Research on Site Selection Planning of Urban Parks Based on POI and Machine Learning—Taking Guangzhou City as an Example

Abstract

Against the background of smart city construction and the increasing application of big data in the field of planning, a method is proposed to effectively improve the objectivity, scientificity, and global nature of urban park siting, taking Guangzhou and its current urban park layout as an example. The proposed approach entails integrating POI data and innovatively applying machine learning algorithms to construct a decision tree model to make predictions for urban park siting. The results show that (1) the current layout of urban parks in Guangzhou is significantly imbalanced and has blind zones, and with an expansion of the search radius, the distribution becomes more concentrated; high-density areas decrease from the center outward in a circle, which manifests as a pattern of high aggregation at the core and low dispersion at the edge. (2) Urban park areas with a service pressure of level 3 have the largest coverage and should be prioritized for construction as much as possible; there are fewer areas at levels 4 and 5, which are mainly concentrated in the central city, and unreasonable resource allocation is a problem that needs to be solved urgently. (3) There was a preliminary prediction of 6825 sites suitable for planning, and the fit with existing city parks was 93.7%. The prediction results were reasonable, and the method was feasible. After further screening through the coupling and superposition of the service pressure and the layout status quo, 1537 locations for priority planning were finally obtained. (4) Using the ID3 machine learning algorithm to predict urban park sites is conducive to the development of an overall optimal layout, and subjectivity in site selection can be avoided, providing a methodological reference for the planning and construction of other infrastructure or the optimization of layouts.

1. Introduction

Urban parks that are open to the public serve primarily recreational purposes but also provide landscape, educational, ecological, and emergency shelter functions [1]. As a crucial component of urban development, they play significant roles in improving urban landscapes, ecosystems, and the quality of life for residents. However, rapid population growth has intensified the strain on public services, posing significant challenges to the development of healthy cities and environments [2,3]. Furthermore, issues such as irrational layouts, waste of resources, failure to meet demand, and management chaos exist in the site selection and planning of urban parks, leading to a severe shortage in the urban park supply and exacerbating the imbalance in urban–rural regional development [4,5,6]. Therefore, comprehensively planning and constructing urban parks are tasks of utmost importance. Effectively avoiding subjectivity and arbitrariness in planning and site selection and enhancing their rationality and scientificity remain important issues that require further contemplation in future research [7,8].

Economic and social transformations have intertwined urban, environmental, and human needs, increasing the demand for urban parks. This has created new challenges for urban spatial planning and renewal [9]. The location and layout of urban parks are influenced by a combination of natural, social, economic, and demographic factors, and, at the same time, they are counterproductive to socio-economic development. Given rising social needs and crowd demands, exploring the relationship between urban park location planning, quality improvement, service pressure, and the current layout is urgent [10,11]. Scholars globally have extensively researched urban parks. In the application of big data technology, layout assessment, spatio-temporal vitality, and usage evaluations in urban park siting have mainly been studied by mining and analyzing data, such as social network data [12], physical activity VGI data [13], cell phone signaling data [14], and network critique data [15], which are combined with layout features to extract the activity trajectory of real user populations and evaluate their post-use satisfaction. Furthermore, intelligent transportation swipe card data [16] have been used to examine the relationship between peripheral transportation facilities and the site planning, maintenance, and management of urban parks. In terms of site selection and planning methods, the IPA–Kano model [17], cognitive map method [18], social network analysis method [19], and two-step mobile search method [20] have been used to assess the performance of urban parks in terms of accessibility and network structure optimization [21,22] to reveal the reasonableness and fairness of urban park siting.

There are five main types of urban park studies in the specific research literature regarding site planning. The first type analyzes landscape accessibility by examining factors such as landscape resistance [16], the impact of landscape characteristics on the urban park cold island effect [23], physical activity types through the use of multivariate data [24], the influence of green visibility on physical and mental health through the construction of an evaluation system [25], and comfortable designs and good management through a comprehensive multiperspective, multidimensional, and multilevel approach to urban park site planning. The second type utilizes the principles of landscape ecology [26] to analyze urban parks as ecological patches with the help of an eco-unit mapping approach to explore various impact indices. The third type is from a planning perspective [27], where accessibility and service radius are used as the planning basis for siting and building in blind zones. The fourth type is based on mathematical models, where artificial intelligence recognition, such as particle swarm and ant colony algorithms, is applied to site selection [28], and the complex problem of site selection is solved through the fusion of multi-objective particle swarm optimization and region shape variation algorithms. The fifth type involves sociological analyses [29], which evaluate the locations of sites through existing city parks and population density or the exercise of subjective judgment and selection with the help of questionnaires or the POE method.

