Changing characteristics, driving factors and future predictions of land use in the Weigan-Kuqa River Delta Oasis, China

Abstract

The oasis serves as the central component of the arid ecosystem and plays a crucial role in supporting human activities. However, the ecological environment in the oasis region is fragile, and even a minor alteration in land use (LU) can significantly impact the stability of the ecosystem. Therefore, it is imperative to undertake comprehensive research on the spatio–temporal patterns of LU change in the oasis, reveal its driving factors, and predict future development. This is crucial for devising scientifically and logically sound land management strategies, upholding the equilibrium between humans and land in arid areas, and attaining sustainable development of the regional ecology and economy. This study focuses on the Weigan–Kuqa River Delta Oasis in China as the research area, analyzes the changes in LU in the oasis from 2010 to 2022 using various methods such as transition matrix, dynamic degree, intensity analysis, and center of gravity shift. The study also investigates the factors influencing these changes using the optimal parameters–based geographical detector (OPGD). Additionally, it predicts the future trends in LU development under four different scenarios using the mixed–cell cellular automata (MCCA), and illustrates distribution characteristics by combining Moran’s I index and hotspot analysis. The results suggest that: (1) Between 2010 and 2022, the LU in the oasis changed rapidly, with consistent increase in the amount of construction land, arable land, and garden land, while the amount of forest-grassland and unused land decreased overall. (2) Population density played a leading role in the changes, but soil type also had a significant impact. Over the course of time, the influence of roads and transportation has progressively increased. (3) Compared with 2022, the acreage of arable land, garden land, and construction land increases under the four future scenarios: natural development scenario (NDS), economic development scenario (EDS), cropland development scenario (CDS), and ecological protection scenario (EPS). However, the acreage of forest–grassland and unused land decrease. From a spatial perspective, large towns, the downstream of alluvial fans, and the central oasis are key areas where the distribution of hot spots and sub–hot spots of each LU type varies significantly among the four scenarios. The EPS provides a certain level of protection for forest-grassland areas and water bodies, making it the most appropriate development model for oases. These findings have the potential to offer valuable academic guidance for oasis land resource management and are crucial for achieving coordinated development at regional level.

Introduction

Land resources play a crucial role in the survival and advancement of humanity, and land use (LU) is a clear representation of how human activities interact with the biological environment^1,2. As a leading factor in global environmental change, LU change not only brings about significant changes in the surface structure but also has an important impact on ecosystem services³. For example, deforestation may lead to biodiversity loss and reduced carbon stocks, similarly, urban expansion can lead to the formation of heat islands and water resource tension, threatening regional ecological security and sustainability. The impact of LU change is very obvious, particularly in ecologically sensitive places. Within this particular framework, the task of reconciling the ecological environment and human necessities poses a formidable challenge. Hence, it is essential to comprehensively comprehend the spatial and temporal patterns of LU change, investigate the factors that drive these changes, and forecast future alterations. These prerequisites are vital for devising evidence-based development policies and management strategies, which play a crucial role in enhancing the ecological environment, ensuring a harmonious relationship between humans and the land, and fostering sustainable development at a regional level.

At present, LU research encompasses a relatively mature system covering category information acquisition, change process analysis, driving factor detection, and predictive modeling. However, scholars continue to seek progress and innovation. Satellite remote sensing technology, with its wide detection range, real–time capabilities, and rich information, is widely used for monitoring LU changes⁴. In past remote sensing research, LU information has been obtained in two ways: through a set of published products such as China land cover dataset (CLCD), 30 m global land cover data product (GlobeLand30), and global land cover product with fine classification system at 30 m (GLC_FCS30)^5–7. When current products are insufficient for research purposes, scholars employ alternative methods, such as supervised or semi-supervised algorithms, to detect LU based on remote sensing picture data⁸. Common classification algorithms include gradient tree boosting (GTB), support vector machine (SVM), minimum distance (MD), random forest (RF), classification and regression trees (CART), and naive bayes (NB)^9–11. Particularly, GTB is an exceptionally precise classification algorithm that effectively reduces variation and bias in classification¹². When analyzing LU transformations, typical methods utilized by experts include transition matrix, dynamic degree, and the center of gravity shift^13–15. Due to the dynamics and complexity of LU transformation, combining several approaches to quantitatively analyze the temporal characteristics and spatial patterns of LU is a vital means to discuss the characteristics of LU transformation. In addition, the analysis of driving factors has received attention since the 1990s¹⁶. In the initial stages of research, qualitative methods such as linear, correlation, and trend analysis were predominantly employed¹⁷. The impact of drivers has been quantified using statistical analysis procedures, including logistic regression, multiple regression, principal component analysis, and others^18–20. While traditional approaches are unable to visually demonstrate the impact of different variables on changes in LU, an emerging geographical detector has been developed to effectively solve this problem²¹. The geographical detector is capable of extracting effective spatial association rules from massive spatial data and revealing the impact of factors by detecting their spatial heterogeneity^22–24. The model is easy to use and highly applicable, showing significant superiority in studying the driving processes of LU transformation. Additionally, numerous research have been conducted on LU prediction. Based on cellular automata (CA), designs like CA–Markov, CLUE–S, FLUS, and PLUS have gradually emerged^25–29. These models demonstrate good modeling capabilities, although they also have certain limitations: each cell contains only one type of LU, disregarding the occurrence of many types mixed within actual cells³⁰. The mixed–cell cellular automata (MCCA) model, proposed by Liang et al.³¹, considers the complexity of LU structure, with each cell containing the coverage ratio of each type, thereby compensating for the shortcomings of the pure cell model. The primary novelty of this study lies in its classification of LU using extensive field survey data, specifically distinguishing the LU category of garden land, which sets it apart from other LU products. In addition, there are relatively few studies using the MCCA model to predict future LU scenarios, so this work somewhat supplements the existing body of research on this model.

The Weigan–Kuqa River Delta Oasis, located in the eastern urban comprehensive development zone of the Aksu region in the Xinjiang Uygur Autonomous Region of China, is currently facing rare historical opportunities for its development due to the Western Development Strategy and regional planning policies. The fast changes in population growth, industrial and agricultural development, and the surrounding environment have rapidly altered the types of LU in the oasis, exerting tremendous pressure on the oasis ecosystem and affecting its sustainable development capacity. This study focuses on analyzing the changes in the LU in the Weigan–Kuqa River Delta Oasis from 2010 to 2022. The analysis includes examining the spatial and temporal patterns of LU through various methods such as transition matrix, dynamic degree, intensity analysis, and center of gravity shift. Additionally, the study explores the factors influencing these changes using the optimal parameters–based geographical detector (OPGD). Furthermore, the study predicts the LU development trends under four scenarios using the mixed–cell cellular automata (MCCA), and illustrates distribution characteristics by combining Moran’s I index and hotspot analysis. The objective of this study is to offer empirical research findings on LU transformations in the oasis. The aim is to give a scientific foundation and academic backing for regional land resource management, with the ultimate goal of achieving sustainable development of oasis ecosystems.

