.24
distances from settlements	0.23
distances from rivers	0.12
digital elevation model	0.34
slope	0.02
aspect ratio	0.05

shows the potential values of the Cramer’s V coefficient for each spatial variable. The Cramer’s V coefficient value indicates that the variables are suitable for modeling the potential for transition, as their values are significant. According to these values, the selection of physical variables, namely: DEM (elevation model): 0.34; slope: 0.02; distance to roads: 0.24; distance to settlements: 0.23.

To assess the reliability of the forecasting model, a validation process was carried out using the MOLUSCE plugin for actual 2021 LULC data, followed by a comparison with the results of 2021 LULC simulations performed based on the CA-ANN. The same approach was used to evaluate the LULC maps for 2030 and 2050. The predictions were validated using four Kappa (K) statistical metrics: Koverall, Khisto, Kloc, and Percentage of Correctness [35].

The validation process was implemented using the Validation module in the MOLUSCE plugin. Kappa (Koverall) evaluates the overall agreement between the classification and prediction, given a random arrangement, and Kappa Location (Kloc) evaluates the ability of the model to locate objects accurately. The Khisto value for each class indicates the level of understanding for that class between the observed and predicted rasters.

3. Results

The results of this study provide a comprehensive overview of land use and land cover changes (LULCC) in the Korgalzhyn District over a significant time period, from 2010 to 2021. The analysis encompasses various aspects of LULCC dynamics and employs advanced cartographic and statistical tools to ensure accuracy and reliability.

3.1. Assessment of the Accuracy of Cartography Products for the Korgalzhyn District

Table 4. Change in LULC classes from 2010 to 2021 in Korgalzhyn District (%) and their accuracy.

Year	Changes					Accuracy
	CL	PE	FT	WB	UL	OA	PA	UA	F1-Score	AD	QD
2010	6.85	56.96	9.31	26.67	0.21	0.91	0.92	0.91	0.91	7.20	1.40
2011	8.83	54.54	9.50	26.91	0.22	0.90	0.90	0.90	0.90	8.40	1.60
2012	10.93	53.20	9.66	25.99	0.22	0.91	0.91	0.91	0.91	6.80	2.40
2013	12.87	50.58	9.77	26.55	0.23	0.89	0.89	0.89	0.89	8.60	2.00
2014	14.95	48.53	9.61	26.68	0.23	0.91	0.91	0.91	0.91	6.20	2.40
2015	16.72	47.23	9.91	25.90	0.24	0.89	0.89	0.89	0.89	9.40	1.20
2016	17.21	46.84	10.26	25.45	0.24	0.90	0.90	0.90	0.90	8.60	1.80
2017	19.80	44.08	10.13	25.75	0.24	0.90	0.90	0.90	0.90	8.40	1.60
2018	23.22	40.59	9.81	26.13	0.25	0.88	0.88	0.88	0.88	10.20	1.60
2019	26.23	36.40	9.67	27.45	0.25	0.91	0.91	0.91	0.91	7.40	1.80
2020	29.51	34.31	9.19	26.74	0.25	0.90	0.90	0.90	0.90	7.40	2.40
2021	30.18	33.00	9.56	27.01	0.25	0.92	0.92	0.92	0.92	6.60	1.60
Average						0.90 ± 0.04	0.90 ± 0.04	0.90 ± 0.04	0.90 ± 0.04	7.93 ± 4.00	1.82 ± 1.2

shows the land allocation of different LULC classes from 2010 to 2021 in Korgalzhyn District and their accuracy. For this set of data, the following results were obtained: OA 0.90 ± 0.04; UA 0.90 ± 0.04; PA 0.90 ± 0.04; and F1-score = 0.90 ± 0.04. Due to the high overall accuracy, the values of AD, which are 7.93 ± 4.00, and QD, which are 1.82 ± 1.2.

For the individual LULC classes (

Table 5. Accuracy assessment of different types of LULC Korgalzhyn District from 2010 to 2021.

