Urban Outdoor Thermal Environment Analysis Based on Semantic Segmentation and Morphology Indicators: A Case Study of Residential Blocks in Wuhan

Abstract

Rapid urbanization has intensified urban heat issues. Previous studies often relied on subjective block selection and rarely integrated vegetation data. This study extracted vegetation from Wuhan’s satellite imagery and combined it with building geometry to generate large-scale 3D block models. Typical blocks were identified by clustering, and thermal environments were simulated using ENVI-met to establish regression models. POI and spatial analyses validated the results. The study found that 1. t-SNE outperforms PCA and UMAP in dimensionality reduction. 2. K-means surpasses GMM, DBSCAN, and Spectral in clustering. 3. ${SVF}_{ave}$, ${FA}_{all}$, ${VD}_W$, $VAR$, and BBA are critical for block morphology classification and block outdoor thermal assessment. 4. The final ridge regression model based on these indices achieved high R² values (0.805, 0.507, and 0.855), indicating excellent model performance. 5. The blocks in Cluster 1 (west of the Yangtze River) exhibit higher mean air temperatures. 6. the blocks in Cluster 2 (new areas) have high vegetation coverage, causing larger temperature differences between the inside and outside of blocks. This study provides a comprehensive workflow for urban block morphology classification and thermal assessment.

1. Introduction

1.1. Background

The accelerating urbanization process in the 21st century is considered to have negative impacts on micro-climatic conditions, resulting in a significant increase in air temperature within urban areas [1,2,3]. This phenomenon is primarily attributed to the following factors: 1. The changing of the natural permeable surfaces with impervious materials leads to the storage and subsequent re-release of substantial solar radiation within urban areas [4,5]. 2. Short-wave and long-wave radiation are repeatedly reflected within street canyons, thereby hindering the urban cooling process [6,7]. 3. The increase in building height and the high roughness of building volumes are regarded as severe issues in modern cities, consequently diminishing the capacity for heat removal by ventilation [8]. 4. Low-albedo materials cause absorption and storage of short-wave solar radiation by building envelopes, resulting in high surface temperatures and subsequent emission of long-wave (thermal) radiation into the outdoor urban environment [9]. 5. Due to the loss of vegetation and green spaces, the latent heat generated through transpiration and the corresponding cooling effect on the surrounding air have been significantly weakened [10,11]. 6. Traffic congestion increases vehicle exhaust emissions, thereby exacerbating urban air pollution and the urban heat-island effect [12,13].

Studies have confirmed that urban block morphology is a key factor causing urban thermal environmental issues, as it affects the urban microclimate by modifying surface heat transfer, solar access patterns, and the outdoor wind environment [14,15,16]. In hot climate zones, unreasonable block morphology can result in greater urban heat-island intensity [17] or more frequent heatwaves [18], potentially making conditions unbearable for residents. A study reported poor outdoor thermal environments in residential blocks in Shenzhen, particularly on the leeward side of the building array [19]. One researcher tested 22 residential areas in Japan using wind tunnel experiments, revealing a strong correlation between building density and the mean wind speed [20]. Unger [21] reported the relationship between UHI intensity and SVF in an earlier study.

Regarding the selection of research blocks, researchers generally conduct studies based on either idealized block models or real-world blocks [22]. Influenced by multiple factors, urban morphology exhibits significant diversity [23], leading to urban blocks which are highly complex and heterogeneous. Idealized models often fail to represent these characteristics of real blocks [24,25]. In contrast, urban blocks with real morphology information have been more widely used in recent research. By selecting typical blocks as samples, researchers aim to reveal universal urban conditions and focus on core issues [26], such as urban energy issues, outdoor wind environment issues and outdoor thermal environment issues. Typically, the selection of typical blocks relies on subjective judgment and expert experience. For instance, Jin [27] investigated 12 real blocks in Harbin without any explanation about the sampling method. However, the limited number of real samples remains insufficient to fully reflect the overall morphology characteristics of whole blocks in a city.

In terms of methodology, numerical simulation is popular in microclimate analysis. A review highlighted the extensive use of Computational Fluid Dynamics (CFD) [28]. Since 2011, the Reynolds-Averaged Navier–Stokes (RANS) model has become the most frequently used model (accounting for 96.0%) in urban microclimate performance studies. Research has confirmed the correlation between wind environmental conditions (e.g., mean wind speed ratio and flow rate) and Building Coverage Ratio [29]. Meanwhile, the influence of wind direction on results has also been emphasized [30]. However, due to the limitations of CFD simulation methods, only relevant studies based on ENVI-met [31] have considered the impact of vegetation on the microclimate of blocks [32,33]. At the same time, due to the difficulty in collecting vegetation data, related studies often manually add vegetation models based on photos or field surveys [34,35].

The manual modeling process is extremely time-consuming, making it difficult to conduct comparative analyses across many cases. Due to the limited availability of vegetation data, some studies have chosen to ignore the impact of vegetation on the urban microclimate, a choice that, in turn, limits the accuracy of their research findings. Other studies have opted for field surveys to determine the location, size, and species of vegetation, followed by manual reconstruction in ENVI-met. Because the location, size, and species of vegetation are all based on the researcher’s subjective judgment, the models are of limited authenticity. This directly limits the accuracy of the research conclusions [36].

With the advancement of digital technologies, a growing number of parametric tools can automatically convert 3D models into ENVI-met-compatible formats. This development enables researchers to conduct large-scale ENVI-met simulations. And the rise of artificial intelligence has introduced semantic segmentation techniques for satellite imagery into urban 3D model generation. This allows researchers to generate richer urban 3D models from multi-source data [37,38,39].

Although existing studies have explored the impact of urban morphology on urban microclimate in different ways, several limitations remain: 1. Few studies have been able to identify residential building types comprehensively considering factors such as blocks scale, building density, building typology, and spatial layout. 2. There is a lack of an effective method for generating 3D block models that include complete 3D information on both buildings and vegetation. 3. Conclusions from numerous studies are presented in mathematical forms, such as regression models. For architects, translating these findings into practical strategies remains a significant challenge.

1.2. Scope of Paper

This study proposes a novel framework to identify typical residential block types within Wuhan. By using multi-source data fusion, the research acquired building and vegetation data for 915 residential blocks in Wuhan and calculated morphological indicators for each block. Based on the indicators filtered by dimensionality reduction methods, cluster analysis was performed on the residential blocks. Subsequently, the study selected the typical blocks of different clusters for ENVI-met microclimate simulation. The thermal indicators including $T_{ave}$, $T_{dif}$, and $T_{RMSE}$ were calculated. Pearson correlation analysis, ridge regression analysis, and comparisons of typical blocks were used to explore the correlations between morphological indicators and microclimate indicators. Furthermore, this study used parametric methods to analyze the spatial distribution of different block types and their surrounding functional patterns, aiming to provide scientific guidance for urban planning and management.

The main highlights of this study are as follows:

• Satellite image semantic segmentation for 915 Wuhan blocks was used to extract vegetation data. Combined with parametric methods, 3D ENVI-met blocks models including vegetation data were generated.
• Twenty-nine morphological indicators for the 915 Wuhan blocks were extracted by parametric methods. Following dimensionality reduction, nine key morphological indicators were screened.
• Through multiple clustering analysis, the 915 blocks were classified into five morphological types.
• The correlations between morphological indicators and thermal indicators were revealed through correlation analysis and ridge regression.
• Spatial distribution characteristics of different block types were analyzed, and their distribution patterns were summarized.

2. Methods

This study analyses the outdoor thermal environment of residential blocks in Wuhan from a morphological perspective through seven main steps, and further explores the spatial distribution characteristics of different block types. The research workflow is illustrated in Figure 1, and the seven steps are as follows:

• Step 1 (Block Data Collection): based on the Baidu Maps database [40], the study acquired building geometry data for 915 blocks, which were saved as SHP and DBF files. Simultaneously, satellite imagery covering a 1024 m area for each block was obtained and stored as JPG files.
• Step 2 (block Data Processing): the study uses semantic segmentation to process block satellite imagery for the extraction of vegetation data. Subsequently, parametric methods are used to generate 3D blocks models by integrating SHP files, DBF files, and semantic segmentation images.
• Step 3 (Morphological Indicator Screening): based on 3D block models, 29 morphological indicators were extracted using parametric methods. Subsequently, the applicability of different dimensionality reduction methods was evaluated, and key morphological indicators were screened.
• Step 4 (Morphology clustering): the study uses multiple clustering algorithms on the key morphological indicators. The optimal clustering result is selected for subsequent research.
• Step 5 (Thermal Environment Simulation): typical blocks from each cluster will be selected, based on the clustering results. Their corresponding 3D models will be converted into ENVI-met models for outdoor thermal environment simulations.
• Step 6 (Regression Analysis): based on the morphological and thermal indicators of typical blocks, the study analyzes the correlation between block morphology and the thermal environment. Furthermore, the influence of mechanisms and the relative importance of key morphological indicators on the outdoor thermal environment are revealed, based on regression analysis.
• Step 7 (Spatial Distribution Analysis): the study conducts spatial distribution analysis on the blocks in each cluster to investigate their spatial characteristics based on cluster labels. The surrounding POI (Points of Interest) of each cluster are compared to analysis the characteristics of the surrounding environment for different block types.

2.1. Step 1 (Block Data Collection)

2.1.1. The Geometry Data

As a historic city in China, the main urban area of Wuhan is divided by the Yangtze and Han rivers into three core regions, known as the “Three Towns of Wuhan.” As shown in Figure 2a, the Hankou region (green) comprises the Qiaokou, Jianghan, and Jiang’an districts; the Wuchang region (yellow) includes the Wuchang, Qingshan, and Hongshan districts; and the Hanyang region (red) consists of the Hanyang district. This study selected 915 residential blocks across these seven administrative districts as the research targets. As illustrated in Figure 2b, the data were obtained from Baidu Maps using BIGEMAP (V9.8.5 X64) software, in the standard Shapefile (.shp) format. The attribute data (the .dbf file) explicitly includes the building height information, which was used to determine the vertical height of buildings. The purple areas in Figure 2b are 915 selected blocks.