Materials and methods

Study area

The Weigan–Kuqa River Delta Oasis (40°57’N–41°48’N, 82°05’E–83°42’E) is located in northwestern China, namely at the northern border of the Tarim Basin in the Xinjiang Uygur Autonomous Region (Fig. 1). The terrain of the region exhibits a distinct fan-shaped plain oasis, with lower elevation in the southern section and higher elevation in the northern region. The region experiences a mild temperate continental arid climate characterized by extremely hot summers, dry and very cold winters, ample sunshine, limited rainfall, and high rates of evaporation. The oasis has a main emphasis on agriculture, including the cultivation of crops like as wheat, corn and cotton, as well as fruit trees including apples, walnuts and dates. The indigenous flora includes Nitraria tangutorum, Karelinia caspia, Alhagi sparsifolia, Tamarix chinensis, Populus euphratica, and various more species. Due to economic development, population growth, and environmental changes, the LU patterns of the oasis have undergone fast transformation, which has had an impact on its ability to achieve sustainable development.

Fig. 1.

Weigan-Kuqa River Delta Oasis (the vector map data of China and Xinjiang are from the National Platform for Common GeoSpatial Information Services: https://www.tianditu.gov.cn/, and the map approval number is GS (2024) 0650; the boundary of the study area is drawn using ArcMap 10.2, link: https://www.esri.com/).

Data sources

This work primarily utilizes data from Landsat, Sentinel, topography, soil type, population, nighttime light, and gross domestic product (GDP), among others (Table 1). All data have been uniformly projected into the WGS_1984_UTM_Zone_44N coordinates.

Table 1.

Information on the data utilized in the study.

Data	Pixel resolutions	Data sources	Derived data
Landsat 5/8	30 m	https://earthengine.google.com/	LULC (2010, 2013)
Sentinel 2	10 m	https://earthengine.google.com/	LULC (2016, 2019, 2022)
SRTM V3	30 m	https://earthengine.google.com/	Elevation
			Slope
Monthly mean temperature dataset for China	1 km	https://data.tpdc.ac.cn/	Annual average temperature
Monthly precipitation dataset for China	1 km	https://data.tpdc.ac.cn/	Annual precipitation
Spatial distribution data of soil types in China	1 km	https://www.resdc.cn/	Soil type
Population data	1 km	https://landscan.ornl.gov/	Population density
Night-time light data	1 km	https://dataverse.harvard.edu/	Light intensity
Real GDP dataset	1 km	https://figshare.com/	GDP density
OpenStreetMap data	–	https://www.openstreetmap.org/	Distance from rivers
			Distance from railways
			Distance from roads

Natural factor data

The topographic data utilized in this study were extracted from SRTM V3 images of Google Earth Engine (GEE), including information on elevation and slope. The monthly mean temperature and monthly precipitation raster data were combined over a period of 12 months to get the average temperature and precipitation for the entire year. Soil type data basically covers all types of soil. Combined with the river distribution vector data, ArcMap 10.2 was used to calculate the Euclidean distance to determine the distances from rivers.

Human factor data

Between 2010 and 2022, the experimental team conducted field surveys in the study area at three-year intervals, resulting in a substantial collection of precise sample data representing various LU types. Therefore, this study opted to acquire LU data from 2010 to 2022 through independent classification. Combining the “Current Land Use Classification (GB/T 21010–2017)”³² and survey realities, the studied region was categorized into six LU categories: arable land, garden land, forest–grassland, construction land, water bodies, and unused land. The classification samples comprised visually interpreted data and field survey data. Accuracy verification was conducted using overall accuracy (OA), producer’s accuracy (PA), and user’s accuracy (UA). Through the GEE platform, Landsat 5/8 or Sentinel 2 data from July 1 to September 30, 2010–2022, with a cloud cover percentage of less than 30%, were accessed. The median composite method was used to generate composite images, and LU classification outcomes were obtained using the Gradient Tree Boosting (GTB) classifier. The OA of the five phases of classification outcomes ranged from 90 to 96%, and the detailed accuracy of GTB classification data is shown in Table 2. Subsequently, manual modifications were implemented to the classification outcomes according to the actual circumstances of the studied region to obtain LU data. The Landscan dataset provided the population density data that was used in the research. The China “DMSP–OLS–like” nighttime light remote sensing dataset, which was available on the Harvard Dataverse platform, was the origin of the light intensity data³³. The density of GDP was derived using actual GDP data obtained from the Figshare platform³⁴. Railway and road distribution vector data were obtained from OpenStreetMap, and the Euclidean distances were calculated to determine the distances from railways and roads, respectively.

Table 2.

Accuracy of LU data.

Indicators	2010	2013	2016	2019	2022
Producer’s accuracy
Arable land	0.98	1.00	1.00	0.91	0.98
Garden land	0.94	0.95	0.95	0.95	0.96
Forest–grassland	0.97	0.94	0.92	0.89	0.94
Construction land	0.78	0.90	0.88	0.88	0.79
Water bodies	1.00	1.00	1.00	1.00	1.00
Unused land	0.87	0.96	0.92	0.84	0.91
User’s accuracy
Arable land	0.98	0.98	0.95	0.98	1.00
Garden land	0.94	1.00	1.00	0.86	0.96
Forest–grassland	0.94	0.97	0.95	0.86	0.85
Construction land	0.88	0.90	0.91	0.91	0.93
Water bodies	1.00	1.00	1.00	1.00	1.00
Unused land	0.80	0.92	0.89	0.81	0.81
Overall accuracy	93%	96%	94%	90%	93%

Research methods

This study comprehensively analyzes the spatio–temporal characteristics of LU transformations in the Weigan River–Kuqa River Delta oasis. It focuses on exploring the driving mechanisms and future development trends, thereby deepening the knowledge of LU transformations and management in the oasis. Figure 2 depicts the research framework and the methodologies employed in this research.

Fig. 2.

Framework for research (the boundary of the study area was created using ArcMap 10.2, link: https://www.esri.com/).