Overall, the existing research on urban park siting focuses primarily on supply-oriented perspectives, such as ecological landscapes, mathematical models, and social analysis, neglecting the needs of residents, resulting in subjective and unscientific site selections. Guangzhou, as a significant port city for China’s foreign trade, is experiencing continuous population growth and territorial expansion, leading to a decrease in per capita urban park green space area amidst high demand. This demand disparity across regions creates pressure on urban park services, resulting in an imbalance between actual demand and supply. Guangzhou serves as an ideal case study for simulating and predicting locations for urban park site selection and planning based on residents’ living needs, interests, and hobbies. Furthermore, amidst the widespread application of smart cities and big data in urban planning and construction, big data have become comprehensive and rapidly accessible. Machine learning, when used as a convenient algorithm with data derived from smart cities [30,31], relies on high-quality, multidimensional data encompassing usage frequency, visitor behavior, environmental indicators, and equipment configurations to drive the decision-making process in urban park site selection prediction. During the decision-making process, machine learning demonstrates real-time learning and optimization capabilities, enabling the rapid identification of key factors influencing site selection decisions based on multi-source data [32,33]. Integrating existing urban park locations ensures that urban park planning aligns with actual spatial environmental needs and reduces arbitrary and subjective human intervention. This approach better satisfies the needs of users, enhances residents’ quality of life, and provides more objective and scientific support for urban park planning and construction, offering unique advantages in site selection prediction [34]. Currently, smart cities and big data, from a machine learning perspective, are vital means and pathways for promoting sustainable urban development and providing a strategic direction for national urbanization development in the field of urban planning. Consequently, big data and machine learning will serve as an essential foundation for future research on evaluating and predicting urban park site selection and layout. In this study, big data on human activity footprints in urban parks and various urban facility POIs in Guangzhou, China, were utilized. By employing machine learning algorithms, we constructed a decision tree model for urban park site selection to enhance objectivity, scientific rigor, and comprehensiveness, aiming to contribute to the improvement and development of urban park site selection and planning methods.

2. Study Area and Ideas

2.1. Study Area

As an international trade center and comprehensive transportation hub, Guangzhou has experienced rapid population growth over the past five years, with the resident population increasing by 3,829,700. In September 2023, the Guangzhou Municipal People’s Government issued the “Green Space System Planning” notice, which stated that the short-term goal is to raise the per capita park green space area to 17.5 m²/person by 2025 and the coverage rate of the service radii of parks and green spaces to 85%. The long-term goal is for the former to be no less than 17.5 m²/person and the latter to reach 90%. However, as of now, the two figures are 17.23 m²/person and 83.1%, respectively, which still fail to meet the per capita demand. This shows that the number of parks in Guangzhou is insufficient, and the coverage is limited, highlighting the pressing need to devise and implement a strategy for the construction of a multitude of parks and green spaces. Furthermore, the planning notice for the park system stipulates that, in accordance with ecological resource protection and utilization requirements, moderate additions should be made to country and urban parks. By 2035, the city’s planning should include a minimum of 2000 parks, including about 120 urban parks. Thus, meeting the travel and life needs of residents at different levels, making the site selection of new urban parks scientific and reasonable, and optimizing the original layout are the key issues addressed in this study.

2.2. Research Ideas

Urban parks are essential city infrastructure with rational layouts and a referenceable status quo. This study’s methodology included (Figure 1) acquiring POI data for Guangzhou city parks via Baidu API and Python, using outdoor tools to collect park usage footprints, and processing these data with ArcGIS 10.2 to determine current layout features and service pressure. Next, Guangzhou was divided into 500 × 500 m grids, and a decision tree model for park siting was developed using the ID3 algorithm based on the existing park distribution. To validate the model, the simulation results were compared with actual parks; a match of over 80% indicates reliability, prompting further optimization if necessary. The model then predicted suitable grids for park construction. Finally, combined with the current layout and service pressure in each area, the grid locations were filtered again to prioritize planning and construction among grids.

3. Data Sources and Methodology

3.1. Data Sources

POI data in this study are mainly divided into two categories: (1) Various types of urban facilities in Guangzhou City. We used Baidu Map API and wrote programs in Python to obtain spatial vector data, such as the names, categories, and longitudes and latitudes of various facilities. The coordinate system of these data was transformed into the WGS_1984 geographic coordinate system through the projection tool in ArcGIS 10.2 software, and duplicates were screened out and eliminated. After processing, a total of 663,749 POI data points were obtained, covering 11 districts in Guangzhou. Subsequently, the various types of facilities were classified according to their relationship with urban parks. This process yielded four primary indicators: basic service facilities, medical and retirement facilities, commercial service facilities, and administrative office facilities. Additionally, 15 secondary indicators were identified: urban parks, life services, leisure and entertainment, shopping centers, retirement facilities, educational facilities, catering, real estate, medical services, companies, transport facilities, hotels, travel sites, financial institutions, and government organizations. Class II indicators are presented in

Table 1. POI data types for various facilities in Guangzhou City.

