Urban green land functional area automatic division method based on combination of deep learning and crowdsourcing data

Abstract

The invention relates to an urban green space functional area automatic division method based on deep learning and crowdsourcing geographic space data. According to the method, an innovative solution is provided for the problems of efficiency and accuracy in urban functional area demarcation. The method comprises the following steps: firstly, integrating remote sensing image data and crowdsourcing geographic space data, including point of interest (POI) data, open street map road network (OpenStreet Map, OSM), Baidu map boundary and the like, and carrying out accurate mapping on social functional areas of urban green space (Urban Green Space, UGS); then, a deep learning algorithm, especially a U-Net model, is adopted to process the remote sensing image so as to extract vegetation distribution information; and then, analyzing POI data by utilizing a natural language processing technology, establishing a semantic relationship between the UGS social function and the POI, and further realizing automatic classification of the social function area. Besides, through accuracy evaluation, high efficiency and accuracy of UGS social function area division are ensured, and powerful data support is provided for urban planning, sustainable development, government decision making and urban geographical research. The application of the method significantly improves the automation level of urban green space functional area division, and provides an important technical means for rapid monitoring, effective management and scientific planning of the urban green space.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of urban green space management, and particularly to an urban green space functional area automatic division method combining deep learning and crowdsourcing data. BACKGROUND

[0002] Urban green space (UGS) constitutes the core of urban ecological space, covering natural and artificial vegetation, and plays a crucial role in ecological balance and sustainable development of the city. Existing researches focus on the physical properties of UGS, such as vegetation types and structure, which reveal the key functions of UGS in mitigating urban heat island effect, regulating climate, purifying air, protecting biodiversity, and hydrological regulation. In addition, the social function of UGS cannot be ignored, as it serves as a link between urban residents and nature, significantly influencing the physical and mental health and happiness of residents.

[0003] Although the physical characteristics of UGS have been extensively studied, data on its social functional areas are relatively scarce. This limits the demand for high-resolution spatial information on UGS social functional areas, which is crucial for urban planning, sustainable development, policy-making, and urban geography research. Therefore, there is an urgent need to develop an efficient and accurate UGS social functional area mapping method.

[0004] Traditional methods usually obtain the physical characteristics of UGS based on medium to high resolution remote sensing images, which are widely used in vegetation type studies. However, the acquisition of UGS social functional areas is complicated due to the limitations of the method, as social functional areas are related to both physical characteristics and human activities, leading to a "semantic gap" between remote sensing images and UGS social functions.

[0005] In addition, the resolution and accuracy of remote sensing images are limited by sensor performance and image processing algorithms. Crowdsourcing geospatial data, such as OSM and POI, although providing rich socio-economic activity information, requires further correction and integration due to inconsistent data quality.

[0006] The present application aims to fill the gap in the prior art and proposes an urban green space functional area automatic division method and system integrating deep learning and crowdsourcing data. This method integrates remote sensing image data, crowdsourcing geospatial data, and applies deep learning algorithms to achieve rapid and accurate division of UGS social functional areas, providing scientific basis for urban planning and green space management. SUMMARY

[0007] In view of the deficiencies of the automatic division of UGS social function zones in the prior art, an innovative method framework is developed to avoid the misjudgment and omission problems in the traditional method and to affect the correct classification of the urban function zones. A kind of urban green space function zone automatic division method based on deep learning and crowd-sourced geographic spatial data:

[0008] The application provides a kind of urban green space function zone automatic division method based on deep learning and crowd-sourced geographic spatial data, the method comprises:

[0009] Using Baidu map platform, OpenStreeetMap platform obtains research data, creates research area data set through Arcgis10.8 software.

[0010] Radiation calibration and atmospheric correction are carried out on the remote sensing image; POI data is cleaned and classified; OSM road network data is processed to remove hanging lines and redundant lines; the above data format is converted and the coordinate system is unified to ensure the consistency and accuracy of the data.

[0011] Using U-Net model, land cover classification is carried out in combination with remote sensing image data, and vegetation distribution

[0012] The processed OSM road network is graded, and a buffer zone is created. The extracted vegetation distribution map is superimposed with the buffer zone based on the OSM road network to obtain the green space function zone of the protective green space.