Gradient tree boosting classifier

GTB is an enhancement algorithm for decision trees, known for its high accuracy and stability. It employs gradient descent to optimize functions, thereby reducing loss. This process iteratively transforms weak learners within the ensemble into strong learners, consequently improving classification accuracy³⁵. The GTB algorithm can be summarized into the following steps:

1. Input training data set:

   1

   whereis the feature vector of each sample, andis its class label.

2. Establish the loss function:

   2

   whereis the fitted function of y.

3. Initialize the model variables:

   3

   where c is a constant for the minimum value of the loss function.

4. Perform the following four steps for each model:

   1. Calculate the residual for each sample:

      4
      Where is the model number.
   2. Use to train the m-th tree, and the area divided by its leaf nodes is.
   3. For each leaf node of tree, calculate its output value:

      5
   4. In order to minimize the loss function and estimate the value of the leaf node area, a linear search is used. Update the tree:

      6
5. Get the final tree model:

   7

   Recent studies have demonstrated that GTB exhibits outstanding performance on LU datasets. Empirical evidence has confirmed that this classifier is more precise compared to other classifiers like CART and RF, demonstrating its high adaptability and stability³⁶.

Transition matrix

The LU transition matrix can be applied to quantify the changes in multiple types of LU from the inception to the completion of the investigation³⁷. The expression is provided below:

8

where T is the acreage of the LU category; a and bare the categories of LU at the inception and the completion of the investigation, respectively; andis the acreage of LU category a at the inception of the investigation that has been converted to LU category b at the completion of the investigation.

Dynamic degree

The LU dynamic degree is an index that represents the velocity and intensity of LU type conversion³⁸. This index can be classified into two categories: single and comprehensive, according to different research subjects.

The single LU dynamic degree focuses on analyzing the changes in a specific LU category within a particular area during a specified period of time. Its calculation formula is:

9

whereandare the acreage of an assigned LU category at the inception and the completion of the investigation, respectively; and S is the time interval between the completion and the inception of the investigation.

The comprehensive LU dynamic degree quantifies the whole extent of change in LU categories within a certain region over a specific period. It measures the annual rate of transformation for different forms of LU in that region. The formula for this index is given below:

10

whereis the acreage of LU category a at the inception of the investigation;is the absolute data of the acreage of LU category a that has changed to other LU categories from the initiation to the completion of the investigation; n is the total quantity of LU categories; and S is the time interval between the start and the end of the investigation.

Intensity analysis

In 2012, Aldwaik and Pontius Jr³⁹. proposed an intensity analysis framework, including three aspects: interval level, category level and transition level. Subsequently, scholars continued to improve the existing framework. This work adopts the intensity analysis framework based on the transition level improvement proposed by Li et al.⁴⁰, analyzing the transformation trends among LU types and their impact on LU patterns, using both absolute and relative intensity viewpoints.

Absolute intensity shows the absolute quantity of mutual conversion between LU types, which can be further divided into two aspects: conversion from a certain LU type to other LU types and conversion from other LU types to the specified LU type. Formulas (11) and (12) compute the absolute transfer-in intensity and the average absolute transfer-in intensity, respectively, while formulas (13) and (14) calculate the absolute transfer-out intensity and the average absolute transfer-out intensity, respectively.

11

12

whereis the absolute transfer-in intensity;is the average absolute transfer-in intensity; i is the LU type code at the beginning of the period; n is the transfer-in LU type code; I is the number of LU types at the start of the period; t is the code of the beginning time node; is the year corresponding to the beginning time node t;is the year corresponding to the end time node;is the land area converted from LU type i to LU type n during the period fromto; andis the area of ​​LU type n that has not changed during the period fromto.

13

14

whereis the absolute transfer-out intensity;is the average absolute transfer-out intensity; m is the transfer-out LU type code; j is the LU type code at the end of the period;J is the number of LU types at the end of the period;is the land area converted from LU type m to LU type j during the period fromto; andis the area of LU type m that has not changed during the period fromto.

Based on the absolute intensity, the relative intensity further analyzes the impact of the conversion intensity between LU types on the LU structure. Formulas (15) and (16) determine the relative transfer-in intensity and the average relative transfer-in intensity, respectively. Formulas (17) and (18) calculate the relative transfer-out intensity and the average relative transfer-out intensity, respectively.

15

16

whereis the relative transfer-in intensity;is the average relative transfer-in intensity;is the land area converted from initial LU type i to final LU type j during the period fromto;is the land area converted from LU type n to final LU type j.

17

18

whereis the relative transfer-out intensity;is the average relative transfer-out intensity; andis the land area converted from initial LU type i to LU type m.

LU change processes that result in a substantial increase in the acreage or proportion of a specific LU type are classified as transformations with absolute or relative propensity. Conversely, processes that lead to a decrease in the acreage or proportion of a LU type are categorized as processes with absolute or relative suppression. The phenomenon of LU change is characterized by two factors: absolute quantity and relative proportion. When, land use type n tends to obtain transfer from the initial land use type i; conversely, transfer from land use type iis suppressed. When, land use type m tends to transfer out to land use type j; conversely, transfer out to land use type j is suppressed. When, land use type n is relatively more inclined to obtain transfer from initial land use type i, and this conversion process has a significant impact on the initial area proportion of land use type i; conversely, transfer from initial land use type i is relatively suppressed, and the impact on the initial area proportion of land use type i is relatively small. When, land use type m is relatively more inclined to transfer out to land use type j, and the conversion process has a relatively large impact on the final area proportion of land use type j; conversely, transfer out to land use type j is relatively suppressed, and the impact on the final area proportion of land use type j is relatively small.

The results of the intensity analysis framework are displayed as an intensity map. Figure 3 is a unit of the LU intensity analysis map, taking the inception LU category i and the completion LU category j as examples. In the intensity map unit, ① and ② represent the absolute transfer-in intensity and the absolute transfer-out intensity, respectively, while ③ and ④ represent the relative transfer-in intensity and the relative transfer-out intensity, respectively. Red fill indicates a tendency and blue fill indicates a suppression nature.

Fig. 3.

Unit of the LU intensity analysis map.

Center of gravity shift

The center of gravity shift can spatially depict the continuous varying processes of various LU categories in a region⁴¹. Its calculation formula is:

19

where x and y are the longitude and latitude coordinates of the centroid of an assigned type of LU; n is the total quantity of patches of an assigned category of LU;is the acreage of the i-th patch of an assigned category of LU; andandare the longitude and latitude coordinates of the centroid of the i-th patch of an assigned category of LU, respectively.

