1. Introduction
Urbanization is one of the main forms of land change worldwide since the 21st century [
1]. It has a significant impact on salient environmental issues such as the hydrological cycle [
2], natural dioxide emission [
3], air pollution [
4], soil degradation, and climate warming [
5]. At present, most parts of Europe and North America have completed rapid urbanization, but this process is still underway in other continents such as Asia and Africa [
6,
7,
8]. China, an Asian country with the largest population in the world, is currently experiencing large-scale population migration from rural to urban regions, although its urbanization rate has exceeded 60% in the year of 2020. This value in China will rise to 70% by 2050 and 80% by the end of this century according to predictions made by other authors [
9,
10,
11], implying a continuous and rapid urbanization in the future. In the process of urban expansion, different land-use structures and their changes within the city usually have different effects on natural-environment issues, such as the photosynthesis of vegetation, and the hydrological process of water, which jointly affect urban ecosystem services [
12,
13], and also the non-point rain scouring pollution and the cold/hot island effect in arid/humid regions from the impervious surface area [
14]. A timely update of the underlying land surface within the city can provide natural background materials for these environmental surveys as well as the description of spatiotemporal heterogeneity of the internal urban landscape. Beijing, as China’s political, economic, cultural, traffic, and talent center, undertakes many functions of the state [
15]. It has an obvious siphon effect for people across whole country due to its advantages in infrastructure, society, the natural environment, education, medical care, and relatively high wages [
15,
16]. The bio-geochemical and bio-geophysical effects of urban change in Beijing have been previous investigated by other authors in the literature [
17,
18,
19], and the survey conducted from the perspective of hierarchical urban-land evolution should be displayed. The hierarchical urban-land mapping strategy can provide a multi-dimensional and new perspective of urban land analysis, and thus, it has a high popularization value, i.e., the urban land can be used for an urbanization analysis, intra-urban land-structure change can be used to analyze the comfort of human settlements and ecosystem services, and urban-land component information at the sub-pixel scale can provide the density change per unit area. In particular, continuous and up-to-date hierarchical urban-land descriptions, such as timely, synergetic land mappings of urban expansion, structural change at the pixel scale, and component change at the sub-pixel scale, are still lacking.
High spatial-resolution remote sensing images, such as Google imagery, Quick Bird, and GaoFen images, were usually conducive to obtain the better spatial heterogeneity of land use classification [
20,
21]. However, the same object with different spectral characteristics and the same spectral characteristic with different objects usually made it difficult to achieve very good results in the automatic classification of high spatial resolution images [
20]. Another limitation was that it was difficult to achieve high spatial resolution images for long-time series, such as 40 years of study on a regional scale. Medium resolution satellites, such as Landsat images, feature suitable resolution, universal spectrum, free download, and good classification accuracy, and have been widely used in land classification at local, regional, and global scales [
20,
22]. Therefore, Landsat images were used for the time series of land use classification in the study area. For the methodological development of urban-land mapping from Landsat images, urban land use can be classified into three types (excluding water) using the vegetation–impervious surface area-soil (VIS) model, identifying areas of impervious surface area, vegetation cover, and bare soil [
23]. On this basis, the combination of a linear spectral mixture analysis model (LSMA) and VIS was applied to the mixed pixel decomposition in the humid urban regions [
24]. Additionally, in the arid urban regions, to solve the problem of a high amount of bare soil in urban regions, the synergetic method of LSMA and decision tree classifier can obtain good results [
25]. Among the processes of methodological development for urban land mapping, the least squares mixed pixel decomposition model replaced the LSMA due to the advantage of more accurate output results [
26]. Meanwhile, a novel urban impervious surface (UIS) index and an improved particle swarm optimization (PSO)-based UIS automatic threshold selection model (UISAT) were proposed to overcome the shortages when detecting UIS derived from multitemporal Landsat images [
27]. To further display the methodological evolution in the UIS mapping, a review of the methods and challenges of UIS detection from remote sensing images has been presented [
28]. The authors also displayed an improved model based on detection of UIS using multiple features extracted from reflective optics system imaging spectrometer data [
29]. With the application of big-data platform and the progress of remote sensing technology, large-scale urban land mapping has become increasingly popular, i.e., the mapping of urban form and functions for the continental US [
30], the 40-Year (1978–2017) human settlement changes in China [
31], and the mapping of global artificial impervious area [
9]. Therefore, urban land mapping has become an increasingly popular area in scientific research.
