Multi-Level Dynamic Analysis of Landscape Patterns of Chinese Megacities during the Period of 2016–2021 Based on a Spatiotemporal Land-Cover Classification Model Using High-Resolution Satellite Imagery: A Case Study of Beijing, China

Abstract

In today’s accelerating urbanization process, timely and effective monitoring of land-cover dynamics, landscape pattern analysis, and evaluation of built-up urban areas (BUAs) have important research significance and practical value for the sustainable development, planning and management, and ecological protection of cities. High-spatial-resolution remote sensing (HRRS) images have the advantages of high-accuracy Earth observations, covering a large area, and having a short playback period, and they can objectively and accurately provide fine dynamic spatial information about the land cover in urban built-up areas. However, the complexity and comprehensiveness of the urban structure have led to a single-scale analysis method, which makes it difficult to accurately and comprehensively reflect the characteristics of the BUA landscape pattern. Therefore, in this study, a joint evaluation method for an urban land-cover spatiotemporal-mapping chain and multi-scale landscape pattern using high-resolution remote sensing imagery was developed. First, a pixel–object–knowledge model with temporal and spatial classifications was proposed for the spatiotemporal mapping of urban land cover. Based on this, a multi-scale district–BUA–city block–land cover type map of the city was established and a joint multi-scale evaluation index was constructed for the multi-scale dynamic analysis of the urban landscape pattern. The accuracies of the land cover in 2016 and 2021 were 91.9% and 90.4%, respectively, and the kappa coefficients were 0.90 and 0.88, respectively, indicating that the method can provide effective and reliable information for spatial mapping and landscape pattern analysis. In addition, the multi-scale analysis of the urban landscape pattern revealed that, during the period of 2016–2021, Beijing maintained the same high urbanization rate in the inner part of the city, while the outer part of the city kept expanding, which also reflects the validity and comprehensiveness of the analysis method developed in this study.

1. Introduction

Currently, more than half of the world’s population lives in urban areas, and almost all countries in the world are becoming increasingly urbanized [1]. It is predicted that, in the coming decades, urbanization will be more rapid than the development of other regions (rural settlements, townships, etc.) [2]. This trend is changing the configuration of human settlements, which will have significant impacts on the living conditions, environment, and development in different parts of the world [3,4]. Urban development has brought many benefits, including promoting regional economic development and industrial agglomeration, providing more jobs, improving the quality of life, and enabling more people to enjoy better living, transportation, health care, and education conditions [5,6]. However, urbanization has also changed the pristine natural landscape of regions and brought about a series of ecological and environmental problems, such as global warming, ecological damage, environmental pollution, traffic congestion, and widening economic disparities [7,8]. These problems exist in different stages of development in different cities and are especially prominent in large cities [9]. The problems brought about by urbanization cause disturbance to people’s lives and also pose threats of different degrees to people’s health [10]. Therefore, the dynamic monitoring and evaluation analysis of the urbanization process has become an important issue and has received a great deal of attention in recent years.

The spatial heterogeneity of urban ecosystems is greater than that of any other type of ecosystem, and the analysis of the spatial pattern of urban landscapes is important for understanding the urbanization process and its ecological consequences [11]. The current research on urban landscape spatial pattern analysis has mainly focused on two aspects: (1) reflecting the urbanization process through urban land use/cover changes and its driving forces [12,13] and (2) constructing urban landscape pattern evaluation indicators to reflect the degree of urbanization through the spatial characteristics and changes in urban landscape patterns [14,15]. Owing to the characteristics of the current rapid urbanization process, it is difficult for the traditional urban landscape collection methods to meet the needs for application [16]. Therefore, it is urgent that a method for rapidly acquiring information and characterizing urban landscape patterns in a large area is developed to reflect the current state of the urban ecology and urbanization process promptly, which is crucial for land-use planning and management, as well as urban ecological environmental protection.

With the rapid development of satellite technology, satellite remote sensing data have become an effective means of rapidly obtaining urban spatial information [17]. Various remote-sensing sensors have been used to extract urban landscape component information and for landscape pattern analysis, such as nighttime light [18], Landsat [19], Sentinel-2 [20], Hyperion [21], and high-spatial-resolution remote sensing (HRRS) images [22]. Compared with medium–low-resolution remote sensing images, HRRS images can describe the urban land-cover information in more detail and more accurately [23]. In addition, the development of HRRS satellites in recent years has resulted in the availability of HRRS images with a large width, high revisit period, and great image quality, which can be used for the dynamic monitoring of large urban areas [24]. In this regard, the GaoFen-1 (GF-1) satellite is the first satellite of China’s high-resolution Earth-observation system (launched in April 2013) [25]. Based on the successful application of GF-1, a series of GF series satellites, such as GF-1 B/C/D, GF-2, GF-6, etc., have been launched one after another, thus further enhancing its overall monitoring capability [26]. Therefore, GF series satellites’ remote sensing images have become an important data source for urban landscape pattern analysis.

In terms of urban land-cover classification methods, although great progress has been made in recent years, fine-scale land-cover mapping of BUAs and their interiors using HRRS imagery remains a challenge [27]. Researchers have designed a series of spatial and structural features (e.g., shape [28], texture [29], and morphological features [30,31]) based on spectral features oriented to urban scenes [32], or they have used deep learning methods to automatically extract high-dimensional features [33], thus improving the separability of spectrally similar objects (e.g., buildings, roads, and impervious surfaces areas (ISA)). On this basis, previous studies have demonstrated the ability to improve the accuracy of urban land-cover classification by improving the classifiers (e.g., multi-classifier combinations and pixel–object union classification strategies) [34,35]. However, previous classification studies have mainly focused on feature extraction and classification methods for single-period remote sensing images, and the correlation between multi-period land-cover information has received less attention, leading to a lack of stability and continuity of multi-period land-cover-oriented classification results [36]. Therefore, in current classification research, how to combine multi-period high-resolution remote sensing information to further develop spatiotemporal land-cover classification methods has become an urgent problem to be solved.

Regarding the methods of urban landscape pattern analysis based on remote sensing images, most of the current methods focus on the analysis and evaluation of urban landscape pattern characteristics through the landscape pattern index [37,38] or the construction of spatial evaluation indices (clustering, shape, density, and structure) at a single scale (e.g., pixel, grid, feature type, city block, BUA, and other scales) [39,40] or on a regular multi-scale grid [41,42]. Although all of these provide a good expression of urban landscape pattern characteristics in one aspect, they all have limitations. Currently, a few studies have constructed a pixel–grid–city block unit based on HRRS images to conduct multi-scale analysis and evaluation of urban landscape patterns, which allows for the comprehensive analysis and evaluation of urban landscape pattern characteristics [43,44]. However, the BUA is the basic element of research on the urbanization process and is an important research scale for urban ecological landscape analysis. Therefore, it is very important to further expand the multi-scale urban landscape pattern analysis methods to include BUAs to better understand the urbanization process and the ecological characteristics of urban landscapes.