Optimal parameters-based geographical detector

The geographical detector is capable of performing detection analysis on LU types and selected impacting variables, discovering spatial differentiation of LU types and revealing their main driving forces⁴². However, the traditional geographic detector relies on experience to discretize continuous factors, leading to subjectivity and poor discretization. Therefore, this study opts for the OPGD for driving factor analysis⁴³. When discretizing continuous factors, the “GD” package in the R4.3.1 platform is used. The effectiveness of discretization is judged by the q value (formula (12)). The ability to explain increases with the q value. The categorization methods include standard deviation, natural breaks, geometric interval, equal interval, and quantile. Through multiple experiments, the total number of categories is set between 6 and 9. The OPGD determines the optimal parameters by choosing the parameter relationship (categorization method and number of breaks) that yields the highest q value.

20

where q is the ability to explain the independent variable;is the LU category or the zoning of influencing variables;and N are the quantities of units in the sub–area h and the entire zone, respectively; andandare the variance of the LU type in sub–area h and the entire zone, respectively.

Mixed-cell cellular automata model

Model introduction

The imitation results of the MCCA represent the percentages of various LU categories (Fig. 4), providing a more accurate depiction of the LU transformation layouts. The model primarily comprises three components: mining quantitative conversion rules, dynamic modeling of LU structure, and evaluation of simulation results. Using LU data from the initial and target years, the random forest regression (RFR) method explores the link between various LU types and impacting variables, helping to predict the developmental potential of various LU types (formula (21)). The temporal and spatial dynamics of LU structure are simulated by integrating macro LU requirements with regional LU competition. Adaptive inertia coefficients guide LU quantities toward target LU demands. In a roulette competition, it is determined whether the percentages of a specific LU category increase and by how much. Subsequently, the number of remaining LU categories converted into this type is determined by a series of quantifiable conversion criteria. The total change probability of land k is expressed as formula (22). When the computer–generated LU quantity equals the target quantity, the simulation results are output. The evaluation methods for simulation results mainly include the sub–pixel confusion matrix (SCM), mixed cell figure of merit (mcFoM), and relative entropy (RE), representing the global accuracy of imitation outcomes, the imitation accuracy of mixed cells, and the likeness of LU architecture, respectively. Additionally, overall accuracy (OA_v), user’s accuracy (UA_v), and producer’s accuracy (PA_v) generated by SCM are considered. The lower the SCM and RE, and the closer the mcFoM, OA_v, UA_v, and PA_v are to 1, the more accurate the imitation outcomes. The calculation equations for OA_v, UA_v, PA_v, mcFoM, and RE are shown in formula (23–32).

21

Fig. 4.

LU structure of mixed–cell CA.

whereis the development potential of land k in mixed cell i;andare the training and prediction processes of the RFR algorithm, respectively;is the sample data;is each driving factor; andis the driving data.

22

whereis the total change probability of land k in mixed cell i at the i-th iteration;is the proportion of land k in mixed cell i; andis the feedback of future demand for land k.

23

where K is the number of LU types, andis each element of the SCM.

24

25

26

where k is a certain land type, whileandare the actual and simulated proportional changes of cell i, respectively.

27

28

29

30

31

whereis the RE of the actual and simulated land structure of the mixed cell i, whileandare the actual and simulated land structures, respectively.

32

where M is the total number of mixing cells.

Simulation accuracy verification

The scale of the mixed cell significantly influences the simulation accuracy of the MCCA model. Liang et al.³¹ tested the simulation accuracy of the MCCA model using mixed pixels of six scales with spatial resolutions of 250 m to 1500 m (250 m intervals). The outcomes showed that the simulation accuracy of 250 m was the highest. In view of this, this study further explored the range of 250 m to 500 m to determine the optimal scale suitable for the study area. The LU data were aggregated into mixed pixels of six scales with spatial resolutions of 250 m to 500 m (50 m intervals) through ArcMap 10.2 software, and the MCCA model was used for LU simulation. The mcFoM values at these six scales were 0.179, 0.174, 0.167, 0.121, 0.132, and 0.119, respectively. It can be concluded that the ideal simulation scale for the MCCA model in this study is 250 m. At a scale of 250 m, LU data for 2016 were used to simulate data for 2022. The OA_v and RE_mean used for result evaluation were 0.836 and 0.974 respectively (see Table 3 for other simulation accuracies), demonstrating that the simulation results are relatively reliable and can be used for prediction research.

Table 3.

Producer’s accuracy and user’s accuracy for simulation verification.

Indicators	Arable land	Garden land	Forest–grassland	Construction land	Water bodies	Unused land
Producer’s accuracy	0.85	0.71	0.80	0.70	0.63	0.73
User’s accuracy	0.86	0.86	0.85	0.88	0.70	0.78

Future scenario settings

In predicting future LU development using the MCCA model, this study established four scenarios based on “The Spatial Planning of National Land in Aksu (2021–2035)” and “The 14th Five–Year Plan for National Economic and Social Development and Vision 2035 of Aksu”, combined with previous research experiences. The scenarios are: natural development, economic development, cropland development, and ecological protection (Table 4).

Table 4.

Scenario settings.

Scenario types	Abbreviation	Description
Natural development scenario	NDS	Under the current development mode, the transition conditions between LU types are expected to continue following historical trends.
Economic development scenario	EDS	Without restrictions on the development of cities and villages, construction land expands rapidly, and the acreage of arable land and garden land also increases to some extent. This scenario assumes the absence of policy constraints.
Cropland development scenario	CDS	The oasis, being the primary agricultural production area outlined in the regional plan, faces a significant increase in the amount of arable land. The development of arable land is not restricted, while its conversion to other LU types is restricted. Garden land and construction land will increase slightly. This scenario does not involve other policy interventions.
Ecological protection scenario	EPS	In accordance with the principle of harmonious coexistence between humans and nature, and under the influence of the “three zones and three lines” policy, there are restrictions on the unchecked expansion of arable land, garden land, and construction land. Additionally, the conversion of forest–grassland and water bodies into other LU categories is controlled.

Both EDS and CDS prioritize production development and the improvement of living conditions, however EDS focuses on urbanization and the expansion of construction land, while CDS focuses primarily on the development of arable land.

Spatial autocorrelation

Spatial autocorrelation is a technique used to determine the distribution characteristics and interrelationships of spatially correlated data. It is divided into two types: global and local. Moran’s I is employed to test the spatial correlation of LU within the study area, and hotspot analysis (G^* index) is utilized to explore the clustering characteristics of local areas. The procedure for calculating spatial autocorrelation is as follows:

33

34

where I is Moran’s I; n is the quantity of units; x_a and x_b are the attribute data of unit a and b, respectively;  is the average data of the attribute; w_ab is the matrix of dimensional weights; and S is the standard deviation of the attribute values.

