Method for Evaluating Urban Building Renewal Potential Based on Multimachine Learning Integration: A Case Study of Longgang and Longhua Districts in Shenzhen

Abstract

With the continuous advancement of urbanization, urban renewal has become a vital means of enhancing urban functionality and improving living environments. Traditional urban renewal research primarily focuses on the macro level, analyzing regions or units, with limited studies targeting individual buildings. Consequently, the unique characteristics and specific requirements of individual buildings during urban renewal have often been overlooked. This study first identified individual buildings undergoing urban renewal in the Longgang and Longhua Districts of Shenzhen, China, from 2018 to 2023 using multisource data such as the 2018 Shenzhen Building Census. A regression analysis based on building characteristics and locational factors was conducted using a stacking ensemble machine learning model. In addition, buildings were categorized into residential, industrial, and commercial types based on their usage, enabling both overall- and category-specific predictions of building renewal. The results show the following: (1) Using the prediction results of multilayer perceptron (MLP) and eXtreme Gradient Boosting (XGBoost) base models as inputs and fusing them with an AdaBoost classifier as the final metamodel, the goodness of fit of the overall building renewal regression model increased by 2.19%. (2) The regression model achieved an overall urban renewal prediction accuracy of 89.41%. Categorizing urban renewal projects improved the goodness of fit for residential and industrial building renewal by 0.14% and 6.13%, respectively. (3) Compared with traditional macro-level evaluation methods, the experimental results of this study improved by 8.41%, and compared with single-model approaches based on planning permit data, the accuracy improved by 29.11%.

1. Introduction

With the advancement of urbanization, many urban centers have experienced issues such as aging buildings and deteriorating infrastructure, necessitating renewals to meet modern demands [1]. China’s urbanization entered its later stages after going through a phase of rapid development with a low starting point. By the end of 2021, the urbanization rate of China’s permanent population reached 64.72%, with the urban population increasing from 170 million in 1978 to 910 million [2]. As China embarks on a new journey to build a modern socialist country, urbanization faces pressing demands for quality improvement and upgrades, necessitating new strategic decisions and work arrangements. The 14th Five-Year Plan and 2035 Vision outline explicitly propose the implementation of urban renewal actions to promote sustainable urban development and meet the growing living demands of the people. Given the rapid pace of urbanization, cities inevitably face increasing competition between growing populations and limited resources [3,4,5].

With the global urbanization rate exceeding 50%, urban renewal is replacing urban expansion as a core issue faced by cities around the world [6]. Urban renewal is a sustainable urban transformation approach that can improve resource utilization, alleviate urban functional decline, and optimize the urban spatial structure [7]. In this context, urban renewal not only strengthens and improves urban functions, preventing and eliminating the aging or decline of urban functions, but also promotes the coordinated development of the economy, society, and environment [8]. Therefore, the study of urban renewal has become a hot topic in current research.

In 1958, the concept of “urban renewal” was first introduced at the Netherlands Urban Renewal Symposium [9]. Jacobs [10] and Mumford [11] criticized the large-scale demolition and reconstruction planning model of time from a humanistic perspective. In 1999, Richard George Rogers, at the request of the UK government, wrote and published the white paper Towards Urban Renaissance, elevating urban renewal to the level of a “renaissance”. Over time, the concept of urban renewal abroad has gradually evolved into a comprehensive urban regeneration strategy, emphasizing the sustainable development of society, economy, and environment [12]. Terms related to urban renewal, such as “urban renewal”, “reconstruction”, and “regeneration”, have also emerged. From the post-World War II period to the 1970s, urban renewal primarily focused on the transformation, demolition, and reconstruction of the physical structure of cities [13]. In the 1990s, during the tenure of the UK’s New Labour government, the concept of “urban renaissance” was introduced, but it was not given a clear substantive meaning. Instead, the term was used more as a metaphor, emphasizing the sustainable development goals in the social, cultural, economic, environmental, and political dimensions [14]. In the 1980s, China began research on urban renewal theory. The concept of “urban renewal” was initially proposed by Chen [15], who also suggested three approaches to renewal: reconstruction, preservation, and maintenance. Subsequently, Wu [16] proposed the concept of “organic renewal”, emphasizing that urban renewal should focus on sustainable development and oppose blind large-scale demolition and reconstruction. This concept promoted the development of urban renewal theory in China, gradually forming a comprehensive theoretical system.

In the field of urban renewal potential assessment, scholars, both domestically and internationally, have engaged in various discussions. Cao et al. [17] proposed a method based on a logistic model to assess the potential of urban renewal projects using planning permit data while predicting the changes in functions before and after the renewal of the units. Li et al. [18] used spatiotemporal real estate data to explore the spatiotemporal patterns and evolution laws of residential floor area ratios in Guangzhou at multiple scales, including the neighborhood, street, and district levels. Wen et al. [19] reviewed the logical evolution of urban renewal policies since 1949 in a systematic analysis of the era background, policy direction, and logical context of “people-centered” urban renewal actions from the perspective of holistic governance. They propose a system governance framework for urban renewal actions from a holistic governance perspective. Wang et al. [20] approached urban renewal potential assessment from the perspective of big data-based quantitative evaluation. They used real estate registration big data to build an evaluation system for the urban renewal potential and performed a quantitative assessment at the street (village, town) scale, enabling spatial analysis at a refined level.

In terms of the scale of urban renewal research, Zhang et al. [21] proposed a multiscale model to support sustainable urban renewal decisions at the city, regional, and community levels, offering comprehensive and sustainable policies and actions for urban renewal from a multiscale perspective. Huang et al. [22], at the community level, evaluated the effects of urban renewal before and after its implementation, using eight communities in Chongqing as the research subjects. Pérez et al. [23] developed a spatial decision support system (SDSS) for community-scale urban renewal projects: URBIUS, which provides adaptable evaluation thresholds for each community through dynamic methods while considering the community’s actual conditions and its anticipated long-term development. Rall et al. [24] conducted a sustainability evaluation of land parcels in three dimensions—economic, social, and environmental—providing support for planning, formulation, and modification.