2.1.2. The Satellite Image Data

This study generates 3D block models through multi-source data fusion, based on two types of data: 1. Geometry Data: building footprints and heights of the selected residential blocks were extracted from .shp and .dbf files by parametric methods. 2. Satellite Imagery: Jilin-1 satellite data [41] for the Wuhan area was used. Satellite images for each block were cropped and saved using parametric methods.

The satellite data were purchased from the official data marketplace of the Jilin-1 Network. The imagery contains multiple standard spectral channels, including blue, green, and red. The overall cloud cover is less than 2%, and over 90% of the imagery used has a resolution of 0.5 m.

2.2. Step 2 (Block Data Processing)

2.2.1. Semantic Segmentation

The study uses the semantic segmentation model from the EarthVQA, a multi-modal multi-task satellite Visual Question Answering (VQA) framework, to process satellite imagery [42]. The EarthVQA dataset was developed by the LoveDA dataset. The LoveDA dataset comprises 5987 pairs of High Spatial Resolution (HSR) satellite images and their corresponding semantic masks from 18 urban–rural areas in Wuhan, Nanjing, and Changzhou. The EarthVQA model was trained for semantic segmentation tasks based on these image pairs.

EarthVQA maps the input imagery ($I\in {\mathbb{R}}^{H\times W\times 3}$) to a semantic category map ($F^v\in {\mathbb{R}}^{H^{\prime }\times W^{\prime }\times C}$), where the number of categories $C$ is 8. The model adheres to the standard semantic segmentation paradigm of “Encoder–Feature Pyramid Network (FPN) [43,44]–Decoder–Classifier Head.” Specifically, a ResNet-50 backbone [45] pretrained on ImageNet serves as the encoder to extract multi-scale features. These features are fused via an FPN and subsequently upsampled by the decoder to restore spatial resolution. Finally, a convolutional layer outputs per-pixel class scores, which are upsampled to the original image scale to generate the prediction.

EarthVQA used the standard semantic segmentation metrics, mean Intersection over Union ($mIoU$) [46] and Overall Accuracy ($OA$) [47], to evaluate model performance. For $mIoU$, each class $C$ is calculated as ${IoU}_C$, $IoU$, $mIoU$ and $OA$ are shown in

Table 1. The semantic-segmentation evaluation indicators.

Nomenclature	Equation	Describe
$IoU$	${IoU}_C={\displaystyle \frac{{TP}_c}{{TP}_C+{FP}_C+{FN}_C}}$	${TP}_c$ (True Positive): pixels or regions that are correctly predicted as belonging to the target category by the model, i.e., the intersection of the prediction and the ground truth for that category; ${FP}_C$ (False Positive): regions incorrectly predicted as this category (actually background or other categories; ${FN}_C$ (False Negative): regions of this category missed by the model (actually this class, but not detected).
$mIoU$	$mIoU={\displaystyle \frac{1}{C}}{\sum }_{C=1}^C{IoU}_C$	$C$: the total number of categories in the dataset (Category Count). ${IoU}_C$: the $IoU$ (Intersection over Union) of the $C$ specific category.
$OA$	$OA={\displaystyle \frac{{\sum }_C{TP}_C}{{\sum }_C({TP}_C+{FN}_C)}}$	${\sum }_C{TP}_C$: total number of correctly classified samples across all categories (i.e., the sum of true positives for all categories). ${\sum }_C({TP}_C+{FN}_C)$: total number of all samples in the dataset.

.

Since training the EarthVQA model requires high-specification GPUs and consumes a considerable amount of time, this study uses the pre-trained model provided by EarthVQA for vegetation extraction. The pre-trained model achieves a mean mIoU of 57.34% and an OA of 78.43% on semantic segmentation tasks, providing a reliable performance for subsequent research.

Because EarthVQA is pre-trained on the LoveDA dataset, which comprises satellite imagery of Wuhan and its corresponding semantic segmentation masks, the pre-trained model exhibits excellent capability in the semantic segmentation tasks targeting Wuhan satellite images. As illustrated in Figure 3, the study found that EarthVQA can perform semantic segmentation on 915 block images with high accuracy.

Figure 3a–c show that the satellite images of the 915 blocks are all high-resolution images containing rich urban elements. As shown in Figure 3d–f, EarthVQA is able to extract the two main urban elements, buildings and vegetation. From Figure 3a,c,d,f, EarthVQA also identifies and extracts smaller urban elements, such as lakes and sports fields, effectively. On the other hand, because the road features in the satellite images are often occluded by vegetation, it is difficult for EarthVQA to identify and extract complete road elements.

To ensure consistency, a rectangular imagery extent of 1024 m × 1024 m was defined for each block, based on its centroid. The cropped satellite image slices were then resized to a fixed resolution of 512 × 512 pixels.

As shown in Figure 4, to facilitate the subsequent 3D-model generation process, semantic segmentation images were processed to extract vegetation pixels using binary masks.

2.2.2. 3D Block Model Generation

As shown in Figure 5, the study imports both 3D building data (.shp and .dbf files) and vegetation data (vegetation binary image) into the modeling software to generate 3D block models. The 3D vegetation models are randomly generated within these boundaries, according to their area. Because Cinnamomum camphora and Koelreuteria bipinnata accounted for 50.7% and 9.5% of the total vegetation in Wuhan [34,48], the study considered them as the dominant urban tree species in Wuhan. Generally, the crown radius of Cinnamomum camphora and Koelreuteria bipinnata range from 3 to 5 m and 3 to 4 m, respectively. To simplify calculations, a uniform crown radius of 3 m was established for both species. Furthermore, to prevent mutual shading between trees, the tree density was fixed at one tree per 30 m².

2.3. Step 3 (Morphological Indicator Screening)

2.3.1. The Morphological Indicators

To comprehensively characterize block morphology, this study uses three categories of morphological indicators: 1. Overall morphological indicators (7); 2. Building morphological indicators (10); and 3. Vegetation morphological indicators (7). The detailed specifications of these indicators are presented in

Table 2. The morphological indicators.

Type	Name	Nomenclature	Equation	Describe	Unit
Overall morphological indicators	Building number	$\mathit{BN}$	None	$\mathit{BN}$ is the total number of buildings in the blocks.	None
	Block Boundary Area	$\mathit{BBA}$	None	$\mathit{BBA}$ is the boundary area of the blocks.	${\mathrm{m}}^2$
	Block Aspect Ratio	${BR}_{axis}$	${BR}_{axis}={\displaystyle \frac{{Axis}_{N-S}}{{Axis}_{E-W}}}$	${Axis}_{N-S}$ is the length of the North–South axis of the blocks, while ${Axis}_{E-W}$ represents the length of the East–West axis.	None
	Compactness Ratio	$CR$	$CR={\displaystyle \frac{{Area}_{Actual}}{{Area}_{Rectangle}}}$	${Area}_{Actual}$ is the actual area of the blocks, while ${Area}_{Rectangle}$ represents the area of its minimum bounding rectangle.	None
	Total Facade Area	${FA}_{all}$	${FA}_{all}={\sum }_{i=1}^n{FA}_i$	$n$ is the total number of buildings in the blocks, and ${FA}_i$ is the total facade area of the i-th building.	${\mathrm{m}}^2$
	Building Area Ratio	$\mathit{BAR}$	$\mathit{BAR}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{\mathit{BA}}_i}{\mathit{BBA}}}$	$\mathit{BBA}$ is the area of the blocks; ${\mathit{BA}}_i$ is the footprint area for the $i$ building.	None
	Floor Area Ratio	$\mathit{FAR}$	${\mathit{BA}}_{\mathit{ave}}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{(\mathit{BA}}_i\text{×}\frac{{\mathit{BH}}_i}{\mathit{FH}})}{\mathit{BBA}}}$	$\mathit{FH}$ is the floor height of buildings, set at 3 m.	None
Building morphological indicators	Block Area	$\mathit{BA}$	$\mathit{BA}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{\mathit{BA}}_i}{n}}$	${\mathit{BA}}_i$ is the footprint area of the $i$ building. $N$ is the number of buildings in the blocks.	${\mathrm{m}}^2$
	Building Shape Coefficient	$\mathit{BSC}$	$\mathit{BSC}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}(\frac{F_i}{V_i})}{n}}$	$F_i$ is the exterior surface area of the $i$ building. $V_i$ is the volume of the $i$ building.	None
	Mean Building Spacing	$\mathit{MBS}$	$\mathit{MBS}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}\mathit{min}{d}_{\mathit{ij}}}{n}}$	$\mathit{min}{d}_{\mathit{ij}}$ is the minimum spacing value between a single building and all its adjacent buildings.	m
	Mean Building-to-Center Distance	$\mathit{BTC}$	$\mathit{BTC}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{\mathit{BTC}}_i}{n}}$	${\mathit{BTC}}_i$ is the distance from the $i$ building to the block center.	m
	Building Height	$\mathit{BH}$	$\mathit{BH}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{\mathit{BH}}_i}{n}}$	${\mathit{BH}}_i$ is the height of the $i$ building.	m
	Average Sky View Factor	${SVF}_{ave}$	${SVF}_{ave}={\displaystyle \frac{{\sum }_{i=1}^{n_T}{SVF}_i}{n_T}}$	A 10 m × 10 m grid of test points was made in the blocks. $n_T$ is the total number of test points, and ${SVF}_i$ is the Sky View Factor (SVF) value of the i-th test point.	None
	Projected Area Ratio North–South	${PA}_{N-S}$	${PA}_{N-S}={\displaystyle \frac{{PB}_{N-S}}{{PR}_{N-S}}}$	${PB}_{N-S}$ is to the projected building area of the blocks on the East–West projection plane; ${PR}_{N-S}$ is the area of the rectangular projection of the blocks on the East–West projection plane.	None
	Projected Area Ratio East–West	${PA}_{E-W}$	${PA}_{E-W}={\displaystyle \frac{{PB}_{E-W}}{{PR}_{E-W}}}$	${PB}_{E-W}$ is to the projected building area of the blocks on the North–South projection plane; ${PR}_{E-W}$ is the area of the rectangular projection of the blocks on the North–South projection plane.	None
	Average Orientation Angle	${\mathit{OAS}}_{\mathit{ave}}$	${\mathit{OAS}}_{\mathit{ave}}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}\left|{\mathit{OAS}}_i\right|}{n}}$	${\mathit{OAS}}_i$ is the angle between the main facade of the $i$ building and true north.	°
	Average Aspect Ratio	${AR}_{ave}$	${AR}_{ave}={\displaystyle \frac{{\sum }_{i=1}^n\frac{\mathit{max}(L_i,W_i)}{\mathit{min}(L_i,W_i)} }{n}}$	$n$ is the total number of buildings; $L_i$ is the length of the i-th building; and $W_i$ is the width of the i-th building.	None
Vegetation morphological indicators	Vegetation Area Ratio	$VAR$	$VAR={\displaystyle \frac{VA}{\mathit{BBA}}}$	$VA$ is the vegetation area in the blocks, while $\mathit{BBA}$ is the total site area of the blocks.	None
	Vegetation-to-Center Distance	$\mathit{VTC}$	$\mathit{VTC}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n_v}}{\mathit{VTC}}_i}{n_v}}$	$n_v$ is the total number of vegetation spaces in the blocks.${\mathit{VTC}}_i$ is the distance from the $i$-th vegetation space to the block center.	m
	Mean Vegetation Spacing	$\mathit{MVS}$	$\mathit{MVS}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n_v}}\mathit{min}{d}_{\mathit{ij}}}{n_v}}$	$n_v$ is the total number of vegetation spaces in the blocks. $\mathit{min}{d}_{\mathit{ij}}$ is the minimum spacing value between a single vegetation space and all its vegetation spaces.	m
	Vegetation Distance to South Boundary	${VD}_S$	${VD}_S={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n_v}}{VD}_S^i}{n_v}}$	$n_v$ is the total number of vegetation spaces in the blocks. ${VD}_S^i$ is the distance from the i-th plant space to the southern boundary of the blocks.	m
	Vegetation Distance to North Boundary	${VD}_N$	${VD}_N={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n_v}}{VD}_N^i}{n_v}}$	$n_v$ is the total number of vegetation spaces in the blocks. ${VD}_N^i$ is the distance from the i-th plant space to the northern boundary.	m
	Vegetation Distance to East Boundary	${VD}_E$	${VD}_E={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n_v}}{VD}_E^i}{n_P}}$	$n_v$ is the total number of vegetation spaces in the blocks. ${VD}_E^i$ is the distance from the i-th plant space to the east boundary of the blocks.	m
	Vegetation Distance to West Boundary	${VD}_W$	${VD}_W={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n_v}}{VD}_W^i}{n_v}}$	$n_v$ is the total number of vegetation spaces in the blocks. ${VD}_W^i$ is the distance from the i-th plant space to the west boundary.	m

