Exploring the Determinants of the Urban–Rural Construction Land Transition in the Yellow River Basin of China Based on Machine Learning

Abstract

Understanding the determinants of urban–rural construction land transition is necessary for improving regional human–land relationships. This study analysed the spatiotemporal pattern of urban–rural construction land transition at the grid scale in the Yellow River Basin (YRB) of China during 2000–2020 by bivariate spatial autocorrelation analysis and further explored its determinants based on a machine learning method, the gradient boosted decision tree (GBDT) model. The results showed that both urban construction land (UCL) and rural residential land (RRL) increased, with an annual growth amount of UCL three times that of RRL, and the proportion of UCL (LUUR) remained stable after 2015. The determinants of UCL, RRL, and LUUR varied. The UCL mainly depended on socioeconomic factors, with their contribution exceeding 50%, while the RRL transition was mainly determined by physical geographic factors, with their contribution decreasing from 67.6% in 2000 to 59.7% in 2020. The LUUR was influenced by both socioeconomic and physical geographic factors, with the relative importance of socioeconomic factors increasing over the years. Meanwhile, the impacts of different determinants were nonlinear with a threshold effect. In the future, optimizing the distribution of urban–rural construction land and rationally adjusting its structure will be necessary for promoting urban–rural sustainability in the YRB.

1. Introduction

The urban–rural construction land transition is an important manifestation of the urban–rural transition [1], which refers to the transformation of the urban–rural construction land form driven by social and economic transformation [2]. Since the reform and opening up in China, a large rural population has migrated to cities, resulting in China’s urbanization rate increasing from 17.9% in 1978 to 63.9% in 2020. During the urbanization process, China has changed from a traditional rural society to a modern urban–rural society [3]. At the same time, urban construction land (UCL), as the main space for housing, production, and leisure, has expanded from 6720 km² in 1981 to 58,353 km² in 2020 due to urban population growth and economic development [4]. It is also worth noting that when the rural population decreased, the rural residential land (RRL) in China did not decrease but increased, with an annual growth rate of 0.12% during 1996–2007 [5]. The expansion of both UCL and RRL can lead to a series of problems, such as ecological damage [6], rural hollowing out [7], land use conflict [8], and food crises [9]. Thus, controlling the rational urban–rural construction land scale and increasing its utilization efficiency have become vital issues for urban–rural sustainable development.

The theory of land use transition is of great significance for understanding urban and rural land use change. It originated from forestland transition and was first proposed by Grainger [10]. Grainger believed that land use transitions referred to the change in land use form in the process of social and economic development during a specific period [11]. Thus, the land use morphology is a quite important concept for the theory of land use transition. Initially, the scholars mainly focused on the structure of land use corresponding to a particular stage of socioeconomic development, especially the spatial morphology and the proportion of a certain land use type. Then, Long and Li argued that the land use morphology should include both dominant and recessive morphology. Specifically, dominant morphology referred to the quantity or spatial structure of land use types, while recessive morphology included the quality, land rights, operation ways, and input and output capacity of the land, which were mainly obtained through surveys, analysis, and tests [12]. Accordingly, the land use transitions were driven by socioeconomic change and innovation, from one dominant and recessive morphology of land use to another morphology of land use, which usually corresponded to the transformation stage of socioeconomic development [13]. Currently, it has become a consensus that land use transition would be a new approach for comprehensive land use/cover change (LUCC) research. With the rapid development of remote sensing and GIS technologies, the dynamic processes of land use change can be monitored and simulated, which has been a key issue of global change research. Many studies have paid attention to the spatial and temporal characteristics of land use change at different spatial scales, including the temporal change of arable land [14,15], the expansion pattern of urban construction land [16,17], the temporal-spatial change of rural residential land [18,19], the degradation and restoration of forest and grassland [20,21], etc. Accordingly, the research about land use change was primarily concerned with the process of quantitative change in different land use types, which focused on the dominant morphology of land use, while land use transition research emphasized the turning point of land use attribute change during the continuous evolution of land use change, which focused on both the dominant and recessive morphology of land use and could be informative for understanding local socioeconomic and ecological development [22]. For the driving mechanisms of land use transitions, Lambin and Meyfroidt divided the influencing factors into internal socioecological feedback and external socioeconomic change. The former is due to the reduction of goods and services provided by the natural ecosystem, resulting in negative socioecological feedback, which is mostly endogenous and local, and the latter is mostly from higher organizational levels, adjacent regions, or local innovations. For instance, socioecological feedback can explain the slowing deforestation rate and forest cover stabilization, and socioeconomic factors explain reforestation and forest transition [23]. According to the urbanization stage theory of Northam, the urbanization process can be divided into three phases, initial stage, acceleration stage, and saturation stage [24], and the transition of the UCL also follows the trend of the urbanization process. Meanwhile, the expansion location of urban land use follows the competitive rent model, which emphasizes the distance to the urban centre [25]. The transition of RRL was manifested in the changes in number, spatial form, and structure, which was in accordance with the rural employment structure, lifestyle, and family structural change. In particular, the evolution of hollow villages has two possible modes, including the “cyclical evolution model” and the “vanishing evolution model” [26]. Thus, studies on the urban–rural construction land transition in China cannot be separated from its socioeconomic transformation.

