Multi-type and fine-grained urban green space function mapping based on BERT model and multi-source data fusion

ABSTRACT

Urban green space (UGS) is important to the urban ecological environment. It has physical characteristics and social function characteristics and plays an important role in urban climate change, sustainable development goals (SDG) and residents’ health. However, existing researches mostly focus on the extraction of UGS physical features, neglecting the importance of UGS social functions, resulting in the unresolved problem of multi-type and fine-grained functional mapping of UGS. Therefore, based on natural language processing (NLP) and multi-source data fusion, this paper proposes a multi-type and fine-grained UGS function mapping method. First, the social functional standards of UGS have been re-established, with a total of 19 categories. Second, the semantic information in the POI data name text is extracted using the deep learning model, and the reclassification of the UGS type of POI data is realized. Then, combined with multi-source data, 18 types of UGS are extracted. Finally, combining multi-source data to extract urban road green spaces (GS), a fine-grained UGS functional map of Shanghai is created. The results show that the overall accuracy rate of the method is 93.6%, and the Kappa coefficient is 0.93, which proves that the method has good performance in large-scale spatial UGS classification.

KEYWORDS:

• Urban green space
• deep learning
• natural language learning
• BERT
• crowd-sourced geographic data

1. Introduction

In recent years, the urban spatial structure has become increasingly complex (Du et al. 2021; Zhang 2020). Impermeable surfaces have replaced the original urban green space (UGS), which refers to urban land with natural and artificial vegetation, leading to a series of problems such as the deterioration of the urban ecological environment and sustainable development goals (SDG) 11,13 and 15 hard to accomplish (Lee and Maheswaran 2011; Shahtahmassebi et al. 2021; Staykova et al. 2021). Among them, fine-grained UGS can express the vegetation coverage information in the city at a micro level, which is helpful for urban refinement research. Large-scale UGS can express the vegetation coverage information of large-scale cities macroscopically, which is helpful for the overall study of UGS. At present, there are few studies on UGS considering both large city scale, fine-grained and multi-type. For example, Chen et al. (2019) used stepwise regression analysis to establish a PM2.5 estimation model based on morphological spatial pattern analysis (MSPA) to analysis the impact of large-scale green space (GS) morphological pattern categories on PM2.5. The above studies analysed the importance of UGS through large-scale remote sensing images, reflecting the impact of UGS on the city as a whole, and ignored the impact of fine-grained UGS on the interior of small-scale urban spaces. However, UGS in small-scale urban spaces has an important impact on the surrounding environment. For example, improving PM2.5, lowering the ambient temperature, and providing oxygen, so fine-grained UGS research is extremely important.

As an important part of urban space, UGS contains physical and social functional attributes (Nor and Abdullah 2019; Tigges, Lakes, and Hostert 2013). Physical attributes include morphological characteristics of GS (such as vegetation coverage type, vegetation height shape, etc.) and vegetation spectral characteristics (Tigges, Lakes, and Hostert 2013). Social function attributes indicate the use of UGS (such as parks and green belts, etc.). Due to the similar spectral properties of vegetation in remote sensing images, traditional remote sensing methods (such as support vector machines and random forests) for extracting vegetation from large-scale remote sensing images have matured (Xu et al. 2020). These methods focus on extracting the physical characteristics of vegetation, but still have limitations in judging the social function types of UGS. UGS function attributes contain a large amount of semantic information, and it has become a feasible method to judge the functional types of UGS through semantic information.

The fine division of urban green space types contributes to more precise research on urban spatial structure and local climate characteristics. Due to the different national conditions, the classification standards of UGS are different in different countries. For example, the U.S. UGS classification standard uses the accessibility and area of residents as the basis for functional classification (Long et al. 2022). Japan and China regard the relationship between GS functions and human activities as the basis for classification. In response to the above problems, it is urgent to establish a multi-type UGS classification system to meet the unified standards of fine classification and research of UGS (Zhou, Liao, and Wang 2022). The social function of UGS reflects the interdependence between human beings and GS, and encompasses a wealth of semantic information, it is necessary to extract semantic information to realize UGS classification.