.

2.3.2. The Indicator Screening

This study uses multiple dimensionality reduction methods to screen morphological indicators. While previous research has relied on Principal Component Analysis (PCA) [49], this method focuses on linear structures and often exhibits limited performance when handling non-linear and complex data relationships. Consequently, in addition to PCA, this study incorporates non-linear dimensionality reduction methods, specifically t-SNE [50] and UMAP [51]. Unlike PCA, which is a linear dimensionality reduction technique, t-SNE (t-distributed stochastic neighbor embedding) and UMAP (Uniform Manifold Approximation and Projection) can capture complex manifold structures in the data, such as curved or folded feature layouts.

The performance of the dimensionality reduction was evaluated using two indicators: Trustworthiness [52] and k-Nearest Neighbors (kNN) [53]. The optimal results from each method were selected based on these indicators.

Trustworthiness quantifies the changes in neighboring rankings for each data point between the high-dimensional and low-dimensional spaces. If a point $i$ is a neighbor of point $j$ in the high-dimensional space but becomes distant in the low-dimensional space, the dimensionality reduction is considered less reliable, resulting in a lower Trustworthiness score. The equation is

$$T(k)=1-{\displaystyle \frac{2}{nk(2n-3k-1)}}{\sum }_{i=1}^n\sum _{j\in v_i}\left(r\right(i,j)-k)$$

(1)

$k$ is the predefined number of neighbors, set to the default value of 15 in this study; $n$ is the total number of data points; $v_i$ is the set of points that are among the $k$ nearest neighbors of point $i$ in the low-dimensional space but not in the high-dimensional space; and $r(i,j)$ is the rank of point $j$ relative to point $i$ in the high-dimensional space.

kNN is a core indicator for quantifying the quality of dimensionality reduction results. This study uses Jaccard similarity ($J\left(i\right)$) [54] to calculate the neighborhood consistency for each data point, and then averages these values across all points. A higher kNN value indicates a greater overlap between the k -nearest neighbor sets in the high-dimensional and low-dimensional spaces, signifying a better dimensionality reduction performance. The equation is as follows:

$$J(i) = {\displaystyle \frac{\left|N_H\left(i\right)\cap N_L\left(i\right)\right|}{\left|N_H\left(i\right)\cup N_L\left(i\right)\right|}}$$

(2)

$$kNN={\displaystyle \frac{1}{n}}{\sum }_{i=1}^{n}J\left(i\right)$$

(3)

$N_H\left(i\right)$ is the set of k nearest neighbors for data point $i$ in the high-dimensional space, while $N_L\left(i\right)$ is the corresponding set in the low-dimensional space.

Due to the inherent randomness in low-dimensional mapping of t-SNE and approximate nearest neighbor search of UMAP, both algorithms exhibit stochastic behavior during dimensionality reduction. To ensure robust results, the study evaluated 2000 random initializations (seeds) for each algorithm. The results were quantitatively assessed using trustworthiness and kNN values. The optimal configuration for each algorithm was then compared against PCA. Finally, the definitive dimensionality reduction method and its corresponding projection were selected, based on these evaluation metrics.

2.3.3. The Importance of Indicators

This study evaluates the importance of morphological indicators, based on the dimensionality reduction results. For PCA, indicator importance was calculated by weighing the factor loadings with their explained corresponding variance. In contrast, unlike PCA, t-SNE and UMAP do not inherently provide importance metrics. Therefore, a post hoc correlation analysis was used to quantify the contribution of each morphological indicator within the t-SNE and UMAP embeddings.

For t-SNE, indicator importance was evaluated using the Pearson correlation coefficients between each morphological indicator and the 3 dimensional axes of the t-SNE. The importance score for a given morphological indicator, denoted as $M_i$, was calculated using the following equation:

$${importance}_i^t= \left|r_i^1\right|+ \left|r_i^2\right|+ \left|r_i^3\right|$$

(4)

where $r_i^1$, $r_i^2$, and $r_i^3$ represent the Pearson correlation coefficients between the $M_i$ and the first, second, and third dimensions of the t-SNE coordinates, respectively.

For UMAP, indicator importance was evaluated using the Pearson correlation coefficients between each morphological indicator and the 3 dimensional axes of the UMAP. The importance score for a given morphological indicator, denoted as $M_i$, was calculated using the following equation:

$${importance}_i^U= \left|p_i^1\right|+ \left|p_i^2\right|+ \left|p_i^3\right|$$

(5)

where $p_i^1$, $p_i^2$, and $p_i^3$ represent the Pearson correlation coefficients between the $M_i$ and the first, second, and third dimensions of the UMAP coordinates, respectively.

For PCA, the minimum number (K) of principal components required to achieve a cumulative variance which was explained, of 85%, was determined. Subsequently, the loading matrix of these top K components was extracted to compute the importance score for each indicator. The PCA importance value for morphological indicator $M_i$ was calculated using the following equation:

$${importance}_i^p= {\displaystyle \sum _{j=1}^K}\left|L_{ji}\right|\times W_j$$

(6)

where $L_{ji}$ is the loading of $M_i$ on the j-th principal component, and $W_j$ is the proportion of variance explained by that principal component.

To achieve a comprehensive assessment of the morphological indicators, this study evaluates and compares their importance rankings derived from the three dimensionality reduction methods: t-SNE, UMAP, and PCA.

2.4. Step 4 (Morphology Clustering)

This study uses multiple clustering algorithms to categorize blocks based on the screened morphological indicators. Since clustering analysis involves a certain degree of randomness, four different types of clustering algorithms were employed to ensure the robustness of the results: 1. K-means [55]; 2. Gaussian Mixture Model (GMM) [56]; 3. DBSCAN [57]; and 4. Spectral clustering [58]. Among them, K-means is a distance-based partitioning method, GMM is a probability-based partitioning method, DBSCAN is a density-based partitioning method, and Spectral clustering is a graph–theory-based partitioning method. Detailed descriptions of these algorithms and their hyperparameter search ranges are provided in

Table 3. The four cluster algorithms’ settings.

Algorithm	Type	Key Mechanism	Hyperparameters
K-means	Centroid-based	Optimized centroid initialization via seeding	Cluster number: $K\in (5,6,7,8,9,10)$ Number of initializations: $n\_init=10$ Maximum number of iterations: $max\_inter =300$
GMM (Gaussian mixture model)	Probabilistic	Expectation–Maximization (EM) fitting of Gaussian distributions	Cluster number: $K\in (5,6,7,8,9,10)$ Covariance type: $[‘full’,‘tied’,‘diag’,‘spherical’]$ Number of initializations: $n\_init = 10$ Maximum number of iterations: $max\_inter =300$
DBSCAN (Density-based cluster algorithm)	Density-based	Eps-neighborhood connectivity with noise filtering	$eps\in [0.1,2.0]$ $min\_samples\in (2,3,\dots ,16)$ Constraint: $K\in (5,6,7,8,9,10)$
Spectral	Graph-based	Laplacian eigen-decomposition for manifold separation	Cluster number: $K\in (5,6,7,8,9,10)$ Affinity: ‘rbf’ or ‘nearest_neighbors’ $gamma\in [{10}^{-3},{10}^1]$ Number of initializations: $n\_init = 10$ Maximum number of iterations: $max\_inter =300$

.

The study optimizes the hyperparameters for each algorithm through 300 iterative runs. To optimize clustering hyperparameters, the study performed 300 iterations, each incorporating 10 random initializations to mitigate the impact of stochasticity on the clustering results. The selection of the optimal hyperparameter configuration is based on the Silhouette Coefficient (SC) [59] and the Davies–Bouldin (DB) Index [60]. The equation of SC (SC∈[−1,1]) is

$$S\left(i\right)={\displaystyle \frac{b\left(i\right)- a\left(i\right)}{\mathit{max}\left(a\left(i\right), b\left(i\right)\right)}}$$

(7)

where $a\left(i\right)$ is the mean intra-cluster distance for sample $i$, and $b\left(i\right)$ represents the mean nearest-cluster distance. Higher Silhouette values approaching 1 indicate better defined and well-separated clusters.