In terms of research methods for urban renewal potential, Fan [25] proposed a geographic information system-based land allocation method for urban renewal to provide a scientific basis for urban planning. Wang et al. [26] adopted the TOPSIS method, which approximates an ideal solution, to assess the potential of redevelopment sites, thereby providing theoretical support for selecting urban renewal projects. Liu et al. [27], using the analytic hierarchy process (AHP) and the fuzzy comprehensive evaluation method, conducted a study on the urban renewal potential of Shanghai using an analytic hierarchy process and a fuzzy comprehensive evaluation method. This study found that transportation accessibility and infrastructure completeness are significant factors influencing the renewal potential. Yildiz et al. [28] conducted a survey to evaluate the contribution of built environment design elements to environmental sustainability and performed an Analytic Hierarchy Process (AHP) analysis. This study found that “natural resource conservation” is the most critical factor influencing the environmental sustainability of urban renewal projects. Omidipoor et al. [29] proposed an integrated system framework based on WebGIS and MCDA to support owners, investors, and urban managers participating in the urban renewal processes. Through a case study of a residential area in Matera, Italy, Manganelli et al. [30] developed a policy model to support urban renewal and sought optimal solutions to balance the varying impacts of different stakeholders. Amorocho et al. [31] employed the TOPSIS method to compare different implementation plans for a five-story building in Spain, thereby providing a comprehensive decision-making approach to the proposed renovation plans. Lee et al. [32] proposed an AHP decision-making model for sustainable renewal solutions from social, economic, and ecological sustainability perspectives and conducted an empirical study to evaluate the feasibility of the model.

Current urban renewal research still has some methodological limitations. First, many studies rely on specific models such as AHP [33], GIS [34], and TOPSIS [35], which, although providing theoretical support, may perform poorly in practical applications owing to data limitations and model assumptions. Second, most studies rely on macro-level and historical data for analysis, which do not fully reflect real-time changes and local details.

The goal of this study is to establish a scientific evaluation system to determine the location of building renewal accurately, avoid large-scale demolition and reconstruction, and optimize resource allocation in a scientific and rational manner to meet the future development needs of the city. This method takes a micro-level perspective on buildings, comprehensively considering factors such as the structural condition of the buildings, the surrounding environment, residential supporting facilities, and social benefits. They classified buildings based on their functional use. Through data analysis and model training, such as machine learning models and GIS spatial analysis, multidimensional data integration and analysis have been conducted to provide strong support for decision-making.

2. Materials and Methods

2.1. Study Area

This study focused on Shenzhen, the smallest megacity in China, with a total area of approximately 1997.27 km², located in the southern part of Guangdong Province, with its geographic coordinates ranging from 113°43′ to 114°38′ east longitude and from 22°24′ to 22°52′ north latitude. Shenzhen has a topography that is higher in the southeast and lower in the northwest, with the highest elevation of 940.7 m. The locations of the study areas are shown in Figure 1. As of 2022, Shenzhen’s permanent population is 17.66 million, with an urban built-up area of 962.17 km², accounting for 48.17% of the total administrative area. This has resulted in limited land for new construction, making urban renewal a crucial means of addressing the conflict between population growth and land resources in Shenzhen.

Shenzhen was the first city in China to formulate official urban renewal policies and urban planning and tends to promote urban renewal by attracting capital and following market principles [36]. Therefore, Shenzhen is ideal for international studies on urban renewal. Moreover, Shenzhen has accumulated extensive project experience in urban renewal [37], providing a rich empirical foundation for analyzing the driving mechanisms behind different types of urban renewal.

As of the end of 2023, 1037 urban renewal units were included in the city’s renewal plan, covering a demolition area of 88.4 km². Among these, 726 units were approved by the Urban Planning Committee, with a demolition area of 58.1 km². From 2018 to 2023, Shenzhen approved 392 urban renewal projects, with 144 projects in the Longhua and Longgang Districts accounting for 36.74% of the total. These districts stand out for their significant urban renewal projects and have accumulated extensive practical experience in urban renewal. Furthermore, based on the proximity of districts, this study focuses on the Longgang and Longhua Districts in Shenzhen as specific research areas.

2.2. Research Data

2.2.1. Data Collection

The data used in this study included vector data, open data from the Internet, and imagery data. The vector data primarily included the administrative divisions of Shenzhen, the city’s road network system (primary and secondary roads), building survey data, census data, and slope information, Gross Domestic Product, Real Estate Price Data. Open data from the Internet include information on the locations of urban renewal units, as well as data on education, parks, hospitals, schools, and subways in Shenzhen. The imagery data included Landsat 8 imagery, historical remote sensing images, and DEM30m elevation data. A summary of the data and sources is presented in

Table 1. Data Summary of the Study Area.

Type	Name	Source
Vector Data	Shenzhen Administrative Division, Shenzhen Road Network System (Primary and Secondary Roads), Building Census Data, Population Census Data, Slope Data, Gross Domestic Product, Building Price	Shenzhen Planning and Natural Resources Bureau (https://pnr.sz.gov.cn) (accessed on 10 May 2024)
Open Internet Data	Urban renewal unit location information Shenzhen’s education, parks, hospitals, schools, and metro location information	Longgang District Urban Renewal and Land Preparation Bureau (https://www.lg.gov.cn) (accessed on 14 May 2024) Longhua District Urban Renewal and Land Preparation Bureau (https://www.szlhq.gov.cn) (accessed on 19 May 2024) Gaode Map API (https://lbs.amap.com)(accessed on 25 May 2024)
Image Data	Landsat 8 imagery data Historical remote sensing imagery DEM30m elevation data	Geospatial Data Cloud (https://www.gscloud.cn) (accessed on 2 June 2024)

.