Level 1 Indicators	Level 2 Indicators	Names of Facilities	Quantities
Infrastructure services	Transport facility	Bus stops, subway stations, bus stations, train stations, etc.	2357
	Educational facility	Science and technology museums, libraries, exhibition centers, central schools, kindergartens, cultural palaces, etc.	31,167
	Leisure and entertainment	Supermarkets, food markets, telecommunication offices, repair shops, photo studios, etc.	3521
	Travel sight	Scenic spots, monuments, museums, churches, etc.	6891
	Life service	Supermarkets, vegetable markets, telecommunication business halls, repair shops, photo studios, public toilets, etc.	101,446
	Urban parks	Comprehensive parks, children’s parks, zoos, botanical gardens, residential parks, strip parks, etc.	870
Medical care facilities for the elderly	Retirement facility	Elderly care centers, social service centers, universities for the elderly, old age apartments, etc.	1369
	Medical service	Clinics, CDCs, nursing homes, general hospitals, specialty hospitals, pharmacies, etc.	16,341
Commercial services	Financial institution	Banks, securities companies, ATMs, insurance companies, trust companies, etc.	12,572
	Shopping center	Department stores, shopping centers, supermarkets, convenience stores, bazaars, building materials and furniture, etc.	225,541
	Hotel	Guesthouse, apartment hotels, express hotels, star hotels, etc.	21,287
	Catering	Chinese restaurants, coffee shops, cake stores, tea houses, fast food stores, snack stores, etc.	72,961
Administrative office facilities	Real estate	Office buildings, commercial buildings, residential areas, industrial buildings, etc.	11,676
	Company	Companies, factories, agriculture, forestry and fishery bases, etc.	126,000
	Government organization	All levels of government, administrative units, welfare organizations, etc.	29,750
Total			663,749

. (2) The traces of people’s activities in various urban parks in Guangzhou. Based on the software Liuzhijiao 4.17.0 (Six Feet: an outdoor sharing platform) and Liangbulu 7.7.9 (Two Steps: an outdoor assistant), a total of 171,958 footprints posted by active people were crawled. The service pressure on urban parks in each region of Guangzhou was calculated using the spatial connection tool in ArcGIS 10.2 software. In addition, according to the service radii of urban parks, Guangzhou was divided into 29,793 grids with a 15 min walking distance as a reference and a scale unit of 500 × 500 m. Each grid has a unique number (Figure 2). ArcGIS 10.2 software and Python code were then used to statistically analyze the distribution of various facilities within the grid, laying the foundation for training the decision tree model after quantification.

3.2. Methodology

3.2.1. Kernel Density Estimation

Kernel density estimation is an important method to assess whether point elements are aggregated, and its value can also reflect the degree of densification. This method was used to reveal the distribution status of urban parks in Guangzhou; the calculation formula can be expressed as follows [1]:

$$\mathrm{f}(\mathrm{x})={\displaystyle \frac{1}{\mathrm{nr}}}{\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{n}}\mathrm{k}{\displaystyle \frac{\mathrm{d}}{\mathrm{r}}}$$

where f (x) represents the kernel density value of urban parks in area x; n is the total number of samples; k is the kernel weight function; d is the distance between the sample point and the measurement point.

3.2.2. Standard Deviation Ellipse

The standard deviation ellipse is a statistical method to express the spatial characteristics of point data. It is used to explore the distribution characteristics of discrete datasets and reveal the directionality and extension direction of urban park development. The calculation formula is as follows [4]:

$$\mathsf{\alpha }={\displaystyle \frac{\mathrm{a}-\mathrm{b}}{\mathrm{a}}}\left(\mathsf{\alpha }>0\right)$$

where α is a flattening parameter, and its value indicates the strength of directionality. The larger the flattening, the stronger the directionality, and vice versa. a is the major semiaxis, and b is the minor semiaxis.

3.2.3. Nearest Neighbor Index

The nearest neighbor index is an important quantitative indicator used to characterize the distance between physical elements in a geographic space. It can reflect the spatial distribution pattern of point elements. The calculation formula is as follows [19]:

$$\mathrm{R}={\displaystyle \frac{\overline{{\mathrm{r}}_1}}{\overline{{\mathrm{r}}_{\mathrm{E}}}}}=2\sqrt{\mathrm{D}}$$

where R is the nearest neighbor index, r₁ represents the actual distance between urban parks in the region, and r_E is the expected distance. When R < 1, the distribution is clustered; when R = 1, the distribution is random; when R > 1, the distribution is uniform.