It is worth noting that the driving mechanism of different influencing factors on urban–rural construction land transition varied greatly. For the physical geographic factors, they were the basic support for construction land and determined the initial morphology of urban–rural construction land [27]. In particular, the elevation, slope, and ecological condition can influence the capacity of population and suitability of construction land, and further affected the site selection of urban and rural settlements, which tended to locate in the areas with low elevation and slope [28]. For the location conditions and socio-economic factors, they were also key influencing factors for urban–rural construction land transition, the role of which became more important than physical geographic factors with the transformation of urban–rural construction land. In particular, the location conditions can influence the land use type, intensity, and value, etc. For instance, the transport development can improve the regional accessibility and location advantage, which further leads to the redistribution of urban–rural construction land. A research about the Huang-Huai-Hai agricultural area of China has found that the possibility of cultivated land transforming into urban–rural construction land was higher in the areas with superior location conditions, and the phenomenon of rural settlement expansion along roads was very common. The socioeconomic factors can also have drastic impact on urban–rural construction land transition and determined the direction, structure, scale, and layout of urban–rural construction land [29]. In China, during the process of rapid urbanization, the urban construction land transition was mainly driven by the flow and interaction of social production factors such as population and capital. The land use demand from population and economic growth was the direct driving factor for urban construction land expansion [30]. Meanwhile, the urbanization process may also have a huge impact on rural construction land. An empirical study about the transect along the Yangtze River found out that the proportion of rural residential land area in the total construction land decreases with the urbanization process [31]. In addition, in the process of urbanization, as a large rural population migrated to the cities, rural hollowing has become more serious, and the function of rural construction land has changed from adapting to agriculture to improving the living environment [32]. There have been many studies on urban–rural construction land transitions. For the UCL, its expansion has received great attention [33,34], which can be described as the ‘diffusion–coalescence’ phase transition and divided into three types, i.e., infilling, edge expansion, and spontaneous growth [35]. Meanwhile, the expansion of different types of UCL, such as industrial, residential, public service, and commercial land, varied greatly in terms of influencing factors [36]. For RRL, its quantitative change [37,38,39], spatial distribution [40], and utilization efficiency [41] have received great attention. However, studies of urban–rural construction land have not received much attention until recent years. For instance, Long et al. argued that with the development of society and the economy, the proportion of rural residential land in the total new construction land has gradually decreased [42]. Nourqolipour et al. predicted the effects of urban development on the urban–rural construction land transition [43]. Zhu et al. adopted the structure index of urban and rural construction land to reflect the urban–rural construction land transition [44]. Considering the different land use intensity, land property rights, and land value for UCL and RRL, it is necessary to take the structure of urban–rural construction land into consideration and explore the different determinants of UCL and RRL.

In terms of research spatial scale, most existing studies about the urban–rural construction land transition have focused on the national, provincial, or city level, and few have studied river basins [45]. For instance, Chen et al. focused on the national level of China and found that the increase in the urban built-up area was faster than that of the urban population, especially after 2000 [46]. Qu et al. took Shandong Province as a research area and found that urban construction land and rural residential land use generally evolve along a coordinated track of change, with fluctuation, resulting in insufficient elasticity and structural transformation of urban and rural construction land [47]. Liu et al. studied the urban construction land expansion in Beijing and found that 80% of its urban growth was at the expense of rural settlements and cultivated land, and the change curve of rural nonagriculturalization intensity along the urban–rural corridor had an inverted “U” shape [48]. However, the river basin is an important region for interpreting human–land relationships but lacks extensive related studies. For instance, the Yellow River Basin (YRB) in China has relatively low urban–rural construction land utilization efficiency [49], varies greatly in its socioeconomic development and land use patterns, and has been one of the regions with the tensest human–land contradiction in China. In 2019, the ecological protection and high-quality development of the YRB was set as a major national strategy of China. The study of the urban–rural construction land transition characteristics and driving mechanisms in the YRB may provide references for its high-quality development. Meanwhile, most existing studies have been based on administrative districts as research units, such as cities, counties or districts, while grid-based studies have been quite limited [44,50], which can provide a more accurate reflection of urban–rural construction land transitions. Furthermore, when detecting the driving factors of urban–rural construction land, most studies have adopted multiple linear regression models [51,52], geographic weighted regression (GWR) models [28], geographic detectors [40], or logistic models [53]. For instance, Huang et al. adopted regression analysis to analyse the driving mechanisms of urban construction land expansion at the prefectural level and found that marketization was a more important factor for overall urban land expansion, while globalization was a more important factor for industrial land expansion [54]. Zhang and Su adopted multiple linear regression to identify the determinants of urban expansion in metropolitan regions and found that economic growth, industrial development, and economic structural transformation were the most important factors [52]. Long et al. used the spatial econometric model on the Huang-Huai-Hai Plain and found that there was a significant negative correlation between the change in per capita construction land area and per capita GDP, which proved that there was a waste of land resources in this region [9]. Maimaitijiang et al. adopted multiple linear regression and GWR models to study the St. Louis Metropolitan Statistical Area and found that urban expansion and population change were generally positively correlated during 1970–2010, but there was a nonlinearity in different periods, which was negative in 1970–1990 and positive in 1990–2010 [55]. In recent years, machine learning has developed and been applied widely to different study areas. However, for the driving factors of urban–rural construction land, the adoption of machine learning has been quite limited, which can reveal the relative importance of different driving factors and the nonlinear relationships between driving factors. To fill the above gaps, this study focused on the YRB and aimed to analyse the spatiotemporal pattern of the urban–rural construction land transition in terms of total amount of UCL and RRL and the proportion of UCL in the total urban–rural construction land area (LUUR) based on the grid level; to apply one of the machine learning methods, the gradient boosting decision trees (GBDT) model; to identify the relative importance of different driving factors of the transition of UCL, RRL, and LUUR; to further explore the nonlinear and threshold impact of influencing factors on UCL, RRL, and LUUR in both 2000 and 2020; and to further discuss the mechanisms of the urban–rural construction land transition in the YRB.

The structure of this paper is as follows. Section 2 introduces the study area, data sources and methodology. Section 3 analyses the spatiotemporal pattern of urban–rural construction land in the YRB and its influencing factors based on machine learning. Section 4 discusses the driving mechanisms and policy implications. Section 5 summarizes the main conclusions.

2. Data and Methodology

2.1. Study Area

The Yellow River Basin (YRB) is located at 32°~42° N and 96°~119° E. It is an important economic zone in China and plays a significant role in China’s economic and social development. The Yellow River originates from the Bayan Har Mountains on the Qinghai–Tibet Plateau and flows across nine provinces in China, namely, Qinghai, Gansu, Ningxia, Inner Mongolia, Shanxi, Shaanxi, Henan, and Shandong, with a total length of 5464 km (Figure 1). The population and GDP of the nine provinces in the Yellow River Basin accounted for approximately 30% and 25% of the national total, respectively. In both 2000 and 2020, the top three dominant land use types in the YRB were grassland, cultivated land and forestland. However, when comparing the average annual change rate of different land use types, the growth rate of urban–rural construction land was the highest. Due to its relatively slow economic development and fragile ecological environment, to alleviate the conflict between humans and land, it is necessary to explore the characteristics and driving mechanisms of the urban–rural construction land transition in the YRB.