2.2.2. Selection of Evaluation Factors

The implementation of urban renewal projects is closely related to factors such as building conditions, regional location, environmental design, and population density. The selection of appropriate evaluation factors is a key step in assessing the potential of urban renewal projects. Lee et al. [38] conducted a study on the factors influencing urban renewal in Hong Kong. The results indicate that factors such as the availability of public service facilities, fulfillment of welfare requirements, accessibility of facilities, urban landscape visual images, and local characteristics are important potential factors in enhancing the social sustainability of local urban renewal projects. Based on previous research and considering the basic conditions of the study area, 21 evaluation factors were selected: Building Data: land area, building perimeter, structural type, building height, above-ground floors, underground floors, volume, and building function; Location Factors: distance to main roads, distance to secondary roads, distance to schools, distance to hospitals, distance to subway, distance to parks, and distance to commercial services; Natural Factors: slope, vegetation index, and altitude; Social Factors: population; and Economic factors: Gross Domestic Product (GDP) and Building Price. The data are shown in Figure 2.

• Building census data

The data used in this study were obtained from the 2018 building census for Shenzhen. The distribution of buildings is shown in Figure 3. Building census data provide a comprehensive survey and statistics on the number, usage, area, height, structure, performance, and other information about buildings in the city. They have been widely applied in urban planning, land use, architectural design, and environmental governance. The building census data are presented in

Table 2. Introduction to Building Census Data.

Indicator Name	Overall Buildings in the Study Area	Updated Buildings in the Study Area
Building Quantity	234,225 buildings	22,733 buildings
Building Perimeter	5.65–3327.55 m	6.0–1178.8 m
Land Area	1.9–193,474.9 m2	2.25–22,385.5 m2
Structure Type	Steel structure, mixed structure, frame structure, tube structure, brick–concrete structure, brick–tile structure.	Steel structure, mixed structure, frame structure, tubular structure, brick–tile structure.
Height	0~200.15 m	0~27 m, accounting for 99.46%.
Number of Above-ground Floors	0~56 floors	0~9 floors; percentage: 99.13%.
Number of Underground Floors	0~4 floors	0~2 floors
Volume	2.7 m2~651,429.7 m2	3.19~205,867.34 m2
Building Function	Office, warehousing, warehousing and logistics, industrial, public facilities, public amenities, shopping centers, transportation, residential, residential amenities, commercial services, commercial streets, mixed-use, municipal facilities, private residences, special, integrated, others.	Warehousing, industrial, residential, residential amenities, commercial services, private residences; percentage: 99.27%.

. Building census information can provide important data and support, particularly for urban renewal and reconstruction. Building census data are difficult to obtain because of the large amount of non-public information included. As a result, few studies have used building census data. Some scholars [39,40] used planning permission data to study the factors influencing building renewal.

2.

Main and secondary road distances

Roads are the backbone of the urban spatial structure, and their layout and accessibility influence the distribution of urban functions and land-use patterns. By analyzing road data, we can understand the transportation convenience, connectivity, and traffic pressure in a region. This study, from the micro-perspective of buildings, adopts the perpendicular distance from the center point of the building to the centerline of the road, as shown in

Table 3. Distance from Updated Buildings to Main Roads.

Indicator Name	Overall Building Straight-Line Distance in the Study Area	Updated Building Straight-Line Distance in the Study Area
Main Road	0.12–4327 m	0.12–600 m, Percentage: 96.1%
Secondary Road	0–3686 m	0–500 m, Percentage: 95.96%

, which better reflects the characteristics of micro-level research on individual buildings.

3.

Regional Support Data

In this study, the indicators selected for supporting facilities included parks, schools, hospitals, commercial services, and metro stations. The point of interest (POI) attributes and geographic coordinates of these facilities were obtained using the Gaode API. The location information is then converted into coordinate data, vectorized, and processed into spatial data. The straight-line distance from the central point of the building to the nearest central point of the support facility was calculated. The supporting data are summarized in

Table 4. Distance from Updated Buildings to Regional Supporting Facilities.

Indicator Name	Overall Building Straight-Line Distance in the Study Area	Updated Building Straight-Line Distance in the Study Area
Park	1.45–4556.9 m	1.15–1000 m, Percentage: 96.27%
Hospital	29–4058 m	47–1300 m, Percentage: 96.98%
School	5.7–6629 m	6.29–2000 m, Percentage: 86.15%
Commercial Facilities	4.3–7868.6 m	28–3000 m, Percentage: 95.97%
Subway	39.7–6233.3 m	28–2100 m, Percentage: 93.97%

.

4.

Raster Data

Population, slope, vegetation, elevation, Gross Domestic Product (GDP), and Building Price data were rasterized into 30 m × 30 m grid cells to ensure consistent spatial resolution. Using the building contour center points within the study area as extraction references, key numerical information from the corresponding raster cells was accurately extracted using the point sampling tool in ArcGIS v.10.6. A summary of the raster data is presented in

Table 5. Raster data of the location of updated buildings.

Indicator Name	Overall Building Location in the Study Area	Updated Building Location
Population	29–107,652 people	20,000–80,000, Percentage: 88.5%
Slope	Level 1–7	Level 1–3, Percentage: 99.47%
Vegetation (NDVI)	0–0.38	0–0.25, Percentage: 96.30%
Elevation	12–314 m	20–80 m, Percentage: 93.4%
Gross Domestic Product	0.32–4744.49 million yuan	1.5–4115 million yuan
Building Price	8351–148,740 yuan	13,912–113,615 yuan

.