Based on this, in this study, urban landscape pattern evaluation units at four spatial scales, i.e., urban administrative boundary–BUA–urban block–urban object type, were constructed based on the development of urban land-cover spatiotemporal mapping. Additionally, a joint evaluation model of the urban landscape pattern at multiple scales was investigated to expand the analysis of urban landscape pattern characteristics. The results of this study provide effective information support for the comprehensive analysis of the current situation and change trend of urban landscape patterns from different perspectives. Therefore, the objectives of this study were to propose a joint feature–pixel–object–temporal knowledge classification method chain to obtain multi-period high-precision land-cover mapping results within BUAs and to construct a joint multi-scale evaluation method for the urban landscape pattern to effectively monitor and analyze the spatial changes in the urban landscape pattern. The spatial changes in the urban landscape pattern were effectively monitored and analyzed.

2. Study Area and Data

The study area was Beijing, the capital of China (Figure 1). China is undergoing rapid urbanization, especially in first-tier cities, where more rapid economic development has led to dramatic changes in urban landscape patterns [45]. In addition, the land-cover types in the study area are characterized by diverse morphologies and types, making it challenging to study the land-cover classification within the city [46]. Therefore, we focused on land-cover classification and dynamic comparative analysis of Beijing.

In this study, we selected images of the Beijing area acquired by GF-1, including four images obtained in 2016 and two images obtained in 2021, and acquired data from https://data.cresda.cn/#/mapSearch (accessed on 26 January 2022). To avoid the influence on the land-cover classification results and dynamic analysis due to the seasonal differences between the two phases of data and image quality factors, such as cloudiness and exposure, the data were selected by data screening for both April and May in this study. The specific details are shown in

Table 1. GF-1 data applied in this study.

No.	Year	Location a/°	Date	Spectrum/nm	Spatial Resolution/m
1	2021	E 116.6, N 40.2	3 April 2021	Blue: 450–520 Green: 520–590 Red: 630–690 Infrared: 770–890 PAN: 500–800	2/8
		E 116.4, N 39.7	3 April 2021
2	2016	E 116.5, N 39.8	21 May 2016
		E 116.6, N 40.2	21 May2016
		E 116.0, N 39.8	16 May 2016
		E 116.1, N 40.2	16 May 2016

^a Central longitude and latitude for GF1-PMS.

. The GF-1 satellite is equipped with two 60 km-wide high-spatial-resolution sensor cameras (GF-1/PMS1 and GF-1/PMS2). The data collected by the panchromatic and multispectral sensors (PMSs) on GF-1 include panchromatic bands with a spatial resolution of 2 m and multispectral bands (blue, green, red, and near-infrared bands) with a spatial resolution of 8 m [25].

The reference data used in this study were the boundary data of the BUA and the OpenStreetMap (OSM) data used to construct the city blocks.

(1)

The boundaries of the BUA were updated to 2016 and 2021 by applying GF-1 HRRS imagery and visual interpretation results based on the National Land Cover Dataset of China, which was produced by the Institute of the Chinese Academy of Sciences (http://www.resdc.cn/Datalist1.aspx?FieldTyepID=1,3 (accessed on 10 February 2022)). The average accuracy was greater than 95%, based on the 2016 and 2021 BUA statistical results provided by the National Bureau of Statistics of China (http://www.stats.gov.cn/tjsj/tjgb/ndtjgb/ (accessed on 10 February 2022)).

(2)

The OSM data were downloaded from the OpenStreetMap Foundation (http://www.openstreetmap.org/ (accessed on 28 January 2022)). The primary and secondary roads were mainly selected to construct the urban neighborhood units.

3. Methodology

Based on HRRS imagery supplemented with multi-source spatial information, in this study, a multi-level urban landscape pattern framework (urban administrative boundary, urban built-up area, urban block, and inner city land-cover objects) was developed. In addition, a multi-scale joint landscape pattern approach for multi-scale landscape pattern dynamic analysis of cities was constructed based on the high-precision urban land-cover mapping results for the two periods. The key steps in this study were (1) GF-1 data pre-processing, (2) high-precision spatiotemporal mapping of inner-city land cover, and (3) multi-scale joint landscape pattern dynamic analysis of the city (Figure 2).

3.1. Pre-processing of GF-1 Data

To ensure the spatial positioning accuracy and to further improve the spatial resolution, the GF-1 image data were preprocessed, including ortho-rectification, panchromatic/multispectral data fusion, and image mosaicking and clipping, using ENVI 5.0 commercial software. In this study, the GF-1 images were ortho-rectified based on the rational polynomial coefficient (RPC) model (using the RPC file in the system) [47] and were fused using the Gram–Schmidt pan-sharpening (GS) method [48]. Compared with fusion algorithms, such as Intensity Hus Saturation (IHS) transformation, Brovey transform, and Principal Component Analysis (PCA) transformation, the GS method improves the problem of the over-concentration of information in PCA and can preserve spatial texture information better without the band limitation; therefore, it has been more widely used in HRRS images [49]. Based on the above processing process, GF-1 ortho-rectification remote sensing images with a spatial resolution of 2 m were obtained. Based on the polynomial model [50], the GF-1 data for 2016 were geometrically corrected using ground control points collected from the 2021 GF-1 images, with an average root mean square error of <0.4 pixels, thus ensuring the relative positioning accuracy between the two phases of data. In addition, to facilitate the application of Beijing BUA images in the analysis, the pre-processed GF-1 images of 2016 and 2021 were mosaicked and clipped based on the corresponding BUA boundary data, respectively, thus obtaining the Beijing BUA images for the two periods (Figure 1b,c).

3.2. In-urban Land-cover Spatiotemporal Mapping

The land-cover changes were studied by obtaining multi-period land-cover results after performing land-cover classification. The post-classification comparison method is widely used because it can effectively reduce the error problem caused by atmospheric and environmental differences between multi-period images. The advantage of this method is that it can show land cover type changes and provide from–to type change information. However, the post-classification comparison method has the disadvantage of error accumulation, which makes it difficult to guarantee accuracy. Therefore, in this study, to ensure the accuracy of the multi-temporal land cover map of BUA, a complex land-cover classification chain was adopted, including multi-type feature extraction, a joint pixel–object classification framework, and object-level spatiotemporal classification post-processing based on expert knowledge.