3.2.4. Service Pressure Evaluation

The service pressure on urban parks in a region refers to the ratio of the number of POIs used by the population within the service range of all urban parks in the region to the total area of all urban parks in the region. It can be used to quantitatively evaluate the service levels of existing urban parks in the region. ArcGIS 10.2 software was used to visualize regions, and the existing urban parks in each region were further used as discrete points to generate the service scope. The total area of urban parks was calculated, and then, through spatial processing, the POIs of the user population were matched with the service scope. By constructing an attribute table, the ratio of the number of POIs within the service range of all urban parks in the region to park area was calculated to obtain the service pressure on the existing urban parks in the region [14].

3.2.5. Decision Tree Model

The ID3 algorithm is a classification prediction algorithm proposed by Prof. J. Ross Quinlan in 1975. The algorithm regards information entropy as the core and information gain as a measure [35]. The ID3 algorithm measures the optimal grouping and splitting points of attributes by calculating the information gain of each attribute (Equations (1) and (2)), and the attribute with the highest information gain is selected as the splitting criterion for each division (Equation (3)); this process is repeated until a decision tree is generated that can perfectly classify the training samples [36]. A decision tree consists of decision nodes, branches, and leaves, with the topmost being the root node and each branch being a new decision node or leaf of the tree. Each decision node represents a problem or decision, usually corresponding to an attribute of the classified object. Each leaf node represents a possible classification result [37]. Along the decision tree, from top to bottom, a test will be encountered at each node, and different questions will lead to different branches of output [38]. Finally, a leaf node will be reached, and a number of variables will be utilized to determine the category to which it belongs. These variables can be used to quickly and accurately reason out a set of classification rules in the form of a tree from a set of unruly and unordered datasets in accordance with their attributes and categories, which are learned through training. After learning through training, a classifier that can be utilized for this kind of sample dataset is obtained, and then the dataset can be classified and determined with the help of the classifier [39]. Scholars in China have used machine learning and decision tree models to explore the age-friendliness of urban street environments [40], pedestrian environment indicators [41], and the spatial characteristics of Chinese classical gardens [42]. These help planners optimize space in a more systematic, refined, and concrete manner and propose corresponding improvement strategies. Therefore, the aim of this study was to establish a decision tree model of Guangzhou urban parks by applying the ID3 algorithm to existing site selection decision problems for 15 types of samples in Guangzhou and to study future planning and construction at the selected sites. The equations for the calculations are detailed below.

Suppose there is a sample set S = $\left\{{\mathrm{S}}_1+{\mathrm{S}}_2,\cdots ,{ \mathrm{S}}_{\mathrm{n}}\right\}$, a decision attribute set D = $\left\{{\mathrm{D}}_1+{\mathrm{D}}_2,\cdots ,{ \mathrm{D}}_{\mathrm{m}}\right\}$, and the j-th attribute D_j = (j = 1, 2, $\cdots$, m). If there are |D_j| samples, the information entropy info(D) of D is [35]

$$\mathrm{info}(\mathrm{D})=-{{\displaystyle \sum }}_{\mathrm{j}=1}^{\mathrm{m}}{\mathrm{p}}_{\mathrm{j}}\log 2{\mathrm{p}}_{\mathrm{j}}$$

(1)

where p_j is the probability of the j-th class, which is commonly evaluated by the ratio of the number of samples to the total number. If attribute A has a data sample set = $\left\{{\mathrm{A}}_1+{\mathrm{A}}_2,\cdots ,{ \mathrm{A}}_{\mathrm{k}}\right\}$ and the i-th attribute takes the value Ai = (i = 1, 2, $\cdots$, k), then |Aj| is the number of samples with the value i, and |Aij| represents the number of samples of the j-th class in Ai. The information entropy of attribute A after partitioning is [35]

$$\mathrm{info}(\mathrm{A})=-{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{k}}{\displaystyle \frac{\left|\mathrm{Aj}\right|}{\mathrm{S}}}·{{\displaystyle \sum }}_{\mathrm{j}=1}^{\mathrm{m}}{\mathrm{p}}_{\mathrm{ij}}·\log 2{\mathrm{p}}_{\mathrm{ij}}$$

(2)

Then, the information gained before and after the division of attribute A is calculated as follows [35]:

Gain (A) = info (A) − info (D)

(3)

Using the ID3 algorithm to calculate the information gain of various attributes before and after division is the core idea of the decision tree model. Each operation takes the attribute with the largest information gain as a node—that is, takes the most effective classification attributes as the specific conditions for classification—and ultimately obtains the optimal decision tree for urban parks through repeated iterations.