2.3. Research Method

Shenzhen’s economy has undergone rapid development over the past 40 years, leading to increasingly severe issues of building aging, with many buildings facing safety risks and deterioration of the urban environment. Therefore, organic renewal is urgently required. However, existing research primarily focuses on macro-level regional or functional area-wide renovations and lacks micro-level analysis and precise assessment of individual building renewal needs. Taking the Longgang and Longhua Districts in Shenzhen as examples, this study constructs a multidimensional urban renewal evaluation system that comprehensively considers the structural conditions of buildings, location information, natural environments, and social elements. The analysis focused on individual buildings to avoid large-scale demolition and reconstruction, thereby optimizing resource allocation. The specific technical approach is presented in Figure 4.

This study utilized RF, SVM, MLP, and XGBoost models due to their complementary strengths in handling complex data structures and diverse prediction tasks. RF was chosen for its ability to manage high-dimensional data with noise. SVM is effective for smaller datasets and capturing non-linear relationships. MLP enables the modeling of intricate interactions among variables, while XGBoost provides efficient computation and strong performance in handling missing data. These models collectively address the multifaceted nature of urban renewal factors, ensuring robust and accurate predictions. Other candidate models, such as KNN and deep neural networks, were not used due to limitations in scalability or their tendency to overfit on small to medium-sized datasets. The selected models strike a balance between accuracy, interpretability, and computational efficiency, making them suitable for the objectives of this study.

The integration of these models leverages their respective strengths, ensuring robust predictions. Additionally, ArcGIS spatial analysis tools were incorporated to precisely identify building update needs, particularly focusing on classifying and predicting industrial, commercial, and residential buildings to determine update priorities. The objectivity and effectiveness of the model were validated through a comparative analysis with historical update data, providing a scientific basis for improving urban renewal efficiency. The final urban renewal model prediction formula ${\hat{y}}_{final}$ can be expressed as follows:

$${\hat{y}}_{final}=g\left(\left\{{\hat{y}}_1\left(x\right),\right.{\hat{y}}_2\left(x\right),..{\hat{y}}_M\left(x\right)\right)$$

(1)

where ${\hat{y}}_1$, ${\hat{y}}_2$,…, ${\hat{y}}_M$ are the prediction results of the base models, and $g$ is the prediction function of the metamodel (AdaBoost classifier). By learning from the outputs of different base models, a weighted integration and adaptive weight adjustment mechanism were used to obtain the final prediction result.

2.3.1. Data-Driven Models

(1) Random Forest

A random forest consists of multiple decision trees, each of which makes a prediction based on the input sample. The final output is the average or majority vote of all tree predictions. The output formula for the random forest is calculated as follows:

$${\hat{y}}_i={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\hat{y}}_l$$

(2)

where $N$ is the number of decision trees, ${\hat{y}}_l$ is the prediction output of the $l$th tree, and ${\hat{y}}_i$ is the final prediction result for the $i$th building. The key steps in implementing the random forest algorithm are as follows:

Data sampling: bootstrapping (sampling with replacement) is performed on the training set to construct multiple decision trees.

Feature selection: at a node of every tree, a subset of features is randomly selected for splitting, reducing overfitting.

Prediction fusion: the final prediction is obtained by majority voting or by averaging the results from all trees.

(2) Extreme gradient boosting

XGBoost is an implementation of the boosting tree model that iteratively optimizes the model by a weighted accumulation of prediction errors of the preceding trees. This model is defined as follows:

$${\hat{y}}_i=\sum _{k=1}^K{f}_k\left(x_i\right)$$

(3)

where ${\hat{y}}_i$ is the predicted value for updating the $i$th building; $f_k$ is the prediction of the $k$th tree for building $x_i$, and $K$ is the total number of trees in the model. As more trees are added, the prediction accuracy gradually improves. The model reduces prediction error by optimizing the objective function $L\left(\varphi \right)$:

$$L\left(\varphi \right)=\sum _{i=1}^{n}l\left(y_i,{\hat{y}}_i\right)+\sum _{k=1}^K\mathsf{\Omega }\left(f_k\right)$$

(4)

where $l\left(y_i,{\hat{y}}_i\right)$ and $\mathsf{\Omega }\left(f_k\right)$ are the loss function and regularization term, respectively. The loss function measures the difference between the actual update demand $y_i$ and the predicted value ${\hat{y}}_i$. The regularization term controls the model complexity to prevent overfitting in the urban renewal prediction. The formula for the regularization term is as follows:

$$\mathsf{\Omega }\left(f_k\right)={\rm Y}T+{\displaystyle \frac{1}{2}}\gamma {\sum }_{j=1}^T{w}_j^2$$

(5)

where ${\rm Y}$ and $\gamma$ are regularization parameters, $T$ is the number of leaf nodes, and $w_j$ is the weight of the leaf.

(3) Multilayer Perceptron

The formula for a multilayer perceptron (MLP): A multilayer perceptron is a feed-forward neural network composed of an input layer, hidden layers, and output layer. The output of each neuron in a layer undergoes a non-linear transformation through the activation function $\sigma$, and the model’s output is as follows:

$${\hat{y}}_i=\sigma \left(W^{\left(L\right)}·\sigma \left(W^{\left(L-1\right)}·\dots \sigma \left(W^{\left(1\right)}·x_i+b^{\left(1\right)}\right)+b^{\left(L-1\right)}\right)+b^{\left(L\right)}\right)$$

(6)

where $x_i$ is the feature vector of the $i$th building; $W^{\left(L\right)}$ and $b^{\left(L\right)}$ are the weight matrix and bias vector of the $l$th layer, respectively; $\sigma$ is the activation function; and ${\hat{y}}_i$ is the predicted update value for the $i$th building.

(4) Support Vector Machine

$${\hat{y}}_i=\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\left(w^T{x}_i+\mathrm{b}\right)$$

(7)

where $x_i$ is the feature vector of the $i$th building, $w$ is the weight vector, $b$ is the bias term, and $\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}$ is the sign function.