The equation of DB (DB > 0) is

$$\mathit{DB}={\displaystyle \frac{1}{k}}\sum _{i=1}^k\underset{j\ne i}{\mathit{max}}\left({\displaystyle \frac{{\sigma }_i+{\sigma }_j}{d_{\mathit{ij}}}}\right)$$

(8)

where ${\sigma }_i$ is the mean intra-cluster distance of cluster $i$ and $d_{\mathit{ij}}$ is the centroid distance between clusters $i$ and $j$. Lower DB values indicate improved cluster compactness and separation.

Furthermore, the study evaluated the final clustering results through stability and sensitivity analyses.

For the stability analysis, we fixed the cluster number ($\mathrm{K}$) derived from the results and performed 300 iterations of random subsampling, extracting 80% of the original dataset ($\mathrm{N}$ samples) in each run [61]. The sampled data were re-clustered 300 times, based on $\mathrm{K}$. The study compared the consistency between the new cluster labels ($\mathrm{V}$) and the original labels ($\mathrm{U}$). The Adjusted Rand Index ($\mathrm{A}\mathrm{R}\mathrm{I}$) [62] and Normalized Mutual Information ($\mathrm{N}\mathrm{M}\mathrm{I}$) [63] were used to quantify the discrepancy between the 300 re-clustering results and the original baseline.

$$ARI(U,V)={\displaystyle \frac{{\sum }_{i,j}(\begin{array}{c}{n}_{ij}\\ 2\end{array})-\frac{{\sum }_i(\begin{array}{c}{a}_i\\ 2\end{array}){\sum }_j(\begin{array}{c}{b}_j\\ 2\end{array})}{(\begin{array}{c}n\\ 2\end{array})}}{\frac{[{\sum }_i(\begin{array}{c}{a}_i\\ 2\end{array})+{\sum }_j(\begin{array}{c}{b}_j\\ 2\end{array})]}{2}-\frac{{\sum }_i(\begin{array}{c}{a}_i\\ 2\end{array}){\sum }_j(\begin{array}{c}{b}_j\\ 2\end{array})}{(\begin{array}{c}n\\ 2\end{array})}}}$$

(9)

$$MI(U,V)={\sum }_{i,j}{\displaystyle \frac{n_{ij}}{n}}\ln ({\displaystyle \frac{n_{ij}n}{a_i{b}_j}}), H(U)=-{\sum }_i{\displaystyle \frac{a_i}{n}}\ln ({\displaystyle \frac{a_i}{n}}), H(V)=-{\sum }_i{\displaystyle \frac{b_j}{n}}\ln ({\displaystyle \frac{b_j}{n}})$$

(10)

$$NMI(U,V)={\displaystyle \frac{2MI(U,V)}{H(U)+H(V) }}$$

(11)

For the sensitivity analysis, morphological indicators were first standardized, and independent noise data was generated for each indicator within every block. As expressed in Equation (8), the noisy indicator value ($x^{noisy}$) was randomly generated using the standardized value as the baseline and the noise intensity ($\sigma$) as the standard deviation. ${\epsilon }_{ij}$ represents the noise for the j th morphological indicator of the i th block.

$$x^{noisy}=x+\epsilon , {\epsilon }_{ij} \sim N(0,{\sigma }^2)$$

(12)

As detailed in

Table 4. The noise intensity settings.

$\mathit{\sigma }$	Description
0.02	Data fluctuation of approximately 2% standard deviation
0.05	5% data fluctuation, similar to standard measurement or computational errors
0.10	Moderate data fluctuation
0.15	Substantial data fluctuation
0.20	Drastic data fluctuation

, five levels of $\sigma$ were defined. For each intensity, 200 noisy datasets were generated through random generation. Using the cluster number $\mathrm{K}$ from the final model, the study re-clustered these 200 noisy datasets. Finally, the ARI was used to quantify the label discrepancies between the 200 noise clustering results and the final cluster result across each noise intensity.

2.5. Step 5 (Thermal Environment Simulation)

2.5.1. ENVI-Met Simulation Settings

This study utilizes the Morpho plugin within the Grasshopper parametric platform [64] to convert Rhino 3D models into ENVI-met simulation models. This study uses ENVI-met 5.7.1 to simulate the thermal environment of target blocks. ENVI-met discretizes the simulation domain into a 3D grid and utilizes iterative calculations to solve the temporal variations of physical parameters—such as air temperature, wind speed, and humidity—at each time step.

As shown in Figure 6, to balance computational efficiency and standardization, the grid dimensions were set to 12 m × 12 m × 6 m along the X, Y, and Z axes, respectively. Consequently, the computational domain of 600 m × 600 m × 120 m is discretized into a 3D grid of 50 × 50 × 20 (totaling 50,000 cells). As shown in

Table 5. The ENVI-met materials settings.

Name	Description	Code
Wall	LEGACY Brick wall (reinforced)	0100B3
Roof	LEGACY Roofing tile	0100R1
Tree	LEGACY Spherical, medium trunk, dense, medium (15 m)	01SMDM
Road	Asphalt road	0200ST
Pavement	Pavement (concrete), used/dirty	0200PP
Green	LEGACY Green, mixed substrate	01NADS
Ground	LEGACY Default unsealed soil (sandy loam)	010000

, uniform material properties were assigned to the elements [65].

Due to the macro-spatial scale of this study, vegetation geometries were abstracted into a uniform type to ensure computational feasibility and prevent ENVI-met convergence failure. As detailed in

Table 6. The tree parameter settings.

Parameters	01SMDM	Cinnamomum camphora	Koelreuteria bipinnata
Crown geometry	Spherical	Spherical	Ovoid
Tree height	15 m	12–18 m	12–15 m
Trunk height	2.0–2.5 m	2.0–2.5 m	2.2–2.6 m
Crown diameter	6.0–8.0 m	6.0–9.0 m	6.0–8.0 m
Leaf Area Index (LAI)	4.5–5.5	4.2–5.6	4.0–5.0
Max Leaf Area Density (Max LAD)	1.5–2.0 (${\mathrm{m}}^2/{\mathrm{m}}^3$)	1.4–1.9 (${\mathrm{m}}^2/{\mathrm{m}}^3$)	1.3–1.8 (${\mathrm{m}}^2/{\mathrm{m}}^3$)
Leaf albedo	0.18	0.16–0.19	0.17–0.20

, the 01SMDM tree model in ENVI-met is similar to the average dimensions of Cinnamomum camphora and Koelreuteria bipinnata. Consequently, all modeled vegetation was standardized using the 01SMDM.

This study simulates a 24 h period from 01:00 to 24:00 on 10 August, based on a typical meteorological year for Wuhan. As shown in Figure 7, the dry-bulb temperature on 10 August reaches 37.36 °C, the highest of the year. The study therefore analyzes the outdoor thermal °C environment of typical Wuhan blocks on this day. The simulation parameters are detailed in

Table 7. The ENVI-met simulation settings.

Name	Description	Value	Unit
Start date	The start date for the simulation is set to August 10.	01.08	None
Start time	The specific start time for the simulation is set to 01:00:00.	01:00:00	None
Duration	The duration of the simulation is set to 24 h.	24	Hour
Wind Speed	The default value was used.	2.5	m/s
Wind direction	The default value was used.	0	°
Roughness	The default value was used.	0.01	None
Initial temperature	The dry-bulb temperature at 01:00 on August 1 from the EPW file was used as the initial temperature.	31.3	°C
Specific humidity	Calculated based on the hourly dry-bulb temperature and relative humidity for the 24 h of August 1 from the EPW file.	20.739672	g/kg
Relative humidity	The relative humidity at 01:00 on August 1 from the EPW file was used as the initial relative humidity.	84	%
Wind limit	The default value was used.	5.0	m/s

. Specific humidity is a critical input parameter for ENVI-met; the value is derived from the hourly air temperature and relative humidity data for 10 August found in the typical year file. Since this study aims to compare the thermal environment performance of different blocks under identical conditions, the outdoor wind environment was not set independently for each block. Instead, the default ENVI-met wind environment setting was applied, with a south prevailing wind of 2.5 m/s at a height of 10 m.

2.5.2. The Thermal Environment Indicators

This study uses multiple temperature indicators to evaluate the thermal environment performance at the pedestrian level (1.5 m) in the blocks. These indicators include the average mean air temperature ($T_{ave}$), the root mean square error of the mean air temperature ($T_{RMSE}$), and the temperature difference between the interior and exterior of the blocks ($T_{dif}$). Specifically, $T_{ave}$ characterizes the overall thermal level in the blocks, $T_{RMSE}$ reflects the fluctuation of the temperature distribution, and $T_{dif}$ measures the thermal disparity between the interior and exterior environments. The study will also calculate ${PET}_{ave}$ and ${UTCI}_{ave}$ for every block. The equations for these five indicators are provided in

Table 8. The thermal indicators.

Nomenclature	Equation	Describe
$T_{ave}$	$T_{ave}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}(\frac{{\sum }_{k=1}^j{T}_{ik}}{j})}{n}}$	$T_{ik}$ is the temperature value at the k-th hour for the i-th probe point in the blocks. j is the total duration of the simulation period. n is the total number of probe points in the blocks. $T_k^a$ is the temperature value at the k-th hour for the a-th probe point outside the blocks.m is the total number of probe points outside the neighborhood. ${PET}_{ik}$ is the PET value at the k-th hour for the i-th probe point in the blocks. ${UTCI}_{ik}$ is the UTCI value at the k-th hour for the i-th probe point in the blocks.
$T_{RMSE}$	$T_{RMSE}={\displaystyle \frac{{\sum }_{i=1}^n\sqrt{\frac{{\sum }_{k=1}^j{(T_{ik}-\frac{{\sum }_{k=1}^j{T}_{ik}}{j} )}^2}{j}}}{n}}$
$T_{dif}$	$T_{dif}={\displaystyle \frac{{\displaystyle {\sum }_{a=1}^m}(\frac{{\sum }_{k=1}^j{T}_k^a}{j})}{m}}-T_{ave}$
${PET}_{ave}$	${PET}_{ave}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}(\frac{{\sum }_{k=1}^j{PET}_{ik}}{j})}{n}}$
${UTCI}_{ave}$	${UTCI}_{ave}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}(\frac{{\sum }_{k=1}^j{UTCI}_{ik}}{j})}{n}}$

.