LULC Classes	UA	PA
CL	0.90 ± 0.10	0.90 ± 0.07
PE	0.90 ± 0.09	0.88 ± 0.07
FT	0.89 ± 0.10	0.89 ± 0.07
WB	0.91 ± 0.11	0.90 ± 0.08
UL	0.91 ± 0.09	0.94 ± 0.08

), from 2010 to 2021, the user accuracy and producer accuracy were relatively high, respectively, amounting to 0.90 ± 0.10; 0.90 ± 0.07 for CL, 0.90 ± 0.09 and 0.88 ± 0.07 for PE, 0.89 ± 0.10 and 0.89 ± 0.07 for FT, 0.91 ± 0.11 and 0.90 ± 0.08 for WB and 0.91 ± 0.09 and 0.94 ± 0.08 for UL.

The results (Table 4 and Table 5) show that the classification exactness of CL, PE, FT, WB, and UL was high, which may be due to the introduction of additional data such as NTL, NDVI, EVI, DEM and MNDWI in this study. Meanwhile, as evident from

Table 6. Confusion matrices for 2010 and 2021.

LULC Classes	CL	PE	FT	WB	UL	Total
2010
CL	18.20%	1.20%	0.60%	0.00%	0.00%	20.00%
PE	1.00%	18.20%	0.40%	0.00%	0.40%	20.00%
FT	0.40%	0.40%	17.80%	1.40%	0.00%	20.00%
WB	0.00%	0.40%	1.20%	18.40%	0.00%	20.00%
UL	0.00%	1.20%	0.00%	0.00%	18.80%	20.00%
Total	19.60%	21.40%	20.00%	19.80%	19.20%	100.00%
2021
CL	18.60%	1.20%	0.20%	0.00%	0.00%	20.00%
PE	0.60%	18.80%	0.20%	0.00%	0.40%	20.00%
FT	0.60%	1.00%	17.20%	1.00%	0.20%	20.00%
WB	0.40%	0.00%	1.20%	18.40%	0.00%	20.00%
UL	0.00%	0.40%	0.20%	0.60%	18.80%	20.00%
Total	20.20%	21.40%	19.00%	20.00%	19.40%	100.00%

pertaining LULC classification in 2010, there remained a certain probability of classification accuracy reduction, as 1.00% of the CL class was prone to be misclassified as PE. A similar trend was observed for FT, with 0.40% being included in the PE category. Additionally, 1.20% of PE was identified as UL and CL.

In 2021, the probability of transitioning to other LULC classes remained practically true for all classes except WB. For instance, 1.20% of PE could be erroneously classified as CL.

Thus, the accuracy of land cover classification in the Korgalzhyn region was relatively high, with a low probability of individual classes being misclassified as other land categories.

3.2. LULC Change Analysis of Korgalzhyn District

An analysis of the dynamics in different classes of LULC was conducted based on the data in Table 4.

3.2.1. Change in Cropland

The area of arable land increased rapidly between 2010 and 2021 (Figure 4).

Changes occurred in two relatively large phases between 2010 and 2015, and 2017 (2.51 times) and 2021 (1.52 times). However, the process of growing acreage during the observation period continued. There was only a relatively small increase in arable land in 2015 and 2016 (1.03%), as well as in the transition from 2020 to 2021(0.67%). In general, the area of cultivated land for growing crops, among which spring wheat predominates, increased by 4.4 times during the observation period and amounted to 30.18% of all LULC classes in 2021.

3.2.2. Change in Pasture

Land changes that fall under the Pasture category are practically a mirror image of Cropland’s dynamics (Figure 4). However, unlike Cropland, a relatively small increase in the development of pastures by other classes of land (mainly arable land) was noted only in 2015 and 2016 (0.39%). In other years, there was an almost permanent loss of land (from 1.31 to 4.19%) allotted for pastures and hayfields. As a result, the territory occupied by the Pasture class decreased by ~1.7 times over the years of observation, and by 2021, it amounted to only 33% of the land structure of Korgalzhyn District.

3.2.3. Change in Forest

Forest-class lands occupy about ten percent of the territory of Korgalzhyn District. They are also subject to annual changes in their area (from 9.31 to 10.26%). Their dynamics during 2010–2021 undulated (Figure 4). The first surge occurred between 2010 and 2014, with a peak in 2013 when class area growth over 2010 was 0.46%. The second wave occurred between 2015 and 2020, with the maximum increase in class area in 2016, where the increase compared to 2010 was 0.95%. In 2021, there was also a notable increase in the area occupied by trees and shrubs compared to 2020 (0.37%). But it is still too early to talk about the third wave. It is interesting to note that the nature of the change in the areas of the Forest class is negatively dependent on the Water Bodies land class, and the linear correlation is −0.72.