The core goal of an SVM is to find an optimal hyperplane that maximizes the margin between classes. The specific optimization objective is to minimize the square norm of weight vector ${‖w‖}^2$ while satisfying the classification constraints for all sample points. The optimization objective (margin maximization) formula is as follows:

$$\underset{w,b}{\mathit{min}}{\displaystyle \frac{1}{2}}{‖w‖}^2$$

(8)

For each sample $\left(x_{i,}{y}_i\right)$, the following classification constraint is satisfied:

$$y_i\left(w^T{x}_i+b\right)\ge 1$$

(9)

where $y_i$ is the true label of the $i$th building, $x_i$ is the feature vector of the $i$th building, and $w^T{x}_i+b$ is the decision function for the $i$th sample.

2.3.2. Spatial Kernel Density Analysis

Kernel density analysis (KDA) is widely used to explore spatial clustering and diffusion trends [41]. This can reveal the hotspots of urban renewal and future development potential. Its calculation formula is as follows:

$$\hat{f}\left(x\right)={\displaystyle \frac{1}{nh}}\sum _{i=1}^{n}K\left({\displaystyle \frac{x-x_i}{h}}\right)$$

(10)

where $\hat{f}\left(x\right)$ is the estimated density at point $x$, $n$ is the total number of data points, $h$ is the bandwidth parameter, $K$ is the kernel function, $x_i$ is each data point in the sample.

2.3.3. Evaluation Metrics Contribution Quantification

Shapley additive explanation (SHAP) values were used to quantify the marginal contribution of each factor to the prediction of urban renewal demand. The SHAP values reveal the direction and magnitude of the influence of each feature on the prediction results by calculating the weighted marginal contribution of each feature to the model output. By comparing the average SHAP values of different features, the key factors that have the most influence on urban renewal decisions can be identified [42].

$${\varphi }_j=\sum _{S\subseteq F\setminus \left\{j\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|F\right|-\left|S\right|-1\right)!}{\left|F\right|!}}\left[f\left(S\cup \left\{j\right\}\right)-f\left(S\right)\right]$$

(11)

where $F$ is the set of all features, $S$ is a subset of features that does not include $j$, $f\left(S\right)$ is the model output using only the features in subset $S$, $f\left(S\cup \left\{j\right\}\right)$ is the model output when feature $j$ is added to subset $S$, $\left|F\right|!$ represents the total number of permutations of all features.

3. Experiment Process

3.1. Sample Selection

By extracting and organizing the updated buildings, a spatial database of urban renewal buildings from 2018 to 2023 was established, comprising 22,733 entries. In urban renewal, features are categorized into two classes: 1 represents updated buildings (totaling 22,733), and 0 represents non-updated buildings. During model training, because only samples of the updated buildings were identified, a combination of stratified and random under-sampling [43] was applied to select an equal number of spatial samples from the non-updated data. These two sample groups were merged to form a total of 45,466 samples for this sampling method. The samples with a value of 1 were fixed, and those with a value of 0 were randomly selected, which may have affected the regression model’s output. To ensure the stability and effectiveness of the parameter estimation, 100 rounds of random spatial sampling were conducted for each model, with the median of the predicted results used as the final output.

This study first considers buildings within the study area as a whole, utilizing GIS spatial analysis methods to explore the robustness of the model and the potential for urban renewal. Second, the buildings were categorized into residential, industrial, and commercial types based on their functions to analyze the robustness and predictive potential of the model further. In addition, the contribution of the influencing factors was quantitatively assessed.

3.2. Data Correlation Analysis

In the process of building and optimizing machine learning models, systematic analysis of the correlations between features can effectively identify and eliminate redundant or irrelevant features, thereby simplifying the model structure, reducing computational costs, enhancing performance and stability, and mitigating the risk of overfitting [44,45]. Therefore, a correlation analysis of the influencing factors was conducted, and the correlation heatmap is shown in Figure 5.

From Figure 5, it can be observed that variables x1 (building perimeter) and x2 (building footprint area) are highly correlated, with a correlation coefficient of 0.881. Similarly, variables x4 (building height) and x5 (number of above-ground floors) showed a high correlation with a correlation coefficient of 0.98. To address multicollinearity and improve the stability of the model, x1 and x5 were removed. The decision to retain x2 over x1 was based on the fact that the building footprint area provides a more direct representation of spatial coverage, which is a critical factor in urban renewal assessments. Likewise, x4 was retained over x5 due to its higher relevance in describing the vertical spatial characteristics of buildings, which are crucial in evaluating urban density and renewal potential.

After performing a multicollinearity check on the indicators and removing highly correlated ones, 19 evaluation factors were selected and were categorized as follows: Building Data: footprint area, structure type, number of above-ground floors, number of underground floors, volume, and building function. Location Factors: distance to main road, distance to secondary road, distance to school, distance to hospital, distance to metro, distance to park, distance to commercial services. Natural Factors: slope, vegetation index, elevation. Social Factors: population. Economic factors: Gross Domestic Product and Building Price.

3.3. Multimodel Fusion: Stacking Ensemble Learning

Ensemble learning is a multialgorithm fusion method based on the statistical learning theory that aims to improve the overall prediction performance by integrating multiple models. The prediction accuracy of a single model typically shows diminishing returns as the training progresses, whereas stacking, as an ensemble learning technique, can combine several different predictive models to create a new model, thus surpassing the performance of any single model [46]. In this study, after training the base models, the MLP and XGBoost models with higher prediction accuracies were selected for stacking integration. By combining different machine learning algorithms, the stacking method can further enhance the overall prediction accuracy and robustness [47]. The prediction method is illustrated in Figure 6.

To validate the reliability of the model, a known urban renewal dataset was divided into training and test sets. The model was trained using data from 2018 to 2021, and its accuracy was validated using data from 2022 to 2023. In addition, a 5-fold cross-validation (CV = 5) was used to evaluate the model. The training dataset was split into five equal parts; in each iteration, four parts were used as the training set, and the remaining part was used as the validation set. This process was repeated five times with a different part chosen as the validation set each time, and the average of the five validation results was used as the model evaluation. GridSearchCV is used to adjust the model parameters and ensure that the output is an optimal solution.