In recent years, crowd-sourced geographic data has been widely used in urban research (Crooks et al. 2015; Ríos and Muñoz 2017; Tu et al. 2017). Compared with traditional remote sensing data, crowd-sourced geographic data can be obtained through personal smart devices, can accurately reflect human social activities and include information such as name, type, and location. The name text of point of interest (POI) data has rich contextual semantic information, which can effectively express the functional information of ground objects (Chen et al. 2018; Yuan et al. 2014). For example, the feature names of city squares usually contain semantic information related to the word ‘square’, and the feature names of sports and fitness parks usually contain semantic information related to ‘sports and fitness’ (Gugulica and Burghardt 2023). Therefore, how to mine and analysis the functional semantic information of fine-grained UGS in POI data name text is very important for this research.

Understanding semantic information in text has always been a key research direction in natural language processing. The traditional language model uncovers shallow semantic information in the text by estimating the likelihood of the text through predicting word probabilities, but it cannot establish contextual relations and has limitations in mining deep semantic information (Cohen, Ravikumar, and Fienberg 2003; Ye et al. 2021). To solve this problem, Mnih and Hinton (2008) proposed the log-bilinear language model (LBL) model, which omits the activation function in the neural network language model (NNLM) and directly transforms the model into a linear transformation, improves efficiency greatly. Mikolov et al. (2013) proposed word2vec, after double optimization of the model and training techniques, semantic relationships can be expressed through vectors. Recently, OpenAI developed generative pre-trained transformer 4 (GPT-4) (Bubeck et al. 2023) using unprecedented computing and data scale, which has general intelligence compared with previous models. In addition to mastering language, it can also solve tasks in fields such as mathematics, vision and medicine, and its performance is closer to humans. The above research provides a basis for mining the deep semantics of POI data names in this paper. However, POI data expresses the point coordinate information of ground objects, but cannot express the spatial range, while UGS data is continuous vegetation coverage data without spatial relationships, and only POI data cannot be used to divide the correct UGS spatial extent. Therefore, the scope and spatial relationship of UGS needs to be divided.

The boundaries and spatial relationships of urban green spaces are extremely important for refined research. Previous work focused on using OSM road data and regular grids to divide urban space. However, using these methods to divide urban green spaces will produce problems such as blurred boundaries and uncertain scales. In recent years, there have been more and more studies on urban functional areas, which have gradually solved the problems of fuzzy boundary accuracy and uncertain scale in small urban areas. This data can be used to accurately divide urban green space boundaries and correctly express spatial relationships. Urban functional zones (UFZs) serve as the fundamental building blocks of urban planning and management, significantly influencing the analysis of urban spatial structure and providing insights into the physical and social attributes of a city (Zhang 2020; Zhang and Du 2015). For example, Zhang, Du, and Wang (2017) employed very-high-resolution (VHR) satellite imagery and POI (Point of Interest) data to successfully classify UFZs in Beijing. Du et al. (2020) used latent Dirichlet allocation (LDA) and support vector machine (SVM) to classify the segmented UFZ by combining the physical characteristics of remote sensing images. The above methods realize the extraction of large-scale urban functional areas, while ignoring the extraction of finer UGS. For this, Cao et al. (2021) proposed a method for fusing urban park mapping using UFZ and crowd-sourced geographic data. However, this method only divides urban park functions into 12 types and is unable to capture detailed information about the fine-grained GSs within the parks. As a result, the fine-grained research of UGS is limited. The studies mentioned above highlight that UGS data is characterized by a fine-grained, continuous patch-like distribution, which lacks proper spatial relationship division. Taking this into account, our current study aims to address this issue by integrating UGS data with UFZ data to precisely delineate the scope and spatial relationships of UGS.

In summary, the existing research on UGS faces four main limitations. Firstly, there is a lack of a standardized classification system for UGS worldwide, which hinders comprehensive research and comparative analysis of UGS across different countries under a unified framework. Secondly, traditional remote sensing methods are limited in their ability to identify UGS social function information, resulting in constraints on the accurate classification of UGS based on their functions. Thirdly, current research heavily relies on geospatial information from POI data, overlooking the valuable social function information present in the text. Thus, the extraction of semantic information becomes crucial for achieving a multifunctional classification of UGS. Finally, UGS data is irregularly and continuously distributed, lacking proper spatial division to effectively represent spatial information. Therefore, incorporating data with precise spatial information is necessary to adequately partition UGS data.

This study has three main contributions:

1. Currently, there is no functional classification system for urban green space that can be used globally. This study is based on the correlation between UGS functions and human activities, a comprehensive social function type system for UGS was developed. Consequently, the functional zones of UGS were re-categorized into 19 distinct types, establishing a standardized and unified global framework for UGS classification research.