Results

LU change analysis

Structure and transition analysis

By integrating remote sensing imagery and utilizing the GTB algorithm for classification, we have generated the LU distribution map spanning from 2010 to 2022 (Fig. 5). Apart from three larger towns, construction land is scattered within the oasis. Garden land primarily surrounds construction land. Arable land exhibits a fan–shaped distribution and is the main component of the oasis. Forest–grassland mainly covers the lower parts of alluvial fans and areas near water sources, while unused land is located on the outskirts of the oasis. Over time, construction land has expanded into clustered formations, garden land encroached on arable land, and arable land has expanded towards the outskirts of the oasis.

Fig. 5.

LU classification map (the boundary of the study area was created using ArcMap 10.2, link: https://www.esri.com/).

The acreage and acreage proportions of various LU categories are displayed in Table 5. The data indicate that the studied region is consistently occupied by forest–grassland. Arable land, garden land, and construction land are increasing year by year, with arable land showing the most significant growth. This increase is connected to agricultural development, urbanization, and population expansion in the studied region. Arable land grew from 20.10% (2010) to 27.30% (2022), garden land grew from 3.92% (2010) to 10.37% (2022), and construction land grew from 0.96% (2010) to 3.74% (2022). The acreage of forest–grassland initially decreased and then increased, from 46.15% in 2010 to 32.53% in 2016, and then to 37.18% in 2022, with an overall decrease of 8.97%. The acreage of unused land initially rose and then fell, from 26.73% in 2010 to 32.76% in 2013, and then to 18.95% in 2022, with an overall decrease of 7.78%. Throughout the research phase, the proportions of water bodies in terms of acreage varied between 1.51% and 3.26%.

Table 5.

Acreage of each LU type from 2010 to 2022.

	Arable land	Garden land	Forest–grassland	Construction land	Water bodies	Unused land	Total
2010	2266.20 (20.10)	442.01 (3.92)	5204.27 (46.15)	108.26 (0.96)	240.97 (2.14)	3013.95 (26.73)	11275.66 (100.00)
2013	2719.17 (24.12)	464.51 (4.12)	4008.72 (35.55)	154.62 (1.37)	234.70 (2.08)	3693.94 (32.76)	11275.66 (100.00)
2016	2867.50 (25.43)	690.28 (6.12)	3667.52 (32.53)	316.85 (2.81)	367.69 (3.26)	3365.82 (29.85)	11275.66 (100.00)
2019	3071.44 (27.24)	965.83 (8.57)	3929.59 (34.85)	334.55 (2.97)	170.48 (1.51)	2803.77 (24.87)	11275.66 (100.00)
2022	3077.99 (27.30)	1169.58 (10.37)	4192.42 (37.18)	421.43 (3.74)	277.46 (2.46)	2136.78 (18.95)	11275.66 (100.00)

The data outside the brackets is the area (unit: km²), and the data inside the brackets is the area ratio (unit: %).

Using the change detection statistics feature in ENVI 5.3 software, we explored the mutual conversion relationships between different LU categories from 2010 to 2022 and displayed the results through a Sankey diagram (Fig. 6). From 2010 to 2013, arable land experienced a net outflow of 17.37%, primarily converting to forest–grassland (0.49%) and garden land (6.49%). Garden land had a net outflow of 51.25%, with most of it converting to arable land (26.93%) and forest–grassland (22.34%). Forest–grassland predominantly converted to unused land (19.83%) and arable land (12.23%). Construction land experienced a net outflow of 17.09% to forest–grassland and 8.30% to unused land, with fewer conversions to the remaining LU categories. Water bodies mainly transformed to forest–grassland, covering an acreage of 33.22%. Unused land had a net outflow of 12.50%, with 9.17% converting to forest–grassland. From 2013 to 2016, 2016 to 2019, and 2019 to 2022, the conversion patterns between LU types were generally consistent with those observed from 2010 to 2013. From 2013 to 2016, there were numerous conversions between arable land, garden land, and forest–grassland. Forest–grassland also experienced some outflow to unused land. Construction land had been transferred to arable land, forest–grassland, and unused land to a certain extent. Water bodies and unused land were primarily turned into forest–grassland. From 2016 to 2019 and 2019 to 2022, there were still significant conversions between arable land, garden land, and forest–grassland. Arable land, garden land, forest–grassland, and unused land were the primary types transferred out for construction land. Water bodies primarily converted to forest–grassland, followed by unused land. Unused land had the highest net outflow to forest–grassland.

Fig. 6.

Sankey diagram of the acreage percentage transfer of each LU category from 2010 to 2022.

Dynamic degree and intensity analysis

According to the statistics of each LU type from 2010 to 2022, formulas (9) and (10) were employed to calculate the single and comprehensive LU dynamic degree (Fig. 7). From 2010 to 2022, the single dynamic degree of arable land, garden land, and construction land in the studied region was consistently positive, indicating a rising trend in the acreage of these three LU categories. Compared to other LU categories, arable land experienced a relatively small change range, peaking at 6.66% (2010–2013). The most significant change in garden land occurred from 2013 to 2016, with its value reaching 16.20%. The value of construction land from 2013 to 2016 was higher than that of any LU type during the four time periods, peaking at 34.97%. The acreage of forest–grassland decreased from 2010 to 2016 but increased from 2016 to 2022, resulting in a single dynamic degree transition from negative to positive. The single dynamic degree of water bodies displayed a fluctuating trend of negative–positive–negative–positive, reaching 20.92% during 2019–2022. Unused land only increased from 2010 to 2013, with a single dynamic degree of 7.52%. The comprehensive LU dynamic degree for the four time periods of 2010–2013, 2013–2016, 2016–2019, and 2019–2022 was 3.55%, 1.98%, 2.24%, and 1.97%, respectively, indicating that LU change is relatively fast, showing a continuous expansion trend of the oasis.

Fig. 7.

Changes in the LU dynamic degree from 2010 to 2022.

Figure 8 depicts the intensity of LU changes. During the four periods of 2010–2013, 2013–2016, 2016–2019, and 2019–2022, the intensity of changes in each LU category in the studied region showed certain differences. However, during the period from 2010 to 2022, some transformations also exhibited consistency: the transformation of arable land into garden land showed a systematic tendency; the transformation of arable land into forest–grassland showed an absolute tendency; the transformations of garden land into construction land and forest–grassland into water bodies showed a relative tendency; the transformations of arable land into water bodies and unused land, garden land into water bodies and unused land, water bodies into arable land, garden land, and construction land, and unused land into garden land showed systematic suppression.

Fig. 8.