4. Results

4.1. Model Accuracy Comparison

The optimal model was selected for integration after training and validation. The integrated model was tested on the 2022–2023 city renewal building test set, with a prediction range of 0–1. A threshold of 60% was set, and the percentage of samples exceeding this threshold was defined as the accuracy of the model. To assess the effectiveness of multimodel integration in building renewal classification accurately, this study used accuracy, precision, recall, and F1 score as the main evaluation metrics. Because the focus of the study was renewed buildings, the accuracy was based on the average of updated buildings (label 1) and non-updated buildings (label 0), whereas the precision, recall, and F1 score were based on updated buildings (label 1).

Table 6. Evaluation Metrics for the Prediction Dataset.

		MLP	RF	SVM	XGBoost	Stacking
Overall prediction	Accuracy	87.22%	80.30%	81.97%	87.20%	89.41%
	Precision	98.79%	99.29%	95.07%	99.45%	99.45%
	Recall	75.37%	61.04%	67.45%	74.81%	79.26%
	F1 Score	85.51%	75.60%	78.91%	85.39%	88.21%
Industrial category prediction	Accuracy	94.33%	91.29%	90.51%	95.40%	95.54%
	Precision	98.01%	95.05%	94.00%	99.08%	98.49%
	Recall	90.51%	87.11%	86.54%	91.64%	92.49%
	F1 Score	94.11%	90.91%	90.12%	95.22%	95.40%
Residential category prediction	Accuracy	88.45%	81.80%	85.14%	89.11%	89.55%
	Precision	97.95%	98.55%	95.81%	99.72%	99.72%
	Recall	78.55%	64.55%	73.50%	78.44%	79.31%
	F1 Score	87.18%	78.00%	83.19%	87.81%	88.36%
Commercial category prediction	Accuracy	84.07%	76.11%	83.63%	82.74%	81.42%
	Precision	96.39%	96.83%	94.19%	96.25%	98.63%
	Recall	70.80%	53.98%	71.68%	68.14%	63.72%
	F1 Score	81.63%	69.23%	81.41%	79.19%	77.42%

presents the results for each metric.

The experimental results show that the ensemble model significantly improves the prediction accuracy for all types of updated buildings, particularly for the evaluation metrics of the test set. The ensemble model outperforms the individual regression models across all metrics. The performance of the model on the test set is shown in Table 6. For the overall urban renewal building prediction, the highest accuracy for a single model was achieved using the MLP model with an accuracy of 87.22%. Using stacking to combine the MLP and XGBoost models, the ensemble model achieved an accuracy of 89.41%, which is a 2.19% improvement over the best-performing single MLP model. This indicates that multimodel fusion effectively integrates the strengths of the individual models and enhances the overall prediction performance.

In classification prediction, the performance of different types of buildings varies. The prediction of industrial building updates exhibited the most significant improvement. The accuracy of industrial building prediction using the stacking ensemble model was 95.54%, which is a 6.13% improvement over the overall building ensemble model accuracy of 89.41%. The prediction of residential building updates has also improved. The accuracy of the residential building prediction using the stacking ensemble model was 89.55%, which is an improvement of 0.14% over the overall building ensemble model accuracy of 89.41%.

However, in the prediction of commercial building renewal, the integrated MLP and SVM model achieved an accuracy of 81.42%, which is a decrease of 2.65% compared with the 84.07% accuracy of the MLP model. During the prediction process of individual models, except for the SVM model, the accuracy of other models all declined. This decline may be due to the more complex update patterns of commercial buildings. Such patterns make it challenging for the model to learn consistent features or predict update needs accurately. Furthermore, the update patterns of commercial buildings are frequently linked to broader redevelopment projects involving multiple building types, introducing complexities that single-category models may struggle to capture. These interdependencies may reduce the model’s predictive accuracy for commercial buildings. Addressing these challenges requires a deeper exploration of spatial interactions and building dependencies, as discussed in the subsequent analysis of spatial distribution.

4.2. Urban Renewal Prediction Results

Prediction results were obtained by integrating the base models, MLP, and XGBoost. Buildings were categorized based on their usage functions into industrial and residential buildings. The experimental results were normalized and classified into three levels: Level I (low potential), Level II (moderate potential), and Level III (high potential), as shown in Figure 7.

According to 2018 census data for the Longhua and Longgang Districts in Shenzhen, there were 234,225 buildings, of which 22,733 were updated between 2018 and 2023. The updated building types include offices, warehouses, industrial buildings, public facilities, shopping centers, transportation, residential buildings, residential support facilities, commercial services, and commercial streets. Among these, warehouses and industrial buildings are categorized as industrial updates; residential buildings, residential support facilities, and private houses are categorized as residential updates; and commercial services and streets are categorized as commercial updates. The remaining office buildings, public facilities, shopping centers, and transportation facilities were classified as other updates. Industrial updates accounted for 18.65%, residential updates 77.38%, commercial updates 3.29%, and other updates 0.68% of the total. This study mainly focuses on industrial, residential, and commercial building updates. Based on the 2018 building census data and after excluding outliers, the total number of non-updated buildings is 206,670, including 44,343 industrial buildings, 148,247 residential buildings, 6991 commercial buildings, and 7089 other buildings. The classification of the building-update potential levels based on the model prediction results is presented in

Table 7. Building Urban Renewal Potential Classification.

	I Level	II Level	III Level	Total (Buildings)
Overall Building Update	124,731	63,243	18,696	206,670
Industrial Building Update	29,265	10,354	4724	44,343
Residential Building Update	98,429	34,167	15,651	148,247

. Commercial buildings were not included in the update prediction because of their abnormal training performance.