2. This study proposes a novel method for UGS functional mapping that integrates deep learning NLP and multi-source remote sensing data fusion. The method effectively identifies semantic information in POI data and accurately extracts various functional types of UGS. Notably, this approach represents the first large city scale and fine-grained UGS classification method that incorporates deep learning NLP techniques, distinguishing it from previous research efforts. The results of this study outperform those of existing methods, demonstrating promising potential for practical applications.

3. UGS data is in a continuous, dense and irregular form, making it difficult to accurately divide regions. This study uses a multi-source data fusion method to accurately segment UGS into independent block units through high-precision UFZ data to ensure the accuracy of the subsequent classification process.

2. Study area and data

2.1. Study area

This study focuses on Shanghai (Figure 1), a super-large city in China with a population of 24.89 million. It is situated in the middle of China's eastern coastline, at the easternmost part of the Yangtze River Delta, bordered by the East China Sea to the east, and Zhejiang and Jiangsu provinces to the west. Shanghai has diverse GSs with abundant vegetation types, resulting in dense coverage. The city comprises 16 municipal districts, and for this study, 15 districts (excluding Chongming District) were chosen as the research area, covering a total of 5,293 km2, which meets the criteria for investigating large city scale cities with complex spatial structures.

Figure 1. The study area of Shanghai, China.

2.2. Data

2.2.1. UFZ data

To aid urban planning and management, the Dolab Laboratory of Peking University has developed Urbanscape Essential Dataset (PKU-USED) (Figure 2). This dataset classifies urban space into 12 distinct types of UFZ based on their social attributes. Each pixel in the dataset corresponds to a resolution of 2.4 meters. Among them, residential-1 is low-rise residences, residential-2 is high-rise residences and residential-3 is shantytowns.

Figure 2. UFZ data for Shanghai.

2.2.2. UGS data

Recently, Shi et al. (Shi et al. 2023) proposed a deep learning framework for UGS mapping, and generated fine-grained UGS maps of 31 major cities in mainland China with a resolution of 1 meter (Figure 3).

Figure 3. UGS Data of Shanghai.

2.2.3. POI data

This study used the application programming interface (API) to crawl the 132,110 POI data of Beijing, Guangzhou, Shenzhen, Hangzhou and Shanghai in 2023 from the Gaode map (Figure 4), of which 83,361 POI data are used as the training set for Bidirectional Encoder Representations from Transformers (BERT) model UGS classification, and 48,749 Shanghai POI data are used for prediction and classification.

Figure 4. Points of interest (POI) data of Shanghai.

2.2.4. OSM road network

OpenStreetMap (OSM) road data is open geographic data that allows users to edit and add data. The OSM road data contains a total of 27 road types. To ensure data quality, this study retains expressways, arterial roads, primary roads, secondary roads, and tertiary roads, and deletes other road types (Figure 5).

Figure 5. OSM road data of Shanghai.

3. Method

This study proposes a UGS function mapping method (Figure 6) that involves several key steps. Firstly, the UGS classification system is redefined by referring to the classification standards of UGS in the United States, Japan, and China. As a result, UGS is reclassified into 19 distinct types. Secondly, the BERT model is utilized to pre-train over 80,000 POI data name texts in Beijing, Guangzhou, Shenzhen, and Hangzhou. Subsequently, more than 40,000 POI data name texts in Shanghai are predicted and reclassified into the 18 UGS types. Next, 18 fine-grained UGS types are extracted by integrating the reclassified POI data, UFZ data, and UGS data. The study also leverages OSM road data to identify and extract urban road GSs, completing the mapping of the 19 UGS functional zones. Finally, the distribution of UGS in Shanghai is analysed, and a precision evaluation is conducted to assess the accuracy and effectiveness of the proposed method.

Figure 6. Workflow of the proposed method.

3.1. UGS function type standard

Currently, there is a lack of a standardized functional classification system for UGS globally, which could impede future research on UGS refinement. For instance, the United States divides UGS into seven categories based on factors like GS area, service population, and service radius. While this method considers the impact of human activities on UGS functions, it overlooks the relationship between UFZ and UGS and fails to fully recognize the functional attributes of UGS. As a result, the classification standards are limited in scope and definition, which may not meet the requirements for future detailed UGS research. Japan, on the other hand, classifies UGS into three levels and 16 types based on social value and function. However, this standard only considers the functionality of UGS and neglects aspects like service scope and residents’ accessibility. Additionally, it fails to consider the relationship between urban regional functions and GS, resulting in a lack of comprehensive criteria. China divides UGS into five types according to the functional attributes of the area where UGS is located, the classification basis mainly comes from GB/T 51346-2019 Urban Green Space Planning Standards (the recommended national standards of China). However, this classification standard suffers from vague function definitions and the problem of overlapping functions among multiple types of UGS (Table 1). To address these challenges, this study draws on the strengths of GS function standards in various countries and uses UFZ attributes and human activity characteristics as classification criteria (Figure 7). This approach leads to the proposal of a more scientifically grounded UGS function type standard system (Table 2), which includes 19 types. This standardized system provides a solid foundation for future research on the refinement of UGS.