2.6. Step 6 (Regression Analysis)

Since correlations among indicators can affect the performance of regression models, the study conducted a Pearson correlation analysis on the indicators before performing regression analysis. Pearson correlation measures the strength and direction of the linear relationship between two continuous variables. It ranges from −1 to 1, where 1 indicates a perfect positive correlation, −1 a perfect negative correlation, and 0 no linear relationship [66].

This study establishes predictive models based on the morphological indicators and the objective functions ($T_{ave}$, $T_{RMSE}$, $T_{dif}$), using regression algorithms. Before the regression analysis, a Variance Inflation Factor (VIF) analysis was performed to detect potential multicollinearity among the indicators. The dataset was split, with 80% of the data allocated for training and the remaining 20% reserved for model validation. Given the correlations among the selected morphological indicators, ridge regression is employed to mitigate multicollinearity issues [67]. Ridge regression is a modified linear regression technique that incorporates L2 regularization (penalty term: λ∑β²) to shrink coefficients and enhance model stability. This method effectively mitigates overfitting while retaining all predictor variables.

Furthermore, the validity of the ridge regression model was evaluated using a 5-fold cross-validation approach. The entire dataset was randomly partitioned into K folds (K = 5), reserving one fold for model validation in each iteration. The regression model was sequentially refitted onto the standardized data of the remaining folds. The $R^2$ scores across all folds were then aggregated to calculate the mean and standard deviation.

Finally, the regression assumptions were assessed based on residual diagnostics. The relationship between the residuals and fitted values was examined to verify linearity and mean bias. Residual normality was evaluated using a Quantile–Quantile (Q-Q) plot, and potential residual issues were further checked using a histogram combined with the Shapiro–Wilk test.

2.7. Step 7 (Spatial Distribution Analysis)

This study collected Point of Interest (POI) data for Wuhan from the Baidu dataset, including six categories: dining services, shopping services, corporate offices, educational and scientific facilities, public amenities, and green spaces. Using a parametric platform, the number of POIs within a 100 m radius of each block is quantified. Based on the cluster labels, the average number of surrounding POIs for each cluster is calculated, to analyze the functional characteristics of different block types. Furthermore, the spatial aggregation of blocks is analyzed based on cluster labels, by calculating the intra-cluster distances. Finally, the spatial distribution patterns of each cluster are investigated, based on the distribution of these groups.

3. Results

3.1. The Morphological Indicators

To ensure the comprehensiveness of the selected indicators, this study performed screening based on three categories. As shown in

Table 9. Trustworthiness and kNN indicators.

Morphology Indicator Types	Dimensionality Reduction Indicators	PCA	t-SNE	UMAP
Overall Morphological Indicators	Trustworthiness	0.8914	0.9830	0.9662
	KNN	0.3474	0.5686	0.4857
Building Morphological Indicators	Trustworthiness	0.9023	0.9721	0.9519
	KNN	0.2693	0.5091	0.4162
Vegetation Morphological Indicators	Trustworthiness	0.9638	0.9839	0.9732
	KNN	0.4846	0.5542	0.4648

, a quantitative analysis of the results from three dimensionality reduction methods yielded the following findings:

• In terms of overall morphological indicators, the PCA method scored lower on both Trustworthiness and kNN indicators (0.8914 and 0.3474) compared to t-SNE (0.9830 and 0.5686) and UMAP (0.9662 and 0.4857). Furthermore, while t-SNE and UMAP achieved comparable scores in Trustworthiness (0.9830 vs. 0.9662), t-SNE demonstrated significantly superior performance in the kNN (0.5686 vs. 0.4857).
• In terms of building morphological indicators, PCA also exhibited lower scores in Trustworthiness and kNN (0.9023 and 0.2693) than t-SNE (0.9721 and 0.5091) and UMAP (0.9519 and 0.4162). Similarly, although t-SNE and UMAP showed similar Trustworthiness, t-SNE (0.5091) significantly outperformed UMAP (0.4162) in the kNN.
• In terms of vegetation morphological indicators, the differences in Trustworthiness among the three methods were negligible (PCA: 0.9638; t-SNE: 0.9839; UMAP: 0.9732). However, in terms of the kNN, t-SNE (0.5542) was higher than PCA (0.4846) and UMAP (0.4648), which showed minimal difference between them.

Generally, this study uses t-SNE to screen indicators based on their importance derived from the dimensionality reduction process across the three categories.

To validate the result of the t-SNE importance scoring analysis, the study compared the result from t-SNE against those from PCA. As shown in Figure 8, the indicator importance evaluations from t-SNE and PCA were compared, where Figure 8a–c and Figure 8d–f present the importance scores for t-SNE and PCA, respectively. A comparison between Figure 8a,d reveals notable similarities between the two methods in analyzing overall morphological indicators, with both identifying $BN$, ${FA}_{all}$, and $BBA$ as critical morphological indicators. Specifically, the importance scores for $BN$ exceeded 1.2 (t-SNE) and 0.3 (PCA), ${FA}_{all}$ exceeded 1.0 (t-SNE) and 0.3 (PCA), and $BBA$ exceeded 1.4 (t-SNE) and 0.3 (PCA). Similarly, both algorithms highlighted ${SVF}_{ave}$, ${AR}_{ave}$, and ${OSA}_{ave}$ as key morphological indicators, with their scores exceeding 1.4, 1.0, and 0.8 for t-SNE, and all surpassing 0.25 for PCA. Furthermore, $VAR$, $VTC$, and ${VD}_w$ were recognized by both methods as the important landscape morphological indicators, with scores exceeding 1.2, 1.0, and 1.2 for t-SNE, and all exceeding 0.35 for PCA. Despite discrepancies in lower-ranked indicators, the substantial consensus among top-ranked indicators validates the t-SNE scoring framework. Consequently, the t-SNE results were used as the dataset for downstream indicator selection.

This study ranked the three categories of morphological indicators based on their importance in the t-SNE dimensionality reduction process. Indicators with an importance score exceeding 1.0 were selected as key indicators for the subsequent cluster analysis. As shown in Figure 8a–c, regarding the overall morphological indicators, three indicators with an importance score exceeding 1.0—$\mathit{BBA}$,$\mathit{BN}$, and ${FA}_{all}$—were identified as the key indicators for characterizing the overall morphology of the blocks. In terms of building morphological indicators, ${SVF}_{ave}$ and ${AR}_{ave}$, with importance scores exceeding 1.0, were identified as key indicators. Additionally, ${\mathit{OAS}}_{\mathit{ave}}$ was incorporated as a significant indicator due to its high importance score of over 0.9. In terms of vegetation morphology, three indicators with importance scores exceeding 1.0—$VAR$, ${VD}_W$, and $\mathit{VTC}$—were considered as the key indicators for characterizing the vegetation morphology of the blocks.

As shown in Figure 9, this study conducted a correlation analysis on the selected morphological indicators. The results showed that among the nine indicators, only two pairs had absolute correlation coefficients exceeding 0.6: ${VD}_W$ with $BBA$, and ${VD}_W$ with $VTC$, both indicating relatively strong correlations. This suggests that the key morphological indicators obtained through the t-SNE dimensionality reduction method do not exhibit strong correlations with one another, and the degree of information overlap among them is low. This further supports the reasonableness of the results obtained through the t-SNE method.

3.2. The Cluster Result

3.2.1. The Cluster Result Comparison

As shown in Figure 10, both DBSCAN and the Spectral clustering algorithms failed to generate effective clustering results. Specifically, the final DBSCAN result contained 97.8% noise points, while 99.8% of the data in the Spectral clustering result belonged to a single cluster. This means these two algorithms are unsuitable for the clustering analysis in this study.

As shown in Figure 11, this study evaluated the clustering performance of K-means and GMM across different K values using the SC and DB indicators. The findings are as follows:

• Regarding K-means, the SC peaked at K = 5 (0.207), while the DB was lowest at K = 7 (1.397). However, at K = 6, the SC (0.204) was only marginally lower than that of K = 5 (0.207), whereas the DB (1.402) was significantly better than that of K = 5 (1.468). Compared to K = 7, K = 6 exhibited a significantly higher SC (0.204 vs. 0.195) and a marginally higher DB (1.402 vs. 1.397). Balancing these indicators, K = 6 was determined to be the most reasonable clustering result for K-means.
• Regarding GMM, the SC was highest at K = 6 (0.176), and the DB was lowest at K = 5 (1.411). At K = 6, the SC was notably higher than that of K = 5 (0.176 vs. 0.161), while the DB (1.414) was only slightly higher than that of K = 5 (1.411). Consequently, K = 6 was identified as the optimal configuration for GMM.
• For comparison at K = 6, K-means achieved a SC of 0.204 and a DB of 1.402, whereas GMM has 0.176 and 1.414, respectively. These results indicate that K-means significantly outperformed GMM at K = 6. Therefore, the K-means clustering result with K = 6 was selected as the final cluster result.

3.2.2. The Cluster Stability

Figure 12 illustrates the distributions of ARI and NMI between the 300 random subsampling runs and the final clustering results. The ARI has a mean of 0.706 and a median of 0.724. The NMI values were slightly higher and more tightly distributed, with a mean of 0.717 and a median of 0.726. Both histograms are left toward high consistency, with no individual run falling below an ARI of 0.43 or an NMI of 0.52. Although the subsampled results do not match the final cluster perfectly (ARI and NMI < 1), most resampled solutions retain the identical overall grouping structure, thereby validating the robustness of the clustering.

3.2.3. The Cluster Sensitivity

Figure 13 shows the ARI values between the noise clustering results and final clustering result, derived from 200 iterations of noise generation and re-clustering per noise level. Across all evaluated fluctuation levels ($\sigma$ = 0.02, 0.05, 0.10, 0.15 and 0.20), the median ARI remained consistently high at 0.864, 0.851, 0.829, 0.806, and 0.777, respectively. The mean ARI decreased from 0.839 to 0.756. Notably, even under the most drastic fluctuation ($\sigma$ = 0.20), the median ARI (0.777) remained well above the random agreement threshold. These findings demonstrate that the final clustering result exhibits low sensitivity to low-to-moderate noise, showing mild sensitivity only under substantial noise fluctuation ($\sigma$ = 0.15) without suffering a collapse in cluster consistency.