2.2. Research Framework

This paper constructed a research framework for studying the urban–rural construction land transition in the YRB from 2000 to 2020. As Figure 2 shows, the data processing consisted of the following four steps.

Step 1: Based on the land use data, the “Create Fishnet” and “Tabulate Area” tools of ArcGIS Pro were used to calculate the UCL and RRL at the grid scale of the YRB in 2000, 2005, 2010, 2015, and 2020. Due to the large area of the YRB, to reduce the influence of the grid amount on the computer processing speed, a grid scale of 5 × 5 km was selected as the spatial analysis unit, and the structural transition index of urban and rural construction land (LUUR) was constructed to reflect the evolution of the urban–rural construction land structure. In particular, the “Create Fishnet” tool was used to divide the whole research area into 5 × 5 km grids, and then the “Tabulate Area” tool was used to calculate the area of different land-use type in each grid based on the 30 m land cover data. The area of UCL referred to the total area of urban built-up land, factory, and mining land, as well as transportation land in each grid, and the area of RRL referred to the area of rural residential land in each grid.

Step 2: According to the urban–rural construction land area and its structural transition index (LUUR) in each grid, both the temporal evolution and spatial pattern of UCL, RRL, and LUUR were analysed in different periods from 2000 to 2020.

Step 3: A bivariate spatial autocorrelation model was adopted to analyse the global spatial autocorrelation and local spatial autocorrelation changes in both the UCL and RRL.

Step 4: The GBDT model was adopted to reveal the relative importance and nonlinear impacts of different driving factors on the urban–rural construction land transition, i.e., UCL, RRL, and LUUR, and analysed the changes from 2000 to 2020.

2.3. Data Sources

Four types of data, including land use data, physical geographic data, transport network data, and socioeconomic data, were collected in the YRB for this study. The land use data in 2000, 2005, 2010, 2015, and 2020 were obtained from the 30-metre resolution national land use data provided by the Resource and Environmental Science Data Center (RESDC) of the Chinese Academy of Sciences. Physical geographic data refer to elevation and slope, which were extracted from DEM data and were also downloaded from the RESDC. The transport network dataset mainly includes the locations of major railways, highways, provincial capitals, and prefecture-level cities, which were obtained from the National Geographic Information Center of China. Socioeconomic data mainly refer to the population and GDP data in 1 km resolution grids and were provided by the RESDC.

2.4. Data Analysis

2.4.1. Structural Transition Characteristics of Urban and Rural Construction Land

This study was based on a 5 × 5 km grid scale and superimposed the land use type map in the YRB onto the 5 km grid to complete the gridding of the whole region. LUUR represents the proportion of UCL in the total urban–rural construction land area and reflects its transition level from the perspective of the urban and rural construction land utilization structure. The calculation formula is as follows:

$${\mathrm{LUUR}}_{\mathrm{i}}=\frac{{UCL}_i}{{UCL}_i+{RRL}_i}\times 100\%$$

(1)

where LUUR_i is the proportion of UCL in total urban–rural construction land in grid i; UCL_i is the urban construction land area in grid i; and RRL_i is the rural residential land area in grid i.

2.4.2. Average Annual Change Amount

The average annual change amount was used to reflect the changes in both the UCL and RRL in the YRB in different periods during 2000–2020. The formulas used are as follows:

$${\mathrm{L}}_{\mathrm{u}}=\frac{{\mathrm{UCL}}_{\mathrm{b}}-{\mathrm{UCL}}_{\mathrm{a}}}{\mathrm{b}-\mathrm{a}}$$

(2)

$${\mathrm{L}}_{\mathrm{v}}=\frac{{\mathrm{RRL}}_{\mathrm{b}}-{\mathrm{RRL}}_{\mathrm{a}}}{\mathrm{b}-\mathrm{a}}$$

(3)

where L_u and L_v refer to the annual change amounts of UCL and RRL, respectively; UCL_a and UCL_b refer to the urban construction land area in years a and b, respectively; and RRL_a and RRL_b refer to the rural residential land area in years a and b, respectively.

2.4.3. Bivariate Spatial Autocorrelation

Bivariate spatial autocorrelation can characterize the spatial correlation of different geographic elements [56,57]. This study adopted bivariate spatial autocorrelation to measure the spatial coupling relationship between UCL and RRL.

First, the bivariate global spatial autocorrelation (Equation (4)) was used to verify the spatial correlation and heterogeneity of UCL and RRL.

$$\mathrm{I}=\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{{\displaystyle \sum }}_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{W}}_{\mathrm{ij}}\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)\left({\mathrm{y}}_{\mathrm{j}}-\overline{\mathrm{y}}\right)}{{\mathrm{S}}^2{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{{\displaystyle \sum }}_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{W}}_{\mathrm{ij}}}$$

(4)

where I is the bivariate global spatial autocorrelation index, that is, the correlation between the spatial distribution of variables x and y in general; n is the total number of spatial units; W_ij is the spatial weight matrix; x_i and y_i are the observed values of variables x and y in spatial unit j, respectively; $\overline{\mathrm{x}}$ and $\overline{\mathrm{y}}$ are the mean values of variables x and y, respectively; and S² is the variance of all samples.

Since the result of global spatial autocorrelation was only one single value, ignoring the instability of the spatial relationship between UCL and RRL at different locations, local spatial autocorrelation was further used to accurately identify their agglomeration and differentiation characteristics. The formula used is shown as follows:

$${\mathrm{I}}_{\mathrm{i}}={\mathrm{z}}_{\mathrm{i}}{\displaystyle \sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{ij}}{\mathrm{z}}_{\mathrm{j}}$$

(5)

where I_i is the bivariate local spatial autocorrelation index and z_i and z_i are the variance normalized values of the observed value in spatial units i and j, respectively. Based on the value of I_i, four clustering patterns can be identified, i.e., “high–high” clustering, “low–low” clustering, “low–high” clustering and “high–low” clustering.

2.4.4. GBDT Model

The gradient boosted decision tree (GBDT) model represents a machine learning method. Compared with most “black box” machine learning algorithms, the model based on decision tree integration can fit the influence of independent variables on the dependent variable in different ranges instead of generating fixed coefficients, which improves the interpretability of the model. Compared with the traditional multiple linear regression, the GBDT model does not need to follow any assumptions and can reduce the influence of outliers. Studies have shown that the GBDT model is superior to regression models [58], random forest models [59], and neural networks [60]. The GBDT model can handle different types of independent variables, identify and rank the influence of each variable on the dependent variable, and output a potential nonlinear relationship graph between the independent variable and dependent variable.