Figure 7. Remote sensing images of UGS of various functional types (a–s).

Table 1. Standards for UGS classification in China, Japan and USA.

China			Japan			America
Level 1	Level 2	Level 3	Level 1	Level 2	Level 3	Class	Area	Serving	Service radius
Park GS	Comprehensive park		Public GS	Public GS	Park green land, playground	Children’s park	200 ∼ 400 m^2	500 ∼ 2500	Neighborhood
	Community park					Small pleasance	200 ∼ 400 m^2	500 ∼ 2500	Neighborhood
	Special Park	zoo				Neighboring park	2 ∼ 8 m^2	2000 ∼ 10000	400 ∼ 800 m
		Botanical garden		Natural GS	River, lake	District park	8 ∼ 40 m^2	10000 ∼ 50000	800 ∼ 5000 m
		Historic garden			Seashore, riverside, lakeside	Large urban park	≥ 0.4km^2	≥ 50000	Riding distance within half an hour (by car)
		Amusement park			Mountain forestry	Regional park	≥1km^2	Serving a larger region	Riding distance within an hour (by car)
		Other special parks		Open GS	graveyard	Specific facility	Including avenues, seashore, square etc.
Protected GS
Square GS					Garden plot of individual facility
Attached GS	Residential land		Private GS	Sharing GS
	Land for public service facilities				Sharing recreation facility
	Commercial service facilities
	Industrial land				School playground
	Logistics and warehousing land			Specific GS	Individual garden plot
	road traffic facilities				Testing land of nursery
	Land for public facilities
Regional GS	Scenic recreation GS	Scenic GS
		Forest park
		Wetland park
		Camping park
	Ecological conservation
	Regional facilities to protect
	Productive GS

Table 2. UGS function classification standard.

Class	Definition
Comprehensive Park GS	This GS provides abundant recreational features, corresponding facilities, and encompasses a large land area, making it suitable for a wide range of public outdoor activities
Zoo GS	This GS serves as a place where wild animals are exhibited for public educational purposes or study
Botanical GS	This GS functions as a scientific research unit that investigates, collects, identifies, and preserves plants
Historical Park GS	A park of great historical value
Amusement Park GS	A park with plenty of recreational facilities
Forest Park GS	This GS is a park designed around a significant expanse of artificial or natural forest, serving as its primary attraction
Wetland Park GS	A park of predominantly wetlands space
Sports And Fitness Park GS	This GS is a unique park featuring a variety of sports and fitness facilities tailored for competitions, training
Green Buffer GS	GS with health, isolation, safety and environmental protection functions
Square GS	To play, memorial, gathering and other functions of the city's public activity venue
Residential GS	GS within residential land
Commercial GS	GS within commercial service facilities
Industrial GS	GS within industrial land
Recreational GS	GS in a football field, stadium, golf course
Logistics Warehousing GS	GS in logistics warehousing land
Educational and Institution GS	Inside the school and institution
Tourist Attraction GS	GS in scenic spot
Ecological Conservation GS	For urban and rural afforestation beautification production, cultivation of nurseries, flower beds, grass beds and other GS
Road GS	Urban road surrounding green belt

3.2. POI data reclassification

BERT is a deep learning NLP model. The BERT model structure has two versions, BASE and LARGE. This study uses the BASE version.

In order to dig deeper semantic relationships between word vectors, which uses a bidirectional Transformer structure (Figure 8) and a Self-attention mechanism to improve the general language model to Mask-language model, uses the Next Sentence Prediction task to learn sentence-level information. (1) $\mathrm{BER}{\mathrm{T}}_{\mathrm{BASE}}:L=12,H=768,A=12,\mathrm{Total}\mathrm{Parameters}=110M$(1) (2) $\mathrm{BER}{\mathrm{T}}_{\mathrm{LARGE}}:L=24,H=1024,A=16,\mathrm{Total}\mathrm{Parameters}=340M$(2) Where $L$ indicates the number of layers, $H$ indicates the dimension of the output, $A$ indicates the number of multi-head attention, and $\mathrm{Total}\mathrm{Parameters}$ indicates the number of parameters.

