Using Landscape Metrics of Pixel Scale Land Cover Extracted from High Spatial Resolution Images to Classify Block-Level Urban Land Use

Abstract

Block-level urban land use classification (BLULUC), like residential and commercial classification, is highly useful for urban planners. It can be achieved in the form of high-frequency full coverage without biases based on the data of high-spatial-resolution remote sensing images (HSRRSIs), which social sensing data like POI data or mobile phone data cannot provide. However, at present, the extraction of quantitative features from HSRRSIs for BLULUC primarily relies on computer vision or deep learning methods based on image signal characteristics rather than land cover patterns, like vegetation, water, or buildings, thus disconnecting existing knowledge between the landscape patterns and their functions as well as greatly hindering BLULUC by HSRRSIs. Well-known landscape metrics could play an important connecting role, but these also encounter the scale selection issue; i.e., the optimal spatial unit size is an image pixel or a segmented image object. Here, we use the task of BLULUC with 2 m satellite images in Beijing as a case study. The results show the following: (1) pixel-based classification can achieve higher accuracy than segmented object-based classification, with an average of 3% in overall aspects, while some land use types could reach 10%, such as commercial land. (2) At the pixel scale, if the quantity metrics at the class level, such as the number of patches, and the proportion metrics at the landscape level, such as vegetation proportion, are removed, the accuracy can be greatly reduced. Moreover, removing landscape-level metrics can lead to a more significant reduction in accuracy than removing class-level metrics. This indicates that in order to achieve a higher accuracy in BLULUC from HSRRSIs, landscape-level land cover metrics, including patch numbers and proportions at the pixel scale, can be used instead of object-scale metrics.

1. Introduction

Block-level urban land use classification (BLULUC) identifies dominant land use functions, such as those shown in Figure 1, including residential, commercial, and industrial, within continuous urban blocks defined by roads, rivers, or other boundaries.

This approach enhances land use management and planning by revealing the spatial distributions and interactions of functional areas [1,2,3]. While social sensing data such as those derived from POIs and social media can reflect human activities, they lack consistent spatial–temporal coverage and long-term reliability [4,5,6]. In contrast, high-resolution remote sensing imagery offers consistent, large-area coverage with the capacity for detailed identification of urban features like buildings, vegetation, and water bodies. This enables the analysis of geometric, morphological, and contextual characteristics within urban blocks, improving the accuracy of land use inference when combined with spatial layout indicators [7,8,9].

The effective expression of spatial patterns in high-resolution remote sensing imagery, and particularly their connections to known relationships between pattern and function, is essential for the broader application of this technology in landscape planning and management. The current methods for spatial pattern representation fall into two main categories: feature engineering, which mimics human visual perception, and deep learning, which models the brain’s neural mechanisms to automatically learn features. Feature engineering is intuitive and interpretable; by designing features such as color, texture, and shape, even non-specialists can understand the classification logic. Common examples include the Normalized Difference Vegetation Index (NDVI) for color and the Gray Level Co-occurrence Matrix (GLCM) for texture, both of which enhance classification accuracy and robustness. These methods also streamline data processing and encourage interdisciplinary collaboration [10,11]. However, they rely heavily on subjective feature design and often fail to capture the full complexity of spatial patterns, which may compromise classification performance.

Deep learning models, inspired by neural networks in the human brain, can autonomously learn high-level feature representations, as a result of which they are widely adopted in remote sensing research. Their multilayer architecture allows for the integration of spatial information across multiple scales, which is particularly effective for capturing complex spatial patterns that traditional feature engineering often misses [12,13,14]. Moreover, deep learning’s vast parameter space enhances its ability to distinguish between similar-looking but different samples as well as dissimilar-looking but related samples. However, these advantages come at the cost of interpretability. Unlike feature engineering, deep learning operates as a “black box”, making its internal decision-making processes difficult to explain and limiting its transparency in practical applications [9,15].

Regardless of whether they adopt feature engineering or deep learning approaches, many currently used methods simulate human visual perception and cognitive mechanisms rather than being grounded in established knowledge of spatial or landscape patterns. This often results in a misalignment between extracted features and the semantic meanings of geographical or landscape metrics, hindering integration with existing knowledge in this domain. On one hand, this misalignment raises the barrier for understanding and acceptance especially considering the wide demand for land use/land cover change (LUCC) analysis and significant regional variations, thus impeding interdisciplinary collaboration and fostering distrust among scholars from different fields [16]. On the other hand, the substantial existing data products and accumulated knowledge cannot be effectively incorporated, limiting further improvements in interpretation accuracy [17,18]. Addressing this gap calls for the construction of explicit feature representations for high-resolution remote sensing imagery, which are grounded in an understanding of geographical or landscape spatial structure. Particularly given the limited access to large-scale, high-quality samples, such application-oriented representations demonstrate growing potential [19].

Landscape metrics are simple, easily understandable, and widely applied quantitative indicators that represent the composition and spatial configuration of landscape structures, showing great potential for use in the explicit representation of spatial pattern features of land use at the urban block level in high-resolution imagery [20,21,22]. At the regional level, landscapes are usually much larger than at the urban block level. Landscape metrics are primarily used to analyze mapped land use or land cover data in order to explore pattern differences between different regional landscapes, thereby revealing their functional differences [23,24,25]. This is because landscape ecology typically posits that pattern determines function, which in turn affects pattern [26]. As the detail-recognition capacity of high-resolution imagery improves, it becomes possible to identify and analyze land cover classification details at the urban block level [27]. As illustrated in Figure 2, a block unit can be viewed as a small landscape system, allowing the use of landscape metrics to measure the spatial pattern features of each block unit and thus infer its landscape function, namely, its land use type [28,29].

Existing case studies have shown that landscape metrics outperform traditional visual features in land use classification at the urban block level [30,31,32]. Compared to visual features and deep learning, landscape metrics are clear in meaning and widely used in geography and landscape studies, indicating that they hold significant potential for application in land use classification research at the urban block level, particularly in the context of diverse application targets.

The issue of scale is a key factor when calculating landscape metrics for land use classification at the urban block level, yet it has not received sufficient attention in previous studies [33,34]. The scale issue here mainly concerns whether to use pixels of high-resolution remote sensing imagery as the basic spatial unit for calculating and expressing spatial heterogeneity within urban block units or to use segmented objects from high-resolution imagery within urban block units. In the absence of image scale transformation, pixels are the smallest spatial unit from which sensors can gather information. However, the ground dimensions corresponding to pixels are typically smaller than the dimensions of buildings, green spaces, water bodies, and other targets within the urban block units [35]. When classifying land cover within urban block units using pixels as the basic spatial unit, challenges often arise in expressing features like texture and topology. Using segmented objects as the basic unit can help to more effectively address these expression challenges. This is also the most common method adopted in current research classifying land use within urban block units [36,37]. However, image segmentation often faces the challenge of optimizing segmentation scales, meaning that the ground sizes of segmented objects may not match the sizes of the corresponding land cover features, which can significantly impact the expression of the segmented object’s features. Although using pixels as the basic unit for classifying land cover within urban block units poses difficulties related to expressing topological features, it can also reflect very fine features. Since landscape metrics also have a similar information-concentrating function as segmented objects, the final classification results produced for urban block units may not necessarily be inferior to those obtained using object-based methods. If the pixel-based method is favored, it can also be used to avoid the time-consuming processes and scale optimization issues associated with object segmentation [38,39].