LU intensity analysis map from 2010 to 2022 (AL is arable land; GL is garden land; FG is forest–grassland; CL is construction land; WB is water bodies; and UL is unused land).

Spatio-temporal migration analysis

Using ArcMap 10.2, we calculated the center of gravity coordinates of several LU categories in the study area from 2010 to 2022, creating a visualized center of gravity shift map (Fig. 9). From 2010 to 2022, arable land moved approximately 17.73 km, experiencing shifts to the southeast, northeast, northwest, and southwest. Garden land moved about 8.72 km, shifting in the southeast, northeast, southwest, and southwest. Forest–grassland migrated approximately 20.36 km, undergoing migration towards the southwest, southeast, northeast, and southwest. Construction land had a larger span, migrating approximately 38.25 km and shifting towards the southwest, southwest, northeast, and southeast. Water bodies had the largest span, with a migration distance of about 54.70 km, moving successively to the southwest, northeast, southwest, and southeast. Unused land migrated approximately 14.49 km, shifting towards the southwest, northwest, northeast, and northeast. Overall, the center of gravity of arable land and unused land exhibited a northwestward migration trend, while the center of gravity of garden land and water bodies showed a southeastward migration trend. The center of gravity of forest–grassland and construction land displayed a southwestward migration trend. During the study period, compared to arable land, garden land, forest–grassland, and unused land, the centers of gravity of water bodies and construction land were more spatially dispersed, indicating that these two LU types had higher mobility within the study area.

Fig. 9.

The trajectory of the center of gravity shift of each LU category in the study area from 2010 to 2022.

LU driving mechanism analysis

Driver selection

Based on principles such as availability, scientific relevance, and representativeness of driving factors, and considering the unique circumstances of the investigation, 11 influencing variables were selected from both natural environment and social economy aspects (Table 6). Using ArcMap 10.2, a 1 km×1 km grid was generated for the studied region, and the raster data of these factors were sampled at the grid centers. Subsequently, the results were imported into the OPGD to investigate the influencing mechanisms of LU change in the studied region from 2010 to 2022.

Table 6.

Influence factors.

Categories	Factors	Variables
Natural environment	Elevation	X1
	Slope	X2
	Distance from rivers	X3
	Annual average temperature	X4
	Annual precipitation	X5
	Soil type	X6
Social economy	Distance from railways	X7
	Distance from roads	X8
	Population density	X9
	Light intensity	X10
	GDP density	X11

Factor detection results and analysis

Using the factor detection module, we detected the influence levels of various factors from 2010 to 2022 (Fig. 10). Based on the detection results and research requirements, we divided the explainable ability (q value) of influencing variables into four levels: weak influence (q < 0.1), moderate influence (0.1 ≤ q < 0.2), strong influence (0.2 ≤ q < 0.3), and significant influence (q ≥ 0.3).

Fig. 10.

Factor detection outcomes of LU change from 2010 to 2022.

From 2010 onwards, it is evident that population density and soil type have the most significant impact on LU changes. The q values for population density and soil type are 0.42 and 0.30, respectively. The distance from rivers has a strong influence, as indicated by a q value of 0.22. Elevation and annual average temperature have q values of 0.15 and 0.14, indicating a moderate level of influence. The latter six factors have weaker explanatory power. By 2013, the q value of population density reached 0.44, becoming the dominant factor in change. The explanatory power of soil type decreased from being a primary influencer to a strong influencer. Annual average temperature also had a strong influence, while distance from rivers and elevation had a moderate level of influence. In 2016, population density remained a major factor in the change, with a q value of 0.39. Next was soil type, which continued to have a strong influence. Average annual temperature, elevation, and distance from the river had q values of 0.19, 0.15, and 0.12, respectively. By 2019, the explanatory power of population density had decreased compared to previous period but still held the most significant influence. Soil type remained a strong influencer. Annual average temperature and elevation had a moderate level of influence. The explainable ability of distance from roads increased to a moderate level, while distance from rivers became a weaker influencer. By 2022, population density has a strong influence on LU changes. Soil type, elevation, light intensity, annual average temperature, and distance from roads had a moderate level of influence.

When comparing the q values of various factors vertically from 2010 to 2022, it can be observed that the influence of these factors on change varies over different periods. However, population density and soil type consistently exhibited strong explanatory power, ranking first and second in each period. The q value for distance from rivers decreased over time, while the q values for distance from railways and distance from roads increased. Elevation, slope, annual precipitation, and GDP density showed relatively stable explanatory power throughout the study period. The average explanatory power of each influencing factor from 2010 to 2022 can be ranked as follows: population density (0.35) > soil type (0.24) > annual average temperature (0.16) > elevation (0.15) > distance from rivers (0.11) > annual precipitation (0.08) > distance from roads (0.07) > light intensity (0.04) > distance from railways (0.03) > GDP density (0.02) > slope (0.01). Overall, LU changes are jointly driven by the social economy and natural environmental variables. Population density is a primary variable in LU change, and soil type is also a significant factor. Annual average temperature, elevation, and distance from rivers have some influence, while the impact of the other six variables is relatively small.

Interaction detection results and analysis

Using the Interaction Detection Module, we investigated the joint effects of 11 driving factors on the change pattern from 2010 to 2022 (Fig. 11). The interaction detection results for each period exhibit nonlinear enhancement or bivariate enhancement, indicating that the integrated impact of both variables is greater than the individual impact of one variable. The LU change is the outcome of the interaction of several variables. From 2010 to 2022, the explanatory power of population density in combination with other variables was relatively high, followed by soil type in combination with other factors. The interactions between various factors primarily concentrated on population density, soil type, annual average temperature, and elevation. Among these, the combined impact of population density and soil type was the most important, with explanatory powers from 2010 to 2022 being 0.49, 0.48, 0.44, 0.36, and 0.31, respectively. The results for each period indicate that the interplay between social economy and natural environmental variables can deepen the impact on LU change patterns.

Fig. 11.

Interaction detection outcomes of LU changes from 2010 to 2022.