The driving forces of physical medium environments affecting urban land expansion include economic development, population size, arid and humid climate environment, topography, and water resources, etc. [
32,
33]. More of the literature has focused on urban land-use change from the perspective of socio-economic development and population mobility. For example, a differential comparison of urban expansion was performed in a total of nine mega cities, scattered across the coastal and inland regions of East Asia; and the economic development and population urbanization were found to have different effects on these cities [
34]. In addition to socio-economic factors, the driving force affecting urban land expansion from the perspective of naturally physical environments has often been concentrated in mountainous regions with undulating terrain [
32]. The trend and gradient of the canyon can significantly affect the pattern of urban expansion, taking into account the huge cost of leveling the top of the mountain for construction, as shown in a study in Lanzhou city, in a mountainous region of Southwest China [
32,
35]. However, the investigation of naturally physical conditions of urban expansion in relatively flat areas was often ignored. Beijing, the capital city of China, is a mega city with more than 21 million people [
36]. The appearance of the city has undoubtedly undergone earth-shaking changes over the past few decades [
37,
38]. Exploring the sub-pixel urban land component change and natural environmental elements, such as slope, aspect, and water resources services, may provide an authentic study of plain area. In particular, the distribution law of sub-pixel impervious surface area/vegetation cover with water resource service needs to be investigated, which will be revealed in this study.
In this study, a multi-scale urban land mapping methodology was established using a time series of environmental imagery and a remote-sensing technique to deal with the following objectives: (1) to display the newly spatiotemporal heterogeneity and the evolution of urban land, including the indicators of urban expansion, land use structure, and its component during the period of 1981–2021; (2) to compare the variables of urban land changes, especially the impervious surface area and vegetation cover, in the old urban land and the newly expanded urban land on the pixel and sub-pixel levels, and analyze the ecological landscape configuration characteristics in the urbanization process; and (3) to reveal the law of multi-scale urban land mappings and physical medium environments, such as the dominant direction and main slope range of impervious surface distribution, and the relationship between water resource services and different components of impervious surface area/vegetation cover. Then, we discussed new reports on multi-scale urban land evolution in the capital city of China during 1981–2021, the findings of the physical medium environments with multi-scale urban land changes, and the potential carbon-cycle effects in the study area. We sincerely expect that these new reports and findings will act as a reference for relevant research in other cities or regions.
2. Methods
2.1. Study Area
Beijing is the capital, megacity, political center, cultural center, international exchange center, and scientific and technological innovation center of China [
39]. It is located on the North China Plain, adjacent to Bohai bay. The Geographic location of Beijing has a longitude ranging from 115.35°E to 117.56°E and a latitude ranging from 39.36°N to 41.14°N (
Figure 1). Beijing has jurisdiction over 16 municipal districts, including Daxing, Fengtai, Shunyi, Xicheng, Haidian, Chaoyang, Tongzhou, Fangshan, Changping, Dongcheng, Mentougou, Yanqing, Shijingshan, Pinggu, Miyun, and Huairou. Its permanent resident population was 21.89 million in the year 2020, and the corresponding gross domestic product (GDP) amounted to 3610.26 billion yuan in the whole year, of which the proportion of Beijing’s first, second, and tertiary industries was 0.4%, 15.8%, and 83.8%, respectively.
From the perspective of natural environments, Beijing has a warm-temperate, semi-humid and semi-arid monsoon climate, featuring a hot and rainy in summer, a cold and dry in winter, and a shorter spring and autumn. The annual frost-free period is 180~200 days and annual average sunshine hours are between 2000 and 2800 h. Beijing’s natural rivers flow through five major water systems (i.e., Juma river, Yongding river, Beiyun river, Chaobai river, and Ji canal river) with the meandering direction from the mountains in the northwest to the southeast. Most of these rivers flow through the plain area and finally merge into the Bohai Sea in Haihe river. The terrain of Beijing is characterised as high in the northwest and low in the southeast, with an average altitude of 43.5-m. In terms of mountains, the western part of Beijing is the west mountain, belonging to the Taihang mountains; Jundu mountain is in the north and northeast, belonging to Yanshan mountains. The Dongling mountain has the highest peak, and is located in the Mentougou district, West Beijing, with an altitude of 2303-m; the lowest ground level is in the southeast boundary of Tongzhou district. The zonal vegetation types in Beijing are mainly warm temperate deciduous broad-leaved forest and warm coniferous forest. Most plain areas are covered by farmland and towns. The vegetation types of low mountains are Quercus variabilis forest, Oak forest, Pinus Tabulaeformis forest, and Platycladus orientalis, the middle mountains are dominated by liaodong oak forest and the high mountains consist of birch trees and mountain miscellaneous grass meadows.