Figure 8. BERT model flow chart.

3.2.1. Encoder and decoder stacks

The encoder in this context is composed of six identical layers (N = 6). Each layer comprises two sublayers: the first sublayer is a multi-head self-attention mechanism, and the second sublayer is a feedforward network. To enable residual connections, which aid in the flow of information, the encoder uses layer normalization around each pair of sublayers. Additionally, to facilitate these residual connections, all sublayers within the model, including the embedding layers, produce output dimensions of $d_{\mathrm{model}}$ = 512.

The decoder also consists of six identical layers (N = 6), similar to the encoder. In each decoder layer, besides the two sublayers present in the encoder layer, a third sublayer is introduced. This additional sublayer performs multi-head attention on the output of the encoder stack, allowing the decoder to access information from the encoder's output during the decoding process. As with the encoder, residual connections are used around each sublayer in the decoder, followed by layer normalization to facilitate smooth information flow.

3.2.2. Scaled dot-product attention

Attention is the process of calculating the semantic weight of word vectors in sentences (Figure 9). First, Self-Attention will calculate the three vectors of Query, Key and Value. These three vectors are the result of multiplying the embedding vector with a matrix. Second, calculate the point product of Query and Key, divide the result by$\sqrt{d_{\mathrm{model}}}$, and perform softmax calculate the attention weight, then multiply the value by the softmax value, and finally get the Self-Attention value of the word. The value indicates the degree of correlation between a word and other words in the sentence. The larger the value, the greater the degree of correlation. (3) $\mathrm{Attention}(Q,K,V)=\mathrm{softmax}\times \left({\displaystyle \frac{\mathrm{Q}{\mathrm{K}}^T}{\sqrt{d_k}}}\right)V$(3) where $Q,K,V$ represent three vectors with a length of 64, $Q$ represents the information being sought in the input sequence, $K$ represents the element information in the input sequence, $V$ represents the information associated with the elements in the input sequence, $\mathrm{Q}{\mathrm{K}}^T$ represents the score score, $\sqrt{d_k}$ represents the square root of the dimension of the key, $\mathrm{softmax}$ represents the normalization function, and $\mathrm{Attention}$ represents the attention score.

Figure 9. Flow chart of self-attention mechanism.

3.2.3. Multi-head attention

In order to improve computational efficiency and word vector correlation, BERT uses the multi-head attention mechanism (Figure 10). The training structure is more efficient and accurate than previous CNN and RNN models. Self-Attention needs to calculate the attention of each word and all words. Between each word maximum distance is all 1, capturing long-distance dependencies. First, Transformer's multi-headed attention initializes 8 sets of Query, Key, and Value matrices for self-Attention calculation, and passes the information captured by each head to the next layer of the network, so that the model can learn more internal word dependencies. (4) $\mathrm{MultiHead}(Q,K,V)=\mathrm{Concat}(\mathrm{hea}{\mathrm{d}}_0,\dots ,\mathrm{hea}{\mathrm{d}}_h)W^O$(4) Where $\mathrm{MultiHead}(Q,K,V)$ represents the multi-head self-attention mechanism, $Q$ represents the information being sought in the input sequence, $K$ represents the element information in the input sequence, $V$ represents the information associated with the elements in the input sequence, $\mathrm{Concat}(\mathrm{hea}{\mathrm{d}}_0,\dots ,\mathrm{hea}{\mathrm{d}}_h)$ represents multiple self-attention mechanisms, $\mathrm{hea}{\mathrm{d}}_0,\dots ,\mathrm{hea}{\mathrm{d}}_h$ represents the output of these individual self-attention heads, $W^O$ represents the weight matrix. (5) $\mathrm{hea}{\mathrm{d}}_i=\mathrm{Attention}({\mathrm{QW}}_i^Q,{\mathrm{KW}}_i^K,{\mathrm{VW}}_i^V)$(5) Where $\mathrm{hea}{\mathrm{d}}_i$ represents the output of the $i$-th attention result calculated separately by this head, $\mathrm{Attention}({\mathrm{QW}}_i^Q,{\mathrm{KW}}_i^K,{\mathrm{VW}}_i^V)$ represents the calculation of each attention head, which will input $Q$, $K$ and $V$ are multiplied by the weight matrices ${\mathrm{QW}}_i^Q$, ${\mathrm{KW}}_i^K$ and ${\mathrm{VW}}_i^V$ respectively associated with the head, and then self-attention calculation is performed.