3.2.1. Multi-Type Feature Extraction

In this section, we consider three types of feature sets (

Table 2. Multiple features considered in this study.

Type	Features	Function	Equation
Spectral index	Normalized difference vegetation index (NDVI) [59]	The NDVI was constructed by exploiting the large differences in the spectral reflectance of vegetation in the near-infrared and red bands.	$NDVI=({\rho }_{NIR}-{\rho }_{red})/({\rho }_{NIR}+{\rho }_{red})$ (1) where ${\rho }_{NIR}$ and ${\rho }_{red}$ are the reflectances in the near-infrared and red bands, respectively.
	Normalized difference water index (NDWI) [60]	Based on the principle that water has strong reflection in the green band and strong absorption in the near-infrared band, the NDWI was used to enhance the characteristics of water.	$NDWI=({\rho }_{green}-{\rho }_{NIR})/({\rho }_{green}+{\rho }_{NIR})$ (2) where ${\rho }_{NIR}$ and ${\rho }_{green}$ are the reflectances in the near-infrared and green bands, respectively.
Spatial features	Differential morphological profiles (DMPs) [61]	DMPs can record image structure information by measuring the slope of the opening and closing curves of the scale parameters in each step. They can be regarded as the shape spectrum of the image.	$DMP=\{\begin{array}{c}DM{P}_O=\left|{\gamma }^{\lambda }(I)-{\gamma }^{\lambda -1}(I)\right|\\ DM{P}_C=\left|{\varphi }^{\lambda }(I)-{\varphi }^{\lambda -1}(I)\right|\end{array}$ (3) where ${\gamma }^{\lambda }(I)$ and ${\varphi }^{\lambda }(I)$ represent morphological opening and closing by reconstruction for an image I, with λ being the scale parameter of the structural element.
	Gray-level co-occurrence matrix (GLCM) [62]	The GLCM is defined as a matrix in which each entry estimates the probability of the occurrence of a pair of gray levels at a certain distance and direction. Then, a set of texture metrics, such as the homogeneity (HOM), entropy (ENT), and contrast (CON), can be extracted from this co-occurrence matrix to delineate the texture statistics of the image.	$HOM={\displaystyle \sum _i{\displaystyle \sum _j\frac{f(i,j)}{1+\left|i-j\right|}}}$ (4) $ENT={\displaystyle \sum _i{\displaystyle \sum _{j}f(i,j)\log (f(i,j))}}$ (5) $CON={\displaystyle \sum _i{\displaystyle \sum _{j}f(i,j){(i-j)}^2}}$ (6) where $f(i,j)$ is the probability of the pairs of gray levels i and j that occur in a specified direction and distance.
Composite index	Morphological building index (MBI) [63]	The MBI is an effective metric that represents the spectral–spatial properties of a building (e.g., brightness, contrast, size, and orientation) through a series of morphological operators to highlight building structures in high-resolution images.	$MBI=\frac{{\displaystyle \sum _{d,s}DM{P}_{W-TH}(d,s)}}{D\times S}$ (7) where $DM{P}_{W-TH}$ denotes the differential morphological profiles of the white top hat and d and s indicate the scale and direction of the structural element, respectively.
	Morphological shadow index (MSI) [63]	The derivation of the MSI is based on the fact that shadows have low reflectance and high local contrast; therefore, the MSI is defined as a dual function of the MBI, i.e., the morphological contour of the black top hat to highlight the shadow structure.	$MSI=\frac{{\displaystyle \sum _{d,s}DM{P}_{B-TH}(d,s)}}{D\times S}$ (8) where $DM{P}_{B-TH}$ denotes the differential morphological profiles of the black top hat in contrast to the white top hat in the MBI.
	Morphological large-area water index (MLWI) [64]	The MLWI represents the spectral–spatial attributes (e.g., spectral, morphological, and size attributes) of large urban water bodies through a series of morphological operators to highlight large urban water body features in high-resolution images.	$MLWI=EMP{s}_{MCR}mean-EMP{s}_{WTH}mean$ (9) $EMP{s}_{MCR}mean=\frac{{\displaystyle \sum {}_i}{\varphi }_{NDWI}^{re}{}^i(r_{\Delta s})}{i}$ (10) $EMP{s}_{WTH}mean=\frac{{\displaystyle \sum {}_i}NDWI-{\gamma }_{NDWI}^i(r_{\Delta s})}{i}$ (11) where ${\varphi }_{NDWI}^{re}(r)$ and ${\gamma }_{NDWI}^{}(r)$ represent the closing-by-reconstruction and the opening of the NDWI image, respectively, and r indicates the radius of a circular structural element (SE).
	Morphological small-area water index (MSWI) [64]	The MSWI represents the spectral–spatial attributes (e.g., spectral, morphological, contrast, and size attributes) of small urban water bodies through a series of morphological operators, thus highlighting small urban water body features in high-resolution images.	$MSWI=EMP{s}_{MCR}mean+EMP{s}_{WTH}mean$ (12) $EMP{s}_{MCR}mean$ and $EMP{s}_{WTH}mean$ (13) are defined as the means of the EMP profiles, as water exhibits large NDWI values at different multi-scales.

) to describe the inner-city land cover. Spectral indices and spatial features (e.g., morphology and texture) were used to reflect the physical attributes and spatial structural attributes of urban land-cover types, respectively. In addition, composite features were used to enhance the separability of the integrated feature types (e.g., urban buildings and urban surface water) in complex urban contexts.

(1)

Spectral index: The remote sensing spectral features can reflect the physical attributes of land-cover types and can visually portray land-cover type information. Spectral indices, such as the normalized difference vegetation index (NDVI) and the normalized difference water body index (NDWI), are indices constructed from spectral information in remote sensing images that better reflect the information about natural land surface attributes, such as water bodies, vegetation, and bare land in the spectral dimension [51,52]. Therefore, the NDWI/NDVI was used in this study to enhance the information about the water bodies, vegetation, and bare land in the inner city.

(2)

Spatial features: In HRRS images, the spatial features provide rich spatial information about the feature types, such as the geometry, texture, and morphological information, which can solve the confounding problem caused by relying only on spectral information [53]. Among them, the texture and morphological information are of greater concern and have been applied by scholars. They play an important role in the classification of features in HRRS images [54]. Therefore, in this study, the texture and morphological features were selected as the spatial features.