2.2. Workflow
To summarize the main process of this study, we provided the workflow of the study area (
Figure 2). In this figure, we present the workflow for urban land development boundary production, the classification of urban land-use structures, and the component mapping of intra-urban lands, together. Additionally, the relationship between the hierarchical urban land mapping and the natural physical medium environments are also shown.
2.3. Data Collection and Preprocessing
Data collection in this study mainly included two aspects, namely, the remote sensing data and the basic geographic data. Remote sensing was performed using Landsat MSS/TM/OLI, high-resolution Google images, and the digital elevation model. The images were mainly used to obtain the development boundary of urban land and its internal land structure as well as land components; and the digital elevation model performed spatial operation on ArcGIS platform to generate the physical medium environments such as slope and aspect. Basic geographic data include administrative divisions, road-loop network, national land-use data (NLUD), and water-resource distribution. Among them, NLUD provided the original boundary for urban land that needs to be revised. A spatial map of water resources was applied to analyze the relationship between urban land components and water resource service radius. All data were preprocessed using the unified coordinates and projections to ensure the accuracy of spatial mapping and statistical analysis. Descriptions of the collected data are provided in the
Table 1 below.
2.4. Establishment of Hierarchical Urban Land Mappings
2.4.1. Urban Land Development Boundary Production
In this study, to identify the entirely urban land boundaries, we referred to the land-use classification system of Resource and Environmental Science and Data Center, Institute of Geographical Sciences and resources, Chinese Academy of Sciences (website:
https://www.resdc.cn) (accessed on 12 March 2021). The land-use classification system in this dataset has been widely used in national, regional, and local land-use change research in China. In this classification system, urban land (i.e., the land use type was coded as 51) was defined as land with a combination of high, medium, and low-density built-up areas. The benefit of this definition was to support a more comprehensive range of urban land by considering the urban and town areas [
40,
41]. In the context of this definition, urban land was defined as the same type of underlying surface of the urban and town areas, which can provide a uniform and spatially continuous land surface, avoiding the hollow and discontinuous regions within the urban and town boundaries.
For the urban land boundaries in 1981, 1991, 2001, 2011, and 2021, we extracted the urban land (shape file format) of the study area in 1980, 1990, 2000, 2010, and 2020 from the National Land Use Data of China, Institute of Geographical Sciences and Resources, Chinese Academy of Sciences. For the city boundaries of 1981 and 1991, Google historical archive images were first considered, but there was a large number of discontinuous boundaries in the study area. The number available Google images in the years of 1981 and 1991 in the study area was too little to carry out urban boundary studies in our investigation. Considering the difficulty of obtaining high spatial resolution and continuous coverage images for the whole study area before 2000, land resources satellites (i.e., Landsat images) from the years of 1981 and 1991 were used. We filtered the images of poor quality with issues such as cloud cover and bad pixels to obtain effective observations in summer (mainly July and August). All the downloaded images were synthesized by false color and displayed on the ArcGIS platform. We superimposed the vector city boundaries on the corresponding Landsat image. Human computer interaction digital technology was used to modify urban land boundaries by identifying the urban land texture, color, patches, and other features through professionally geographical knowledge. In the process of digitization, we controlled the polygon error within two pixels and obtained the urban-land boundaries in 1981 and 1991. Then, to reduce the inconsistency of urban boundaries caused by the use of Landsat image (i.e., 1981 and 1991) and Google image (i.e., 2001, 2011, and 2021), the early urban hand drawn map in the study area was digitally scanned and used as a reference to correct the urban boundaries. Although the scanned map had the problem of spatial geometric correction, it provided an essential reference for urban boundaries in the difficulty of obtaining sufficiently high spatial resolution images. As a result, the urban land boundaries were further modified to improve the accuracy of the boundaries in the years of 1981 and 1991. For the urban land boundaries after 2000 (i.e., 2001, 2011, and 2021), we obtained the 2 m resolution provided in Google’s historical archives by the authorized 91 bitmap platform, as high spatial-resolution remote sensing images have been effective since 2000. Similarly, using the high-resolution Google data for the natural background image, we obtained the urban land boundaries in 2001, 2011, and 2021 by modifying the urban land boundaries with manual digitization technology. Additionally, the administrative-division materials of the study area were also used as a reference in consideration of the administrative division change factors in the past four decades. Then, we obtained the required urban land boundaries for the years of 1981, 1991, 2001, 2011, and 2021, respectively.