3.2.4. Change in Water Bodies

Lands of the Water Bodies class occupy approximately one-quarter of the territory of Korgalzhyn District. Their changes are also of a wavy type, with three pronounced rises in 2011, 2014, and 2019 (Figure 4). The first relatively weak increase in the area of water bodies was noted at the beginning of the observation, during the transition from 2010 to 2011 (only 0.24%), which ended with a drop in the water surface level in 2012 to −0.92%. The second wave began in 2012, which led to an increase in the water surface in 2013 by 0.56% and in 2014 by 0.69%. After which, in 2015 (0.78%) and especially in 2016 (1.23%), there was again a decrease in the size of the water class. Then, over four years (from 2016 to 2019), the surface of water bodies increased by 1.7%. However, in 2020, it again decreased slightly by 0.71%, followed by an upward trend in 2021 (0.27%), which may be the beginning of the fourth wave. Changes in areas of the Water Bodies class are also negatively related to the Forest land class. In the years of increasing areas occupied by water, there was a decrease in the space occupied by Trees and Shrubs.

3.2.5. Change in Urban Land

A magnification characterizes Korgalzhyn District in the space of built-up areas. However, this occurs relatively slowly, varying only from 0.21 to 0.25% of the total AOI (Figure 4). That is, the total change is only 0.04% of the total area of the district. At the same time, from 2010 to 2018, there were four stages of increase in the size of the areas occupied by urban lands. If we take 2010 as a basis, then in 2011 and 2012 the built-up area increased by 4.46%, in 2013 and 2014—by 9.52%, in 2015–2017—by 14.29%, and from 2018 to 2021—by 19.05%. However, this process has stabilized in the last four years and remains at the same level.

Thus, the consequences of the change analysis show that from 2010 to 2021, Korgalzhyn District mainly experienced a steady magnification in the space of crops (from 6.85% to 30.18%) primarily due to a decrease in grassland occupied (from 56.96% to 33.00%). The Sankey diagram of LULC confirms these changes for 2010, 2016, and 2021 (Figure 5).

According to the chart, the predominant class for all three years was PE, which occupied 56.96% to 33.00% of the area of the district from 2010 to 2021, followed in order by WB, CL, FT, and UL, ranging from 26.27% to 27.01%, from 6.85% to 30.18%, from 9.31% to 9.56%, and from 0.21% to 0.25%, respectively. CL has been increasing its share in the area for three years, displacing PE; WB has a slight tendency to move into FT and PE and back, which is typical for a particular wetland area. At the same time, the share of CL in 2021 was almost equal to that of PE (30.18% for CL and 33.00% for PE), in comparison to the apparent predominance of PE in 2010 (6.85% for CL and 56.96% for PE) and 2016 (17.21% for CL and 46.84% for PE).

3.3. LULC Prediction Based on CA-ANN for Korgalzhyn District

To evaluate changes in LULC classes in the future with CA-ANN, we used the data obtained based on the LULC classification of Korgalzhyn District for 2010 and 2016. First, a transition probability matrix was created using the MOLUSCE plugin (

Table 7. Matrix of transition probabilities obtained from LULC maps of Korgalzhyn District for 2010–2016 (%).

LULC Classes	CL	PE	FT	WB	UL	Total	Loss
CL	73.72	26.02	0.21	0.04	0.00	100.00	26.28
PE	21.06	76.72	1.06	1.12	0.04	100.00	23.28
FT	0.03	1.44	97.66	0.85	0.02	100.00	2.34
WB	0.40	5.96	2.58	91.04	0.02	100.00	8.96
UL	0.00	0.18	1.00	4.84	93.98	100.00	6.02
Total	95.22	110.32	102.51	97.89	94.06
Gain	21.49	33.60	4.85	6.85	0.08

). In Table 7, the diagonal data indicate the probability that the territory of individual classes will remain unchanged, which turned out to be relatively high and, depending on the LULC class, varied from 73.72 to 97.66. The information in the rows indicates the possibility of loss in favor of another type, and the data presented in the columns indicate the probability of gain by a separate category from other LULC classes. As seen in Table 7, the biggest gain and loss probabilities exist for CL and PE, which exceed 20%.