The GBDT model can be expressed as:

$${\mathrm{F}}_{\mathrm{M}}\left(\mathrm{x}\right)={\displaystyle \sum }_{\mathrm{m}=1}^{\mathrm{M}}\mathrm{T}\left(\mathrm{x};{\mathsf{\theta }}_{\mathrm{m}}\right)$$

(6)

where $\mathrm{T}\left(\mathrm{x};{\mathsf{\theta }}_{\mathrm{m}}\right)$ is a decision tree, ${\mathsf{\theta }}_{\mathrm{m}}$ is the parameter of the tree, and M is the number of trees. ${\mathrm{F}}_{\mathrm{M}}\left(\mathrm{x}\right)$ is linearly summed by multiple weak classifiers $\mathrm{T}\left(\mathrm{x};{\mathsf{\theta }}_{\mathrm{m}}\right)$; the loss of $\mathrm{T}\left(\mathrm{x};{\mathsf{\theta }}_{\mathrm{m}}\right)$ is expressed as L(y,${\mathrm{F}}_{\mathrm{M}}\left(\mathrm{x}\right)$), and the loss function in the GBDT model is the squared error function. The current decision tree is denoted as T_m−1(x), and the GBDT model determines the parameter ${\hat{\mathsf{\theta }}}_{\mathrm{m}}$ of the next decision tree by minimizing the loss function.

$${\hat{\mathsf{\theta }}}_{\mathrm{m}}=\mathrm{argmin}{\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{N}}\mathrm{L}\left[{\mathrm{y}}_{\mathrm{i}};{\mathrm{T}}_{\mathrm{m}-1}\left({\mathrm{x}}_{\mathrm{i}}\right)+\mathrm{T}\left(\mathrm{x};{\mathsf{\theta }}_{\mathrm{m}}\right)\right]$$

(7)

The GBDT model also has certain limitations. For instance, the local correlogram generated by the model may produce abnormal noise values in an interval with few samples, which can weaken the unreliable estimation interval according to the distribution of samples. The GBDT model cannot test whether the difference between the relative importance of the respective variables is statistically significant. Nevertheless, GBDT is still better than most models that assume linearity by default or any predefined relationship.

This study applied the GBDT model to reveal the dynamic impact of each driving factor on the expansion of urban and rural construction land. Land use transitions may be caused by the interactions of natural factors and socioeconomic factors. Both UCL and RRL can be affected by various factors, and ten indicators were selected, including elevation, slope, and distance from major rivers in regard to natural environmental conditions; distance from provincial capitals, distance from prefecture-level cities, distance to the main railway, and distance to main roads in regard to location conditions; and GDP, per capita GDP, and population density in regard to socioeconomic conditions (

Table 1. Influencing factors of urban–rural construction land.

Category	Variable Code	Variable Name (Unit)	Definition	2000–2020
				Mean	SD
Physical geographic factors	Elevation	Elevation (m)	Average of different elevations within the grid	1219	785.1
	Slope	Slope (°)	Average of different slopes within the grid	2.2	1.9
	DisRV	Distance from major rivers (km)	Distance to the nearest major river	7.9	10.6
Location conditions	DisPC	Distance from the provincial capital city (km)	Distance to the nearest provincial capital city	137.8	75.3
	DisRC	Distance from the prefecture-level city (km)	Distance to the nearest prefecture-level city	48.8	30.5
	DisRW	Distance from major railroads (km)	Distance to the nearest major railroad	45.4	54.7
	DisHW	Distance from major roads (km)	Distance to the nearest major highway	16.5	19.8
Socioeconomic conditions	GDP	GDP (CNY 104 yuan)	GDP in 2000 or 2020, based on 2000 constant price	8373	30,911
	PGDP	Per capita GDP (CNY 104 yuan)	Per capita GDP in 2000 or 2020, based on 2000 constant price	1.0	1.1
	PD	Population density (person/km²)	Population density in 2000 or 2020	322	632.2

). The collinearity among the influencing variables were tested by estimating their VIFs, all of which were less than 10, except that the VIFs of population density (PD) were slightly larger than 10 in the model of UCL and LUUR in 2020. Thus, low colinearity existed among the influencing variables of urban–rural construction land. Meanwhile, based on existing literature, the GBDT model was regarded as a method that can deal with multicollinearity very well [61,62,63]. Its algorithm can take the interaction between different influencing variables into account and accurately evaluate the effects of different influencing variables that may have collinearity on dependent variables. Accordingly, the model estimation results in this study should be reliable.

The dataset was split into two different subsets containing 50% of the data. The final model was set to a maximum of 10,000 trees, the learning rate was kept at 0.001, the three-way interaction was selected, and the developed GBDT model adopted a triple cross-validation procedure.

3. Results

3.1. Spatiotemporal Pattern of the Urban–Rural Construction Land Transition

3.1.1. Temporal Change in the Urban–Rural Construction Land Transition

Figure 3a shows the temporal change in the area of UCL and RRL in the YRB of China. From 2000 to 2020, the area of UCL increased from 3954 km² to 13,016 km², with an annual average growth rate of 6.1%, and the area of RRL increased from 15,000 km² to 18,000 km², with an annual average growth rate of 0.9%. The proportion of UCL in total urban–rural construction land changed from 21% in 2000 to 42% in 2020. The expansion of UCL experienced a slow period from 2000 to 2005, a rapid period from 2005 to 2015, and then entered into a stable period after 2015, whereas the expansion of RRL maintained a relatively steady growth rate during the study period.