(3)

Composite features: Urban surface information has a certain degree of comprehensiveness and complexity due to features such as urban buildings and urban surface water [55,56]. To be able to enhance the extraction accuracy of urban surface information, many scholars have incorporated spatial features, such as the multi-scale angle, scale, and morphology of different urban feature types based on spectral features, thus forming composite features to enhance the separability of the urban surface information [57,58]. Therefore, in this study, the morphological building index (MBI) was used to enhance the urban buildings, the morphological shadow index (MSI) was used to enhance the shadow types, and the morphological large/small area water index (MLWI/MSWI) was used to enhance the information on urban surface water types.

3.2.2. A Pixel- and Object-Based Classification Framework

In this study, we adopted a joint pixel–object-based classification framework for land-cover mapping to resolve the uncertainty of the results brought about by using only a single classification method, thus improving the accuracy of land-cover mapping and the reliability of the multi-period land-cover classification results [65]. The pixel-based classification method can give full play to the significance of the spectral index features described in Section 3.2.1 for the spectral attributes of water bodies, vegetation, bare land, impervious surfaces, and other types, and can enhance the differentiability of the urban land surface types in the spectral dimension. The object-based classification method can fully utilize the spatial features described in Section 3.2.1 to increase the separability of the different types of features, such as buildings and roads. In addition, object-level mapping can avoid the problem of pretzel noise caused by pixel-based classification, and it can also unify them to the same scale for subsequent decision-level fusion processing. In the classification model used in this study, the random forest (RF) algorithm was used as the classification model because it is characterized by overfitting prevention, high stability, and strong generalization ability, and it can automatically perform feature selection among high-dimensional features [66]. Finally, the two strategies were fused at the decision level at the object-level scale, and land-cover mapping results with a higher accuracy were obtained through an iterative method of continuous optimization.

The steps of the joint pixel–object-based land-cover classification method are as follows: (1) Stack the multi-source features extracted in Section 3.2.1 to obtain a stacked feature F. (2) Use a multi-scale segmentation method to obtain an image segmentation object O. (3) Use an RF model to classify F and obtain the pixel-based classification results, pixel-based class map (pixel). (4) Use the pixel-based classification result R (pixel) and maximum voting for each segmentation object O to obtain the classification result, class map (PO). (5) Use the RF model to classify segmentation object O using feature F to obtain the object-based classification result, class-map (O). (6) Perform consistency judgment of class-map (PO) and class-map (O). When the classification result is consistent, obtain the initial class-map result. When the judgment result is inconsistent, repeat steps (3–6) for iterative computation until the process is terminated when the consistency judgment result is unchanged. (7) Obtain the final inconsistent segmentation object O using the maximum voting method for all of the results of the iterative computation in step (6) to obtain the maximum probability of the class, thus confirming the final type and finally obtaining the initial class-map results. In this study, the parameters of the random forest model were obtained by the trial-and-error method, i.e., the overall number of trees was 100 and the depth of each tree was 10. The details are shown in Figure 3.

3.2.3. Expert Knowledge-Based Post-Processing for Object-Level Spatiotemporal Classification

To ensure the accuracy of the spatiotemporal mapping of the internal urban land cover, in this study, expert knowledge-based object-level spatiotemporal classification post-processing rules were used to further check and refine the initial classification results. In this study,

Table 3. Post-processing of classification based on spatiotemporal expert knowledge.

Mixed classes Class 1 and Class 2	Principles	Rules1 (T a/TA b)	Rules2 (T/TA)	Correction Results
Shadow and water	Buildings and shadows are spatially adjacent, but cannot be adjacent to water bodies.	ClassT(O) = water and T: RB to O (building) > 0	ClassTA(O) = water and TA: RB c to O (building) > 0	ClassT(O) = water-> d shadow and ClassTA(O) = water-> shadow
			ClassTA(O)≠water and TA: RB c to O (building) > 0	ClassT(O) = water-> shadow
			ClassTA(O) = water and TA: RB to O (building) = 0	ClassT(building) -> TA Class(O)
			ClassTA(O)≠water and TA: RB to O (building) = 0	ClassT(O) = water-> shadow
		ClassT(O) = shadow and T: RB to O (water) > 0	ClassTA(O) = shadow and TA: RB to O (water) > 0	ClassT(O) = shadow -> water
			ClassTA(O) ≠ shadow and TA: RB to O (water) > 0	ClassT(O) = shadow -> water
			ClassTA(O) = shadow and TA: RB to O (building) > 0	ClassT(O) = water-> shadow
Building and shadows	Buildings and shadows are spatially adjacent	ClassT(O) = building and T: RB to O (shadow) > 0	ClassTA(O) ≠building and TA: RB to O (shadow) > 0	ClassTA(O) -> building
Building and road	Buildings and shadows are spatially adjacent	ClassT(O) = road and T: RB to O (shadow) > 0	ClassTA(O) = building and TA: RB to O (shadow) > 0	ClassT(O) = road -> building
			ClassTA(O) = road and TA: RB to O (shadow) = 0	ClassT(shadow) -> TA Class(O)
Building and bare/ISA	Buildings and shadows are spatially adjacent	ClassT(O) = bare/ISA and T: RB to O (shadow) > 0	ClassTA(O) = building and TA: RB to O (shadow) > 0	ClassT(O) = bare/ISA -> building
			ClassTA(O) = bare/ISA and TA: RB to O (shadow) = 0	ClassT(shadow) -> TA Class(O)

^a T: One time; ^b TA: another time; ^c RB: Relative border; ^d ->: changed to.

was used to check the results of the unreasonable land cover objects between two phases of features based on the spatial correlation or interrelationship between the feature types. Based on this, a multi-temporal joint expert knowledge correction method was constructed to further automatically correct the unreasonable object results and obtain the final land-cover classification results. The advantage of this method is that it takes into account the spatial correlations between the features in the two periods and combines the spatial knowledge of the multi-temporal phases to correct the unreasonable object types, thus reducing the problem of the uncertainty of the feature types introduced in the post-processing phase based only on a single time.

The steps were as follows: (1) To ensure that the images of different moments were analyzed at the same scale, the initial land-cover classification result for moment TA was segmented according to the initial land-cover classification result for moment T, and the initial land-cover classification result for moment TA with the same object boundary as moment T was obtained. (2) The unreasonable objects for moments T and TA were detected according to the rule set. (3) The unreasonable objects were automatically corrected according to the formula in Table 3 to obtain the results.