Selecting effective landscape metrics is a key influencing factor at play in the scale issue, which has not been systematically explored in previous research. There are numerous landscape metrics, each reflecting different spatial patterns and configuration information. From the perspective of the hierarchical structure, the metrics can be divided between the patch level, the class level, and the landscape level [40,41]; based on the attributes of the metrics themselves, Fu et al. summarized them into four main categories—size, shape, type, and proportion [42]. When specifically selecting landscape metrics, we typically face two types of questions: Which types of landscape metrics usually achieve better recognition performance? For a specific type of urban block, which types of landscape metrics perform better? No previous has conducted systematic empirical research, categorized by type, to answer these questions.

Given the potential use of landscape metrics in spatial pattern expression, this study uses 2 m resolution satellite images of Beijing, China, to devise a case study to systematically explore the impacts of scale issues on block-level land use classification through regional empirical research. The research will address the following questions:

(1) When classifying land cover within urban block units, is it better to use pixels or objects as the basic spatial unit?

(2) Under the determined optimal spatial unit, which types of landscape metrics can most effectively improve the classification results?

2. Materials and Methods

2.1. Study Area and Data

This study covers an area of approximately 2267 km² within the sixth ring road of Beijing, which is located in an arid to semi-arid climate zone with very little cloud cover. This is conducive to obtaining high-resolution remote sensing images, enabling strict comparative experiments under consistent meaning conditions and reducing the impacts of spatiotemporal changes on spatial patterns. As the political, cultural, and economic center of China, Beijing is highly urbanized and features a rich variety of land use types, including residential, commercial, industrial, urban green, and farmland. The diversity of its land use and the significant heterogeneity of the spatial layout within urban blocks make the region an ideal location for evaluating the effectiveness of landscape metrics in capturing urban spatial patterns. For example, taking residential land as an example, the spatial arrangement of building density and structure exhibits significant heterogeneity and diversity, as shown in the attached figure (see Appendix A, Figure A1). In addition, the research area is large enough to cover different heterogeneous regions between the second ring and the sixth ring, and the continuity in these regions gives rise to spatial sample errors in model training and validation. The large-scale and contiguous nature of the study area enables rigorous comparative experiments under consistent imaging conditions, thereby avoiding the subtle impacts of temporal and spatial variations on spatial patterns. Furthermore, the use of full-coverage ground truth data for validation, as opposed to sampling-based approaches, enhances the reliability and reproducibility of the results. These factors all contribute to the potential transferability of the proposed methodology and findings to other urban contexts.

The high-resolution images used in this study were acquired by the Gaofen-1 satellite in 2021, consisting of one scene from February and two scenes from April, which were synthesized into multispectral image data. To ensure data quality, a series of preprocessing steps was applied to the raw images. First, radiometric calibration was performed using ENVI 5.3 software to eliminate the effects of sensor and atmospheric conditions on the images. Second, geometric correction was applied to ensure the accuracy of geographic coordinates and projection information. Finally, the images were cropped and mosaicked based on the study area’s boundaries with the careful selection of cloud-free or minimally cloud-covered scenes from the 13 available Gaofen-1 satellite images. The spatial resolution of the processed images is approximately 2 m with four spectral bands: red, green, blue, and near-infrared. At this resolution, detailed information on urban features, such as buildings within city blocks, vegetation, water bodies, and shadows, can be well captured.

In this study, road network data from OpenStreetMap (OSM) have served as the primary source for block division. The fine landscape unit division OSM data are freely available and have extensive global coverage, thus contributing to the possibility of replicating and promoting this experiment. Importantly, the attributes of the OSM road network include classification information, allowing us to select the road levels that are most suitable for dividing block units. This helps prevent the use of internal roads within block units for division, thus maintaining their integrity. Specifically, the road classifications selected for this study include highways; primary, secondary and tertiary roads; and trunk roads, as shown in Figure 3.

2.2. Experimental Methods Process

The research methodology flowchart, as illustrated in Figure 4, comprises six modules, which are grouped into three sequential stages for greater conceptual clarity: (a) the creation of a standardized land use classification dataset at the block level for the entire study area (data preparation); (b–d) land cover classification using pixels as the basic spatial unit, object-based classification with different segmentation scales, and the calculation and grouping of landscape metrics treating block units as micro-landscape systems (feature engineering); and (e–f) block-level land use classification using landscape metrics as features along with accuracy evaluation followed by a comparative analysis of the classification results (model training and evaluation).

2.2.1. Creation of a Standard Dataset for Block-Level Land Use Classification Across the Entire Region

Unlike the sampling methods used in most previous studies, this research employed an entire-region coverage approach to create a standardized block-level land use classification dataset. This dataset is used to train the classification model and evaluate classification accuracy, covering the entire area with Beijing’s sixth ring road. The full-area coverage approach effectively avoids accuracy evaluation biases caused by spatial sampling inconsistencies, resulting in a classification that more accurately reflects real-world conditions in the region, such as sample balance and spatial differentiation issues [43].

The vector data of urban roads from the OSM (OpenStreetMap) are primarily used as a source for delineating block units. Typically, the land use and functional attributes within these areas are relatively consistent. Due to some errors in the open-source road data and the fact that certain blocks are not fully divided by road networks, preprocessing is required during the division process. This includes deleting duplicate or incomplete sections, using water systems and rivers as auxiliary references, and even correcting the road network with high-resolution satellite images to ensure the generated block units are closed and accurate. During manual interpretation, high-resolution remote sensing images and online maps were mainly used for visual interpretation. The greatest challenge encountered during visual interpretation was in areas where blocks had mixed functions, such as commercial and residential mixed-use areas. In these regions, the primary function was determined by analyzing the structure, density, and surrounding environments of buildings in the images and maps along with urban view images and field surveys. The corresponding imagery for each land use type is shown in Figure 5, and the classification system is based on a similar study conducted in the same region [44], as outlined in

Table 1. Classification system.

Land Use Type	Definition
Commercial	Mainly includes financial centers, retail centers, shopping centers, office buildings, and other areas primarily used for commercial activities.
Villa	High-end residential areas.
Ordinary residence	Multi-story and high-rise residential buildings.
Intensive residence	Dense residential areas and rural settlements.
Institution	Mainly includes educational, medical, cultural, administrative offices, and public services.
Industrial	Mainly includes light and heavy industrial factories and warehouses.
Transport	Transportation hubs, train stations, and airports.
Undeveloped	Vacant land, bare land, and land under construction.
Urban Green	Urban parks, botanical gardens, zoos, golf courses, and other artificial grasslands.
Woodland	Forests and shrubs in natural vegetation.
Farmland	Mainly vegetable plots, arable land, orchards, and other agricultural land.
Water	Natural and artificial water bodies.