LU prediction

Multi–scenario prediction

The prediction outcomes of the MCCA for the LU layout of the studied region in 2028 are displayed using RGB images (Fig. 12), and the acreage proportions of all categories under the four scenarios are calculated (Fig. 13). Under the NDS, all LU types maintain their original development trends, with an increase in the proportions of arable land, garden land, and construction land, and a decline in the proportions of unused land, forest–grassland, and water bodies. In the EDS, the acreage of construction land rapidly increases, with its proportion rising from 3.74% in 2022 to 5.60%. Cities like Kuqa, Xinhe, and Shaya continue to expand on their existing bases, and many townships experience rapid clustered development. While the proportions of arable land and garden land also increase, there is a significant decline in the proportions of forest–grassland, dropping from 37.18% in 2022 to 30.86%. This indicates that economic development is encroaching on forest–grassland. Under the CDS, the proportion of arable land increases from 27.30% in 2022 to 33.29%, with a further increase in the degree of contiguousness. As arable land expands and urban development continues, it inevitably leads to a reduction in the acreage of other LU categories. In the EPS, the extents of water bodies and forest–grassland are greater than in the other three scenarios, while the extents of construction land and arable land are smaller. While arable land, garden land, and construction land are being developed, the degree of encroachment on forest–grassland and water bodies is relatively low, promoting the preservation of forest–grassland resources and the restoration of watershed water ecology, so that ecological security is maintained to some degree. Compare the acreage proportions of each LU category in 2028 under four scenarios: the most significant rise in arable land is under the CDS; the greatest quantity of garden land is under the EDS; the largest share of forest–grassland and water bodies appear in the EPS; and the smallest proportions of construction land and unused land occur in EPS and EDS, respectively. Across all four scenarios, there is relatively minor variation in the proportion of water bodies, while arable land, garden land, and construction land show increasing trends.

Fig. 12.

Actual structure in 2022 and predicted structure for 2028 under different scenarios (data generated using the MCCA model, link: https://github.com/HPSCIL/Mixed_Cell_Cellullar_Automata).

Fig. 13.

Acreage percentages of different LU categories in 2028 under different scenarios.

Multi-scenario spatial analysis

In four scenarios, the Moran’s I values for arable land, garden land, forest–grassland, construction land, water bodies, and unused land are greater than 0.75, 0.76, 0.82, 0.68, 0.64, and 0.82, respectively, indicating significant spatial clustering of each LU type. Among them, forest–grassland and unused land exhibit the most pronounced clustering effect, while the clustering effect of water bodies is relatively weaker than the other types.

In order to conduct a more detailed examination of the level of spatial clustering, the hotspot analysis tool was used to obtain distribution maps of hot and cold spots for all LU types in future scenarios (Fig. 14). The degree of spatial clustering was categorized into five levels from weak to strong under the natural breaks method: cold spot, sub–cold spot, insignificant, sub–hot spot, and hot spot. The oasis is the main location where sub-hot spots and hot spots of arable land are predominantly concentrated. However, there are relatively fewer hot spots in the oasis within the EPS. Furthermore, the downstream districts of alluvial fans in the EDS and CDS exhibit a greater number of sub-hot spots and hot spots of arable land. The hot spots in the garden land show the spatial clustering features with more hotspots in the north and fewer in the south within the oasis, and the sub–hot spots near Shaya have noticeable differences under the four scenarios. Hot spots of construction land are concentrated in Kuqa, Xinhe, and Shaya, and the extent of hotspot areas in Kuqa under EDS is larger than in other scenarios. Hot spots of water bodies are primarily situated in the southern portion of the region, with additional sub–hot spots in the southeast corner of the region under EPS compared to previous scenarios. The region contains sub-hot spots and hot spots of forest-grassland in its eastern and southern sections, while unused land is present in the western and northern regions. The distribution of these areas is influenced by factors such as soil, water supplies, and topography.

Fig. 14.

Hotspot detection of all LU types under various scenarios in 2028 (data generated using ArcMap 10.2, link: https://www.esri.com/).

Discussion

Accuracy of LU classification

In this study, we established the LU category for garden land by considering the actual development status of the study area, which differs from existing LU products. In order to fulfil the research requirements, we integrated remote sensing imagery with survey sample data and employed the GTB algorithm for LU classification. The precision of LU classification is crucial for subsequent land–related research⁴⁴. The U.S. Geological Survey recommends achieving a classification accuracy of over 85%⁴⁵. In this work, the OA of the five phases of classification data was greater than 90%, demonstrating a certain level of accuracy. These results are sufficiently appropriate for further research. When comparing the classification results with existing products on the regional map (Fig. 15), it is evident that our study effectively differentiated between arable land and garden land and improved the identification of construction land and unused land. Orieschnig et al.⁴⁶ also used the GTB algorithm for land cover mapping in the Mekong Delta of Cambodia and achieved good classification results. However, there were some fragmented patches in the classification outcomes, which had some effect on accuracy. In future work, combining the GTB algorithm with methods for shape and spatial information extraction should significantly improve classification performance. In addition, due to the large size of the sample plots used in the field survey, arbors, shrubs, and herbaceous plants were present simultaneously within the plots. As a result, we set up a large category of forest–grassland, making it difficult to distinguish between woodland and grassland. In future studies, more precise sample plot surveys should be conducted, and higher spatial resolution imagery should be utilized to enable more detailed classification research.

Fig. 15.

Comparison of LU data (regional boundaries were created using ArcMap 10.2, link: https://www.esri.com/).

The process of LU change

In the process of analyzing LU change, this study simultaneously used different methods such as transition matrix, dynamic degree, and intensity analysis. The dynamic degree quantifies the rate of LU change, although its informational scope is limited. The tool is unable to provide information on the specific number of changes in each LU type, nor can it demonstrate the effect of LU change on the overall structure⁴⁷. The transition matrix can specifically reflect the magnitude and direction of LU change, addressing the deficiency of dynamic degree in quantity. However, it shares the same limitations as dynamic degree when it comes to assessing the impact of land change on the global structure. Intensity analysis can intuitively determine the transfer trend between LU types and their impact on LU structure⁴⁰. However, it has shortcomings in displaying the rate and specific amount of LU change. Dynamic degree, transition matrix and intensity analysis can compensate for each other’s shortcomings, making the change analysis more comprehensive and complete. Among them, intensity analysis is particularly noteworthy as it draws an important conclusion: there is no direct correlation between absolute intensity and relative intensity. For example, during the study period, there was a small amount of garden land that was converted into construction land. However, this conversion had a significant impact on the proportion of these two types of land in the region. This suggests that the absolute intensity of the conversion does not directly limit the relative intensity of the change. The systematic tendency, absolute tendency, relative tendency and systematic suppression of intensity analysis can help understand the quantity and regional proportion of conversions between LU types in the process of land change. After a comprehensive analysis of the dynamics of LU transformation by combining dynamic degree, transition matrix and intensity analysis, this work mainly found that the rate of LU change in the study area is relatively fast, and the significant quantitative transformation between two LU types may not necessarily significantly change the LU structure in the region.