2.4.2. Component Mapping of Intra-Urban Lands
A collaborative methodology of “satellite images—vegetation impervious surface area soil model—endmember decomposition model—multiple index—decision tree classifier—unsupervised classification—spatial analysis” was created to obtain the land use structure and its component data for the urban land in this study. Specifically, all the Landsat images for the summer of 1981, 1991, 2001, 2011, and 2021 were evaluated and obtained in the weather conditions of sunny and windless. The method of radiometric calibration was firstly applied to these images on the Environment for Visualizing Images Platform (ENVI). The fast line-of-sight atmospheric analysis spectral hypercubes (FLAASH, [
42]) were used for atmospheric correction for the images obtained for the years of 1981, 1991, 2001 and 2011, so as to standardize the spectrum of each land type in urban areas. Furthermore, the Landsat OLI images acquired in 2021 were atmospherically corrected by a land surface reflectance code (LaSRc) algorithm from USGS earth resources observation and science (EROS) data center. After that, minimum noise fraction rotation (MNF Rotation) concentrated these land surface spectra in the first-three spectral features (generally higher than 90%). A visual interpretation of the MNF-based pairwise scattergrams from the first-three features was performed to obtain pure spectral endmembers to solve the vegetation–impervious surface area using the soil model (V-I-S, [
23,
43]), which was critical and directly affected the accuracy of the urban-land classification structure and its component products. In this process, a combination of automatic model extraction and manual recognition was used to filter pure endmembers. We obtained several sets of pure endmembers for each land class, which was input in the fully constrained least-squares solution model (FCLS, [
44]) to generate the component images of the vegetation object, soil object, low albedo object, and high albedo object from Landsat images in the years of 1981, 1991, 2001, 2011, and 2021, respectively. Spatial analysis operation was used to produce the component image of impervious surface area. To further obtain better results, we filtered these density maps through field investigation data and a digitization of the land types of impervious surface area, vegetation cover, and bare soil using historical high spatial resolution images from the years of 2001, 2011, and 2021. Furthermore, we used a combination of Landsat images and the early urban hand drawn map from the years of 1981 and 1991, respectively, to eliminate the interference of the mixed pixel spectrum issues. If the error of the decomposition result was high, we continued to optimize the endmembers until a better density map was obtained. Then, the density maps of impervious surface area, vegetation cover, and bare soil with the sub-pixel values of 0.01~100.00% were obtained for the years of 1981, 1991, 2001, 2011, and 2021, respectively.