.

2.2.2. Land Cover Classification Within Blocks Using Pixels as Basic Spatial Units

Pixel-based land cover classification methods directly classify each pixel in remote sensing images, mainly based on the spectral information of the pixels. Compared to classification based on segmentation objects using multiple pixels as the basic units, pixel-based classification can reflect the most detailed spatial-scale information obtained by sensors. However, due to the lack of comprehensiveness, pixel-based classification has limitations in reflecting cross-pixel spatial combination information in remote sensing images. Considering that the subsequent calculation of landscape metrics in this study also needs spatial combination information derived from pixels, from the perspective of overall land use classification, pixel-based land cover classification methods may not necessarily be worse than object-based classification methods at the current stage. This is also the main focus of discussion in this study.

To ensure the reproducibility of the experiment, we adopted an index-threshold-based layer-by-layer thematic extraction method for pixel-based land cover classification within block units. Specifically, we have chosen four main types that have a significant impact on the spatial pattern of the urban area as the land cover classification system: namely, buildings, water bodies, vegetation, and other types (remaining land cover types) [8]. First, we used the Normalized Difference Vegetation Index (NDVI) [45] to extract vegetation features from the imagery. Based on the actual distribution of land cover observed in the remote sensing image, we manually selected the optimal threshold that corresponds to the real distribution of the land cover types, classifying pixels below the threshold as vegetation and those equal to or above the threshold as non-vegetation areas. Next, within the non-vegetation areas, we employed a similar strategy to distinguish between water bodies and non-water bodies, using Normalized Difference Water index (NDWI) [46] feature extraction. Following this, in the non-water areas, we applied a similar strategy to differentiate non-building areas into buildings and others, using the Morphological Building Index (MBI) [47] as the feature. The specific methodological process is illustrated in Figure 6.

2.2.3. Land Cover Classification Within Blocks Using Different Segmentation Scale Objects as Basic Spatial Units

The object-based land cover classification method combines pixels into larger objects or regions to capture the spatial relationships and contextual information between features, achieving higher classification accuracy. The segmentation of objects directly influences their feature expression and the scale of information acquisition, making it a critical issue in object-based classification. We employed the widely used MRS segmentation algorithm and selected three relatively optimal scale parameters for segmentation using the ESP method—58, 95, and 137. The shape and compactness parameters were set to default values of 0.1 and 0.5, respectively. Notably, this segmentation was conducted on the imagery within the block units, and it was constrained by the boundaries of those units.

2.2.4. Landscape Metrics Calculation and Grouping with Block Units Treated as Micro-Landscape Systems

To date, numerous landscape metrics have been proposed in the field of landscape ecology, and some of these metrics exhibit correlations. A systematic grouping and understanding of these metrics’ features is crucial in order to comprehend their expressions. According to the widely used Fragstats software, landscape metrics can typically be categorized into three levels from a hierarchical perspective—individual patch metrics, patch-type metrics, and overall landscape metrics [48]. Individual patch metrics primarily reflect the geometric features of single land cover patches, while patch-type metrics and overall landscape metrics focus on the spatial distribution, complexity, and interrelationships of land use types or land covers at the regional scale. When classifying land use at the block level, the comprehensive spatial pattern information derived from the land cover classification within block units is mainly utilized. Therefore, this study aims to select metrics from both the class level and the landscape level for calculation and analysis. In selecting metrics at the landscape metric level, this research references the summary by [42,49] of five independent dimensions of landscape metrics’ features—size, shape, number, type, and configuration—selecting representative metrics for each dimension in sequence. Following this, metrics representing the dimensions of size, shape, number, and configuration are also chosen at the class level. Consequently, a total of nine groups of landscape metrics have been selected, comprising 24 specific landscape metrics. Detailed information on the groupings, names, formulas, and meanings of these metrics is presented in Appendix A 

Table A1. Class-level landscape metrics grouping.

Group Name	Selected Metrics	Calculation Formulas	Metrics Meaning
CLASS_SIZE	Mean Building Area	$M_{i}A={\displaystyle \frac{1}{n_i}}{\displaystyle \sum _{j=1}^{n_i}}{a}_{i,j}$	This metric reflects the distribution density of building patches. A larger average area may suggest a more continuous building distribution.
	Mean Vegetation Area	$a_{i,j}$ is the area of the j-th patch in class i and $n_i$ is the number of patches in class i.	This metric reflects vegetation patch density, indicating coverage and distribution within the block.
	Mean Water Area		This metric reveals water body patch density, assessing coverage and distribution within the block.
	Mean Other Area		This metric reflects the distribution density of patches from other categories, helping to understand their coverage extent and spatial distribution within the block.
CLASS_SHAPE	Mean Building Shape Index	$SHAPE{\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}}_i={\displaystyle \frac{0.25p_{i,j}}{n_i\sqrt{a_{i,j}}}}$ $p_{i,j}$ is the perimeter of the $j$-th patch in class i, $a_{i,j}$ is the area of the $j$-th patch in class i, and $n_i$ is the number of patches in class i.	This metric reflects building patch complexity, including fragmentation and edge shapes, aiding in spatial structure analysis.
	Mean Vegetation Shape Index		This metric reflects the complexity of vegetation patches, aiding in analyzing their spatial structure and morphology.
	Mean Water Shape Index		This metric reflects water body patch complexity, aiding in evaluating spatial structure and morphology.
	Mean Other Shape Index		This metric reflects the complexity of patches from other categories, aiding in the analysis of their spatial structure and morphology.
CLASS_AMOUNT	Building Number of Patches	$N{P}_i=n_i$ $n_i$ is the number of patches in class i	This metric indicates the number of buildings, aiding in assessing distribution density and spatial patterns within the block.
	Vegetation Number of Patches		This metric reflects the building count, aiding in the assessment of distribution density and spatial patterns.
	Water Number of Patches		This metric reflects the water body patch count, aiding in the assessment of distribution density and spatial patterns within the block.
	Other Number of Patches		This metric reflects the patch count of other categories, aiding in the assessment of distribution density and spatial patterns within the block.
CLASS_COMBINATION	Mean Building Euclidean Nearest Neighbor	$ENN{\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}}_i={\displaystyle \frac{1}{n_i}}{h}_{i,j}$ $h_{i,j}$ is the distance to the nearest neighboring patch of the same class as vegetation patch j; $n_i$ is the number of patches in class i.	This metric reflects the average nearest neighbor distance between building patches, aiding in assessing clustering or dispersion within the block. A smaller distance suggests higher clustering.
	Mean Vegetation Euclidean Nearest Neighbor		This metric reflects the average nearest neighbor distance between vegetation patches, aiding in the analysis of clustering or dispersion. A smaller distance suggests higher clustering.
	Mean Water Euclidean Nearest Neighbor		This metric reflects the average nearest neighbor distance between water body patches, aiding in assessing clustering or dispersion within the block. A smaller distance suggests higher clustering.
	Mean Other Euclidean Nearest Neighbor		This metric reflects the average nearest neighbor distance between patches of other categories, aiding in assessing their clustering or dispersion within the block.

and

Table A2. Landscape-level landscape metrics grouping.