Figure 3b–d further shows the urban–rural construction land change in the upper, middle and lower reaches of the YRB. For the change in UCL, the upper, middle and lower reaches all experienced a trend from slow expansion to rapid expansion and then to steady expansion, with the lower reach entering a period of stable growth earlier (in 2010) than the upper and middle reaches. For the change in RRL, the upper reach showed a trend of slow expansion to substantial reduction and then to rapid growth during the study period. The decrease in RRL in 2005–2010 was probably due to the implementation of ecological migration in Bayannur city in Inner Mongolia and other places in the upper reach [64]. The middle reach had a trend of slow expansion to fast expansion and then to slow expansion. Compared with the UCL, the RRL in the middle reach entered the slow expansion stage earlier. The lower reach of the YRB had a trend from rapid expansion to slight decrease, then to slow expansion and finally rapid expansion. The slight decrease in RRL was mainly due to the policy of the urban and rural construction land “increase- and decrease-linked” policy [65], in which the government promoted combining villages and residential areas in the downstream areas of the YRB [4,66].

3.1.2. Spatial Pattern of Urban–Rural Construction Land Transition

Figure 4 reflects the spatial pattern of urban–rural construction land based on the 5 km grid in the YRB of China. As shown in Figure 4a, compared with 2000, the UCL increased in 2020, which was driven by urbanization and industrialization. The newly added UCL was mostly distributed in urban agglomerations, with a more developed economy and greater demand for UCL. Figure 4b shows the spatial pattern of RRL in the YRB of China. Compared with 2000, the newly added RRL in 2020 was mostly distributed around the original rural settlement concentration areas, mainly on the Hetao Plain, Weihe Plain, Fenhe Valley, and North China Plain, which are important agricultural areas with superior natural conditions.

In terms of the average annual growth of UCL and RRL in the YRB of China during 2000–2020, there was great spatial disparity (Figure 5). The areas with high average annual growth amounts of UCL (>0.3 km²) were mainly distributed in the city centres and were driven by economic development and population increases [67]. The areas surrounding city centres had moderate annual growth amounts of UCL (0.0~0.1 km²). For the annual average growth of RRL, most areas were in a range of 0.0~0.1 km², indicating that the RRL generally grew at a relatively low rate due to a large rural population moving to the cities.

3.1.3. Structural Change in Urban–Rural Construction Land

The proportion of UCL in total urban–rural construction land increased significantly from 2000 to 2020 in the YRB of China (Figure 6), with great disparity at the grid level. In 2000, the grids with a high proportion of UCL (>50%) were concentrated in Dongying city in Shandong Province, Taiyuan city in Shanxi Province, and Wuhai city in Inner Mongolia, the distribution of which was quite scattered. In 2020, some concentrated contiguous areas began to form, with concentrations in Yulin city in Shannxi-Ordos city in Inner Mongolia and Wuhai city in Inner Mongolia-Shizuishan city in Ningxia. Additionally, there were some scattered distributions around Dongying city in Shandong Province, Zhengzhou city in Henan Province, Xi’an city in Shaanxi Province, Taiyuan city in Shanxi Province, Yinchuan city in Ningxia Province, etc. Meanwhile, mixed urban and rural construction land at the grid level became more common in 2020 and was mainly distributed in the middle and lower reaches of the YRB, around the major cities and along the transport corridor, some of which were similar to those in the Desakota region [68].

3.2. Bivariate Spatial Autocorrelation between UCL and RCL

To reveal the spatial coupling relationship between UCL and RRL in the YRB, their spatial distribution relationship was further analysed by bivariate spatial autocorrelation. In 2000, the Moran’s I of UCL and RRL was 0.136, and there was a positive correlation between them, which suggested a relatively concentrated distribution of urban and rural construction land. As Figure 7 shows, “low–high” clustering was the most dominant type, followed by “high–high” clustering. The former was mainly distributed in Inner Mongolia, the Weihe River Plain, and the Fen River Valley, as well as the lower reaches where the area of UCL was small and the RRL was large. The latter was mainly scattered in the middle and lower reaches of the Yellow River, such as Xi’an city in Shaanxi Province, Tai’an city in Shandong Province, Yuncheng city in Shanxi Province, and Luoyang city in Henan Province, with large amounts of both UCL and RRL.

In 2020, the Moran’s I of UCL and RRL was 0.199, with the trend of agglomeration of urban–rural construction land strengthened. The “low–high” clustering was still the most dominant type, followed by the “high–high” clustering. Compared with 2000, the number of grids representing “low–high” types decreased, and the number of “high–high” types increased, mainly distributed in Xi’an and Xianyang in Shaanxi Province, Zhengzhou and Luoyang in Henan Province, Jinan and Tai’an in Shandong Province, and Hohhot City in Inner Mongolia, where the UCL and RRL increased simultaneously.

3.3. Determinant of the Urban–Rural Construction Land Transition

3.3.1. Relative Importance of Influencing Factors

This study used 50% of the data as the training data to establish a GBDT prediction model and test the corresponding robustness of the model estimation by calculating the indicators of RMSE, MAE, and R². The RMSE and MAE can reflect the gap between the prediction value and the real value, which suggested that all of the models had relatively small differences between the prediction value and the real value. Meanwhile, the R² of all models was larger than 0.3, and the R² of UCL reached 0.63 in 2000 and 0.50 in 2020 (

Table 2. Relative importance of influencing factors of urban–rural construction land.

Year	Category	Variable	UCL			RRL			LUUR
			Rank	Relative Importance (%)		Rank	Relative Importance (%)		Rank	Relative Importance (%)
2000	Physical geographic factors	Elevation	7	4.9	13.0	1	42.0	67.6	1	20.5	39.2
		Slope	6	4.9		2	22.6		5	10.5
		DisRV	8	3.3		6	2.9		6	8.2
	Location conditions	DisPC	4	6.7	34.6	4	6.6	10.5	7	7.4	30.1
		DisRC	2	23.7		8	1.3		3	13.4
		DisRW	9	2.9		7	1.6		9	5.2
		DisHW	10	1.3		10	1.0		10	4.1
	Socioeconomic conditions	GDP	1	29.5	52.4	3	15.0	21.9	8	5.8	30.7
		PGDP	5	5.2		9	1.1		2	13.7
		PD	3	17.8		5	5.9		4	11.1
	RMSE		2.01			0.80			29.02
	MAE		1.24			0.57			24.31
	R2		0.63			0.53			0.35
2020	Physical geographic factors	Elevation	5	6.5	19.1	1	37.6	59.7	1	27.0	38.5
		Slope	4	9.5		2	19.9		4	7.7
		DisRV	8	3.1		9	2.2		8	3.8
	Location conditions	DisPC	6	5.7	26.7	4	7.2	14.0	7	3.9	15.5
		DisRC	3	16.4		8	2.2		5	6.4
		DisRW	9	2.3		7	2.6		9	3.2
		DisHW	10	2.3		10	2.0		10	2.0
	Socioeconomic conditions	GDP	1	27.5	54.3	6	5.1	26.4	6	5.3	46.1
		PGDP	7	3.2		5	5.5		3	18.2
		PD	2	23.5		3	15.8		2	22.6
	RMSE		2.41			0.94			29.70
	MAE		1.44			0.64			24.97
	R2		0.50			0.49			0.35

). Thus, both the robustness and the accuracy of the models in this study were tested to be high, which was in accordance with other studies that argued that the GBDT model generally had high robustness [69]. Table 2 also summarizes the relative importance and ranking of all independent variables in predicting the UCL, RRL, and LUUR. It is worth noting that the total contribution of all independent variables was 100%.