[0013] The extracted vegetation distribution map is superimposed with the Baidu map boundary and road network buffer zone, the UGS social function zone index is calculated, and the square green space, park green space and accessory green space are divided by constructing a lookup table, wherein the UGS social function zone index has four items, namely: area index, shape index, width index and vegetation proportion index.

[0014] In one embodiment, the research data obtained using the Baidu map platform and the OpenStreeetMap platform is created by Arcgis10.8 software to create a research area data set, which includes:

[0015] CBERS-04A high-resolution remote sensing image data is obtained, OpenStreetMap (OSM) road network and Baidu map boundary data, and point of interest (POI) data are collected, and the cutting tool in Arcgis10.8 is used as input element, and the vector data of the research area is used as cutting element, to cut the above data in the research area and create corresponding element data set.

[0016] In one of the embodiments, the remote sensing images are radiometrically calibrated and atmospherically corrected; the POI data is cleaned and classified; the OSM road network data is processed to remove hanging lines and redundant lines; the above data is converted in format and the coordinate system is unified to ensure the consistency and accuracy of the data, including:

[0017] First, the CBERS-04A remote sensing image data is radiometrically calibrated to correct sensor bias and atmospherically corrected to eliminate the influence of atmospheric conditions. Second, according to the Baidu Map POI industry classification standard, the attribute selection tool in Arcgis10.8 is used to classify the points of interest (POI) data, and the hanging lines and redundant lines in the OpenStreetMap (OSM) road network data are removed by combining topological analysis and manual assistance to ensure the spatial consistency of the road network data. Finally, all data is converted to a unified format, and all data sets use the CGCS2000 coordinate system to achieve seamless integration and analysis between different data sources.

[0018] In one of the embodiments, the U-Net model is used in combination with remote sensing image data for land cover classification and vegetation distribution extraction, including:

[0019] First, the U-Net model based on convolutional neural network (CNN) is used for land cover classification. This model effectively extracts deep semantic information of images and identifies spectral and spatial features of different scales through its encoder-decoder architecture and skip connection. The U-Net model is trained using the deep learning module in ArcGIS Pro 3.1 software, which provides a user-friendly interface and powerful algorithm support. In this study, the backbone network of the U-Net model is modified to ResNet-101, which uses its residual structure to enhance feature extraction capability and uses ImageNet pre-trained weights through transfer learning to speed up model convergence. In model training, the original image slice size is set to 512x512, the batch size is set to 4, and 5 convolutional layers and early stopping techniques are used for 25 iterations to optimize the training process and prevent overfitting. In addition, remote sensing indexes such as normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and soil-adjusted vegetation index (SAVI) are introduced to enhance the extraction accuracy of vegetation features. The entire process is completed on a Windows 10 operating system equipped with an NVIDIA GeForce RTX 2070 GPU, ensuring the computational efficiency of the training and the generalization ability of the model.

[0020] In one of the embodiments, the processed OSM road network is classified and buffer zones are created. The extracted vegetation distribution map is overlaid with the buffer zones based on the OSM road network to obtain the green space function area of the protective green space, including:

[0021] The processed OSM road network data was hierarchically classified and buffered, with the following settings: Using the `fclass` field describing the OSM road network type, roads were reclassified into expressways, arterial roads, secondary arterial roads, and local roads according to the standards of the Ministry of Housing and Urban-Rural Development of China. Simultaneously, based on actual road conditions, roads were reclassified into primary, secondary, and tertiary levels. Using a buffer tool, buffer zone distances were set to 40m, 20m, and 10m respectively, with the end type set to FULL and the merging type selected as ALL. Finally, the data was merged into a single whole. The extracted vegetation distribution map was then overlaid sequentially with the buffer zones of primary roads, secondary roads, and tertiary roads to obtain the green space functional zones of the protective green areas.