Group Name	Selected Metrics	Calculation Formulas	Metrics Meaning
LANDSCAPE_SIZE	Area Mean	$AM={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^n}{A}_i$ $A_i$ is the area of class i N is the total number of patches in the landscape.	This metric reflects the average patch area, aiding in assessing size distribution, continuity, and fragmentation within the block.
LANDSCAPE_SHAPE	Mean Shape Index	$MSI={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{\displaystyle \frac{{0.25P}_i}{A_i}}$ N is the total number of patches in the landscape; $A_i$ is the area of class i; $P_i$ is the perimeter of class i.	This metric reflects the average patch shape complexity, helping assess morphological patterns and spatial structure in the block.
LANDSCAPE_AMOUNT	Number of Patches	$N{P}_i=N$, where N is the total number of patches in the landscape.	This metric reflects the number of patches in the block, aiding in assessing fragmentation, structural features, spatial distribution, complexity, and continuity.
LANDSCAPE_PROPORTION	Building Proportion	$PLAND={\displaystyle \frac{A_i}{TA}}$ $A_i$ is the total area of patches of class i, and $TA$ is the total area of all patches.	This metric reflects the relative distribution of buildings in the block, aiding in the analysis of their spatial distribution patterns.
	Vegetation Proportion		This metric reflects the relative distribution of vegetation in the block, aiding in assessing its spatial distribution pattern.
	Water Proportion		This metric reflects the relative distribution of water bodies in the block, aiding in assessing spatial patterns for water resource management and environmental protection.
	Other Proportion		This metric reflects the relative distribution of patches from other categories, aiding in the analysis of their spatial distribution patterns within the block.
LANDSCAPE_COMBINATION	Shannon Diversity Index	$SHDI=-{\displaystyle \sum _{i=1}^m}{p}_i\ln \left(p_i\right)$ $P_i$ is the proportion of class i in the landscape, and m is the total number of classes in the landscape.	This metric measures the richness and evenness of patch classes, reflecting land cover diversity. A higher value indicates greater diversity.

.

In landscape ecology, the software FRAGSTATS is commonly used as an effective tool for calculating landscape metrics; however, it is limited by the size of the dataset [50]. Typically, calculating metrics for large areas requires clipping using the vector boundaries of a significant number of land cover information from block units, which results in low computational efficiency [51]. Therefore, this study utilizes the zonal analysis module from the landscape metrics calculation library PylandStats in Python to develop a program for the fast computation of the summarized metrics [52]. Finally, the calculated feature attributes are linked to the block units to achieve feature extraction from the dataset.

2.2.5. Land Use Classification at the Block Level Based on Landscape Metrics as Features and Accuracy Evaluation

Considering the diversity of land cover types within blocks, the uneven spatial distribution of different land use types, and the landscape definitions of certain features, null values may arise during calculations, leading to data sparsity. For example, certain blocks may lack specific land cover types, such as water bodies, resulting in null values for shape features. A typical example is the Mean Water Shape Index, which may produce null values when the denominator (i.e., the number of patches) is zero due to the absence of water bodies in certain blocks. To address this, the study employs XGBoost, which is a model that is well suited for handling sparse data and missing values. Its built-in sparsity-aware algorithm automatically detects patterns of missingness and optimizes split directions during training, enhancing model robustness and efficiency. This capability is especially critical when dealing with landscape metrics that inherently contain nulls due to spatial heterogeneity. The focus of this study is to compare the classification performance of pixel-based versus object-based landscape metrics at the block scale rather than to assess differences among classification algorithms. Therefore, a single, reliable, and high-performance model was adopted to ensure consistency in comparative evaluations. In the experiments, the model parameters were optimized through a grid search approach to ensure optimal performance. The final parameter settings were as follows: random_state = 100 for reproducibility, n_estimators = 100 to build 100 decision trees, max_depth = 10 to limit tree depth and prevent overfitting, learning_rate = 0.1 to control the contribution of each tree, and reg_alpha = 1 to apply L1 regularization, which penalizes large coefficients and further reduces overfitting. These settings were selected to balance model accuracy and complexity, and their impacts on classification results were carefully evaluated. While minor adjustments to these parameters (e.g., increasing n_estimators or max_depth) could lead to slight improvements in accuracy, the chosen configuration provided a robust and computationally efficient solution for the task at hand.

The model classification prediction process involved random sampling, where 5% of samples from each category were selected as the training set, ensuring that at least one sample from each category was chosen. The remaining data were used for testing. The experiments were repeated 100 times to avoid issues where insufficient sample sizes in certain categories could manifest inadequate predictive capabilities in the model. Furthermore, through multiple iterations, the model’s performance stability and consistency across different datasets could be comprehensively assessed to validate its generalization ability. To explore the performances of different types of landscape metrics across various land use types in detail, this study chose single-class evaluation metrics that could exclude sample size imbalance for accuracy evaluation, specifically mean precision, mean recall, and F1 score, to assess the model’s performance. The calculation formulas can be found in (1) to (3):

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(1)

In this context, TP (True Positive) is the number of samples correctly predicted as positive. For example, in the context of residential land use, TP represents the number of blocks correctly identified as residential. FP (False Positive) is the number of samples that are actually negative but predicted as positive. For example, FP represents the number of non-residential blocks incorrectly classified as residential. Precision quantifies the proportion of blocks predicted as residential that are actually residential. For multi-class classification tasks, the mean precision is the average of the precision values across all categories.

Recall measures the proportion of correctly predicted positive samples among all samples that are actually positive. Its calculation formula is

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(2)

In this context, TP (True Positive), as defined above, is the number of samples correctly predicted as positive. FN (False Negative) is the number of samples that are actually positive but predicted as negative. For example, FN represents the number of residential blocks incorrectly classified as non-residential.

Recall quantifies the proportion of actual residential blocks that are correctly predicted as residential by the model. Additionally, to gain a more detailed understanding of the model’s performance across different categories, the mean precision and mean recall for each category are also calculated. The F1 score, as the harmonic mean of precision and recall, better balances these metrics in object detection, is sensitive to extreme values, suits imbalanced data, and has wide applications. Its calculation formula is

$$\mathrm{F}1=2\times {\displaystyle \frac{\mathit{Precision}\times \mathit{Recall}}{\mathit{Precision}+\mathit{Recall}}}$$

(3)

2.2.6. Comparison of Block-Level Urban Land Use Classification

The primary objective of this study was to explore the optimal scale for landscape metric feature representation through comparative experimental results. To achieve this goal, three sets of experiments were designed. First, pixel-based and multi-scale segmentation-based block-level land use classifications have been compared using all nine selected groups of 24 landscape metric features to determine the optimal scale for landscape pattern representation.