For the influencing factors of UCL, socioeconomic factors contributed the most, followed by location factors and physical geographic factors in both 2000 and 2020. In particular, the GDP ranked first among the socioeconomic factors, indicating that economic development was the main driving force for urban construction land expansion. The more developed the economy was, the greater the proportion of nonagricultural industries, the greater the demand for industrial land and the greater the expansion of UCL [67]. The relative importance of population density ranked third and second in 2000 and 2020, respectively, indicating that population was also an important factor for the expansion of UCL. The higher the population density in urban areas was, the more demand for living and production land and the greater the expansion of UCL. The finding in the YRB was in accordance with other studies about the whole of China [70] and specific regions of China, including the Beijng–Tianjin–Hebei region [71], Yangtze River Delta [44], and Shandong province [47], all of which believed that urban construction land transition was mainly driven by urban–rural population and economic interaction.

For the influencing factors of RRL, the contribution of physical geographic factors was the largest, followed by socioeconomic factors and location factors. The elevation and slope in regard to physical geographic factors ranked first or second in both 2000 and 2020, which means that the distribution of RRL was closely related to the natural environment and probably because agricultural development was greatly affected by natural conditions. The finding was similar to the study about RRL transition in the transect of the Yangtze River, which found that the transition stage of RRL in each section of the transect along the Yangtze River coincided with the economic development level of each section in the transect [31]. However, it was different from the finding in the Huang-Huai-Hai agricultural area of China, which argued that location condition was dominant in the RRL transition, instead of physical geographic factors and socioeconomic factors [29].

For the influencing factors, LUUR, i.e., the proportion of UCL in total urban–rural construction land, the highest relative importance was the physical geographic factors, followed by the socioeconomic factors and then the location factors in 2000. In 2020, the highest relative importance was attributed to socioeconomic factors, followed by physical geographic factors and then location factors. Compared with 2000, it can be found that the relative importance of physical geographic factors decreased, which was consistent with existing research that believed that the influence of physical geographic factors on UCL expansion decreased over time [72], while the relative importance of socioeconomic factors increased. In particular, the ranking of GDP and population density greatly increased, indicating that socioeconomic factors became more important in the structural transition of urban–rural construction land. It means that social and technological improvement can reduce the restrictive effect of natural factors on the location of urban–rural construction land [73]. However, this study suggested that the role of location conditions was weaker than physical geographic factors and socioeconomic factors in urban–rural construction land transition in YRB. It was different from the findings in the Bohai Rim region that emphasized the determinant role of location conditions in urban–rural construction land transition [70]. The weaker role of location conditions in the YRB has indicated that the leading role of core cities and secondary central cities in the YRB was not as strong as the coastal developed regions in China.

3.3.2. Nonlinear Associations between UCL and Its Determinants

Due to the highest relative importance of socioeconomic factors to the change in UCL in both 2000 and 2020, this part only focused on the nonlinear association between UCL and socioeconomic factors (Figure 8). First, the association between GDP and UCL was positive and nonlinear. This means that when GDP grew, the UCL tended to expand [67,74]; however, when the GDP reached a certain threshold, the UCL would not further expand due to the limitation of natural conditions and increasing land use efficiency. The threshold value increased from 1250 million in 2000 to 3000 million in 2020. Meanwhile, within these thresholds, the association between GDP and UCL was also not linear, which was different from most existing studies with linear assumptions. In 2000, the impact of GDP on UCL could be divided into two stages with 500 million as the cut-off point, which means that within 500 million, when GDP grew, the growth rate of UCL was quite low, while when GDP was more than 500 million, the growth rate of UCL increased as GDP grew. In 2020, the impact of GDP on UCL could also be divided into two stages with 800 million as the cut-off point, which means that when GDP was lower than 800 million, the growth rate of UCL was high, and when GDP was larger than 800 million, the growth rate of UCL decreased as GDP grew.

Second, the association between per capita GDP and UCL was also positive and nonlinear. In 2000, when per capita GDP was within CNY 5000, it had little impact on UCL; when per capita GDP was over CNY 5000, it had a significant impact on UCL, which corresponded to the Chinese economy from primary product production Stage I to Stage II [75]. In 2020, when the per capita GDP was within CNY 10,000, it had little impact on UCL; when the per capita GDP was CNY 10,000–40,000, the UCL increased when per capita GDP grew. The influencing threshold of per capita GDP to the UCL was CNY 35,000 in 2000 and CNY 46,000 in 2020.

Third, the association between population density and UCL was also positive and nonlinear. The threshold of population density to the change in UCL was 4700 person/km² in 2000 and 5000 person/km² in 2020, which means that when the population density was higher than 4700 person/km² in 2000 and 5000 person/km² in 2020, the UCL would not have expanded when the population density further expanded due to limitations from natural conditions.

3.3.3. Nonlinear Associations between RRL and Its Determinants

Due to the highest relative importance of physical geographic factors and socioeconomic factors to the change in RRL in both 2000 and 2020, this part focused on the nonlinear association between RRL and physical geographic factors and socioeconomic factors (Figure 9).

In terms of physical geographic factors, the association between elevation and RRL was negative and nonlinear. In both 2000 and 2020, when the elevation was lower than 1000 m, the RRL tended to expand as the elevation decreased, and when the elevation was higher than 1000 m, the area of RRL tended to be small and stable, which emphasized that the plain areas were the priority areas for rural settlements [76]. The association between slope and RRL was also negative and nonlinear. When the slope increased, the RRL first decreased rapidly, then decreased slowly, and finally tended to be stable and remained at a low value.