4. Current Urban Park Layout in Guangzhou

4.1. Distribution Characteristics of Urban Parks

Using analysis tools such as kernel density estimation, the standard deviation ellipse, and the nearest neighbor index in ArcGIS 10.2 software, the current urban park layout in Guangzhou was visually expressed at multiple scales to portray its spatial heterogeneity. As 1000 m represents the accessible range for residents’ daily walking, 2000 m corresponds to the accessible range for cycling or short-distance public transportation, and 3000 m covers a wider range of transportation modes, such as the subway, the search radii in this study were set to 1000 m, 2000 m, and 3000 m to better analyze the current layout of urban parks based on residents’ actual usage patterns. The pixel size was adjusted to 50, and the density values were classified into five levels to analyze the spatial characteristics of urban parks in Guangzhou. Furthermore, the standard deviation ellipse was employed to elucidate the distribution center, extension direction, and discrete degree of Guangzhou urban parks. A nearest-neighbor analysis using the Euclidean platform distance method was used to calculate the expected distance and the actual distance between urban parks in Guangzhou. The aggregation state is determined by the ratio of the two distances, and the smaller the ratio, the stronger the agglomeration.

The kernel density analysis map of Guangzhou city parks (Figure 3) reveals a significant imbalance in the overall distribution, with clear local clustering in the middle and dispersion in the periphery. There are numerous distribution blind spots, showing a pattern of “large clustering and small dispersion.” Urban parks with a search radius of 1000 meters exhibit the most high-density areas, with a more balanced distribution. Furthermore, as the search radius continues to increase, the distribution of urban parks in Guangzhou becomes more concentrated, with high-density areas gradually becoming more prominent. This is evidenced by a decreasing circle from the western part of the city, which demonstrates a “core–edge” pattern. The distribution of high-density urban parks is concentrated in the Yuexiu, Haizhu, Tianhe, Panyu, and Baiyun Districts. The nuclear density contours of the Baiyun and Panyu Districts decrease gradually from the center to the periphery, indicating that the number of urban parks in these areas changes dramatically with the increase in distance; the nuclear density contours of the Tianhe and Haizhu Districts show spatial alienation characteristics, reflecting a slow change in the number of urban parks with the increase in distance; the Huadu, Conghua, Zengcheng, and Nansha Districts have sparse urban park distributions, forming low-density spatial clusters with weak homogeneity, indicating that the above areas are considerably lacking in urban parks. According to the standard deviation ellipse analysis chart of Guangzhou urban parks, the average center of their distribution is located in the Tianhe District. The major semiaxis of the ellipse extends from the southeast to the northwest, and the difference between the major and minor semiaxes is large. This phenomenon indicates that Guangzhou city parks have a more significant directionality from southeast to northwest, and their distribution is more clustered. The nearest neighbor calculation results of Guangzhou’s urban parks show that the expected distance is much smaller than the actual distance. The nearest-neighbor ratio is 0.58, which is much smaller than 1, indicating that the overall distribution of urban parks shows significant clustering characteristics at the 0.01 level. In summary, the overall layout of Guangzhou’s urban parks can be described as exhibiting a “large clustering, small dispersion” pattern with a southeast–northwest direction. There are blind spots in many areas, a weak spatial balance, and significant differentiation.

4.2. Service Pressure Evaluation of Urban Parks

In addition, with the help of the extraction tool of ArcGIS 10.2 software, the areas with better service capacity were screened out from the areas with medium service pressure values, and their service pressure values were taken as the third level. On the basis of this criterion, the service pressure in each region was evaluated and graded. The greater the pressure, the higher the level.

The results of the service pressure evaluation are shown in Figure 4, with higher tiers indicating a greater service pressure on the region’s urban parks. This observation indicates relatively few areas with level 1 and level 2 service pressure, suggesting that the number and coverage of urban parks in such areas are relatively reasonable. The standard level 3 service pressure area is large, covering 11 districts in Guangzhou, and future planning should include as many urban parks in this type of area as possible. In addition, level 4 and level 5 areas are fewer in number and smaller in size and are mainly concentrated in the centers of economically developed and densely populated built-up areas with high activity densities. This pattern clearly indicates that the number of urban parks in Guangzhou is unable to share the high recreational pressure and that the overall layout fails to meet the needs of the population. In conclusion, the number, current layout, and service pressure level of Guangzhou’s urban parks fail to meet the population demand. This has led to an inefficient use of resources and an urgent need to increase the number of urban parks.

5. Decision-Making on the Location of Urban Parks in Guangzhou

The construction of urban parks represents a significant aspect of urban development, with the potential to enhance the quality of life and meet the leisure needs of the population. The siting of completed urban parks is rational, offering insights that can inform future predictions and planning. In this study, future siting was analyzed based on the existing urban park siting characteristics in Guangzhou, and the process was divided into three parts. First, the ID3 algorithm was used to train a decision tree model suitable for the site selection of urban parks in Guangzhou. Second, with the help of this model, simulation prediction was carried out for each 500 × 500 m grid in Guangzhou to determine whether the new urban parks are reasonable. Finally, the site selection results of the priority layout of urban parks were screened and finalized.