Based on this, two additional experiments have been designed to further explore the impact of pixel scale on classification accuracy. Experiment 1 involves training a baseline XGBoost model with all 24 features and then recording its classification accuracy. Then, features are systematically removed one group at a time, the model is retrained, and the classification accuracies of the baseline and the modified models and compared. Experiment 2 involves training models separately using class-level and landscape-level features and then comparing their overall precision and category-specific accuracies. These experiments aim to elucidate the classification capabilities of landscape metric features at different levels and identify effective feature combinations, providing strategies for optimizing classification accuracy in future research.

3. Results

3.1. Region-Wide Coverage Standard Dataset of Block-Level Land Use

Figure 7 and

Table 2. Detailed information on land use types in the dataset.

Class	Amount	Amount Percentage	Area Percentage
Commercial	1431	16.28%	5.20%
Villa	159	1.81%	2.20%
Ordinary residence	2258	25.69%	13.61%
Intensive residence	741	8.43%	6.53%
Institution	372	4.23%	3.60%
Industrial	1103	12.55%	12.78%
Transport	50	0.57%	2.08%
Urban green	853	9.70%	17.42%
Undeveloped	663	7.54%	7.40%
Woodland	63	0.72%	5.05%
Farmland	1000	11.38%	21.49%
Water	97	1.10%	2.65%

collectively serve to visualize and quantify the spatial heterogeneity of urban land use patterns inside Beijing’s sixth ring, establishing a geographic and statistical foundation for subsequent classification accuracy validation. Figure 6 shows the overall spatial distribution of various land use types within the sixth ring road area of Beijing along with a zoomed-in view of regions radiating outward from the center. The block-level land use classification pattern in Beijing follows a typical “Von Thünen Ring” model [53,54]. Within the second ring, intensive residence, commercial, and institution dominate, reflecting the urban core. The third ring features concentrated commercial and intensive residence areas, indicating significant urban activities. The fourth ring road has more ordinary residences, commercial, and institutions, transitioning to high-density land types in the city center. The fifth ring mixes ordinary residences, farmland, industrial land, and some urban green spaces, primarily in outer regions. The sixth road mainly consists of ordinary residences, farmland, urban green space, and undeveloped land, reflecting suburban agriculture and open spaces.

Table 2 presents the numbers and proportions of patches for each category of block-level land use inside the sixth ring road of Beijing along with their area proportions. Overall, residential areas dominate the patches in Beijing’s block units, with over 3000 patches identified, including more than 2000 high-rise residential communities. This reflects Beijing’s urban population, which is almost 20 million (https://www.beijing.gov.cn/gongkai/shuju/sjjd/202105/t20210519_2392886.html accessed on 7 July 2024). As China’s capital, Beijing has been designed as a political, economic, cultural, and international exchange center, thus exhibiting a higher number of commercial, industrial, and institution patches compared to second-tier cities. The large area inside the sixth ring road, located in the urban suburb area, also includes numerous agricultural and undeveloped patches. The category with the fewest patches is transportation centers, totaling about 50; this is related to the fact that this study classifies land use regions as delineated by major road networks with transportation centers typically featuring larger transportation hub patches.

3.2. Land Cover Distribution Based on Pixel and Object Methods Within Blocks

The land cover classification operation covers the entire sixth ring road, and it is categorized into buildings, water bodies, vegetation, and others. To compare the two methods employed, a central sample area is depicted in Figure 8.

Figure 8a presents the classification results using the pixel-based method, while Figure 8b–d depict the classification results of the multi-scale object-based method at varying segmentation scales. The pixel-based method demonstrates superior capability in spatially expressing land cover details, allowing for more precise extraction, particularly in the central region (e.g., lakes and squares), where classification accuracy is high and details are abundant. However, as the sample area transitions from the center to the peripheries (building-dense zones), the granularity of classification diminishes, resulting in a reduced resolution for smaller features.

In the classification results obtained using the multi-scale object-based method, increasing the segmentation scale results in a gradual reduction in detail with smaller features being aggregated into larger classification units. In the larger-scale image shown in Figure 8b, the boundaries of buildings and vegetation become increasingly indistinct, leading to larger and smoother classification units. The multi-scale object-based method effectively identifies prominent land features (such as water bodies and large buildings) in the central portion of the sample area; however, during the transition toward the edges, particularly in building-dense areas, smaller structures and other details may be merged or classified differently.

In summary, the pixel-based method excels in capturing fine details, while the multi-scale object-based method tends to enhance classification smoothness with increasing segmentation scales, resulting in a gradual loss of detailed information.

3.3. Results of Pixel-Based and Object-Based Block-Level Land Use Classification

We utilized nine groups of 24 landscape metrics to measure features in block units based on both pixel-based and object-based land cover classifications. Subsequently, we conducted 100 iterations of bootstrapped sampling with replacement to train the models and execute classifications on the block units. To compare overall model performance and analyze the effectiveness and performance differences between the two feature extraction methods, the classification accuracy results are presented in Figure 9 and Figure 10. Finally, predictions were made across the entire dataset, selecting the result with the highest total sum of mean precision and mean recall from all experiments as the best classification result, and ultimately generating land use classification maps by use of the different methods (Figure 11).

Figure 8 presents boxplots illustrating the overall recall, precision, and F1 score for land use classification, comparing the performance of the pixel-based method with those of object-based methods across different segmentation scales. By examining the median values of these boxplots, it is evident that the pixel-based method outperforms the object-based methods across all evaluation metrics. Specifically, the pixel-based method achieves recall, precision, and F1 scores that are approximately 3 percentage points higher overall compared to the object-based methods. The performances of the object-based methods vary with segmentation scale, with Scale58 and Scale95 demonstrating better stability, as indicated by their shorter whiskers, while Scale137 exhibits the greatest fluctuation, indicating lower stability. Furthermore, as the segmentation scale increases, the overall performance of the object-based methods tends to decline, further highlighting the robustness and superior performance of the pixel-based method.

Figure 10 illustrates the F1 scores for various land use classes across different methods. The results indicate that commercial, urban green, farmland, and ordinary residence classes achieve relatively high F1 scores across all methods. Among them, the pixel-based method outperforms the best object-based method (Scale58) across multiple categories, and it further improves upon Scale95 and Scale137 with an overall improvement range of 0.05 to 0.16.

Specifically, for the commercial category, the pixel method achieves an F1 score that is approximately 0.10 higher than that of Scale58, 0.14 higher than that of Scale95, and 0.16 higher than that of Scale137, demonstrating strong classification performance. In the ordinary residence category, the pixel method achieves an F1 score about 0.03 higher than those of Scale58 and Scale95 and 0.05 higher than that of Scale137, maintaining a relatively stable advantage. In the urban green and farmland categories, the pixel method shows improvements of about 0.03–0.04 over Scale58, 0.05–0.06 over Scale95, and 0.07–0.10 over Scale137, indicating a stronger ability to capture spatial characteristics and avoiding the accuracy fluctuations observed in object-based methods across different segmentation scales.