2.4.3. Classification of Urban Land Use Structure
Spectral features were often used for urban land classification; however, the issue of mixed pixels at the pixel level led to interference in the classification accuracy. Generally, bright high albedo objects were disturbed by bare soil, while low albedo objects were disturbed by building shadows and water bodies. To weaken the interference of these factors, in this study, we obtained urban classification products by the process of “pixel—sub-pixel—pixel”. Firstly, for the pixel level, we downloaded all Landsat images that needed to be classified in the study area from the USGS official website. We performed image preprocessing on these images at the pixel level. Secondly, the process for images from pixel to sub-pixel levels was performed. The main processes of MNF transformation, endmember selection, vegetation—impervious surface area—soil model and fully constrained least-squares solution model were applied to obtain the sub-pixel images of impervious surface area, vegetation cover, and bare soil, with the sub-pixel values of 0.01~100.00%, respectively. A detailed description of sub-pixel extraction is provided in
Section 2.4.2. Finally, the process of sub-pixel to pixel levels was undertaken. In this process, the interference of mixed pixels was reduced at the sub-pixel scale, to obtain more accurate classification results. In this study, we carried out the sub-pixel to pixel process through the combination of a decision-tree classifier and manual observation. The process is detailed below. The modified normalized difference water index (MNDWI, [
45]) was firstly used to extract water bodies through the combination of green and NIR from Landsat images, which was applied to mask the interference signal from dark high albedo objects. Then, the synergism of the vegetation cover component data and the normalized difference vegetation index (NDVI, [
46]) was used to extract vegetation types. The low-albedo and soil difference index (LSDI) proved to be an effective method to eliminate the spectral interference of bare soil in the bright high albedo objects. The combination of LSDI and soil component data was used for the extraction of bare soil. After that, the low and high albedo objects were applied to obtain the cover of impervious surface area. Then, the remaining unclassified surface types (usually less than 5%), were addressed through unsupervised classification. In this study, the unsupervised classification divided the unclassified region into 50 types (i.e., the 50 images) in each Landsat image. The manual observation of the spatial position as well as the optical characteristics from each unsupervised classified image were used to identify the land use types. Finally, all the classified land types were integrated into four categories, including the impervious surface area, vegetation cover, bare soil, and water body in the urban land regions.
2.5. Selection of Urban Land Landscape Index
The landscape index has always been applied to describe the spatial organization form and spatial allocation information of land use structure, acting as an effective way to explore the fragmentation, advantages, connectivity, aggregation, mosaic, and diversity of land surface [
47]. Generally, a single index was not adequate to express these characteristics of land types, but too many indexes would become repetitive and redundant. We selected as few indexes as possible to express the required characteristics, referring to the guidance manual of landscape meaning. Finally, the patch density, aggregation index, connectivity index, largest patch index, landscape shape index, and Shannon’s diversity index were obtained from the two scales of landscape type level and landscape diversity level to explore these characteristics [
48]. The name, abbreviation, formula, and definition of all landscape indexes are provided in the following table (
Table 2).
2.6. Division of Physical Medium Environments
In this study, we focused on the physical medium environments of slope, aspect, and water resource service, because these natural conditions were considered significant and may intuitively overlap with the urban land component data using the spatial clustering analysis, to form a geographical map of spatial visualization. It provided the advantage, in this study, of exploring the issue at the sub-pixel scale, namely, the relationship between sub-pixel urban land component and the physical medium environments (slope, aspect, and river distribution) was investigated, to provide a new evaluation.
The fluctuation of slope often affected the spatial distribution of buildings, which was an indispensable factor that must be considered in the building construction. To obtain the variable of slope, a spatial analysis module from ArcGIS was opened first, following which we navigated the surface analysis tools. Inside the surface analysis group, the slope analysis tool appeared and was applied to identify the slope (gradient or steepness) from each cell of a raster using the input of digital elevation model (DEM), for generating the spatial slope map. To facilitate the spatial analysis of slope and urban land component data, the classification tool was applied to divide the slope map variable into five levels, with values of [0~5°), [5°~10°), [10°~15°), [15°~22.5°), and (>22.5] from level 1 to level 5, according to the harmfulness reference of the slope to buildings, etc. Similarly, the building density in each pixel of the aspect always reflected the orientation of the building. During the preprocessing of aspect data, the aspect variable was divided into eight directions at an interval of 45 degrees clockwise through a similar method to that used to produce the slope map. The direction of aspect map included the North [0–22.5°), Northeast [22.5°–67.5°), East [67.5°–112.5°), Southeast [112.5°–157.5°), South [157.5°–202.5°), Southwest [202.5°–247.5°), West [247.5°–292.5°), Northwest [292.5°–337.5°), and North [337.5°–360°), respectively.
Water resources within the city such as the rivers, lakes, and ponds played a considerable role in regulating the hydrological process of the urban ecosystem and the comfort of human settlements. The evaporation from the water body was conducive to increased air humidity and reduced local polar heat. The distance from the river became an important index to measure this regulation of water resources to urban residential and living areas. The service radius of water resources was set as 0.5 km, 0.5–1 km, and 1–2 km according to the survey of the buildings, roads, and squares in the study area. This buffer service radius not only ensured that the patches did not overlap as much as possible, but also ensured the full coverage of impervious surface area components. Additionally, although the road was considered to be a result of planning, we still explored the distribution characteristics and spatiotemporal evolution law of urban land and its structure in different loops (mainly the typical loops from 1 to 5), considering the urban land development characteristics of the loop line in the past 40 years in the study area.