As illustrated in Figure 14, the K-means clustering results indicated that Cluster 5 comprised only 16 samples, accounting for merely 1.8% of the total population. This suggests that the blocks in Cluster 5 are not representative of typical residential blocks in Wuhan. Consequently, these blocks were excluded from the dataset.

The study compared the mean values of key morphological indicators across the five clusters derived from K-means clustering (as shown in Figure 15). The main findings are as follows:

• Cluster 0: Exhibits the highest ${AR}_{ave}$, indicating a prevalence of slab-type buildings.
• Cluster 1: Shows the highest ${SVF}_{ave}$ and the largest ${OAS}_{ave}$, suggesting a loose and irregular building layout.
• Cluster 2: Exhibits the largest $VAR$, indicating superior greening conditions.
• Cluster 3: Shows a relatively large ${OAS}_{ave}$ but a significantly lower ${SVF}_{ave}$ compared to other clusters, suggesting an irregular but high-density building layout.
• Cluster 4: Contains the highest $\mathit{BN}$, the largest $\mathit{BBA}$, and the greatest ${FA}_{all}$, indicating large-scale blocks with high building density.

3.3. The Thermal Environment Simulation Result

3.3.1. The ENVI-Met Simulation Results

Due to the computationally intensive ENVI-met simulations, this study selected 25 typical blocks closest to the cluster center from each of the five clusters (totaling 125 blocks) for simulation. As shown in

Table 10. The thermal environment visualization of five block clusters.

Cluster Number	The Simulation Visualization of Five Clusters
0
1
2
3
4
	$T_{ave}$	$T_{RMSE}$

, the study found that

• Significant variations in $T_{ave}$ were observed across different spatial locations in the cases of each cluster.
• Significant differences in $T_{RMSE}$ values were evident at different spatial locations in the cases of each cluster.

3.3.2. The Simulation Results Validation

To validate the accuracy of the ENVI-met simulation, field measurements were conducted in Wuhan from 13 June to 21 June 2026. As illustrated in Figure 16, three residential blocks in three districts in Wuhan were selected for field measurement: Community A (Dongting), Community B (Changqing Garden 4), and Community C (Haihongjing). As summarized in

Table 11. The photos of the measurement points.

Point	Photo	Point	Photo	Point	Photo
A1		B1		C1
A2		B2		C2
A3		B3		C3

, three monitoring points were established within each block. Specifically, points A1, B1, and C1 were positioned in outdoor plazas; A2, B2, and C2 were located in semi-outdoor spaces, such as building stilts or outdoor pavilions; and A3, B3, and C3 were situated alongside pedestrian pathways.

Environmental data were recorded at 10 min intervals using high-precision instruments, including the Kestrel 5400 and TES-1333R. As detailed in

Table 12. The monitored parameters of experimental instrument.

Environmental Parameter	Instrument	Measurement Range	Resolution	Accuracy	Logging Mode
Air temperature	TES Hot-wire Anemometer	−10 to 60 °C	0.1%℃	±0.4%	Automated logging every 10 min
Relative humidity	TES Hot-wire Anemometer	10% to 95%	0.1%	±3%	Automated logging every 10 min
Wind speed	TES Hot-wire Anemometer	0 to 30 m/s	0.01	±3%	Automated logging every 10 min

, the monitored parameters including air temperature, relative humidity, wind speed, solar radiation, and surface temperature were recorded.

To establish hourly temperature data series, the 10 min measured data were converted into hourly averages for each point, which were subsequently averaged to the mean hourly temperature data for each block. Figure 17 shows that both measured and simulated temperature curves for all three blocks displayed a consistent unimodal diurnal pattern. The peaks consistently materialized at approximately 14:00, consistent with the characteristic summer microclimate of Wuhan. The simulation effectively captured the morphological trends of the observed data.

The simulation error during the peak heat period (12:00–16:00) was maintained within ±1.0 °C, meeting the accuracy standards for block-scale microclimate studies. Larger peak errors in Communities B and C were likely caused by the presence of complex tall trees, whose geometric variations are difficult to replicate by standardized vegetation models. Furthermore, the simulation underestimated temperatures by 0.2 to 0.5 °C during 08:00–10:00, potentially due to the simplified parameterized handling of anthropogenic heat in the ENVI-met model. Because the study focuses on the daytime outdoor thermal environment, this systematic bias exerts a limited effect on the study.

The hourly observed temperature series from all monitoring points were compared against the corresponding hourly simulated values extracted from the ENVI-met outputs. To comprehensively evaluate the simulation accuracy, three statistical metrics were used: mean absolute error (MAE), root-mean-square error (RMSE), and mean relative error (MRE).

Table 13. Simulation error metrics in three blocks.

Statistical Metrics	Block A	Block B	Block C	All
Sample size	21	21	21	63
MAE (°C)	0.71	0.37	0.66	0.58
RMSE (°C)	0.85	0.43	0.79	0.7
MRE (%)	2.12	1.12	2.01	1.75

presents the comparison between observed and simulated air temperatures, revealing a composite MAE of 0.58 °C, which validates the model’s accuracy. Block B achieved a minimal MAE of 0.37 °C, satisfying the accuracy standards. Conversely, the MAE for block A and C both surpassed 0.60 °C; this increased error is caused by simplified parameterized modeling and the exclusion of anthropogenic heat sources. The small discrepancy of 0.12 °C between the RMSE and MAE shows satisfactory model stability. Moreover, the temporal alignment of simulated temperature peaks and field data confirms that the ENVI-met model reliably captures spatial microclimate variations in Wuhan’s residential blocks. Ultimately, these validation analyses verify the validity of using ENVI-met for outdoor thermal environment analysis.

3.3.3. The Simulation Results Analysis

This study analyses the average mean air temperature ($T_{ave}$), the root mean square error of the mean air temperature ($T_{RMSE}$) and the temperature difference between the interior and exterior of the blocks ($T_{dif}$) for the typical blocks in each cluster. As illustrated in Figure 18, the findings are as follows:

• In terms of $T_{ave}$, the average value of Cluster 1 was significantly higher than that of the other clusters. The approximate mean values were 31.6 °C for Cluster 0, 33.4 °C for Cluster 1, 31.6 °C for Cluster 2, 32.0 °C for Cluster 3, and 31.8 °C for Cluster 4.
• Regarding $T_{RMSE}$, the average value of Cluster 1 was also significantly higher than that of the other clusters. The mean values were approximately 0.30 °C for Cluster 0, 1.15 °C for Cluster 1, 0.39 °C for Cluster 2, 0.30 °C for Cluster 3, and 0.35 °C for Cluster 4.
• Regarding $T_{dif}$, Cluster 2 exhibited the highest average value, while Cluster 1 was significantly lower than the others. The mean values were approximately 0.18 °C for Cluster 0, −0.12 °C for Cluster 1, 0.40 °C for Cluster 2, 0.28 °C for Cluster 3, and 0.28 °C for Cluster 4.

As illustrated in Figure 18 and Figure 19, $T_{ave}$ was compared against two thermal exposure indices (${PET}_{ave}$ and ${UTCI}_{ave}$). Across the 125 blocks, $T_{ave}$, ${PET}_{ave}$ and ${UTCI}_{ave}$ ranges were 31.14–35.39 °C (mean: 32.26 °C), 43.92–51.24 °C (mean: 46.62 °C), and 38.52–43.46 °C (mean: 40.32 °C), respectively. All three indicators peaked in Cluster 1, demonstrating the capacity of morphological clustering to differentiate overall thermal environments. Cluster 0 maintained high ${PET}_{ave}$ and ${UTCI}_{ave}$ levels and experienced moderate $T_{ave}$, underscoring the fact that air temperature is insufficient to capture actual human thermal exposure [68,69,70,71]. These findings align with those of the established literature [72] indicating that outdoor thermal comfort depends not only on air temperature, but is also significantly shaped by mean radiant temperature, shading, and ventilation.

3.4. The Regression Analysis

As illustrated in Figure 20, Pearson correlation analysis [73] was conducted to investigate the relationships between the morphology of blocks and thermal environment based on the nine key morphological indicators and three outdoor thermal environment indicators for the 125 typical blocks. The study found the following:

• In terms of $T_{ave}$, a significant positive correlation was observed with ${SVF}_{ave}$, with an absolute correlation coefficient of 0.94. A strong positive correlation was also found with ${\mathit{OAS}}_{\mathit{ave}}$, with an absolute coefficient of 0.61.
• In terms of $T_{dif}$, a strong negative correlation was identified with ${SVF}_{ave}$, with an absolute correlation coefficient of 0.62. Additionally, a positive correlation was observed with $VAR$, with an absolute coefficient reaching 0.53.
• In terms of $T_{RMSE}$, a significant positive correlation was exhibited with ${SVF}_{ave}$, with an absolute correlation coefficient of 0.93. A positive correlation was also noted with ${\mathit{OAS}}_{\mathit{ave}}$, with an absolute coefficient of 0.56.

Before the regression analysis, the multicollinearity among morphological indicators was evaluated. As illustrated in Figure 21, the Variance Inflation Factor (VIF) values for $BBA$ and ${FA}_{all}$ exceeded 5, indicating significant multicollinearity and necessitating the adoption of a ridge regression model.

This study established regression models based on morphological indicators and outdoor thermal environment indicators using ridge regression. To compare the importance of each morphological indicator, they were standardized. To ensure the generalizability of the ridge regression model, this study randomly selected 80% of the samples from 25 typical blocks in each of the five clusters, forming a training set of 100 blocks (20 blocks per cluster). The remaining 20% of typical blocks were taken as the validation set, with 5 blocks per cluster and a total of 25 blocks. The performance of the ridge regression model was evaluated according to the coefficient of determination R² on the validation set.