Factors influencing LU change

This work utilized the OPGD as a tool to examine the impact of social economy and natural environmental variables on LU changes in the oasis. From a single–factor perspective, population density is a major driver in changes in the LU structure of the oasis. A global–scale investigation on LU also reveals that 60% of LU change is attributed to human socio–economic activities, while the remaining 40% is induced by indirect causes such as climate change⁴⁸. The role of population expansion on LU change has been widely recognized⁴⁹. In areas with higher population density, urbanization tends to be faster, leading to a corresponding increase in the demand for land. LU is generally characterized by its higher level of intensity and efficiency. At the same time, high population density may cause greater pressure on the environment, including air and water pollution, noise, and reduction of green space. In order to meet these challenges, urban planners may need to take measures such as increasing green space, improving public transportation systems, and implementing environmental protection policies. From 2010 to 2022, the population of the oasis experienced continuous growth, leading to the expansion of construction and agricultural land to meet housing, commercial, infrastructure, and production needs. It is essential to raise knowledge about the impact of human activities on land resources in ecologically vulnerable arid regions such as oasis, in order to promote harmony between humans and the environment⁵⁰. Soil type is a secondary component that influences the shift in LU in this investigation. Soil type plays a key role in LU change, and its physical and chemical properties directly affect the suitability, development difficulty and economic benefits of the land. The oasis is characterized by dark semi–hydromorphic soil and anthrosol. Dark semi–hydromorphic soil is relatively moist, and anthrosol is fertile, making them suitable for arable land and garden land development. Outside the oasis, desert soil, skeletol primitive soil, and saline–alkali soil are widespread, with LU types primarily being forest–grassland and unused land. Throughout the duration of survey, the impact of proximity to roads and railways continued to increase. As the infrastructure of internal roads and transportation in the oasis progressed and enhanced, the ease of reaching different areas rose, driving the development and utilization of land along these transportation routes. Due to the relatively flat terrain in the oasis, the slope had a consistently minor impact on LU change. From a dual–factor perspective, LU change is strongly impacted by population density and soil type, and the comprehensive impact is larger than the individual impact of each factor. This indicates that LU changes are influenced by several factors working together, and the interaction between social and natural variables might intensify the impact of LU change. High population density brings greater pressure on land resources, while soil type determines the development potential and challenges of these land resources. Effective LU planning needs to take these two factors into consideration to achieve a balance between economic, environmental and social benefits. In order to thoroughly and comprehensively assess the causes of LU changes, future study should perform a more extensive examination of relevant policy documents and broaden the range of factors considered.

LU prediction and sustainable development recommendations

This study utilized the MCCA to reveal the future trends of LU development in the Weigan–Kuqa Delta Oasis under different scenarios, including NDS, EDS, CDS, and EPS to explore a sustainable development model for the oasis. Under NDS, the LU situation in this study is basically consistent with the conclusions of Wang et al.⁵¹ in the Hetian Oasis. The ongoing expansion in the extent of arable land, garden land, and construction land and the decline of forest–grassland and water bodies indicate that the pressure on the oasis ecosystem is constantly increasing, which is not conducive to its healthy development. The expansion of construction land under EDS is unregulated, leading to the unbridled exploitation of oasis land resources, thereby limiting the long–term development of the oasis. Under CDS, the rapid expansion of arable land and the increased water demand for crops will worsen the issue of water scarcity in the oasis with limited water resources. However, the development status of LU in the oasis under EPS is somewhat satisfactory. The expansion scale of arable land, garden land, and construction land is limited, and efforts are made to safeguard and partially repair the ecological environment. To ensure the long–term sustainability of the Weigan–Kuqa Delta Oasis, regarding agricultural development, it is necessary to strengthen arable land improvement, optimize arable land layout, improve its quality and yield, enhance field support projects, strengthen irrigation and drainage facilities, and encourage the creation of high–quality farmland. In terms of the ecological environment, it is essential to prioritize protection, implement strict environmental protection systems, strengthen the preservation of forest–grassland resources, and promote watershed water ecological governance. Urban and rural development must ensure the rationality of the scale and development boundaries of construction land in cities and villages, save unique villages, and direct the systematic growth of rural areas.

Conclusion

This study analyzed the LU change characteristics in the Weigan–Kuqa River Delta Oasis from 2010 to 2022, employing the OPGD to identify the driving factors behind these changes. Additionally, the MCCA was used to forecast future LU development scenarios. Finally, we elucidated the spatial characteristics of all LU categories in different scenarios through Moran’s I index and hotspot analysis. The main conclusions are summarized as follows:

1. The region experienced an annual increase in construction land, arable land, and garden land from 2010 to 2022, while unused land and forest–grassland decreased overall, and water bodies remained relatively unchanged. The comprehensive dynamic degree exceeded 1% each time, indicating a relatively fast rate of LU transformation. There is a tendency for conversions from arable land to garden land, arable land to forest–grassland, garden land to construction land, and forest–grassland to water bodies. The center of gravity for arable land and unused land shifted northwestward, while garden land and water bodies moved southeast, and forest–grassland and construction land shifted southwest.

2. The analysis of drivers revealed that LU transformation is influenced by both social economy and natural environmental variables, with population density being the primary influencing factor and soil type having a strong impact. The synergistic effects between population density, soil type, annual temperature, and elevation are particularly significant. Roads and transportation infrastructure have an increasing impact on LU change.

3. Future predictions indicate that the extent of arable land, garden land, and construction land will expand compared to the previous period, while unused land and forest–grassland will shrink. All LU types maintain historical transformation trends under the NDS; arable land increases most obviously under the CDS; construction land expands significantly in clusters under the EDS; and forest–grassland and water bodies are well protected under the EPS. Significant spatial clustering effects are observed across all scenarios, with forest–grassland and unused land showing the highest clustering.

Given that the Weigan–Kuqa River Delta Oasis is located in a fragile ecological environment in an arid region, and as part of the comprehensive development zone in the eastern cities of Aksu, research on the spatio–temporal patterns, driving mechanisms, and future predictions of LU can provide valuable insights for oasis land rehabilitation, ecological restoration, and sustainable development.

Author contributions

Conceptualization, Software, Methodology, Writing–original draft. B.A.; Data curation, Supervision, Writing–review & editing. X.W.; Investigation. X.H. All authors reviewed the manuscript.

Funding

This study was supported by the Natural Science Foundation of Xinjiang Uygur Autonomous Region, China (2023D01A44), the National Natural Science Foundation of China (41561051), and the Graduate Research Innovation Project of Xinjiang Normal University (XSY202301002).

Data availability

The data used to support the findings of this study are included within the article.

Declarations

Competing interests

The authors declare no competing interests.