5.1. Decision Tree Establishment for Urban Park Siting

The following criteria were used to determine whether each grid in Guangzhou is suitable for an urban park. All grids and facilities were connected, matched, and classified using the connection tool in ArcGIS 10.2 software. If a grid contains an urban park, it means that the objective environment of this grid is suitable and meets the site selection requirements for the urban parks. In this case, the grid is classified as positive and assigned a value of +1. Otherwise, it is classified as negative and assigned a value of −1, thereby realizing the binary classification of potential sites of urban parks. Subsequently, based on the above categorization and assignment method, other facility elements in the grid were also classified through binary classification.

The fundamental process in urban park siting decision-making involves continuously putting forward requirements for the surrounding environmental conditions of alternative spaces and then removing spaces that do not meet these requirements. This process is consistent with the idea of the decision tree algorithm. The presence or absence of various types of facilities in the site selection decision is analogous to each bifurcation node of the decision tree model. Whether it meets the requirements for becoming an urban park site is the decision layer of the decision tree. In this study, the types of existing facilities in all grids in Guangzhou are considered independent variables, and the dependent variable is the presence or absence of an urban park. The ID3 algorithm was used to train the decision tree, thus producing the decision tree model of Guangzhou’s urban parks (Figure 5). The first four layers of the decision tree cover 11 of the 14 types of POI data other than urban parks. Among them, the left and right branch points represent whether the city grid contains this type of POI. According to the principle of the ID3 algorithm, the higher the level of the POI type, the greater the impact on the urban park site selection decision. Life services are located at the first level of the decision tree model, indicating that they are the most important factor in the site selection of Guangzhou’s urban parks. Meeting people’s life needs has prompted most existing urban parks to be located within grids containing life services. The second layer is leisure and entertainment, as well as shopping centers. The leisure and entertainment description on the left shows that if there are no life services within the grid, the priority will be to build a park near leisure and entertainment facilities such as cinemas and amusement parks, which is consistent with the site selection principles for urban parks. The shopping center on the right side shows that if an urban park is built within a grid that contains life services, then there will also be shopping facilities within the grid, which is more in line with people’s shopping needs. The presence of retirement facilities and educational facilities on the left side of the third level indicates that if an urban park is built near leisure and entertainment facilities, there will be retirement facilities or science education and culture facilities in the surrounding area, which will support elderly users of urban parks by facilitating their entrance to and exit from elderly care institutions or promote popular science learning for children. On the right-hand side, the presence of catering and real estate indicates that there are also catering facilities, such as hot-pot restaurants or milk tea shops, near urban parks. In addition, real estate is also included. For example, urban parks are often located near office buildings and employee dormitories. The fourth layer and the three types of facilities that do not appear in the diagram have a weaker influence on location decisions for urban parks, such as transportation facilities and hotels, indicating that some urban parks will also be built in more remote suburban wetlands or rivers. Urban parks are not highly dependent on financial institutions, government organizations, or travel sites, which do not appear in the figure.

5.2. Prediction of Urban Park Site Selection

In order to effectively avoid an overlap between the training set and the prediction set, 29,793 valid grids in Guangzhou were screened. These grids were then subjected to random sampling with equal positive and negative proportions, with the final result being the selection of 1396 grids as the training sample set. The urban parks in each grid were designated as the dependent variable, while the remaining 14 types of POIs were set as independent variables. The decision tree prediction model was then trained, and 6825 grids suitable for the construction of urban parks in Guangzhou were finally obtained through calculation (Figure 6). The existing 870 urban park were utilized in conjunction with the connection and overlay coupling tool in ArcGIS 10.2 software to conduct a fitting test. Through a comparative analysis, it was determined that 93.7% of the existing urban parks in Guangzhou overlapped with the prediction results, thereby indicating that the prediction results are relatively reasonable.

The preliminary predicted site selection results (Figure 6) show that, among the 6825 predicted grids suitable for the site selection of Guangzhou urban parks, the central urban area, suburban areas, and remote suburbs account for 47.6%, 36.9%, and 15.5%, respectively. Among the existing urban parks in Guangzhou, 32.5% are located in the central urban area, showing a central concentration feature; 51.7% are distributed in the suburbs, showing a multi-core pattern; 15.8% are located in the far suburbs, showing an edge dispersion form. Among them, 93.7% of the existing urban parks fall within the predicted grids, and there are 5391 predicted grids suitable for building urban parks where no urban parks exist. This indicates that future urban parks can be built within these grids.