[0022] In one embodiment, the step of overlaying the extracted vegetation distribution map with the Baidu Map boundary and road network buffer zone, calculating the UGS social functional zone index, and dividing plaza green space, park green space, and ancillary green space by constructing a lookup table includes:

[0023] The obtained vegetation distribution map was spatially overlaid with the boundary data provided by Baidu Maps to obtain the boundaries of UGS social functional zones. Simultaneously, according to the "Urban Green Space Classification Standard" CJJ / T 85-2017, four UGS social functional zone indicators were calculated for each zone. The area indicator was divided into three categories: greater than 10 hectares, greater than 1 hectare, and less than or equal to 1 hectare. The shape indicator was divided into two categories: linear and non-linear. The width indicator was divided into two categories: greater than 12 meters and less than or equal to 12 meters. The vegetation proportion indicator was divided into three categories: greater than or equal to 65%, greater than or equal to 35%, and less than 35%.

[0024] Based on the "Classification Standard for Urban Green Space" CJJ / T 85-2017, and combining the aforementioned social functional area indicators and POI data classifications, a lookup table is constructed to classify plaza green spaces, park green spaces, and ancillary green spaces in sequence. The classification order is as follows: First, based on the lookup table, plaza green spaces are extracted; then, comprehensive parks, community parks, gardens, theme parks, and residential area green spaces are extracted. After removing these types of green spaces and protective green spaces, the remaining green spaces are classified as other ancillary green spaces. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart of an automatic urban green space functional zone delineation method based on deep learning and crowdsourced geospatial data, according to an embodiment of the present invention.

[0027] Figure 2 This is a vegetation distribution map based on U-net, according to an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the basic implementation process of the functional area in an embodiment of the present invention;

[0029] Figure 4 This is a result diagram of urban green space functional areas in an embodiment of the present invention; Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Urban green space, a key aspect of urban ecological space, refers to urban land covered by both natural and artificial vegetation, and is crucial for supporting urban ecosystems and sustainable development. Previous studies have extensively examined the physical characteristics of unmanned ground stations (UGS) (such as trees, shrubs, and grasses), demonstrating their significant role in mitigating the urban heat island effect, regulating microclimates, improving air quality, protecting biodiversity, reducing runoff, and controlling floods. Furthermore, UGS, as a bridge connecting urban residents and nature, possesses important social functions, influencing not only residents' physical and mental health but also enhancing human well-being. However, unlike the numerous open-source land cover products used for UGS physical characteristic analysis, data on UGS social functional zones is severely lacking. To meet the growing demand for high-resolution spatial information on UGS social functional zones and to provide valuable data and information support for rational urban planning, sustainable urban development, government decision-making, and urban geography research, an accurate and effective UGS social functional zone mapping method is urgently needed.

[0032] Traditional classification methods based on acquiring physical features of UGS from medium to high spatial resolution remote sensing images have been widely used, typically involving studies of vegetation types such as the species, quantity, structure, and abundance of various trees, shrubs, and grasses. However, because the social functions of UGS are related to both the external environment and internal human social activities, these methods are insufficient to directly obtain the social functional zones of UGS. Physical features and internal human social activities can be described separately using remote sensing images and socioeconomic data. This leads to a so-called "semantic gap" between raw remote sensing images and raw image pixels, hindering a high-level understanding of the social functions of UGS. Therefore, relying solely on remote sensing data is insufficient to accurately identify the different social functional zones of UGS.

[0033] The following is combined with Figures 1-4 This invention describes an automatic delineation method for urban green space functional zones that combines deep learning and crowdsourced data.

[0034] like Figure 1 As shown in one embodiment, an automatic delineation method for urban green space functional zones utilizing a combination of deep learning and crowdsourced data includes the following steps:

[0035] Step S110: Obtain OSM road network data of the study area from the open source platform, perform topology analysis using ArcMap topology tool, and detect and delete duplicate lines in batches.

[0036] Specifically, we acquired CBERS-04A high-resolution remote sensing image data, collected OpenStreetMap (OSM) road network and Baidu Map boundary data, as well as Points of Interest (POI) data. Using the cropping tool in ArcGIS 10.8 as input features, and the vector data of the study area as the cropping feature, we cropped the data within the study area to obtain the aforementioned data and created the corresponding feature dataset.