3.3. Accuracy Evaluation

In this study, internal urban land-cover mapping was generated in each computational step. The accuracy of the results was verified qualitatively through visual interpretation and quantitatively by a statistical confusion matrix containing the overall accuracy (OA), kappa coefficient (Kappa), producer’s accuracy (PA), and user’s accuracy (UA) [67]:

$$OA=\frac{TN+TP}{TN+TP+FN+FP}$$

(14)

$$Kappa=\frac{T\times (TP+TN)-\sum }{T\times T-\sum }$$

(15)

$$PA=\frac{TP}{TP+FN}$$

(16)

$$UA=\frac{TP}{TP+FP}$$

(17)

$$\sum =(TP+FP)\times (TP+FN)+(FN+TN)\times (FP+TN)$$

(18)

where TP, FN, FP, and TN are the areas of correct extraction, undetected water bodies, incorrect extraction, and non-water bodies that were correctly rejected, respectively; and T is the total area of the study area.

3.4. Analysis of Quantitative Changes in Multi-Scale Urban Landscape Patterns

To overcome the limitations caused by single-scale urban dynamic analysis, four levels of association were considered in our multi-level dynamic analysis study (Figure 4): intra-urban land-cover types, urban blocks, BUA, and urban administrative scales. First, using the information about the BUA boundary and inner urban land-cover types obtained from the HRRS images, we analyzed the urban expansion through the multi-scale union of the urban administrative boundary, BUA, and inner urban land-cover object layers (buildings, roads, and ISA types). Second, we obtained the fine-grained changes within the BUA by calculating the trajectories of land-cover types within the BUA. Finally, a joint landscape pattern analysis of the urban block and land-cover-type scales was conducted, focusing on the changes in the building and vegetation coverage and the configuration (landscape configuration indicators) that constitutes the urban-block scale.

3.4.1. Multi-scale Driven Analysis of BUA Levels

In this section, we use a method that combines the urban administrative area (UA), built-up urban area (BUA), and all-urban impervious surface area (AISA) to construct a multi-level combined BUA expansion evaluation index to analyze the urban expansion. Among them, the AISA is the artificial surface within the city, which specifically includes the sum of three types of feature information, i.e., buildings, roads, and ISA, in the urban land cover. The specific structural relationship is shown in Figure 5. The AISA is the internal component of the BUA, while the BUA is the internal component of the UA. In addition, the ratio of the AISA to the BUA and the ratio of the BUA to the UA were calculated to obtain the multi-level joint BUA expansion evaluation index. The equations are as follows:

$$BUR=\frac{BUA}{UA}$$

(19)

$$ABR=\frac{AISA}{BUA}$$

(20)

where BUA, AISA, and UA are the area of the built-up urban area, the area of the all-urban impervious surfaces, and the area of the urban area, respectively. BUR and ABR are the ratio of the BUA to the UA and the ratio of the AISA to the BUA, respectively. The value of the BUR reflects the level of external expansion of the urbanization process. A higher value indicates a higher level of external expansion of the urbanization process, and a lower value indicates a lower level of external expansion of the urbanization process. The value of ABR reflects the density of the internal urban development. A higher ABR value indicates a higher density of internal urban development, and a lower value indicates a lower density of internal urban development.

3.4.2. Analysis of in-Urban BUA Land-Cover Dynamics

Within the BUA of the city, the analysis was conducted from two perspectives: the proportion of the landscape components and the landscape pattern. In the landscape component analysis, the area of each landscape component and its percentage were quantified based on the spatial distribution of the landscape components to analyze the urban landscape characteristics [68]. Based on this, landscape pattern analysis was conducted at the patch scale. Among the many landscape pattern analysis methods, the landscape index is the most widely used index type. Landscape indices synthesize and generalize landscape information as quantitative composite indicators, which can be used to quantitatively evaluate the spatial combination, structure, and configuration of the landscape pattern and other characteristics.

In this study, the patch density (PD), maximum patch index (LPI), landscape shape index (LSI), aggregation index (AI), sprawl index (CONTAG), and Shannon diversity index (SHDI) were used to quantify the landscape pattern of the city and to analyze the changes in the landscape pattern of the study area based on the relevance of previous research. The concepts, calculation methods, thresholds, and ecological significance of these indices were described in previous studies. The FRAGSTATS 4.2.1 software was used to extract the landscape indices (

Table 4. Detailed information about the landscape indices implemented in FRAGSTATS 4.2.1 [69].

Names	Abbreviations	Formulas		Description of the Land Cover Configuration
Plaque density	PD	$\mathrm{PD}=\frac{ni}{\mathrm{A}}\times 10,000$	(21)	Number of patches per unit area
Maximum plaque	LPI	$\mathrm{LPI}=\frac{{\max }_{j=1}^n(aij)}{\mathrm{A}}\times 100$	(22)	Percentage of the total landscape area represented by the largest patch
Landscape shape	LSI	$\mathrm{LSI}=\frac{ei}{\min ei}$	(23)	Degree of deviation between the shape of a given patch and a circle or square with the same area
Aggregation index	AI	$\mathrm{A}\mathrm{I}=[\frac{g_{ii}}{\max \to g_{ii}}](100)$	(24)	Degree to which a landscape type is adjacent to other types of surrounding landscapes
Sprawl indicator	CONTAG	$\mathrm{C}\mathrm{O}\mathrm{N}\mathrm{T}\mathrm{A}\mathrm{G}=\left[1+\frac{{\displaystyle \sum _{i=1}^m{\displaystyle \sum _{j=1}^m{P}_{ij}In(P_{ij})}}}{2In(m)}\right](100)$	(25)	Degree of aggregation or extension of different patch types in the landscape
Shannon’s heterogeneity index	SHDI	$\mathrm{SHDI}=-{\displaystyle \sum _{i=1}^m(pi\times \ln pi})$	(26)	Reflects the degree of the uniform distribution of each type of patch in the landscape

Note: $a_{ij}$—plaque area of $ij$; $A$—total landscape area; $ij$—plaque area; $N_i$—number of plaque types; $pi$—plaque type; $i$—total proportion of landscape; $e_i$—total length of type $i$; and $\min e_i$—probability of minimum value of $e_i$.

).

3.4.3. Dynamic Analysis of Urban Block-Level Landscape Patterns

In this study, inner-city land-cover analysis was conducted based on primary and secondary roads from OpenStreetMap (OSM) to construct the BUA block results. Among them, the building coverage ratio (BCR) and vegetation fraction (VF) are the two main landscape composition parameter values that have been studied and received attention. Namely, the BCR (VF) of each block is the ratio of the area covered by buildings (vegetation) to the total area. Thus far, fewer studies have focused on the variations in the building density and vegetation ratio at the urban block scale. Therefore, in this study, the BCR and VF at the urban block scale were constructed to dynamically evaluate and analyze the degree of human activity aggregation and vegetation greenness in the block patches. The BCR and VF were calculated as follows:

$$d(B_i)=\frac{{\displaystyle {\sum }_{j=1}^n{A}_{B_i}(O_j)}}{A(B_i)}$$

(27)

$$dv(B_i)=\frac{{\displaystyle {\sum }_{j=1}^n{A}_{B_i}(v_j)}}{A(B_i)}$$

(28)

where $B_i$ is the ith block base; A is the area; $A(B_i)$ is the area of the ith block base; $O_j$ and $v_j$ are the jth buildings and vegetation types of the ith block base, respectively; d is the density; and $d(B_i)$ and $dv(B_i)$ are the building coverage and vegetation proportion of the ith block base, respectively. Based on this, the BCR and VF were counted in 0.2 steps in the interval of 0–1 and are divided into the following intervals for quantitative analysis: [0, 0.2], [0.2, 0.4], [0.4, 0.6], [0.6, 0.8], and [0.8, 1.0].