Figure 10. Multi-headed attention mechanism.

3.2.4. Sample collection and model training

In this study, a sample set comprising 83,361 POI name texts from Beijing, Guangzhou, Shenzhen, and Hangzhou was created based on the UGS function type standard. The results revealed that the training accuracy reached 93.15%, and the Kappa coefficient was 0.92, indicating that the method performed well in reclassifying POI data. Subsequently, predict the name texts of 48,749 POI data entries in Shanghai, enabling the successful reclassification of POI data in Shanghai (as shown in Figure 11).

Figure 11. POI data reclassification result.

3.3. Multi-source data fusion

Indeed, POI data alone cannot adequately represent the spatial extent of UGS. To address this limitation, this study integrated UFZ data, OSM data, and UGS data, and performed attribute linkage and overlay analysis of the multi-source data to accurately delineate the spatial boundaries of the UGS. This allows for the correct spatial range division of UGS. Next, the reclassified POI data, with their social function attributes, is mapped to the corresponding spatial units, thus creating a fine-grained representation of different GSs in the city based on their social functions.

3.3.1. Attribute mapping

First segment the urban green space data (Figure 12). POI data is point-like information, which cannot effectively represent planar spatial information. The data of UFZ is planar information, which represents the coverage of functional zones and cannot express the fine-grained UGS information within them. UGS data is distributed in fine-grained patches, which can express fine-grained GS information but lacks correct coverage and spatial relationships. In view of this, this study uses the attribute mapping method to integrate POI data, UFZ data and UGS data to extract various functional GSs. The attribute mapping process counts the number of POI data of each type in the functional zone unit, and then determines the type of UGS in the zone unit. In view of the situation that there is no POI data or a large number of POI data of different types in the functional zone unit, this experiment uses the attribute of the functional zone unit to assist the judgment, as shown in Figure 13. For example, define the functional zone unit as the POI type with the largest number of attributes, and refer to the functional attributes of the functional area unit, if there is the same type of industrial GS POI data in the industrial zone, the industrial GS POI data will be given a higher weight. This method reduces the impact of too many POI types, and ensures the accuracy of extracting UGS types.

Figure 12. UGS segmentation process.

Figure 13. The workflow of attribute mapping.

3.3.2. Road GS extraction

Urban road GSs are distributed in the middle and on both sides of urban roads. Since POI data is point information, urban road information cannot be located, and the above method cannot be used to extract road GSs. OSM road data is linear data, which can accurately express urban road information. This study uses OSM road data to establish a buffer zone and integrate UGS data to extract urban road GS data (Figure 14). According to China's CJJ37-2012 Urban Road Engineering Design Code, the design width of urban main roads is the spatial distance from the road center line to the building setback line. The width is approximately 18 meters. This width will vary due to different road conditions, so the buffer zone width is set to 18 meters ensures the accuracy of green belt extraction.

Figure 14. Results of road GS extraction.

3.4. UGS analysis

The proportion of urban green area serves as a vital evaluation index for assessing the sustainable development of the urban ecological environment. It provides insight into the overall level of urban greening, with higher proportions indicating a better urban ecological environment. Moreover, the proportion of UGS area for each specific function represents a crucial evaluation index in the analysis of refined GSs. It reveals the percentage of UGS for each functional type relative to the total area, shedding light on the urban environment's characteristics and the structural diversity of UGS. In this study, the area and proportion of UGS for various functional types in Shanghai are calculated, enabling an analysis of the distribution and structural characteristics of UGS throughout the city. (6) $\mathrm{UGSP}={\displaystyle \frac{\mathrm{UGA}}{\mathrm{TA}}\times 100\text{\%}}$(6) where $\mathrm{UGSP}$ represents the proportion of UGS, $\mathrm{UGA}$ represents the area of UGS, and $\mathrm{TA}$ represents the area of urban built-up area. (7) $\mathrm{USF}{\mathrm{P}}_i={\displaystyle \frac{\mathrm{US}{\mathrm{F}}_i}{\mathrm{UGA}}\times 100\text{\%}}$(7) where $\mathrm{USF}{\mathrm{P}}_i$ means the proportion of UGS function type $i$, $\mathrm{USF}{\mathrm{P}}_i$means UGS type $i$, and $\mathrm{UGA}$ means UGS area.