5.3. Refiltering According to the Current Layout of Urban Parks and Service Pressure

In the notice of the General Office of Guangzhou People’s Government on the issuance of the Guangzhou Green Space System Plan (2021–2035), it is emphasized that, by 2035, the city plans to build approximately 120 urban parks. The preliminary forecast of 6825 grids is considerably higher than the number of parks planned for future construction in Guangzhou. Consequently, it is imperative to re-screen the results of the preliminary prediction with regard to the intensity of near-term construction demand. This information can then be prioritized by the government in the decision-making process.

Based on the service pressure on urban parks in various regions of Guangzhou, we first screened out towns and streets with pressure values at levels 3–5. These areas failed to meet the daily needs of the people and have reached a saturated state. Similarly, the number and size of existing urban parks in a region are also important criteria for measuring the construction of urban parks. According to the nuclear density distribution of Guangzhou’s urban parks, the number of regional urban parks with service pressure levels of 3–5 in each township and street in Guangzhou exceeds 0.25. Therefore, the towns and streets with service pressure levels 3–5 and a number less than 0.25 were screened out. This screening produced a total of 103 towns and streets. With the help of ArcGIS 10.2 software, the grids predicted for site selection in 103 towns and streets were extracted, and the buffer tool was used to eliminate the grids of existing urban parks in Guangzhou according to location attributes. This operation resulted in 1537 grids suitable for priority planning, the specific distribution of which is shown in Figure 7. According to Figure 7, all 103 towns and streets in Guangzhou with high service pressure and no urban parks contain grids predicted to be suitable for building urban parks, indicating that the results are relatively reasonable. Based on the overall distribution, a large number of Guangzhou’s central urban areas have high service pressure. The existing urban parks are relatively concentrated, and the predicted number of grids is still large. Land resources are scarce in the central urban areas, but transportation is convenient. Consideration can be given to integrating small street green spaces and underground space construction to achieve the goal of saving land. In addition, we should make full use of the city’s existing public resources and build different types of urban parks in combination with riverside green belts, sports facilities, and botanical gardens in various regions. The economic development of suburban areas is lower than that of central urban areas. If constructed using the traditional site selection method, these streets are the most likely to be overlooked, so planning departments should pay special attention to them. The economy in the remote suburbs is underdeveloped, but the land prices are not high. They have good ecological environments and are rich in natural resources. In the later planning and construction stages, we can combine regional characteristics and advantages to build ecological urban parks and forest parks to meet the needs of people for leisure, exercise, and suburban travel.

6. Discussion

This study combines the current urban park layout in Guangzhou with the service pressure on urban parks in each region. The results reveal problems in urban park siting and layout in Guangzhou, ranging from the aspects of aggregation characteristics, distribution centers, residents’ needs, and service pressure, providing an important basis for re-assessing the priority planning of urban parks. The approach in this study can more scientifically satisfy residents’ demands for urban parks and green spaces while also pinpointing suitable sites for urban parks in each grid. At the same time, the site selection is accurate for each township and street, which is more helpful in promoting the balanced development of the region and optimizing the allocation of resources. Previous studies have focused on geographic information [39], social information [40], population distribution [41], and traffic conditions [42]. The data sources for geographic information primarily encompass terrain, landforms, hydrology, population distribution, traffic flow, socio-economic factors, and land use, among others. While these sources facilitate the intuitive display of the spatial relationship between parks and their surroundings, thereby aiding decision-makers in their choices, they fail to integrate residents’ needs, resulting in a lack of rationality in prediction outcomes. Furthermore, the accuracy of geographic information visualization interpretation relies, to a certain extent, on the professional knowledge and experience of analysts. In this study, by incorporating POI data related to residents’ interests, the service pressure on urban parks was effectively calculated in various regions, assessing whether there is an urgent need for urban park construction in a given area based on residents’ demands and habits. Machine learning algorithms, with their high degree of automation, can automatically learn from data and discover patterns, thereby addressing the issue of manual intervention caused by insufficient professional knowledge or experience. Social information primarily stems from social surveys, resident feedback, and statistical data. Being heavily influenced by residents’ subjective perceptions, such information may contain biases and misunderstandings, which can subsequently impact the site selection for urban parks. Additionally, collecting social information necessitates significant human, material, and time costs. In contrast, POI data in this study, which are based on residents’ interests, offer the advantage of convenient collection. Moreover, these data can be updated in real-time as social information, geographic information, and attribute information evolve, ensuring the accuracy and timeliness of the information. In addition to its high degree of automation, machine learning can effectively reduce the interference caused by subjective factors during the site selection process. As the volume of data increases and algorithms continue to be optimized, the predictive capability of machine learning will continue to improve, enabling it to uncover hidden patterns and trends within a dataset quickly. Population distribution data, primarily sourced from censuses and population sampling, encompasses population size, age structure, and gender, presenting a level of complexity. Meanwhile, traffic condition data, derived from traffic flow, road data, and congestion levels, can be limited when used solely to predict urban park locations, as they overlook other crucial site selection factors, such as the surrounding infrastructure. In this study, comprehensively gathered POI infrastructure data were categorized into four major and fifteen minor types in Guangzhou. By training the model and optimizing its parameters, a predictive model tailored to urban park site selection in Guangzhou was developed, enhancing the scientific rigor of the site selection process. This research seamlessly integrates POI data with machine learning, leveraging resident footprint points and various urban facilities to establish a decision tree model for urban park site selection. It delves into the influencing factors from all aspects, perspectives, and levels, specifically mapping these factors to each 500 × 500 m urban grid. This approach enables a more holistic understanding of the demand and urgency for urban park construction across different regions, allowing for targeted and rational park development that effectively addresses regional needs. Other similar studies have used more traditional methods, such as statistical analysis [43] and evaluation [44], for park site planning, which may achieve good results in specific situations but may be inadequate when dealing with large-scale and complex data [45]. At the same time, this kind of research also entails some subjectivity. The sites predicted in this study are more reasonable, objective, and scientific for the planning of urban parks in Guangzhou. In addition, to remain consistent with the Green Space System Planning policy of Guangzhou City, the predicted locations of urban parks were re-screened and evaluated to identify the townships and streets to be prioritized for planning, which provides a good reference for the government’s decision-making. Previous studies have been conducted more from the macro level, failing to effectively consider the targeted number of urban parks planned for future construction, with a large number of predicted points and a wide range [46]. The approach used in this study is able to quickly identify areas in urgent need of construction through secondary screening. Machine learning has been proven to facilitate the construction of smart cities and digital cities [47], and this study has an important reference value for the planning and layout of urban facilities such as elderly care, medical care, and sports.