4. Results and Analysis

4.1. Urban Land-Cover Accuracy Validation

Accuracy assessment of the newly developed 2 m urban land-cover product was performed using a randomly selected sample (pixels). A total of 4000 samples were randomly collected in both 2016 and 2021, resulting in a total of 8000 randomly and equally distributed samples. As the area of urban land-cover types differed from year to year, stratified random sampling was used in this study to randomly select samples of land-cover types within cities during the different periods. We ensured the accuracy of the sample points through manual interpretation and field validation. The OA, Kappa, PA, and UA were used to conduct the accuracy evaluation [70].

As shown in

Table 5. Confusion matrix for Beijing land-cover mapping in 2016 and 2021 (UA—user’s accuracy; PA—producer’s accuracy).

Year	Classified Data	Ground Truth Data (Pixels)
		Shadow	Road	Vegetation	Building	Bare land	ISA	Water	Total	UA/%
2016	Shadow	144	4	6	0	3	2	6	165	87.2
	Road	0	627	0	14	9	14	0	664	94.4
	Vegetation	6	0	846	0	40	0	3	895	94.5
	Building	0	0	0	683	1	14	1	699	97.7
	Bare land	0	2	8	0	803	49	8	870	92.3
	ISA	4	30	11	20	32	338	0	435	77.7
	Water	10	1	16	0	5	4	235	271	86.7
	Total	164	665	887	719	932	381	253	4000
	PA (%)	87.8	94.3	95.4	95.1	86.1	88.8	92.8
	Overall accuracy = 91.9%; kappa coefficient = 0.90
2021	Shadow	180	0	11	8	2	0	4	205	87.6
	Road	0	546	8	12	18	0	0	585	93.4
	Vegetation	6	0	947	14	18	16	4	1005	94.2
	Building	0	6	54	860	17	31	2	970	88.6
	Bare land	0	7	11	18	554	43	2	635	87.2
	ISA	12	2	4	18	12	340	4	392	86.7
	Water	6	6	4	2	0	0	189	208	91.1
	Total	205	567	1039	932	621	431	205	4000
	PA (%)	88.0	96.3	91.1	92.3	89.2	78.9	92.2
	Overall accuracy = 90.4%; kappa coefficient = 0.88

, the overall accuracies of the land-cover classifications in 2016 and 2021 for the BUA of Beijing were 91.9% and 90.4%, respectively, and the kappa coefficients were 0.90 and 0.88, respectively. The overall accuracy of the land-cover results for the BUA in Beijing in both periods exceeded 90%, and the kappa coefficients were greater than 0.88, indicating the validity and stability of the classification method developed in this study. The validity and stability of the classification method developed in this study were demonstrated. Except for the impervious surface type, the mapping accuracy and user accuracy of the land-cover results of the remaining categories in 2016 and 2021 were greater than 86%, indicating that the misclassification and omission rates of the classification results were low. In summary, the land-cover classification framework for the BUA developed in this study can ensure the accuracy of the urban land-cover classification. Its overall accuracy was maintained at >90%, which meets the overall requirements of mapping and analysis.

4.2. Multi-scale Spatial Analysis of Urban Expansion

The results obtained in this study are presented in

Table 6. Area values from UA–BUA–AISA.

Beijing	AISA/km2	BUA/km2	UA/km2	BUR/%	ABR/%
2016	718.3	1312.2	16,411	8.0	54.7
2021	868.1	1460.3	16,411	8.9	59.4

. The area of the BUA increased from 1312.2 km² to 1460.3 km², an increase of 148.1 km². The BUR increased from 8.0% to 8.9%, an increase of 0.9%. In addition, the AISA increased from 718.3 km² in 2016 to 868.1 km² in 2021, an increase of 149.8 km². The ABR increased from 54.7% to 59.4%, an increase of 4.7%. These results indicate that Beijing’s urbanization process was in rapid development within the city and the outward urban expansion was dramatic.

4.3. Analysis of Land-Cover Dynamics within Object-Level Cities

Figure 6 and Figure 7 show the statuses of the land cover within the BUA of Beijing in 2016 and 2021, respectively. Although the outward expansion of the BUA was significant, the roads and water bodies did not significantly change in terms of the overall patterns of their spatial distributions, indicating that the urban corridor landscape (road and water body elements) was relatively stable. During the period of 2016–2021, there was a significant increase in the areas of buildings and bare land patches, indicating that the BUA still had a high level of urbanization. In contrast, the areas of vegetation and the ISA significantly decreased during this period, indicating that the accelerated urbanization process compressed the vegetation and ISA.

As can be seen from

Table 7. Dynamic changes in the urban land cover types between 2016 and 2021.

Types	2016		2021		Change
	A/km2	F/%	A/km2	F/%	△A/km2	△F/%
Road	45.8	3.5%	53.2	3.6%	7.4	0.1%
Water	17.8	1.4%	24.0	1.6%	6.3	0.2%
Vegetation	404.8	30.8%	240.5	16.5%	−164.3	−14.3%
Building	231.6	17.7%	271.6	18.6%	39.9	0.9%
Shadow	68.6	5.2%	168.8	11.6%	100.2	6.4%
ISA	372.2	28.4%	374.3	25.6%	2.1	−2.8%
Bare	171.4	13.1%	327.7	22.4%	156.3	9.3%

, the areas of the water bodies, roads, and ISA increased from 17.8 km², 45.8 km², and 372.2 km² in 2016 to 24 km², 53.2 km², and 374.3 km² in 2021, with increases of 6.3 km², 7.4 km², and 2.1 km², respectively. Their proportions increased from 1.4%, 3.5%, and 28.4% in 2016 to 1.6%, 3.6%, and 25.6% in 2021, with increases of 0.2%, 0.1%, and −2.8%, respectively, indicating the stability of the urban corridor pattern. The areas of the buildings and bare land increased from 231.6 km² and 171.4 km² to 271.6 km² and 327.7 km², with increases of 39.9 km² and 156.3 km², respectively; and their proportions increased from 17.7% and 13.1% to 18.6% and 22.4%, with increases of 0.9% and 9.3%, respectively, indicating that the urbanization process remained at a significant level. In contrast, the area of vegetation decreased from 404.8 km² to 240.4 km², a decrease of 164.4 km², and its proportion decreased from 30.8% to 16.5%, a decrease of 14.3%, indicating that the urbanization process occupied the previously vegetated areas, to some extent.