4. Experimental result and analysis

4.1. UGS classification result

Based on the experimental results, a map showcasing the functional types of UGS in Shanghai has been produced, as depicted in Figure 15. Further details of the functional classification of UGS in Shanghai are illustrated in Figure 16. The comprehensive parks in Shanghai are numerous, widely distributed, and cover significant zone unit areas, indicating a high degree of urban greening. However, due to Shanghai's relatively high level of urbanization, there are few virgin forests and water systems, resulting in relatively smaller GSs for forest parks and wetland parks. Additionally, the smaller areas of other park types make their distribution less prominent in the figure. In the built-up area of downtown Shanghai, there is a large number of residential zones and commercial zones, with the associated residential GSs and commercial GSs being the predominant types of GSs in central Shanghai. On the other hand, the surrounding areas exhibit a relatively lower degree of urbanization, with an abundance of industrial zones, plantations, and farmlands, accompanied by vast areas of industrial GSs, leisure GSs, and ecological conservation GSs. Due to the concentration of numerous colleges and universities, educational and institutional GSs have a wide distribution, albeit with relatively smaller areas in each zone unit. This map serves as a valuable scientific reference for future research on refined UGS and as a basis for informed government decision-making.

Figure 15. Map of functional types of UGS in Shanghai.

Figure 16. Local details of Shanghai UGS function classification (A-H) and comparison before and after classification (I–L).

4.2. Evaluation results

In this study, a total of 280 zones from various types of Gaode maps were randomly selected for analysis. The classification results were evaluated by comparing them with the visual interpretation of the Gaode map, and the accuracy of the classification was established using the confusion matrix. The findings indicate that the overall classification accuracy was 93.6%, with a Kappa coefficient of 0.93, demonstrating that this method achieves excellent results in UGS classification. The heat map of the confusion matrix (Figure 17) reveals that the UGS extraction results from data fusion were impacted by misclassifications of residential GS and commercial GS in POI data. However, the attribute mapping process successfully mitigated this effect, resulting in overall good classification results. In response to the problem of a small number of samples of certain types of UGS, since this study uses multiple city samples for centralized training, the number of samples of each type of UGS meets the training requirements and ensures the accuracy of training. The types of a small number of UGS have been determined, but there are a large number of UGS that do not have clear types. During the evaluation, a small number of UGS of different types conflict with the determined types, so corrections will be made based on the determined types.

Figure 17. UGS function classification confusion matrix heat map (CP, Comprehensive Park GS; ZP, Zoo GS; BP, Botanical GS; HP, Historical Park GS; AP, Amusement Park GS; FP, Forest Park GS; WP, Wetland Park GS; SFP, Sports and Fitness Park GS; GB, Green Buffer; SG, Square GS; RG, Residential GS; CG, Commercial GS; IG, Industrial GS; REG, Recreational GS; LWG, Logistics Warehousing GS; EG, Educational and Institution GS; TAG, Tourist Attraction GS; ECG, Ecological Conservation GS; RDG, Road GS).

4.3. Shanghai UGS classification result analysis

The area of UGS in Shanghai is 1299.4 km². The proportion of UGS calculated in Table 3 is 24.5%, which shows that Shanghai's urban greening and overall ecological environment are relatively good. Figure 18 shows the ratio of 19 kinds of urban green areas in Shanghai, among which the proportion of residential green areas is the highest, which is 21.97%, indicating that the living environment of Shanghai residents is relatively good. There are a large number of comprehensive parks in Shanghai, with a total area of 18.91%. Comprehensive parks contribute to the sustainability of the urban microclimate and the freedom of travel for tourists. The spatial structure of Shanghai is complex, the urban road network is dense, and the urban road GS is 13.14%. Good urban road greening ensures the sustainable development of Shanghai's urban ecological environment. Due to the high level of urbanization in Shanghai, the area of virgin forests and wetlands is relatively small, resulting in the lowest proportion of forest park GS and wetland park GS.

Figure 18. The proportion of 19 kinds of UGS in Shanghai.

Table 3. The urban green area ratio in Shanghai.