When training the neural network to consider the spatial interactions between adjacent cells, we took a neighborhood of 3 × 3 pixels in size. We carried out 1000 iterations with a learning coefficient of 0.1. In each iteration, we adjusted the weights and biases of the neural network to reduce the error between the forecast and observed LULC categories. The neural network learning curve is shown in Figure 6.

In the next step, based on CA-ANN, a forecast of the change in LULC classes for 2021, 2030, and 2050 was generated, as shown in

Table 8. Comparison of actual LULC class areas 2021 with forecast data obtained from CA-ANN (%) for 2021, 2030, and 2050.

LULC Classes	Actual 2021	Predicted by CA-ANN and Deviation from Actual Data (%)
		2021		2030		2050
CL	30.18	28.73	−1.45	41.13	+10.95	48.76	+18.58
PE	33.00	37.54	+4.54	23.10	−9.84	16.82	−16.18
FT	9.56	12.02	+2.46	11.40	+1.87	8.43	−1.13
WB	27.01	21.50	−5.51	24.12	−2.89	25.68	−1.33
UL	0.25	0.21	−0.04	0.25	0	0.31	+0.06
Total	100	100	0.00	100	0.00	100	0.00

and Figure 7, Figure 8 and Figure 9.

First of all, the nature of the change in the actual and forecast values for CL and PE should be noted. If the areas of CL and PE in 2021 were approximately equal, then by 2030, the regions occupied by agricultural plants will increase to 43.16%, and the size of land allocated for pastures will decrease by 69.71%. This trend will continue in 2050, when the share of arable land will be nearly 50% of the AOI territory and the percentage of pastures and hayfields will be less than 17%. The area occupied by FT, WB, and UL will be relatively little susceptible to changes.

Finally, the Korgalzhyn District LULC change prediction model was tested using CA-ANN, the results of which are presented in

Table 9. Model validation using K indices for 2021, 2030 and 2050.

K Indexes	2021	2030	2050
Koverall	0.88	0.81	0.84
Khisto	0.88	0.81	0.84
Kloc	0.92	0.93	0.92
% of Correctness	87.24	85.17	88.22

.

As in the previous steps, the forecast validation using the studied methods is based on comparing the actual LULC map for 2021 with forecast maps for 2021, 2030, and 2050. The values of the K indices reflect the forecast estimate. For 2021, the values of Koverall, Khisto, and Kloc were relatively high (>0.85). The same was true for Correctness (87.24%). The importance of Koverall, Khisto, and Kloc obtained for 2030 and 2050 varied between 0.81 and 0.93, and that of the Percentage of Correctness varied between 88.22 and 85.17. This shows that the forecast models derived from CA-ANN generally satisfactorily describe the future state of LULC in Korgalzhyn District.

4. Discussion

Land use changes carried out by people on land worldwide play an essential role in food security, ecosystem services (ESs), biodiversity conservation, and climate change management. Thus, combining RS data with statistical information from 1960 to 2019 showed that global changes in LU affect about 32% of the land area. This means they are approximately four times larger than previously thought [71]. However, changes in LU occur differently in different regions of the world [72]. For example, a detailed analysis of ESs in Europe showed [73] that regional decisions can lead to entirely different positive and adverse events. The creation of protected areas led to increased regulatory and cultural ESs, and the reduction in pressure on agriculture led to comparable, although attenuated, changes. In contrast, uncontrolled and massive urban sprawl and the conversion of pastures to arable crops have led to a decline in many ES indicators. In general, changes in LU depend on many factors [24,72,74,75], one of which is agriculture. Agricultural activities provide jobs, food, and necessities for much of the world’s population [76]. Increasing demand for food also increases the need for agricultural land [77]. At the same time, there are also negative consequences of agriculture that create severe problems for the environment [78]. For example, cropland expansion and intensification are significant strategies for increasing agricultural production but are major drivers of decreases in biodiversity and ESs [79]. The main reason for the decrease in biodiversity and ESs during the intensive use of agricultural land for crop production is, first of all, the high level of use of agrochemicals for their cultivation and many other equally important factors [80].