Figure 8 shows the landscape pattern indices of the BUA in Beijing between 2016 and 2021. In particular, Figure 8a assesses all land-cover types as a whole at the landscape level, and Figure 8b–f evaluate the landscape pattern index for each land-cover type at the patch level. The landscape pattern indices used in the analysis were defined in Section 3.4.2.

The LPI decreased from 6.2 to 1.0, with the largest patch decreasing by 5.2, reflecting an overall increase in the utilization of development within the BUA in Beijing. The LSI and SHDI increased from 455.4 and 1.6 in 2016 to 508.9 and 1.7 in 2021, with increases of 53.5 and 0.1, respectively, indicating that the diversity of the landscape types in Beijing increased during these 5 years. The CONTAG and AI increased from 25.9 and 57.4 to 30.7 and 55.5, with increases of 4.8 and 3.3, respectively, indicating that the connectivity and aggregation of the patches increased further; thus, the land use in the BUA in Beijing increased the connectivity of the land types during these 5 years. The development and utilization of the BUA in Beijing increased the connectivity and aggregation of the land types. The AI of the bare land increased from 50.2 to 54.7, an increase of 4.5, which was the most significant increase among all of the types, indicating that there were fewer major changes within the original BUA, and the significant changes were brought about by the expansion of the BUA.

4.4. Multi-Scale Joint Analysis at the City Block Level

As can be seen from Figure 9, in the outward expansion area of the BUA from 2016 to 2021, the block building density was higher in the northwest and southeast areas of the BUA, and it had a certain scale initially. However, the overall block building density in the outward expansion area of the BUA remained in the early stage of construction during these 5 years. Within the part of the BUA that remained unchanged during the 5 years, the urban center did not change much, and the increase in the block building density was more significant in the periphery of the adjacent urban center. The overall building density increased steadily with fewer peaks, indicating that Beijing, which had a high level of urbanization, was still in the process of rapid and stable urbanization during the period of 2016–2021. As shown in Figure 10, the overall vegetation density decreased at the block scale, while the vegetation density remained high in the northern part of the BUA. This indicates that the accelerated urbanization process had a certain impact on the urban greening in the city center and adjacent areas, while it had less of an impact on the northern part of the BUA, maintaining a high greening rate at the block scale.

As can be seen fromFigure 11a,b, the significant increases in the BCR were mainly concentrated within the intervals [0.2, 0.4] and [0.4, 0.6], the significant decreases were in the interval [0, 0.2], and there were no significant changes in the other intervals. The number of blocks within intervals [0.2, 0.4] and [0.4, 0.6] increased from 364 and 594 in 2016 to 454 and 801 in 2021, with increases of 90 and 347, respectively, and their percentages increased from 11.7% and 19.1% to 13.6% and 23.9%, with increases of 1.9% and 4.8%, respectively, exhibiting a general trend of increasing building coverage, thus reflecting the continued gradual increase in the urbanization process. Within interval [0,0.2], the number of blocks decreased from 1458 in 2016 to 1377 in 2021, a decrease of 81, and the percentage decreased from 47% to 41.1%, a decrease of 5.9%, with some undeveloped and underutilized plots or low-developed plots undergoing gradual development and utilization. In summary, it can be concluded that the urban area of Beijing City experienced steady urbanization and a stage of increasing utilization from 2016 to 2021.

As can be seen from Figure 11c,d, the number of blocks with VF in intervals [0, 0.2] and [0.2, 0.4] increased from 678 and 579 in 2016 to 1440 and 1006 in 2021, with increases of 762 and 427, respectively, and their percentages increased from 21.9% and 18.7% to 43% and 30.1%, with increases of 21.1% and 11.4%, respectively. Meanwhile, the number of VFs in intervals [0.4,0.6], [0.6,0.8], and [0.8,1.0] decreased from 908, 592, and 342 to 481, 276, and 144, with decreases of 566, 316, and 198, and their percentages decreased from 29.3%, 19.1%, and 11% to 14.4%, 8.2%, and 4.3%, with decreases of 14.9%, 10.9%, and 6.7%, respectively. The significant increase in interval [0, 0.4] and the significant decrease in interval [0.4, 1] reflect the increase in VF as contiguous vegetation blocks were exploited. This also reflects the direct change in and occupation of the urban greenspace landscape pattern due to the increase in urbanization processes from 2016 to 2021.

5. Discussion

5.1. Comparison and Analysis of the Accuracy of Urban Land-Cover Mapping

In this section, the classification accuracy, applied methods, and important sources of errors are discussed. First, the land-cover classification results are consistent with the area of the visual interpretation and validation data results. Therefore, the remote sensing classification method proposed is credible in this study area. In the classification and dynamic analysis of urban land-cover at the scale of a BUA, medium-resolution remote sensing images are still more widely used. Compared with medium-resolution remote sensing images used for the land-cover classification mapping of BUA [71,72], the overall accuracy of the method developed in this study is higher (1.0–2.0% higher), the land-cover types are richer, and the spatial patterns of the different types of land-cover are portrayed better. For example, HRRS images can reflect the different spatial patterns of buildings, such as single buildings, continuous buildings, and large factory buildings. It can be concluded that the land-cover mapping results based on HRRS images are more reasonable for constructing a multi-scale analysis framework that includes rich object types, urban blocks, and BUAs. At present, most of the studies on HRRS image land-cover mapping focus on method testing or dynamic analysis of local typical urban areas [73], but the classification method constructed in this study for BUA land-cover and dynamic analyses was applied to the complete BUA in two time periods, indicating that this study has a certain degree of application demonstration value and reference significance. In addition, in terms of accuracy, although the classification accuracy was similar to that of other typical urban areas (within 2.0%) [74,75], the larger the selected study area is, the more complex the land-cover type characteristics reflected in the HRRS images are, according to the comprehensive and complex characteristics of cities. Therefore, the classification accuracy can be maintained under more complex application scenarios, further indicating the effectiveness of this research method.