In terms of socioeconomic factors, the association between GDP, per capita GDP, and population density and RRL was also nonlinear. First, in 2000, when GDP was lower than 25 million, the RRL increased rapidly as GDP grew; when GDP was between 25 and 600 million, the impact of GDP on RRL became small; and when GDP exceeded 600 million, the RRL tended to stabilize as GDP further grew. In 2020, 500 million was the cut-off point when analysing the association between GDP and RRL. When GDP was lower than 500 million, the RRL increased as GDP grew; when GDP exceeded 500 million, the RRL showed a decreasing trend, which means that high economic development may lead to urbanization transition and change the construction land from rural areas to urban areas. Second, the association between per capita GDP and RRL was negative, and the threshold value of per capita GDP to RRL was CNY 20,000 in 2000 and 2020, which means that when per capita GDP was higher than CNY 20,000 in 2000 and 2020, the RRL remained stable at a low level. Third, in 2000, when the population density was lower than 1000 person/km², the RRL increased rapidly as the population density increased, and when the population density was higher than 1000 person/km², the RRL had a fluctuating trend. In 2020, when the population density was lower than 2500 person/km², the RRL had a fluctuating increasing trend, and when the population density was higher than 2500 person/km², the RRL decreased due to the urbanization transition. The threshold value of population density to RRL was 4000 person/km² in 2000 and 2020. In sum, the nonlinear association between RRL and socioeconomic factors was evident in the YRB, which was different from most existing studies that treated the relationship between the socioeconomic factors and RRL as linear and neglected their nonlinear and threshold effect.

3.3.4. Nonlinear Associations between LUUR and Its Determinants

Figure 10 shows the nonlinear association between LUUR and its determinants. First, for the elevation, the growth rate of LUUR was low as the elevation increased within 1000 m. However, when the elevation reached between 1000 and 3000 m, the LUUR decreased in 2000 and remained stable in 2020 when the elevation increased. This means that during the past two decades, the increase in the urban construction land proportion was mainly in an area with elevations between 1000 and 3000 m. Second, the association between distances from prefecture-level cities and LUUR was negative, which was the same as the finding in Yangtze River Delta [44], and the threshold in the YRB was 7 km in 2000 and 10 km in 2020, which means that the influencing scope of prefecture-level city centres expanded, which was correlated with the expansion of urban construction land to its surrounding areas. Third, the association between GDP and LUUR was positive, and the threshold value of GDP to LUUR was 1100 million in 2000 and 2000 million in 2020, which means that when GDP was lower than the threshold value, the proportion of urban construction land increased when GDP grew. Fourth, the per capita GDP and LUUR were also positively correlated, and the threshold value was CNY 35,000 in 2000 and CNY 31,000 in 2020. Finally, in 2000, when the population density was within 600 persons/km², the proportion of urban construction land showed a downwards trend as population density grew, and when the population density was between 600 and 4700 persons/km², the proportion of urban construction land increased as population density increased; in 2020, when the population density was within 800 persons/km², the proportion of urban construction land showed a downwards trend as population density grew, and when the population density was between 800 and 4000 persons/km², the proportion of urban construction land increased as population density increased. This means that the areas with higher population density could directly affect the growth of urban construction land, resulting in a higher proportion of LUUR [77].

4. Discussion

4.1. Mechanisms of the Urban–Rural Construction Land Transition in the YRB

According to the empirical results of the YRB, the urban–rural construction land transition was influenced by multiple factors, including physical geographic factors, location condition factors, and socioeconomic factors, and the influencing factors varied by UCL and RRL. The finding is in accordance with studies in other regions of China, such as the Beijng–Tianjin–Hebei region [71], Yangtze River Delta [44], and Shandong province [47]. In the YRB, the UCL was mostly affected by socioeconomic factors, and economic development greatly promoted the transformation of nonurban land to urban construction land, thereby promoting the expansion of urban construction land [52]. During the period of 2000–2020, the GDP in the YRB increased by 855 billion, and the UCL increased by 9062 km². This study suggested that GDP ranked first in all influencing factors of UCL expansion, which emphasized the far-reaching influence of GDP on UCL transition. RRL was mostly affected by physical geographic factors, which would choose areas with superior natural environments in the early stage of formation, and the area of RRL was negatively correlated with elevation and slope, which reflected the restriction of the natural environment on the distribution of rural settlements. However, the contribution of physical geographic factors to RRL has decreased over the years. The finding was different from the research about the Huang-Huai-Hai agricultural area of China that emphasized the dominant role of location conditions [29]. The structural transition of urban–rural construction land was mainly influenced by both socioeconomic and physical geographic factors, with the relative importance of socioeconomic factors increasing from 30.7% in 2000 to 46.1% in 2020, indicating that with urban population growth and rural population migration to cities, the demand for urban construction land would inevitably have been strengthened, and the proportion of urban construction land would have increased. However, the relative importance of physical geographic factors to LUUR has decreased from 39.2% in 2000 to 38.5% in 2020. During the study period, the proportion of urban construction land within an altitude range of 1000–3000 m in the YRB increased significantly, mainly due to local industrialization development [78].

In sum, the construction land transition in the YRB can be regarded as the combined effects of socioeconomic change and socioecological feedback [79,80], which is driven by both the market and government. As Figure 11 shows, the pathway of socioeconomic change was manifested as the change in construction land due to socioeconomic development, including industrialization, urbanization, and globalization. Socioeconomic development can directly affect population migration and industrial structural changes, regional market participation, and urban–rural integration, and can indirectly lead to urban construction land expansion due to the production and living demand for land in cities. Meanwhile, due to the large rural population floating between cities and villages, even when they migrated to the cities, they still intended to keep rural Hukou and their rural landholdings, which caused the rural residential land expansion as well. However, the disorderly expansion pattern of urban–rural construction land may lead to low efficiency of construction land and to negative effects on the socioecological system, such as ecological damage and food crises, which will cause feedback to socioeconomic change and induce the government to adjust its policies. Policies such as new urbanization, linked urban and rural construction land increases and decreases, and three control lines, control of ecological protection boundaries, permanently protected farmland, and urban development boundaries, may promote the coordinated transition of urban–rural construction land. Meanwhile, socioeconomic development may also actively induce the transformation of planning policies towards intensive and economic land use. In summary, the transition mechanisms of urban–rural construction land can be reflected as the effects of socioeconomic change and socioecological feedback in the YRB of China.