The spatial distribution of the urban renewal potential for all, industrial, and residential buildings exhibits a gradient pattern. Level I (low-potential) buildings are mainly located in the peripheral areas of the two districts, with low-density buildings and weak renewal demand. Level II (moderate-potential) buildings are distributed in secondary development zones or transitional areas with a certain renewal potential. Level III (high-potential) buildings are concentrated in five major clusters: Pinghu Street, Longhua Street, Buji Street, Henggang Street, and Longgang Street. Industrial buildings in urgent need of renewal are primarily located near major urban roads for convenient transportation. Residential buildings in urgent need of renewal are concentrated in urban village clusters with high population densities and underdeveloped infrastructure, such as Bantian Street, Longhua Central District, and its surrounding areas.

a.

Urban Renewal Evaluation Analysis

Based on the prediction results, a kernel density analysis was conducted, as shown in Figure 8. Figure 8a shows the number of buildings updated between 2018–2023. The kernel density analysis of the spatial distribution of the updated buildings, shown in Figure 8b, clearly identified concentrated areas of renewal on the northern and northeastern edges of Longhua District, as well as in the central and northern areas of Longgang District.

An overlay map (Figure 8(e1)) compared the kernel density analysis of building updates from 2018–2023 (Figure 8(c1)) with the predicted update density analysis (Figure 8(d1)). The predicted hotspots closely matched the actual hotspots, indicating that the model effectively captures the spatial patterns of urban renewal, particularly in key areas with high potential for future development. This provides valuable reference data for future urban renewal planning. In classification prediction, industrial building updates showed notable performance. The kernel density analysis of industrial updates (Figure 8(c2)) and predictions (Figure 8(d2)) in the overlay map (Figure 8(e2)) demonstrated a high correlation, reflecting the model’s ability to capture urban renewal patterns. For residential buildings, overlaying the update analysis (Figure 8(c3)) and predictions (Figure 8(d3)) in Figure 8(e3) revealed an expansion from original areas to surrounding regions, with concentration trends in central Longhua District and the southwest intersection of Longgang and Longhua Districts. These areas are likely to become key urban renewal zones, as indicated by the predictions.

In predicting commercial building updates, anomalies in model evaluation metrics were observed. Figure 8(c4) shows the spatial distribution of building types, while Figure 8(d4,e4) highlight the updated distribution of commercial, industrial, and residential buildings. Unlike industrial and residential buildings, which exhibit clear spatial concentration, commercial updates appear scattered and lack clustering. This dispersion, often linked to comprehensive renovation efforts involving other building types, increases prediction complexity and evaluation instability.

b.

Evaluation Metrics Contribution Quantification

To better understand the influence of each factor on urban renewal predictions, the SHAP values are analyzed, as shown in Figure 9. Figure 9a shows the average SHAP values for each factor, while Figure 9b presents a SHAP summary plot (bee swarm plot), which further illustrates the positive and negative contributions of each factor to the prediction for individual samples, along with their distribution.

To predict urban renewal demand, SHAP value analysis was used to evaluate the contributions of various influencing factors. From the figure, it can be observed that “x3-1” (brick and tile structure) has the highest contribution, with a SHAP value of 2.0691, indicating that this factor has the most significant positive impact on the model’s prediction. This suggests that brick and tile structures, owing to their age and poor risk resistance, urgently need to be updated and prioritized in terms of urban renewal demand. Next, “x16” (population density) and “x12” (distance from the building to the hospital) are also important influencing factors, with large SHAP values indicating that their influence on urban renewal demand is also significant.

From the distribution in the SHAP summary plot, the SHAP values of “x16” (population density) are relatively evenly spread, meaning the model’s predictions are not significantly concentrated in one direction based on population density. Both high and low population densities have relatively balanced positive and negative influences on the predictions. This suggests that high population density may increase the demand for updates due to regional pressure and resource shortages, but low population density areas may also need updates due to insufficient infrastructure. Both “x20” (GDP) and “x21” (Building Price) exhibit a strong negative relationship, with higher GDP regions and lower housing price areas showing a more pronounced negative effect. This may be because high GDP areas are commercial centers, and areas with low housing prices have fewer people, making updates more difficult. In contrast, “x3-1” (brick and tile structure) and “x12” (distance from the building to the hospital) have SHAP values more concentrated in the positive direction, indicating that these factors consistently drive the need for updates. For location-based factors, being closer to major roads, schools, subways, parks, and commercial areas significantly influences building updates. Natural Factors like “x17” (slope of the building’s location) have less correlation with building updates. Buildings located in areas with higher “x18” (vegetation index) and lower “x19” (elevation) indicate a better environment, with more development potential and investment return, positively affecting building updates. For building attributes, lower “x4” (building height) typically reflects older buildings, which aligns with the natural progression of building development, positively influencing building updates. “x7” (building volume) and “x2” (land area) may have a weaker correlation, likely due to larger industrial buildings. Building structures like “x3-2” (frame structure) and “x3-3” (mixed structure) show less intensity in driving updates compared with “x3-1” (brick and tile structure), as these structures are relatively newer. Through data-driven analysis, key factors and their specific impacts can be identified, allowing for more targeted updated planning, such as prioritizing brick and tile buildings, densely populated areas, and buildings near hospitals.

5. Discussion

5.1. Prediction Results and Spatial Kernel Density Validation

This study adopts a data-driven approach to predict building renewal potential by integrating multiple machine learning models (XGBoost and MLP), achieving an accuracy of 89.41%. Compared with traditional methods based on expert experience or qualitative analysis, the data-driven approach minimizes subjective interference and improves the accuracy and objectivity of the predictions. Notably, the prediction results for industrial and residential buildings demonstrate high accuracy, aligning well with actual renewal needs.