Our study aimed to identify the dynamics of LU over the past 12 years and model the forecast of its changes by 2030 and 2050 in Korgalzhyn District. The relevance of this goal is that Korgalzhyn District covers the territory of a biosphere reserve, where active production activities can seriously impact the conservation of natural biodiversity and ESs. Typically, an assessment of changes in LULC is performed based on two or more decades [81,82,83]. We adopted the relatively short study period between 2010 and 2021 because 2010 marks the beginning of the region’s economic recovery after the financial crisis of 2007–2009 [84], which had a noticeable impact on the growth of Kazakhstan’s economic development [85]. The years 2030 and 2050 are frequently used periods for LULC forecasting [86,87,88] since achieving the SDGs [89] and mitigating climate change [90], as well as accomplishing important strategic goals of the Republic, are associated with them in Kazakhstan [91,92].

Classification of space images is a rather complex process, the methodology of which is constantly being improved [93,94,95], and the division into the number of classes depends on the predictors that prevail in the space under study [96,97]. Considering the results of these studies, we used one of the most reliable and, therefore, widely used algorithms—Random Forest—to classify Landsat satellite images. To achieve our goals, we considered five LULC classes characteristic of Korgalzhyn District: first of all, arable land and pastures, ensuring the efficiency of the economic activity of the district; water and forest resources of the Tengiz-Korgalzhyn Lake System, which play a decisive role in the conservation of biodiversity and changes in ESs; and built-up areas where human resources and industrial infrastructure are located, which are some of the driving forces for the development of the region. At the same time, we did not provide a detailed classification of individual crops grown in the AOI that could be used to estimate and predict the yield of particular crops. For example, Kazakh researchers [98] used machine learning (ML) and neural network methods based on RS data to predict crop yields. Other researchers [99] have used comparative correlation analysis between yield data and regional NDVI temporal profiles based on a linear regression (LR) model to estimate the yield of individual crops. Deep learning (DL) can also accurately represent complex features needed for crop mapping and yield forecasting by accounting for nonlinear relationships between variables [100]. Our work’s lack of mapping and crop forecasting limits our research. However, this drawback will be eliminated in our further study using the advantages of the above-mentioned methodological approaches (LR, ML, and DL) based on the use of statistical data and RS over a long period of time.

The results of monitoring changes in LU convincingly showed that from 2010 to 2021 in Korgalzhyn District, there was a process of intensive expansion of arable land, dominated by spring grain crops, for which these lands were once ploughed during the development of virgin lands [101]. For example, in 2023, in Korgalzhyn District, grain crops are expected to be grown on 151 thousand hectares, oilseeds on 6.9 thousand square meters, potatoes on 45.4 hectares, and vegetables on 31.8 hectares [48]. As we can see, according to this plan, more than 95% of the AOI’s arable land will be sown with grain crops, which will practically lead to monoculture with all the ensuing consequences [102]. At the same time, the increase in crop area has been accompanied by an almost equivalent decrease in land occupied by pastures, which play a fundamental role in mitigating the effects of climate change, especially in carbon storage and sequestration, as well as in the conservation of biodiversity [103]. Changes in the areas of land occupied by water and forest resources did not have a unidirectional trend, and their fluctuations from year to year were wave-like within a particular corridor. In Korgalzhyn District, water bodies occupy 25% to 28% of the total territory. In turn, most of these objects belong to the shallow fresh—salt Tengiz-Korgalzhyn Lake System, the average depth of which is 2.5 meters. The lake basin also includes a network of small lakes east of the main lake. They are susceptible to annual precipitation, and precipitation during the growing season, and temperature conditions [104,105]. This part of the AOI is represented by wetlands that experience an explosive peak in vegetation growth in mid-summer [43]. As a result of the coupling of the drying out of shallow lakes with the overgrowth of marsh vegetation, the dynamics of water bodies negatively correlate with the dynamics of the Forest class (−0.72), which includes wetland vegetation. The built-up area of the region as a whole grew, but in small quantities. Perhaps this is due to the relatively low attractiveness of the territory of Korgalzhyn District for the mass relocation of people since the AOI is located far from large cities on the edge of Akmola Oblast and borders the distant regional center of the Karaganda region. For example, the population of Korgalzhyn District by 1 February 2022, was only 8415 people [106] and showed a downward trend (97.3% in 2021). It is noteworthy that our LULC AOI classifications are highly accurate. The main reason for the increase in all types of assessments of the accuracy of the LULC classification of the Korgalzhyn region (UA, PA, total, F1-score, QD, and AD) was the approach we used with the additional integrated use of EVI, NDVI, NDWI, NDBI, NDTI, NDMI, and MNDWI, as well as SRTM3 and NTL. The usefulness of these indicators has been demonstrated by other researchers [107,108,109,110,111,112]. However, the objectives of our research did not include a comparative study of the effectiveness of different LULC classification algorithms [113] and modern approaches to increasing their classification accuracy [114].