[0037] Step S120: Perform radiometric calibration and atmospheric correction on the remote sensing image; clean and classify the POI data; process the OSM road network data to remove dangling lines and redundant lines; convert the above data formats and unify the coordinate system to ensure data consistency and accuracy.

[0038] Specifically, radiometric calibration was performed on the CBERS-04A remote sensing image data to correct sensor bias, and atmospheric correction was performed to eliminate the influence of atmospheric conditions. Secondly, according to the Baidu Maps POI industry classification standard, the attribute selection tool in ArcGIS 10.8 was used to classify the Points of Interest (POI) data. Simultaneously, dangling and redundant lines in the OpenStreetMap (OSM) road network data were removed using a combination of topology analysis and manual assistance to ensure spatial consistency of the road network data. Finally, all data was converted to a unified format, and all datasets were ensured to use the CGCS2000 coordinate system to achieve seamless integration and analysis between different data sources.

[0039] Step S130: Use the U-Net model to classify land cover and extract vegetation distribution from remote sensing image data.

[0040] Specifically, firstly, land cover classification was performed using the U-Net model based on a convolutional neural network (CNN). This model effectively extracts deep semantic information from images and identifies spectral and spatial features at different scales through its encoder-decoder architecture and skip connections. The U-Net model was trained using the deep learning module in ArcGIS Pro 3.1 software, which provides a user-friendly interface and powerful algorithm support. This study specifically modified the backbone network of the U-Net model to ResNet-101, utilizing its residual structure to enhance feature extraction capabilities, and used ImageNet pre-trained weights through transfer learning to accelerate model convergence. During model training, the original image slice size was set to 512×512, the batch size to 4, and 25 iterations were performed using 5 convolutional layers and early stopping to optimize the training process and prevent overfitting. Furthermore, remote sensing indices such as Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Soil Adjusted Vegetation Index (SAVI) were introduced to enhance the accuracy of vegetation feature extraction. The entire process was completed on a Windows 10 operating system equipped with an NVIDIA GeForce RTX 2070 GPU, ensuring computational efficiency during training and the model's generalization ability, such as... Figure 2 The image shown is a vegetation distribution map based on U-net according to an embodiment of the present invention.

[0041] Step S140: The processed OSM road network is classified and buffer zones are created. The extracted vegetation distribution map is overlaid with the buffer zones based on the OSM road network to obtain the green space functional areas of the protective green space.

[0042] Specifically, the processed OSM road network data is classified and buffered, with the following settings: Using the `fclass` field describing the OSM road network type, roads are reclassified into expressways, arterial roads, secondary arterial roads, and local roads according to the standards of the Ministry of Housing and Urban-Rural Development of China. Simultaneously, based on actual road conditions, roads are reclassified into primary, secondary, and tertiary levels. Using a buffer tool, buffer zone distances are set to 40m, 20m, and 10m respectively, with the end type set to FULL and the fusion type selected as ALL. Finally, the data is merged into a single whole. The extracted vegetation distribution map is then overlaid sequentially with the buffer zones of primary roads, secondary roads, and tertiary roads to obtain the green space functional areas of the protective greenbelts.

[0043] Step S150: Overlay the extracted vegetation distribution map with the Baidu Map boundary and road network buffer zone, calculate the UGS social functional area index, and divide the square green space, park green space, and ancillary green space by constructing a lookup table. The UGS social functional area index has four items: area index, shape index, width index, and vegetation ratio index.

[0044] Specifically, the obtained vegetation distribution map is spatially overlaid with boundary data provided by Baidu Maps to obtain the boundaries of UGS social functional zones. Simultaneously, according to the "Urban Green Space Classification Standard" CJJ / T 85-2017, four UGS social functional zone indicators are calculated for each zone. The area indicator is divided into three categories: greater than 10 hectares, greater than 1 hectare, and less than or equal to 1 hectare. The shape indicator is divided into two categories: linear and non-linear. The width indicator is divided into two categories: greater than 12 meters and less than or equal to 12 meters. The vegetation proportion indicator is divided into three categories: greater than or equal to 65%, greater than or equal to 35%, and less than 35%. Figure 3 This is a schematic diagram of the basic implementation process of the functional area in an embodiment of the present invention.