There are two reasons for the decrease in accuracy. The first reason is the insufficient spectral information. Although spatial features, such as the morphology and texture, and comprehensive indices were used to supplement the spectral features in this study to improve the overall classification accuracy, the complexity of the scenes and the diversity of the surface types in Beijing, a mega-city in China, made the same object with different spectra or different objects with the same spectrum phenomena more notable. Thus, supplementary spatial information cannot completely replace spectral information, which affects the classification accuracy. The second reason is that, to ensure the classification accuracy and avoid the influence of noise in the incoming bands, such as clouds and fog, some of the selected image time phases were not in the vegetation growing season, which led to lower extraction results for some vegetation. However, because spatial information was used to enhance the vegetation, the confusion between vegetation and bare land was greatly reduced. In addition, the high accuracy of the extraction results of the buildings and shadows also indirectly reflected the correct trend of the reduction in the vegetation types. Therefore, for HRRS images, the non-growing-season images will cause a certain degree of error in the extraction of the vegetation type information, but this will not cause a notable effect.

5.2. Comparison and Analysis of the Effects of Urban Landscape Patterns

The results of this study were consistent with the general trend of the results of the analysis of urban landscape patterns at a single scale (e.g., BUA, urban type scale, and city block scale) in the current study [76,77,78], indicating the validity of this study’s analysis. Among them, in terms of urban expansion, compared with other similar studies [79,80], the joint multi-scale evaluation method constructed by this study’s methodology can quantitatively describe the external expansion of cities while also depicting the internal expansion of cities, thus reflecting the degree of urbanization more accurately. In particular, for cities whose external expansion is gradually stabilizing while their internal expansion is still on the rise, this research method can more accurately reflect the urbanization process of such cities than that when relying on a single external expansion evaluation index. In addition, the BUA is widely studied and applied as a complete urban ecosystem unit. Compared with most studies that use medium-resolution remote sensing images to obtain complete information regarding the extent of BUAs and their internal land cover [81,82], the urban landscape pattern evaluation indexes constructed based on HRRS images in this study can provide high-dimensional and more comprehensive characteristics of urban sprawl and internal spatial patterns, further reflecting their rationality and necessity. In analyzing the spatial pattern characteristics within the city, the multi-scale composite index constructed in this study, based on refined land-cover mapping and using city blocks as the landscape analysis unit, is less likely to segment the spatial characteristics of the landscape pattern and is able to present the ecological functional characteristics of the city more comprehensively than the regular grid [38] or city block scales [83] used in previous studies.

The impact of this study on the landscape pattern analysis method is mainly focused on two aspects: On one hand, it is focused on the construction of landscape pattern evaluation units. Although this study adopted the city blocks constructed by the modified urban road network data to express the urbanization process and the evaluation units were more reasonable, the lack of unified standards for the construction of city block units for remote sensing applications caused the results of landscape pattern analysis under different evaluation units to fluctuate, to a certain extent. On the other hand, it is focused on the limitations of landscape pattern evaluation indexes. Although the refined statistics of land-cover types, the construction of typical landscape pattern analysis indicators, and the expansion of multi-scale joint analysis methods in this study can effectively express the landscape pattern characteristics of cities, the comprehensive and complex nature of cities makes it difficult for the current indicators to fully and comprehensively portray the urban landscape pattern, thus leading to a certain degree of variation in the studies of different scholars, but, in general, the trends were generally consistent and did not have a significant impact on the final analysis.

5.3. Limitations and Prospects of the Study

The advantage of the classification method developed in this study is that the automatic mapping of land cover was satisfactorily achieved by combining multi-source features according to the characteristics of different urban land-cover types and using the object–pixel–spatiotemporal expert knowledge method. The shortcoming of this method is that the insufficient spectral information from HRRS images leads to inevitable mixing and omission problems, which make the classification results less accurate. Therefore, in subsequent research, we will try to introduce data sources with the advantages of spectral or temporal dimensions and will construct a multi-source data fusion extraction method to further improve the accuracy of urban land-cover classification. In addition, in this study, the update of BUA boundaries was performed by the manual interpretation of HRRS data, which can guarantee accuracy but is also more time-consuming. In contrast, the current study on the application of night-light data was able to identify the BUA boundaries better at a large scale. Therefore, we will further study the fusion method of night-light data and HRRS image data for developing an automatic extraction method of BUAs at a detailed scale.

In terms of urban landscape pattern analysis methods, the multi-scale joint landscape pattern analysis method constructed in this study can provide a comprehensive analysis of the external expansion, internal expansion, and internal spatial pattern characteristics of a city, but due to the limitation of the data sources of high-definition remote sensing images, the current studies have all been focused on horizontal dimensional expansion and have not yet involved the study of the vertical dimension of cities. At present, with the development of satellite technology, there have been research advances in urban building height [84,85] or vegetation height monitoring [86,87]. Therefore, in the subsequent research, we will further try to combine satellite data with altimetry capability loads to obtain refined land-cover height information and be able to further construct 3D urban landscape pattern evaluation indexes, such as the volume, density, and volume ratio, to deeply characterize the urban landscape pattern from higher dimensions.

6. Conclusions

The focus of this study was to make full use of multi-period HRRS images to extract land-cover information for BUAs and their internal areas. Based on this, we constructed a framework for multi-scale joint analysis and evaluation methods, such as urban administrative district–BUA–urban block–land type elements to achieve the joint macro–micro dynamic change analysis and evaluation of cities. In this study, we took Beijing, the capital of China and also a mega-city in China, as an example to test and verify the effectiveness of the newly developed method. The main conclusions of this study are as follows.

(1)

From the perspective of multi-temporal urban land-cover mapping, the spatiotemporal extraction method for intra-city land cover developed in this study can ensure the accuracy of the multi-period land-cover extraction results. Therefore, it provides a guarantee for the analysis of multi-temporal urban spatial pattern changes.

(2)

The joint multi-scale urban spatial pattern dynamic analysis method developed in this study can better overcome the analytical limitations caused by the use of single-scale evaluation methods. In terms of analyzing the expansion of BUAs, the joint analysis method of the object–BUA–urban administrative area layer developed in this study can jointly analyze the expansion of BUAs from the two levels of their external expansion and internal expansion in a refined way, overcoming the problem of a single data source only expressing the expansion of BUA. In addition, a multi-level joint analysis method of BUA–block–object was constructed within the BUA to analyze the spatial patterns of and changes in land-cover types within the city, which makes up for the problem that a single scale can only reflect the change trajectory of land-cover types.

(3)

We found that 2 m-resolution remote sensing images, whether used in the high-precision and high-resolution land-cover results provided or in the evaluation method developed in this study, were applicable. However, when 30 m Landsat data were applied for this purpose, it was difficult to guarantee the evaluation effect, such as the spatial information provided, due to the resolution.