In contrast, categories such as woodland, villa, and water exhibit significant fluctuations in F1 scores across different object-based segmentation scales, suggesting weaker classification stability. For example, in the woodland category, even the best-performing object-based method, Scale58, still lags behind the pixel method by 0.05, while Scale95 and Scale137 further decline by 0.08 and 0.13, respectively, indicating lower classification stability. Similarly, in the water category, the pixel method achieves F1 scores that are approximately 0.08, 0.10, and 0.13 higher than those of Scale58, Scale95, and Scale137, respectively, highlighting the substantial impact of segmentation scale on object-based classification performance.

Additionally, all methods exhibit relatively low F1 scores in the transport category, suggesting the possibility of challenges in accurate identification, larger margins of error, and lower classification stability. Meanwhile, institution, industrial, and intensive residence categories show relatively small differences in F1 scores across methods, indicating a stable classification performance with minimal impact from segmentation scale.

Overall, the pixel method outperforms all object-based segmentation scales in high-accuracy categories (commercial, ordinary residence, urban green, farmland), achieving the highest F1 scores with notable improvements. In contrast, object-based methods show considerable fluctuations in F1 scores for certain categories (such as woodland and water), indicating that classification performance is highly sensitive to segmentation scale. Therefore, optimizing segmentation scale selection for different land use categories is crucial to reducing classification errors resulting from object-based methods.

Despite differences in detail, all four methods reflect a consistent overall trend and accurately represent the land use patterns inside Beijing’s sixth ring road.

A comparative analysis of the land use classification spatial distribution maps reveals a gradual shift in land use types from the urban core to the outer regions inside Beijing’s sixth ring road. The area inside the second ring is predominantly characterized by high-density residential, commercial, and institutional land use, showcasing the highly developed nature of the city center. Similarly, the third ring area contains significant amounts of commercial and high-density residential land, continuing the intense trend of development of the central region. Around the fourth ring, ordinary residential, commercial, and institutional land use dominate with a noticeable reduction in land use density.

By the time we reach the fifth ring, land use types diversify, blending ordinary residential areas, farmland, industrial land, and urban green spaces, and marking the transition from urban to suburban landscapes. In the outermost sixth ring region, ordinary residences, farmland, urban green spaces, and undeveloped land are predominant, reflecting suburban agricultural activities and the presence of large open spaces.

The pixel-based method excels in detail recognition, especially in high-density areas (such as residential and commercial land inside the second ring), delivering superior accuracy. The Scale58 method also performs well in identifying fine-grained land use categories, though it falls short in classifying larger areas. For example, Scale95 strikes a balance between recognizing and distinguishing various land use types, showing accuracy in identifying industrial land and farmland but struggling with commercial land classification.

Scale137, with its broader scope, effectively classifies large-scale land use types, such as farmland, urban green spaces, and undeveloped land in suburban areas. While providing smoother classification results, its large segmentation scale sacrifices detailed recognition, potentially oversimplifying complex land use in the city’s core. Small land use types may be overlooked or merged, reducing classification granularity.

3.4. Results of Different Landscape Metrics Feature Combinations

This study selected the pixel-based method, which exhibited slightly higher classification accuracy, as the dataset for experiments grouped by different landscape metrics features. The aim is to explore the classification capabilities of various landscape metrics features for different land use types. We created boxplots of the classification accuracy metric F1 score and heatmaps of the F1 score for each category under different feature combinations (see Figure 12 and Figure 13).

Figure 12 illustrates the impacts of feature removal on overall F1 scores. Removing the LANDSCAPE_PROPORTION feature significantly reduces the median F1 score, dropping it below the Q1 value (25th percentile) of the ALL-features combination, indicating that this feature plays a crucial role in maintaining classification performance, particularly for lower-performing cases.

In contrast, removing CLASS_SHAPE and CLASS_COMBINATION slightly raises the Q3 value (75th percentile) above the ALL-features baseline, suggesting that their exclusion may improve classification accuracy for higher-performing cases. The Q3 increase implies that at least 25% of the model’s results benefit from their removal.

Feature importance varies across different land use categories. Analyzing per-class accuracy is essential to identifying key features for specific categories, enabling more effective feature selection and improving model robustness. The F1 score considers both precision and recall, providing a detailed assessment of model performance in the context of class imbalance. Through F1 score analysis, we examined the impacts of each feature combination on the classification accuracy of various categories, as presented in a heatmap (Figure 13).

In Figure 13, the white numbers indicate cases where removing specific feature groups significantly improved classification accuracy, suggesting their removal benefits performance. Feature groups with a green background (without white numbers) show minimal changes in accuracy after removal, indicating these features contribute little and can be considered for elimination.

In most cases, the impact of removing specific feature groups on the F1 score is relatively small; however, some removals can lead to notable changes. For instance, CLASS_SHAPE, LASS_AMOUNT, and LANDSCAPE_PROPORTION demonstrate significant impacts across categories, affecting classification outcomes both positively and negatively. Notably, removing CLASS_SHAPE improved accuracy for most categories, indicating its substantial negative influence. Conversely, removing LANDSCAPE_PROPORTION decreased accuracy across categories, underscoring its importance, while other feature groups had a smaller effect. Due to the differences in classification tasks, the emphasis on accuracy evaluation methods varies, and the strategies for feature selection differ as well. Therefore, this study also presents heatmaps for two additional indicators that can be used for reference in future research (see Appendix A, Figure A1, Figure A2 and Figure A3).

To further explore differences in feature classification at the class and landscape levels in the context of land use classification, overall accuracy and individual class accuracy figures were plotted (Figure 14 and Figure 15).

Figure 14 shows that the class-level features generally exhibit weaker classification capabilities compared to landscape-level features. However, the classification abilities of these two levels of features may differ for different land use types. Figure 15 shows that both the farmland and urban green categories have relatively high F1 scores under both levels of feature groups. In most categories, the landscape-level classification accuracy is higher than that of the class level; nevertheless, some class-level features demonstrate better accuracy, such as farmland and industrial. There are certain differences in the F1 scores of different categories under both levels of feature groups, indicating that different categories have varying sensitivities to different features, leading to differences in classification performance. Therefore, specific feature selection and model adjustments are necessary for each category.

4. Discussion

4.1. Effectiveness of Land Use Classification Based on Pixels and Different Scales of Segmented Objects

The pixel method demonstrates superior overall precision and performs better for most categories, while the object-based method shows a gradual decline in overall precision as the segmentation scale increases.

The pixel method likely outperforms the object-based method in classification accuracy because it captures details and local features more effectively when the spectral characteristics of objects are distinct. In contrast, while the object-based method captures spatial relationships between objects, determining an optimal segmentation scale for high classification accuracy is challenging, as the ideal scale varies across different cities. Moreover, as the segmentation scale increases, object contours tend to become more generalized and boundaries become smoother, which may lead to the loss of critical information or the misclassification of boundary details. Additionally, while aggregated object features can mitigate individual pixel classification errors, they may also obscure local detail errors, ultimately affecting overall classification accuracy [55,56]. Previous studies [28,30,57] have highlighted the potential application of the object-based method in land use classification, whereas this study emphasizes the pixel method’s potential benefits in terms of detail capture and classification accuracy, especially as remote sensing image resolution improves.