2.7. Accuracy Evaluation Scheme
Urban land mappings in this study were created according to spatial expansion dimension, the internal land structure change dimension as well as its component dimension, displaying a hierarchical urban land classification system. For the land expansion, urban land boundaries were based on the coordination of 2 m Google images, historical city maps, and high spatial resolution land products. Thus, we focused on the accuracy evaluation of the urban structure and its components. For the land use structure, facing the dilemma of validation in 1981 and 1991 due to the difficulty in obtaining the high resolution image, we designed a cross validation scheme from Landsat images, namely, the verified Landsat images were separated from the images used to obtain a time series of land-use structures and the land use components in the study area, i.e., we used some images to obtain land use structures and their components; furthermore, we employed other images in the same year to conduct accuracy verification. After that, a stratified random sampling technique was applied to generate a total of 7500 sampling points, following which the Landsat images for the years of 1981 and 1991 from the USGS website and the 2 m Google image in the years of 2001, 2011, and 2021 from the authorized and professional 91 bitmap platform were obtained. Taking these images as the background data, we superimposed all the samples to the imagery and manually checked the samples one by one to identify the true and false values. An accuracy evaluation matrix was used for accuracy calculation, along with the indicators of classified samples, reference samples, the number of correct samples, ground truth samples, producer’s and user’s accuracy, using the verified samples. For the land-use component validation, a 3 × 3 pixels window (pixel) method was applied to calculate the actual land surface area, with a total number of 100 in each land type. To facilitate the verification, we divided each land component data into 10 levels, with a 10% component interval from level 1 to level 10. Then, we compared the actual values with the reference values (i.e., the land component generated in this study) to calculate using the combination of fitting coefficient (R2) and mean square error (RMSE).
5. Conclusions
A multi-scale methodology was established to demonstrate the spatiotemporal heterogeneity and the evolution mappings of urban land, land use structure, and its component during the period of 1981–2021.We further investigated the relationship between hierarchical urban lands with using physical medium environments (i.e., slope, aspect, water resources service) in the study area. The multi-scale urban land dataset displayed good results with a comprehensive accuracy of over 90%. During 1981–2021, an intense urban land expansion occurred from 467.13 km2 in 1981 to 2581.05 km2 in 2021, along with a total growth rate of 452.53%. The spatial distribution of urban land expanded across all directions at the beginning of the urbanization period and then focused on the directions of north, northwest, and southwest. For the urban land structure, the most dramatic extension occurred in the ISA, with a sharp growth rate of over 650.00% (i.e., +1649.54 km2). A greening study area was observed, with a vegetation coverage rate of 8.43% in the old urban regions (OUR) and 28.42% in the newly expanded urban regions (NEUR)). The conclusions of this work are consistent with current frameworks/viewpoints, i.e., that the urban land was also greener in the arid regions of China and along the Belt and Road. In the process of land structure change, the integrity of the urban landscape (SHDI = −0.164, PD = −8.305) has been enhanced over the past 40 years. For the land component, the component changes of different land use types can elucidate the advantages of the lands’ temporal and spatial heterogeneity at sub-pixel scale. Additionally, a lower component of ISA (0.637 vs. 0.659) and a higher component of VC (0.284 vs. 0.211) were observed in the new urbans compared to the old, implying the promotion of a better living environment. Furthermore, the dominant aspect of low, medium, and high density ISA was captured with the north–south orientation considering the sunlight conditions and traditional house construction customs in North China; and over 92.00% of the ISA was distributed in the flat region. When the water resource service distance changed from 0.5 km to 0.5–1 km and 1–2 km, high-density ISA was relatively low but high-density vegetation was relatively high in the buffer zone closest to water resources. This study provides a new report on the evolution of multi-scale urban land during 1981–2021 and revealed its relationship with physical medium environments. The findings related to the land structure and its compositional changes, as well as the physical medium environments, reveal the important societal implications of the Sustainable Development Goals (SDGs) and of urban design policies, with Beijing acting as a good testbench, thereby providing an essential reference for related research.