In this work, to predict changes in LULC by 2030 and 2050, the combined CA-ANN model was used, and studies were carried out using the MOLUSCE plugin of QGIS software, the effectiveness and reliability of which have been confirmed in many works [32,33,34,35,36,37]. In our case, the results of the validation of the forecast model also turned out to be relatively high (≥0.85), which gives grounds to consider the effects of modeling future changes in LULC quite reliable since 0.75 is regarded as an acceptable threshold [36]. Modeling and forecasting the development of LULC using variables such as the distance from roads, settlements, rivers, slope, aspect ratio, and DEM level showed that the trends in the growth of arable land and the decrease in grasslands in Korgalzhyn District by 2030 will be preserved. But by 2050, the rate of increase in sown areas and decrease in grasslands will decrease somewhat, reaching 48.76% and 16.82% of the district’s area, respectively. According to some researchers, using the CA-ANN model for predicting LULC has advantages and disadvantages [37]. The advantage of CA-ANN is its higher forecasting accuracy compared to other models, which allows you to model various scenarios using input variables of multiple natures. The negative side of CA-ANN is the need for a large amount of data and their high sensitivity to input parameters. In this regard, another omission of our research is the fact that the input variables were limited to only six factors, without considering many other spatial, socio-economic, climatic, and political variables. At the same time, covering many predictors was not part of the objectives of this study due to the lack of systematic collection of the necessary information by the relevant services of Korgalzhyn District over a long period of time. But, despite this, in subsequent work, we will expand the range of predictors as much as possible.

Thus, our research goals have been achieved, and all the assigned tasks have been solved. We have shown that in Korgalzhyn District, from 2010 to 2021, there was a process of increase in arable land, dominated by grain crops, with an adequate decrease in pastures. Modeling and forecasting future LULC transformations using one of the currently advanced combined CA-ANN methods showed that grassland-occupied lands in the category of cropped areas will continue in 2030 and 2050, but at different rates in specific periods.

5. Limitations of the Study and Prospects

Our research has several limitations. First of all, it was not the scope of our research to perform long-term, large-scale mapping of AOI crops using imagery at a larger scale than Landsat imagery. As a result, we cannot present detailed changes in the yield of individual crops over a long period of time. In addition, our prediction of LULC change was highly limited by the predictors. These limitations are because, at the first stage of the research, we needed to determine whether there were changes in the composition of LULC classes in the past period and how they would vary. Therefore, the above disadvantages are problems we will solve in future research.

6. Conclusions

The main pattern of the spatial distribution of land in Korgalzhyn District from 2010 to 2021 was a steady increase in the share of arable land, more than 95% of which was occupied by spring grain crops, with approximately the same decrease in the percentage of arable land from year to year. Changes in water and forest resources in the AOI did not have a clearly defined direction and, over the years of research, changed in a relatively small corridor, while the territories occupied by settlements and industrial infrastructure tended to grow, but at a low rate. The use of the CA-ANN method to model future changes in the composition of LULC showed that in the Korgalzhyn region, in 2030 and 2050, the trend of an increase in the area of land intended for growing crops will continue, making up more than half of the AOI’s territory. This process will lead to a decrease in pasture areas by about one and a half times. At the same time, there will be no drastic changes in the composition of other land classes. The results obtained can help in the development of land management plans and policies for the socio-economic development of Korgalzhyn District, considering the conservation of biodiversity and ecosystem services as well as climate change mitigation.