Although this study involved innovatively using POI data and machine learning to predict and plan urban parks in Guangzhou, there are still some limitations. The POIs of urban parks and various facilities are the latest data, but the jurisdictions of various regions, towns, streets, and communities in Guangzhou originated in 2020. There may be a poor match between them, resulting in deviations in the calculated service pressure values of urban parks in various regions of Guangzhou. To ensure the accuracy of the results, the area of each region in 2020 was compared to that in the present, and the changes are small, so the deviation caused has little impact on the research results. Additionally, population demand and accessibility also affect the site selection and planning of urban parks. In this study, only the layout status and service pressure value were used for the secondary screening, and the predicted grids still outnumber the planned number. Therefore, more screening conditions should be considered in future studies to streamline the prioritization of planned sites and better facilitate their referencing by relevant departments. Finally, urban parks include comprehensive parks and specialized parks, among other types. This study only provides a reference for the scope of urban park site selection and planning and lacks a distinction between the site selection and planning of various types of parks. With the development of technology, channels should be broadened for further optimization and deepening.

7. Conclusions

In this study, POI data and the ID3 machine learning algorithm were used to predict sites for the selection and planning of Guangzhou urban parks. The results show the following: (1) The current layout of Guangzhou’s urban parks is significantly imbalanced, with a large number of blind spots in many areas that urgently need development and construction. The spatial characteristics of Guangzhou’s urban parks with search radii of 1000 m, 2000 m, and 3000 m were analyzed. The results show that as the search radius continues to expand, the distribution of urban parks becomes more concentrated, especially in the high-density area in the west. It decreases in a circle from the center to the outside, and this is more obvious along the southeast–northwest direction. The core area shows a clustering feature, while the marginal areas show a small, scattered pattern. The distribution of urban parks is quite different. (2) The service pressure in the first- and second-level areas is relatively low, and the number and distribution of parks constructed in these areas are fairly reasonable. Third-level areas cover the entire city, and future planning should prioritize construction in these areas as much as possible. The fourth- and fifth-level areas are relatively small in number and concentrated in central urban areas with developed economies, dense populations, and high activity densities, resulting in irrational resource allocation. (3) A preliminary forecast of 6825 suitable site locations for planning was made, and the degree of fit was tested using 870 existing urban parks, with a fitting degree of up to 93.7%. This prediction result is reasonable, indicating that the machine learning algorithm has high feasibility. After the secondary screening, 1537 sites suitable for priority planning were finally obtained, mainly distributed in suburban areas. (4) POI data have the advantages of rich content, a large amount of information, timeliness, and convenient collection, which are of great significance to the optimization and construction of urban spaces. This study represents an attempt to use the machine learning algorithm to select urban park sites based on data mining and quantification, which can optimize the overall layout, avoid subjective site selection, and make urban park site selection and planning more scientific and reasonable. This method is extensible and can serve as a reference for the planning and site selection of other facilities.