The results of kernel density analysis show that the updated buildings are highly concentrated in the northern and northeastern edges of Longhua District and the northern central area of Longgang District. The model’s prediction results are also concentrated around these updated clusters, indicating that these areas still hold significant potential for further renewal. In the update potential prediction for the south-central area of Longgang District, the results show a significant update potential in this region. Further analysis revealed that the area is located near Shenzhen East Station and Longgang Avenue, offering a prime geographical location with convenient transportation, particularly enhanced by the seamless connection between the Buji subway station and Shenzhen East Station, strengthening the area’s transportation advantage. The region is characterized by many urban villages, high building density, significant aging, and a dense residential population, reflecting the urgent need for updates in line with past urban renewal trends. This further confirmed the reliability of the model. The model accurately predicted the update potential in this area, demonstrating that under multidimensional data analysis and model integration techniques, the model not only precisely identifies urban renewal needs but also provides scientific support for future urban planning and renewal decisions.

5.2. SHAP-Driven Policy Recommendations

Based on SHAP value analysis, the demand for building renewal in Shenzhen is primarily influenced by the condition of old buildings, population density, and the distribution of medical resources. The aging infrastructure and long service life of old buildings, such as brick and tile structures, are the key factors driving the need for renewal, especially in areas like urban villages, where these buildings pose significant public safety risks. Therefore, it is recommended to prioritize the evaluation and renewal of areas with older buildings and deteriorating infrastructure, providing financial support and policy incentives to promote infrastructure upgrades and functional optimization, ensuring the safety of residents’ lives and property. Buildings close to hospitals have a higher renewal demand, indicating that residents in these areas rely more on medical resources and enjoy better living convenience. It is recommended to prioritize the renewal of old buildings around hospitals, increase supporting facilities such as community health service centers and rehabilitation care centers, and improve transportation conditions to enhance the overall service level of these areas. High population density areas show an especially urgent need for renewal, reflecting residents’ expectations for improved living conditions and quality of life while also highlighting the inadequacy of public service facilities. For these areas, it is recommended to prioritize building renewal, increase public service facilities such as schools, hospitals, and community service centers, and improve transportation infrastructure to enhance overall quality of life. These recommendations will help improve the living environment in various regions of Shenzhen, enhance the accessibility of public service facilities, and promote the sustainable development of the city.

5.3. Uncertainty Analysis

Although the model’s predictions are highly aligned with the actual updated areas, in rapidly developing urban environments, building updates are influenced by various factors such as economic development, policy adjustments, and market demand. These dynamic changes may fall outside the scope of the model predictions. Therefore, future research should consider incorporating more dynamic data sources, such as real-time economic data and policy changes, to improve the accuracy of predictions of future renewal trends. In addition, the multimodel fusion technique used in this study demonstrated high accuracy and robustness in building update predictions, particularly for industrial and residential buildings. However, the prediction metrics for commercial buildings showed anomalies, primarily because commercial buildings are often part of comprehensive redevelopment efforts within urban renewal, and their update patterns are highly dependent on the renewal of other types of buildings. This interdependence makes it challenging for the model to predict commercial building updates independently. Thus, future research should explore fusion strategies tailored to different building characteristics to improve prediction stability and accuracy and better address the complex challenges of urban renewal.

6. Conclusions

This study conducted an in-depth analysis of the spatial distribution and renewal potential of buildings in the Longhua and Longgang Districts of Shenzhen from 2018 to 2023. Multimodel integration, GIS spatial analysis, and other technical methods have been used to systematically evaluate the renewal demand and spatial characteristics of different types of buildings, providing a scientific basis for urban renewal decision-making.

The model training and integration results show that by using the predictions from MLP and XGBoost as inputs and integrating them through the AdaBoost classifier, the overall goodness of fit of the regression model improved by 2.19%, reaching a prediction accuracy of 89.41%. In the building classification analysis, the model’s goodness of fit for residential buildings improved by 0.14%, whereas that for industrial buildings increased significantly by 6.13%. The prediction accuracy of this study is 8.41% higher than the macro-level urban renewal unit prediction accuracy (81%) based on the AET model proposed by Yao et al. [48]. Furthermore, compared with the single-model accuracy (60.3%) based on planning permit data by Cao et al. [17], the improvement was 29.11%, validating the superiority of the micro-level multimodel fusion approach in urban renewal prediction.

The following building data were selected: plot area, structural type, building height, number of underground floors, volume, and house function; Location Factors: distance to main roads, secondary roads, schools, hospitals, subways, parks, and commercial areas; Natural Factors: slope, vegetation index, and altitude; and Social Factors: population, with a total of 19 evaluation factors. From the micro-level perspective of buildings and by considering the surrounding factors, various machine learning methods in data-driven models have been used to effectively avoid the influence of subjective human judgment, ensuring the scientific and rational nature of potential assessment.

This study found significant differences in the updated demand for industrial, residential, and commercial buildings. Industrial building updates were mainly concentrated in the northern edge and northeastern central areas of the study region, close to the main roads, indicating a strong demand for functional upgrades and environmental improvements. Residential building updates are relatively uniform but show a trend of expansion into surrounding areas, with the main update focusing on densely populated urban village areas. Commercial building updates are unique and primarily occur alongside industrial and residential building updates.

The model’s prediction results align closely with the actual update areas, validating the effectiveness of multimodel integration techniques in identifying update demands. This method combines the strengths of different models, significantly improves prediction accuracy, and provides an accurate foundation for the positioning and prioritization of urban renewal projects. This helps decision-makers identify areas in urgent need of updates more effectively, thus enabling the formulation of more precise update strategies and promoting efficient urban development.

The applicability and generalization of this study still need further validation, as different cities vary in their update demands, policy environments, and socioeconomic contexts. Therefore, the promotion and application of the research findings should be adjusted according to local conditions. Additionally, future research could expand the sample areas and increase comparative analyses of data from different periods, further enriching and improving building renewal potential evaluation systems.