The pixel classification method processes each pixel directly, capturing subtle changes and small-scale features. For instance, it can identify small buildings between larger structures in urban areas. Conversely, large-scale segmentation methods may depict general contours and smooth boundaries, but they can overlook or misclassify boundary details, thus losing significant information. Individual errors might also be masked by overall object features, reducing actual classification accuracy.

This study analyzed three specific areas (commercial, ordinary residence, and farmland), and we present image structure maps, extracted land cover maps, and specific feature values in the appendix (Figure A4, Figure A5 and Figure A6,

Table A3. Commercial block sample landscape metrics’ calculation results.

Metrics	Pixel Method	Scale 58	Scale 95	Scale 137
Mean Building Area	0.049	0.135	0.102	0.135
Mean Vegetation Area	0.068	0.314,	0.427	0.314
Mean Water Area	0.003	0.000	0.000	0.000
Mean Other Area	0.090	1.146	0.525	1.146
Building Mean Shape Index	1.668	2.236	2.089	2.236
Vegetation Mean Shape Index	1.365	3.344	2.268	3.344
Water Mean Shape Index	1.329	None	None	None
Other Mean Shape Index	1.574	1.917	1.782	1.917
Building Number of Patches	42.000	7.000	12.000	7.000
Vegetation Number of Patches	7.000	1.000	1.000	1.000
Water Number of Patches	15.000	0.000	0.000	0.000
Other Number of Patches	36.000	4.000	8.000	4.000
Building Mean Euclidean Nearest Neighbor	5.073	19.792	6.834	19.792
Vegetation Mean Euclidean Nearest Neighbor	32.774	None	None	None
Water Mean Euclidean Nearest Neighbor	13.619	None	None	None
Other Mean Euclidean Nearest Neighbor	4.763	5.249	5.907	5.249
Mean Area	0.058	0.162	0.278	0.487
Mean Shape Index	1.562	1.748	1.981	2.222
Number of Patches	100.000	36.000	21.000	12.000
Building Proportion landscape	35.419	6.559	20.862	16.199
Vegetation Proportion of Landscape	8.207	24.663	7.299	5.374
Water Proportion of Landscape	0.825	0.355	0.000	0.000
Other Proportion of Landscape	55.477	68.243	71.839	78.428
Shannon Diversity Index	0.942	0.807	0.759	0.645

,

Table A4. Ordinary residence block sample landscape metrics’ calculation results.

Metrics	Pixel Method	Scale 58	Scale 95	Scale 137
Mean Building Area	0.024	0.043	0.057	0.043
Mean Vegetation Area	0.010	0.004	0.024	0.004
Mean Water Area	0.000	0.000	0.000	0.000
Mean Other Area	0.236	6.406	3.962	6.406
Building Mean Shape Index	1.465	1.958	1.996	1.958
Vegetation Mean Shape Index	1.243	1.256	1.382	1.256
Water Mean Shape Index	None	None	None	None
Other Mean Shape Index	1.417	2.505	2.852	2.505
Building Number of Patches	176.000	22.000	32.000	22.000
Vegetation Number of Patches	42.000	13.000	4.000	13.000
Water Number of Patches	0.000	0.000	0.000	0.000
Other Number of Patches	39.000	2.000	3.000	2.000
Building Mean Euclidean Nearest Neighbor	4.964	21.453	13.421	21.453
Vegetation Mean Euclidean Nearest Neighbor	15.433	47.989	32.635	47.989
Water Mean Euclidean Nearest Neighbor	None	None	None	None
Other Mean Euclidean Nearest Neighbor	5.119	3.721	3.868	3.721
Mean Area	0.054	0.261	0.354	0.373
Mean Shape Index	1.421	1.994	1.999	1.741
Number of Patches	257.000	53.000	39.000	37.000
Building Proportion of Landscape	30.221	2.297	13.219	6.806
Vegetation Proportion of Landscape	3.088	20.746	0.702	0.386
Water Proportion of Landscape	0.000	0.000	0.000	0.000
Other Proportion of Landscape	66.692	76.958	86.079	92.808
Shannon Diversity Index	0.741	0.616	0.433	0.273

and

Table A5. Farmland block sample landscape metrics’ calculation results.

Metrics	Pixel Method	Scale 58	Scale 95	Scale 137
Mean Building Area	0.002	0.037	0.036	0.037
Mean Vegetation Area	0.782	0.004	0.024	0.004
Mean Water Area	0.000	0.000	0.000	0.000
Mean Other Area	0.080	0.182	0.139	0.182
Building Mean Shape Index	1.195	1.095	1.095	1.095
Vegetation Mean Shape Index	1.315	1.719	1.520	1.719
Water Mean Shape Index	None	None	None	None
Other Mean Shape Index	1.502	2.180	1.759	2.180
Building Number of Patches	117.000	1.000	1.000	1.000
Vegetation Number of Patches	27.000	8.000	8.000	8.000
Water Number of Patches	0.000	0.000	0.000	0.000
Other Number of Patches	46.000	17.000	28.000	17.000
Building Mean Euclidean Nearest Neighbor	8.042	None	None	None
Vegetation Mean Euclidean Nearest Neighbor	6.280	5.169	4.420	5.169
Water Mean Euclidean Nearest Neighbor	None	None	None	None
Other Mean Euclidean Nearest Neighbor	10.654	25.738	10.730	25.738
Mean Area	0.132	0.706	0.678	0.965
Mean Shape Index	1.286	1.996	1.689	1.996
Number of Patches	190.000	33.000	27.000	26.000
Building Proportion of Landscape	1.077	85.455	0.143	0.148
Vegetation Proportion of Landscape	84.179	0.143	84.384	87.544
Water Proportion of Landscape	0.000	0.000	0.000	0.000
Other Proportion of Landscape	14.744	14.474	15.473	12.308
Shannon Diversity Index	0.475	0.422	0.439	0.382

). In the commercial area, the pixel method exhibits a higher building shape index (1.668) and a higher building ratio (35.491%). However, the closest neighboring distance of buildings in the commercial area is greater under the object-based method compared to the pixel method (e.g., the closest neighboring distance of buildings for Scale58 is 19.792), indicating a deficiency in describing the distribution discreteness of buildings in this method. In the ordinary residence area, the pixel method exhibits a smaller building area (0.024), a lower shape index (1.421), and a shorter Euclidean nearest neighbor distance (4.964), which better reflect the dense and regular layout of residential areas. In contrast, the object-based method shows relatively consistent building area and shape index at different scales, indicating a relatively stable description of building features in residential areas at different scales. In the farmland area, the Pixel method yields a relatively small landscape shape index (1.286) while displaying a significant Euclidean nearest neighbor distance (6.280), indicating the more accurate capturing of the vegetation distribution pattern in regular regions like farmland. Overall, using the pixel-based method, rare and small land features can be reflected, thereby exhibiting a stronger ability to express structural composition and spatial pattern differences, and improving the classification accuracy of block units.