4.2. Policy Implications of the Urban–Rural Construction Land Transition in the YRB

Under the urban–rural dual structure of China, the urban–rural construction land transition was the main manifestation of urban–rural transition. The quantity and proportion of UCL and RRL can reflect the transition characteristics of urban–rural construction land. The transition characteristics of urban–rural construction land can be seen as a reflection of social and economic development. In the YRB during 2000–2020, driven by rapid urbanization and industrialization, a large amount of the population migrated to the urban areas, which led to the rapid expansion of UCL, while the RRL did not decrease, but increased at a low growth rate, even when the rural population migrated to the cities. It can be explained by both the rural residential land institution in China and the hometown sentiments in Chinese traditional culture. Accordingly, the problems of rural hollowing out and rural land resource waste have been acute. Meanwhile, since the growth rate of UCL decreased after 2015, while the growth rate of RRL increased after 2015, the proportion of RRL in urban–rural construction land remained stable in the YRB after 2015. Since the expansion of RRL may lead to a food crisis and rural hollowing-out problems [9], the control of RRL is quite necessary, and the potential of hollow villages may be much higher than that of urban construction land. It is necessary to carry out a voluntary return mechanism for rural residential land in an orderly manner and gradually reduce land use in rural residential areas. At the same time, mining areas in rural areas need to return abandoned industrial and mining land to reduce the waste of construction land [81].

Meanwhile, although the upper, middle, and lower reaches of the YRB have shown similar urban–rural construction land transition stages in terms of the proportion of UCL, from fast expansion to slow expansion, the transition time was different for the upper, middle, and lower reaches. The upper and middle reaches turned to slow expansion in approximately 2015, and the lower reach turned to slow expansion in approximately 2010. In the future, when the urban–rural integrated development and rural revitalization strategy is further promoted, the exchange of urban and rural land, population, and capital factors will become smoother, and the urban–rural construction land transition will develop in a more balanced way, which requires urban–rural integrated planning to optimize the rational spatial distribution of urban–rural construction land and promote the adjustment of the urban–rural construction land structure.

Furthermore, grid-based analysis can help the YRB implement more accurate strategies at a finer resolution than traditional administrative unit-based analysis. For the locations with a great amount of urban–rural construction land expansion, it is necessary to increase its utilization efficiency by industrial upgrading, agricultural modernization and service industry development, which will eventually lead to regional high-quality development. Additionally, the linkage incentive system for urban and rural construction land and the transaction mechanisms for construction land in the whole river basin should be established so that the indicators of construction land can flow into areas with high development potential, especially areas with better locations, which will promote regional integrated development. With the facilitation of information technology and remote sensing technology, an intelligent management and monitoring platform for urban and rural construction land should be developed to improve the level of urban and rural spatial governance.

This study inevitably has a few limitations. For instance, specific urban–rural construction land was not analysed in detail, such as housing, industry, transportation, and green land. Meanwhile, the transition mechanisms of urban–rural construction land were closely related to the transition mechanisms of urban–rural development, which needs deep exploration in the future. Furthermore, the studies about recessive morphology of urban–rural construction land such as its quality, land rights, operation ways, input, and output are required in the future. Nevertheless, the spatiotemporal pattern of the urban–rural construction land transition was analysed based on the grid scale in the YRB, and the nonlinear relationship of different influencing factors on the urban–rural construction land transition was explored, which may provide some policy implications for land use and regional high-quality development in both the YRB in China and other watersheds in developing countries.

5. Conclusions

This study utilized the YRB in China as the research area and analysed the spatiotemporal characteristics of the urban–rural construction land transition, further exploring the relative importance and nonlinear impacts of various factors affecting the urban–rural construction land transition based on the GBDT model. The main conclusions are as follows.

First, during the past two decades, both urban and rural construction land increased in the YRB, and the annual average growth amount of urban construction land was almost three times that of rural construction land, with the proportion of urban construction land in total urban–rural construction land increasing rapidly before 2015 and then increasing slowly after 2015, which was in accordance with the regional urban–rural integrated development and rural revitalization strategy. In terms of the spatial distribution, compared with 2000, the newly added urban construction land was mainly distributed in the metropolitan regions, while the newly added rural residential land was mainly distributed around the original rural settlements and concentrated in the traditional agricultural regions.

Second, the urban and rural construction land change in the YRB was influenced by different driving factors, and the contribution of different influencing factors varied by UCL and RRL. For the transition of UCL, it was most affected by socioeconomic factors including GDP, per capita GDP, and population density, followed by location conditions and physical geographic factors. In particular, the economic growth represented by the GDP growth ranked first among all influencing factors of UCL. The transition of RRL was mainly determined by physical geographic factors, including elevation, slope, and distance to rivers, which means that the constraint of the natural environment was still large for the expansion of RRL in the YRB. The proportion of urban construction land was influenced mostly by physical geographic factors in 2000 and by socioeconomic factors in 2020, which means that socioeconomic development has played a more important role in the urban–rural construction land transition. It means that the improved social and technology factors have led to the urban–rural construction land transition gradually overcoming the limitations of natural geographical conditions in the YRB.

Third, the impact of different determinants on the urban–rural construction land transition was nonlinear with a threshold effect. For the transition of UCL, socioeconomic factors played the most dominant role and were positively related to UCL, with the threshold of GDP, per capita GDP, and population density increasing from 2000 to 2020, which suggested that the impact scale of socioeconomic factors on UCL was broadened. For the transition of RRL, the natural geographic factors played the most dominant role and were negatively related to RRL. For the LUUR, the elevation played the most dominant role, with a low growth rate of LUUR as the elevation increased within 1000 m, while the LUUR decreased in 2000 and remained stable in 2020 when the elevation increased between 1000 and 3000 m. This means that during the past two decades, the increase in the urban construction land proportion was mainly in areas with elevations between 1000 and 3000 m in the YRB. Meanwhile, from the influence of distance to prefectural cities on LUUR, the threshold increased from 7 km to 10 km during the study period, which means that the impact scale of cities became broader in the YRB.