In terms of $T_{ave}$, as shown in Figure 22a, the ridge regression model has an R² value of 0.861 on the test set, indicating excellent performance. The model equation is as follows:

$$\begin{gathered} T_{ave}=32.1656 + 0.0992BN-0.0480BBA-0.2205{FA}_{all}+0.5470{SVF}_{ave}- 0.0522{\mathit{OA}}_{\mathit{ave}} \\ -0.0460{AR}_{ave} -0.0796VAR-0.0960VTC+0.1084{VD}_W \end{gathered}$$

(13)

Notably, the absolute coefficients for ${SVF}_{ave}$, ${FA}_{all}$, and ${VD}_W$ were significant, reaching 0.5470, 0.2205, and 0.1084, respectively. As shown in Figure 22b, a simplified ridge regression model was developed based on these three key indicators, achieving an R² value of 0.805 on the test set, which also demonstrates excellent performance. The equation is as follows:

$$T_{ave}=31.5642+0.7729{SVF}_{ave}-0.4412{FA}_{all}+0.3269{VD}_W$$

(14)

The equations and coefficients suggest the following:

• An increase in ${SVF}_{ave}$ shows a stronger capacity for solar radiation reception and weaker thermal buffering structures. This leads to significantly enhanced daytime heat absorption, where the advantage of night-time cooling is insufficient to offset the cumulative heating effect throughout the day.
• An increase in ${FA}_{all}$ represents a larger building facade area, suggesting a higher density of buildings and stronger shading effects. This reduces solar penetration into the blocks, ultimately resulting in lower outdoor temperatures.
• An increase in ${VD}_W$ indicates that green spaces are located further from the western boundary. This reduces effective afternoon shading (western sun exposure) and allows direct solar heating of ground and west-facing walls. Consequently, heat accumulation increases significantly, raising the outdoor temperature.

Regarding $T_{dif}$, as shown in Figure 23a, the R² value of the ridge regression model is 0.531, indicating moderate performance. The model equation is as follows:

$$\begin{gathered} T_{dif}=0.2019 -0.0267BN-0.0089BBA+0.0209{FA}_{all}-0.0704{SVF}_{ave}-0.0203{\mathit{OA}}_{\mathit{ave}}+0.0146{AR}_{ave} \\ +0.0414VAR-0.0149VTC-0.0141{VD}_W \end{gathered}$$

(15)

The absolute coefficients for $BN$, ${SVF}_{ave}$, and $VAR$ were significant, reaching 0.0267, 0.0704, and 0.0414, respectively. As illustrated in Figure 23b, a simplified ridge regression model was developed, based on these three key indicators, achieving an R² value of 0.507, which also demonstrates moderate performance. The equation is as follows:

$$T_{dif}=0.3101 + 0.2630VAR-0.1262BN-0.4473{SVF}_{ave}$$

(16)

The equations and coefficients suggest the following:

• As ${SVF}_{ave}$ increases, the blocks transitions from a “closed body” with strong thermal buffering capacity into a “permeable body” that freely exchanges heat with the external environment. This reduces the temperature difference with surrounding areas, leading to a decrease in $T_{dif}$.
• An increase in $VAR$ shows a larger green space area, which forms a local “cool island.” Through mechanisms such as shading, evaporate cooling, and suppression of surface thermal radiation, the green space significantly lowers temperatures, while non-green external areas remain hot. This widens the temperature difference between the interior and exterior, resulting in an increase in $T_{dif}$.
• An increase in BN indicates a rise in the number of buildings within the blocks and a larger block size. Since the simulation domain in this study is fixed at 600 m × 600 m, the larger the target blocks’ area is, the more its physical properties will converge with those of the background area, leading to a decrease in $T_{dif}$.

The study found that the R² value of the regression model based on morphological indicators is low, suggesting that the temperature difference between the interior and exterior of the blocks cannot be reliably predicted using morphological indicators. This is mainly because the urban morphology of the blocks and their surrounding environment is relatively similar, and no significant temperature difference exists between the inside and outside.

In terms of $T_{RMSE}$, as shown in Figure 24a, the R² value of the ridge regression model is 0.849, indicating excellent performance. The model equation is as follows:

$$\begin{gathered} T_{RMSE}=0.5216+0.0328BN+0.0000BA-0.0535{FA}_{all}+0.2208{SVF}_{ave}-0.0061{\mathit{OA}}_{\mathit{ave}}+0.0529{AR}_{ave} \\ -0.0156VAR-0.0458VTC+0.0293{VD}_W \end{gathered}$$

(17)

The absolute coefficients for ${SVF}_{ave}$, $BA$, and ${FA}_{all}$ were significant, reaching 0.267, 0.161, and 0.062, respectively. As illustrated in Figure 24b, a simplified ridge regression model was developed based on these three key indicators, achieving an R² value of 0.855, which also demonstrates excellent performance. The equation is as follows:

$$T_{RMSE}=0.5088 +0.6274{SVF}_{ave}+0.1375{AR}_{ave}-0.1137{FA}_{all}$$

(18)

The study found the following:

• A higher ${AR}_{ave}$ value indicates that buildings within the blocks tend to be slab-type buildings. compared to point-type buildings, slab-type buildings exert a stronger shading effect and create a larger north-facing shaded area. The expansion of both the south-facing sunlit area and the north-facing shaded area leads to an increase in $T_{RMSE}$.
• As ${SVF}_{ave}$ increased, shading in the blocks decreased, leading to increased direct input of solar radiation. Simultaneously, the unobstructed output of long-wave radiation is enhanced, resulting in increased temperature fluctuations in the outdoor space.
• An increase in ${FA}_{all}$ indicates a larger building surface area. A larger surface area enhances the efficiency of heat exchange with the atmosphere, causing faster daytime heat absorption and more rapid night-time cooling. This amplifies the amplitude of air temperature fluctuations.

As illustrated in Figure 25, 5-fold cross-validation was performed to evaluate the ridge regression models for the two thermal indices ($T_{ave}$ and $T_{RMSE}$). The models have mean $R^2$ values of 0.64 ($\pm 0.10$) for $T_{ave}$ and 0.39 ($\pm 0.07$) for $T_{RMSE}$, with corresponding mean RMSE values of 0.43 and 0.28, respectively. Models reconstructed using the filtered indicators exhibited comparable performance to the final models, achieving mean $R^2$ values of 0.67 for $T_{ave}$ and 0.38 for $T_{RMSE}$. Throughout the cross-validation process, the $R^2$ scores remained strictly positive, ranging within [0.48, 0.74] for $T_{ave}$ and [0.30, 0.48] for $T_{RMSE}$. The results demonstrate the robust predictive stability of the regression models.

As shown in Figure 26a, residual diagnostics for the $T_{ave}$ validation set showed a mean bias near zero (−0.05 °C) and an RMSE of 0.27 °C. No obvious heteroscedasticity was shown, with only a single standardized residual exceeding $\left|2\right|$. The Shapiro–Wilk test had a $p$-$\mathrm{value}$ of 0.60, confirming that the regression assumptions are statistically acceptable.

For $T_{RMSE}$ (Figure 26b), the residuals exhibited a tighter distribution, characterized by a mean bias of −0.04 and an RMSE of 0.14. No standardized residuals exceeded $\left|2\right|$, and the scale–location correlation remained weak (0.27). A Shapiro–Wilk test $p$-$\mathrm{value}$ of 0.31 indicates that the regression assumptions were well satisfied on the independent test set.

In summary, ${SVF}_{ave}$ and ${FA}_{all}$ are critical morphological indicators that influence the three thermal indicators. Additionally, ${VD}_W$ primarily affects $T_{ave}$, and ${AR}_{ave}$ is the key determinant for $T_{RMSE}$.

4. Discussion

4.1. Impact of Block Morphology on Outdoor Thermal Environment

Urban block morphology has a dual influence on the outdoor thermal environment by modulating both the absolute thermal magnitude and the fluctuation amplitude. This mechanism is primarily governed by the interplay between solar radiation interception and the thermal storage properties of the built environment.

4.1.1. Impact on Outdoor Mean Temperature ($T_{ave}$)

Blocks morphology dictates the outdoor temperature primarily through shading effects and solar accessibility:

• Shading vs. Exposure: a positive correlation exists between ${SVF}_{ave}$ and outdoor temperatures. High $SVF$ values signify reduced geometric obstruction, leading to increased short-wave radiation gain during diurnal hours. The study indicates that the resulting heat accumulation surpasses the efficiency of nocturnal long-wave radiative cooling, leading to a net thermal surplus. In contrast, an increase in the Total Facade Area (${FA}_{all}$) serves as a cooling driver; the expanded vertical surfaces provide significant mutual shading within street canyons, effectively mitigating solar penetration and lowering the sensible heat flux.
• Vegetation Spatial Configuration: the spatial configuration of cooling buffers is equally critical. The increase in Vegetation Distance from the West (${VD}_W$) exacerbates thermal stress. The absence of vegetative shading during peak western solar exposure allows for direct radiative forcing on building envelopes and ground surfaces, triggering rapid localized heating.

4.1.2. Impact on Outdoor Temperature Dynamics ($T_{RMSE}$)

The morphology also determines the temperature fluctuation dynamics of the urban blocks by affecting heat exchange efficiency and storage capacity:

• Shading Heterogeneity: higher ${AR}_{ave}$ values indicate a shift from point-type buildings toward plate-style buildings. Compared to point-type buildings, plate buildings create more extensive shading, particularly in north-facing areas—while maintaining significant south-facing sunlit zones. This expansion of both high-exposure and high-shade areas increases the $T_{RMSE}$ value.
• Exchange Efficiency: in contrast, high $SVF$ and $FA$ serve as “thermal amplifiers.” A high $SVF$ allows for direct and rapid radiate exchange with the sky, while a large $FA$ increases the contact interface for convective heat exchange with the atmosphere. Both configurations accelerate the heating and cooling cycles, leading to more volatile temperature swings throughout the day [74,75].