In conclusion, the pixel method is more precise in capturing the complexity of building and vegetation distribution, while the object-based method simplifies and standardizes feature descriptions across scales. The effectiveness of different scales varies by land use type, making the selection of appropriate classification methods and scales crucial for accurate feature representation.

4.2. Effectiveness of Land Use Classification with Different Types of Landscape Metrics Features

This study aims to explore the significant differences in how various landscape metrics impact classification accuracy. By designing comparative experiments, we evaluate the effectiveness of these features and identify optimal combinations for specific tasks.

When the proportion of landscape type is removed, the accuracy of all categories is significantly reduced, underscoring its importance. In contrast, removing certain feature groups (e.g., those marked with white numbers) was shown to improve some categories’ accuracies, suggesting they may contain redundant or overfitted information that hinders model generalization. Conversely, other groups (represented by a green background) have minimal effects on accuracy; thus, their contributions are relatively small, and they can be adjusted during selection. Although light-yellow background features positively contribute to classification performance, their influence is limited; careful consideration is thus needed in balancing model complexity and effectiveness (Figure 13).

Previous studies [30,51,58,59] have analyzed the importance scores of specific features in classification models using statistical metrics, which can help in evaluating the contributions of these features to classification effectiveness and in identifying key influencing factors. These studies indicate that features such as the landscape shape index (Shape Index), which describes the complexity of landscape patches, shows higher values for irregular shapes (e.g., natural areas) and lower values for more regular shapes (e.g., urban buildings). The proportion of patch types (PLAND) provides the effective differentiation of different land use types within a given area, while the average proximity index (ENN) indicates that buildings in structurally similar areas tend to be spaced more regularly, potentially aiding in distinguishing ordinary residence categories. The mean patch size (MPS) measures the degree of aggregation among different land use types, such as high-density residential versus typical residential areas. Furthermore, Ref. [60] indicated that landscape-level metrics are highly predictive of changes in land use types with indices such as the LPI (largest patch index), NP (number of patches), LSI (landscape shape index), and SHDI (Shannon diversity index) highlighting the discriminative capacity of landscape-level metrics when applied to different land use types.

The findings of this study are consistent with those of previous research, confirming the effectiveness of using landscape metrics as features in land use classification. Unlike Herold et al.’s approach, we employed a single-factor controlled experimental design to analyze the separate impacts of different feature categories on classification performance. This yielded a more accurate evaluation and understanding of how these features relate to the outcomes.

In summary, optimizing feature selection strategies is crucial for improving block-level land use classification models. This involves retaining key contributing features while managing potentially noisy or complex ones carefully. This ensures that the model can be used to accurately interpret and predict different land use categories, leading to improved accuracy and practicality.

4.3. Uncertainty and Future Work

The primary objective of this study is to validate the potential use of pixel-scale landscape metrics in block-level urban land use classification, demonstrating their advantages in overcoming the inherent scale dependency and detail loss issues of traditional object-based segmentation methods, and thereby improving classification accuracy at the block scale. Compared with EULUC-China [61], the first national urban land use dataset in China, our method shows more merit. EULUC-China merges multiple data sources, such as high-resolution remote sensing imagery, POI locations, and road networks, and uses an object-based classification framework. But despite its large capacity for data integration and complex modeling, its overall accuracy when used in typical plain cities like Beijing and Chengdu does not exceed 50%.

In contrast, this study relies solely on 2-meter-resolution remote sensing images. By extracting pixel-scale landscape pattern metrics, it achieves much higher accuracy without extra data. This confirms the effectiveness of pixel-scale measures in showing urban block spatial heterogeneity. Also, the method shows significant advantages in improving classification performance and reducing reliance on multi-source data, showing broad applicability and strong generalization potential. However, certain limitations remain, leaving room for further improvements in classification accuracy [62,63]. First, this study extracts only the major land cover types (buildings, vegetation, water, and others), with landscape metrics primarily reflecting the dominant land use patterns of blocks, and it does not further refine the classification of land cover features. This may reduce the model’s ability to distinguish certain block types. Therefore, future research could optimize the land cover classification system, enabling landscape metrics to more comprehensively represent block-level spatial structures. Additionally, although this study employs a ground truth dataset covering the entire study area, avoiding sampling bias arising from manual data selection, the dataset is primarily based on the dominant land use type of each block. Since real-world blocks often contain mixed land use functions, and the proportions of different land use types vary significantly (e.g., commercial is relatively scarce, while urban green is abundant), this imbalance may affect the model’s generalizability.

Furthermore, this study relies solely on high-resolution remote sensing imagery, using physical characteristics for classification. In complex urban environments, spectral and structural features alone may not be sufficient to distinguish functionally similar blocks, such as commercial and administrative areas. Integrating economic datasets such as POI, nighttime light, and human mobility data could help to extract additional functional information and enhance classification stability and accuracy.

Lastly, the recognition accuracy of remote sensing images at different resolutions, particularly for 10-meter-scale images, should be studied in depth. This will help achieve a more comprehensive understanding of variations in landscape metric features across different geographical contexts, ensuring the broad applicability and scientific validity of the research conclusions [3]. Zhang et al. [64] noted that data-driven deep learning methods have gained widespread attention due to their powerful feature extraction capabilities. However, the “black box” nature of deep learning methods results in weak interpretability, making it difficult to integrate existing knowledge, which presents a significant challenge. To address this issue, a knowledge-guided framework for land pattern depiction has been proposed. This framework extracts land use patterns by analyzing the spatial pattern modes of different land cover types.

In the future, learning from the automated feature extraction capabilities of deep learning is expected to further improve the classification accuracy of models, as deep learning could be applied to learn any cross-scale pattern from the samples [18,65,66,67]. This integration will significantly enhance the model’s practicality and transparency, making it more valuable and applicable in real-world scenarios.

5. Conclusions

This study focuses on the issue of scale and systematically evaluates the performance of pixel-based methods and object-based methods in utilizing landscape metrics for block-level urban land use classification. Through a systematically designed comparative experiment, the study identifies the optimal scale (pixel level) that can be used to achieve the highest classification accuracy, and it further explores the roles of different types of landscape metrics in classification.

Compared with previous studies that primarily employed object-based methods combined with landscape ecology metrics for block-level urban land use classification, this paper has the following key findings:

(1) Pixel-based methods outperform object-based methods in classification accuracy. The results show that pixel-based methods can capture fine details and spatial pattern information in high-resolution remote sensing images more accurately, which significantly enhances classification accuracy. In contrast, object-based methods, which rely on segmentation scales, exhibit certain limitations in adapting to and integrating complex spatial patterns;

(2) At the optimal pixel scale, landscape-level metrics are superior to class-level metrics in capturing spatial pattern characteristics. Through systematically designed feature grouping comparative experiments, the study finds that landscape-level metrics more effectively represent spatial pattern characteristics, providing a more reliable feature set for classification. This finding addresses a gap in previous research, which often lacks systematic analyses of feature selection, and offers practical guidance for improving the accuracy of block-level urban land use classification.