The Urban Green Area Ratio
Total urban GS area (km^2)	1299
The total study area (km^2)	5293
The proportion of urban green area (%)	24.5

5. Discussion

5.1. Comparison of text classification models

The names of POI data are typically short in length. In the process of natural language classification, unlike longer sentences, short texts present a challenge due to their limited word count, making it challenging to extract their semantic information. Semantic classification of short texts is more demanding compared to longer ones. Although the BERT model excels in classifying longer texts, there is limited research verifying its performance in short text classification. Table 4 provides a comparative analysis of the ERNIE model, known for its excellence in short text classification, to assess BERT's performance in classifying POI data within the context of short texts. This study employed the same training dataset to train both the BERT and ERNIE models in Google Cloud Platform Colab. The results indicate that the BERT model achieved a training accuracy of 93.2% in just 21 min, surpassing the ERNIE model, which achieved a training accuracy of 92.7% in 24 min. These findings underscore the BERT model's superiority in terms of accuracy and efficiency, reaffirming its capability for short text classification.

Table 4. Comparison of POI reclassification accuracy between the BERT model and ERNIE model.

Class	BERT	ERNIE
Comprehensive Park GS	0.922	0.910
Zoo GS	0.886	0.890
Botanical GS	0.826	0.818
Historical Park GS	0.785	0.795
Amusement Park GS	0.901	0.889
Forest Park GS	0.611	0.619
Wetland Park GS	0.886	0.864
Sports And Fitness Park GS	0.733	0.732
Green Buffer	0.986	0.984
Square GS	0.700	0.689
Residential GS	0.939	0.936
Commercial GS	0.879	0.877
Industrial GS	0.969	0.965
Recreational GS	0.915	0.899
Logistics Warehousing GS	0.982	0.979
Educational and Institution GS	0.989	0.988
Tourist Attraction GS	0.858	0.848
Ecological Conservation GS	0.910	0.906
Overall Accuracy	0.932	0.927

5.2. Uncertainties regarding the UGS mapping

5.2.1. Impact of POI classification

The classification accuracy of UGS is largely affected by the reclassification of POI data. In order to verify the accuracy of POI data reclassification, a confusion matrix is established to calculate the classification accuracy. The result shows that the overall accuracy is 93.4%, and the Kappa coefficient is 0.92. The result of POI data reclassification is excellent. According to the confusion matrix heat map (Figure 19), the types with large classification errors are mainly residential GSs and commercial GSs. This is because the names of some residential zones and commercial buildings are relatively close, and there will be misclassification during the training process, and training texts of its are much more than other types, so the number of misclassifications is more than other types.

Figure 19. POI name text reclassification confusion matrix heat map.

5.2.2. The impact of data accuracy

During the creation of UFZ data, the risk of over-fitting or incomplete segmentation exists, which can subsequently impact the accuracy of data division. Therefore, directly using the original UFZ data might lead to inaccuracies in UGS classification. In order to solve this problem, use OSM data to re-divide the UFZ data, eliminate blocks that are not completely separated by UFZ data, and then eliminate blocks that are too small to be of practical significance (Figure 20).

Figure 20. Data preprocessing of UFZ.

6. Conclusions and future work

In recent years, the reduction of UGS has had adverse effects on the urban ecological environment and residents’ well-being. UGS are essential elements of UFZs with diverse functions. Therefore, there's a pressing need for multi-type and detailed mapping of UGS to support in-depth research. To address this, our study proposes a novel method framework that leverages semantic information from POI data for accurate and fine-grained mapping of UGS with multiple functionalities. The results demonstrate high performance, with an overall accuracy of 93.6% and a Kappa coefficient of 0.93, affirming the effectiveness of our approach in UGS classification. This research offers valuable insights for future investigations into refined UGS and contributes to the sustainable urban development planning process.

The contributions are as follows:

1. In this study, a comprehensive set of scientific UGS function type standards is proposed, resulting in the classification of 19 distinct types of UGS function types.

2. The study employs the deep learning natural language processing BERT model, in combination with multi-source data fusion, to extract fine-grained UGS from large city scale maps.

3. The proposed method for reclassifying POI data is not affected by complex spatial relationships, ensuring robustness in identifying the type of UGS. Additionally, the attribute mapping correction further enhances the accuracy of UGS type identification, making the classification process more reliable and precise.

Future research should consider using UFZ data with higher precision to reduce the impact of segmentation accuracy from UGS data. Second, the semantic information of other data should be considered to improve the accuracy of UGS classification. Third, when reclassifying POI data in the future, more types of data need to be added to reduce the impact of misclassification. Finally, factors such as the utilization rate, accessibility, and average area of various UGSs should be considered.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are openly available in ‘Zenodo’ at https://zenodo.org/record/8211083.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China [grant number 42201512], the National Natural Science Foundation of China [grant number 41930650] and the National Natural Science Foundation of China [grant number 42371412], the China Postdoctoral Science Foundation [grant number 2021M703511].