4.2. The POI Analysis

POI data [57] in Wuhan were used from the Baidu Maps, encompassing six categories: food services, shopping services, companies, science and education services, public facilities, and landscape. Using a parametric method, the study calculated the number of POIs within a 100 m radius of each block and computed the average POI counts for each cluster, based on their labels. As illustrated in Figure 27, the findings are as follows:

• Food services: blocks in Cluster 1 and Cluster 3 are surrounded by a high density of dining facilities, with average counts of 10.75 and 11.93, respectively. In contrast, blocks in Cluster 2 and Cluster 4 have relatively scarce dining amenities, with averages of only 2.01 and 2.17.
• Shopping services: blocks in Clusters 0, 1, and 3 are surrounded by a high density of shopping facilities, with average values of 13.94, 18.46, and 20.47, respectively. Conversely, blocks in Clusters 2 and 4 have limited access to shopping services, with averages of 3.4 and 3.9.
• Companies: many companies concentrate around blocks in Clusters 1 and 3, with averages of 3.71 and 3.39. In comparison, blocks in Clusters 2 and 4 host fewer companies, with averages of 1.1 and 0.71.
• Science and education: blocks in Clusters 0 and 4 are surrounded with a higher number of educational and research institutions, with averages of 2.32 and 3.07, respectively. On the other hand, Cluster 3 has fewer such institutions nearby, with an average of only 1.21.
• Public facilities: blocks in Clusters 1 and 3 possess a higher number of public facilities, with averages of 0.38 and 0.39. In contrast, Cluster 2 is characterized by a scarcity of public facilities, with an average of only 0.039.
• Landscape: blocks in Clusters 0 and 2 contain a high number of landscapes, with average values of 0.766 and 0.838, respectively.

In summary, the study found distinct functional characteristics among the different clusters. Blocks in Cluster 0 are characterized by a higher presence of science and education institutions and landscape. Blocks in Cluster 1 exhibit a high concentration of food, shopping, company, and public facilities, but suffer from a scarcity of landscape. Blocks in Cluster 2, while having a considerable amount of landscape space, lack POI points in food, shopping, company, educational, and public aspects. Cluster 3 displays a profile like Cluster 1, surrounding a dense mix of food, shopping, company, and public facilities, yet similarly lacking landscape. Finally, blocks in Cluster 4 are associated with a significant number of research institutions, but show a deficiency in food, shopping, company, and public facilities.

4.3. The Spatial Distribution Analysis

As illustrated in Figure 28, this study visualizes the distribution proportions of different clusters across the seven major administrative districts. The findings are as follows:

• Cluster 0: blocks in this cluster are mainly located in Hongshan District, accounting for 58.22% of the total. Furthermore, 69.62% of these blocks belong to “Wuchang town” (comprising Wuchang, Hongshan, and Qingshan districts), indicating that Cluster 0 blocks are mainly situated east of the Yangtze River.
• Cluster 1: blocks in this cluster are mainly set in Jianghan, Jiang’an, and Qiaokou districts, with proportions of 24.71%, 18.53%, and 18.54%, respectively. Additionally, “Hankou town” (Jianghan, Jiang’an, and Qiaokou) accounts for 61.80% of these blocks, suggesting that Cluster 1 is mainly distributed within the Hankou area.
• Cluster 2: blocks in this cluster are mainly found in Hongshan and Hanyang districts, representing 37.98% and 21.07%, respectively. The analysis reveals that only 20.93% of these blocks are in “Hankou town” (Qiaokou, Jiang’an, and Jianghan), indicating that Cluster 2 blocks are mainly distributed to the east of the Yangtze River.
• Cluster 3: blocks in this cluster are mainly distributed in Jiang’an and Jianghan districts, accounting for 24.22% and 20.41%, respectively. The data shows that “Wuchang town” (Wuchang, Hongshan, and Qingshan) comprises only 28.03% of these blocks, meaning that Cluster 3 is mainly located west of the Yangtze River.
• Cluster 4: blocks in this cluster are mainly situated in Hongshan District, with a proportion of 55.97%. Moreover, 64.93% of these blocks belong to “Wuchang town” (Wuchang, Hongshan, and Qingshan districts), indicating that Cluster 4 blocks are mainly located east of the Yangtze River.

In summary, the study found that the distribution of blocks across different clusters exhibits distinct spatial patterns. Blocks in Clusters 0 and 4 are primarily located east of the Yangtze River. Blocks in Cluster 1 are mainly concentrated in the Hankou area. Cluster 2 blocks are primarily distributed in the regions outside of Hankou, while those in Cluster 3 are primarily situated west of the Yangtze River.

This study used a parametric approach to calculate the distances between blocks in each cluster. Blocks with a spacing of less than 200 m were grouped into clusters. As illustrated in Figure 29, the groups exhibit distinct spatial distribution patterns, detailed as follows:

• Cluster 0 (Red): these groups are mainly located in Hongshan District, with a concentration in the Nanhu and Optics Valley (Guanggu) areas. These areas consist largely of residential compounds developed after 2000, characterized by uniform slab-type buildings, which contribute to the higher ${AR}_{ave}$ values of these blocks.
• Cluster 1 (Yellow): these groups are mainly distributed across the riverside areas of Qiaokou, Jianghan, and Jiang’an districts. Due to the north–south direction of the Yangtze River, buildings in these riverside blocks generally face the river (oriented east–west), resulting in higher ${\mathit{OAS}}_{\mathit{ave}}$ values. Additionally, these areas contain numerous older blocks with generally lower building heights and a lack of tall trees, leading to higher ${SVF}_{ave}$ values.
• Cluster 2 (Green): these groups are mainly found in Wuchang and Hongshan districts, concentrated around inland lakes such as Huangjia Lake and Nanhu Lake. Due to the abundant vegetation surrounding these lakes, the blocks in these areas exhibit higher $VAR$ values.
• Cluster 3 (Blue): these groups are similarly distributed in the riverside areas of Qiaokou, Jianghan, and Jiang’an districts. Influenced by the river’s orientation, the buildings are mainly east–west facing, leading to higher ${\mathit{OAS}}_{\mathit{ave}}$ values. Conversely, due to the high commercial value of these riverside zones, some older blocks have been redeveloped into dense high-rise residential areas, resulting in lower ${SVF}_{ave}$ values.
• Cluster 4 (Purple): these groups are mainly distributed in Hongshan and Qiaokou districts, specifically in the Shuangdun, Baisha 1st Road, and Optics Valley areas. Representing a new cycle of urban development in Wuhan, these blocks are characterized by their large scale and high building densities, resulting in higher $\mathit{BN}$ and $\mathit{BBA}$ values.

5. Conclusions

Based on multi-source data, this study constructed 3D models of 915 residential blocks in Wuhan, and extracted their morphological indicators. Using these indicators, the blocks were classified, and typical blocks from each cluster were selected for thermal environment simulations for outdoor thermal indicators. Subsequently, regression models were developed using the morphological and thermal indicators of these typical blocks. Finally, the spatial distribution and surrounding POI of each cluster were analyzed. The main findings are as follows:

• The study compared PCA, UMAP, and t-SNE algorithms for screening morphological indicators. Results indicate that t-SNE outperforms PCA and UMAP in terms of Trustworthiness and KNN indicators. As nonlinear dimensionality reduction methods, t-SNE and UMAP effectively eliminate “confusing indicators” that interfere with local neighborhood judgment, clarifying the manifold structure and facilitating the capture of true local data relationships, thereby significantly improving Trustworthiness and KNN values.
• Nine key indicators were screened from 29 indicators. Among these, ${SVF}_{ave}$, ${FA}_{all}$, ${VD}_W$, $VAR$, and $\mathit{BBA}$ play a critical role, not only in the classification of block morphology, but also in the assessment of the outdoor thermal environment.
• The study found that $T_{ave}$ of Cluster 1 blocks was significantly higher than that of other clusters. $T_{dif}$ of Cluster 2 was significantly higher than that of other clusters, whereas Cluster 1 exhibited a significantly lower $T_{dif}$. Additionally, $T_{RMSE}$ of Cluster 1 was notably higher than that of other clusters.
• The study demonstrates that $T_{ave}$ can be accurately predicted using ${SVF}_{ave}$, ${FA}_{all}$, and ${VD}_W$, with an R² of 0.89. $T_{RMSE}$ can be well predicted using ${SVF}_{ave}$, $\mathit{BBA}$, and ${FA}_{all}$, with an R² of 0.879.
• Cluster 0 blocks are mainly located on the east side of the Yangtze River, surrounded by educational institutions and green spaces. Cluster 1 blocks are mainly distributed in Hankou town and surrounded by food, commercial, and corporate POI. Due to the history of Hankou, these blocks (Cluster 1) typically consist of older, low-rise buildings with high ${SVF}_{ave}$ values, resulting in higher outdoor average temperatures. Cluster 2 blocks are mainly found in Hanyang and Wuchang town, and have a large number of green spaces, which leads to larger temperature differences.

6. Limitation and Future Research

This study generated detailed 3D models for 915 blocks in Wuhan based on multi-source data and classified blocks morphology using morphological indicators. Furthermore, typical blocks were selected to simulate the outdoor thermal environment based on their 3D models. Regression models were then developed to identify key morphological indicators influencing the outdoor thermal environment. The rationality of the morphological classification and the thermal performance differences among block types were validated through POI and spatial distribution analyses. Despite these contributions, certain limitations remain, which will be addressed in future research, as follows:

• More detailed 3D models: this study extracted vegetation information from satellite semantic segmentation images but did not incorporate other elements, such as water bodies, into the 3D models. Previous studies indicate that urban water bodies significantly affect the surrounding thermal environment. Future research will extract diverse urban element information based on semantic segmentation and incorporate it into 3D models for further analysis.
• More detailed vegetation information: while vegetation was extracted, it was not categorized; instead, all vegetation was uniformly modeled as trees, lacking fine-grained classification. This limitation constrains the precision of the thermal environment simulations. Future studies will use refined semantic segmentation datasets to extract specific vegetation types and construct more detailed 3D models.
• More accurate thermal environment simulations: this study focused on the outdoor thermal environment using ENVI-met software. To enhance computational efficiency and ensure consistent boundary conditions, specific wind directions and initial wind speeds were not defined for each block. However, wind environments significantly influence thermal conditions. Future research will calculate wind environment using CFD, using the calculated average wind speed and dominant wind direction as initial parameters for ENVI-met simulations.
• More comprehensive thermal environment simulations: this study was limited to the summer thermal environment of Wuhan, neglecting winter conditions. As a city in a hot-summers-and-cold-winters region, the winter thermal environment of Wuhan should be considered. Future research will simulate both summer and winter thermal environments for target blocks and develop predictive models for both seasons, based on morphological and thermal indicators.
• More comprehensive research methods for the urban thermal environment: this study focused primarily on the influence of block morphology on outdoor temperature, and did not address human thermal comfort indicators such as PET or UTCI. Future research will integrate these thermal comfort indicators with outdoor temperature metrics to further analyze the impact of block morphological indicators on both outdoor temperature and resident comfort.