[0045] Based on the "Classification Standard for Urban Green Space" CJJ / T 85-2017, and combining the aforementioned social functional area indicators and POI data, a lookup table was constructed to classify plaza green spaces, park green spaces, and ancillary green spaces in sequence. The classification order is as follows: First, based on the lookup table, plaza green spaces are extracted; then, comprehensive parks, community parks, gardens, theme parks, and residential area green spaces are extracted. After removing these types of green spaces and protective green spaces, the remaining green spaces are classified as other ancillary green spaces. Figure 4 This is a result diagram of the urban green space functional area in an embodiment of the present invention.

Claims

1.A method for automatic division of urban green space functional zones based on deep learning and crowdsourced geospatial data, characterized in that, The method comprises: Using Baidu map platform, OpenStreeetMap platform to obtain research data, through Arcgis10.8 software, create research area data set. Radiation calibration is performed on the remote sensing image data to correct sensor bias, and atmospheric correction is performed to eliminate the influence of atmospheric conditions. Secondly, according to the industry classification standard of Baidu map POI, using Arcgis10.8 in attribute selection tool, the POI data is classified, and the suspended line and redundant line existing in the OpenStreetMap (OSM) road network data is removed combined with topological analysis and manual assistance, to ensure the spatial consistency of the road network data. Finally, all data is converted into a unified format, and all data sets adopt CGCS2000 coordinate system, realizing seamless integration and analysis between different data sources. Using U-Net model, combined with remote sensing image data for land cover classification, extraction of vegetation distribution. The processed OSM road network is graded, and a buffer zone is created. The extracted vegetation distribution map is superimposed with the buffer zone based on the OSM road network to obtain the green function area of the protective green land. The extracted vegetation distribution map is superimposed with the Baidu map boundary and road network buffer zone, and the UGS social function area index is calculated. Through the construction of lookup table, the square green land, park green land and accessory green land are divided, and the UGS social function area index has four items, namely: area index, shape index, width index and vegetation proportion index. 2.The method according to claim 1, wherein the data of the study area is obtained, and wherein, Using Baidu map platform, OpenStreeetMap platform to obtain research data, through Arcgis10.8 software, create research area data set, including: CBERS-04A high-resolution remote sensing image data is obtained, OpenStreetMap (OSM) road network and Baidu map boundary data, and point of interest (POI) data are collected, combined with the clipping tool in Arcgis10.8, which is used as input element, and the vector data of the study area is used as the clipping element. The above data in the study area is clipped to obtain the corresponding feature data set. 3.The method according to claim 1, wherein the data preprocessing of the study area is characterized by, Radiation calibration is performed on the remote sensing image data to correct sensor bias, and atmospheric correction is performed to eliminate the influence of atmospheric conditions. Secondly, according to the industry classification standard of Baidu map POI, using Arcgis10.8 in attribute selection tool, the POI data is classified, and the suspended line and redundant line existing in the OpenStreetMap (OSM) road network data is removed combined with topological analysis and manual assistance, to ensure the spatial consistency of the road network data. Finally, all data is converted into a unified format, and all data sets adopt CGCS2000 coordinate system, realizing seamless integration and analysis between different data sources. Using U-Net model, combined with remote sensing image data for land cover classification, extraction of vegetation distribution, including: 4.The method according to claim 1, wherein the green space distribution extraction is characterized in that, ​ First, the U-Net model based on convolutional neural network (CNN) is used for land cover classification. The model effectively extracts deep semantic information of images and identifies spectral and spatial features of different scales through its encoder-decoder architecture and skip connection. The U-Net model is trained using the deep learning module in ArcGIS Pro 3.1 software, which provides a user-friendly interface and powerful algorithm support. In this study, the backbone network of the U-Net model is modified to ResNet-101, which enhances the feature extraction capability through its residual structure and uses ImageNet pre-trained weights through transfer learning to speed up model convergence. In the model training, the original image slice size is set to 512×512, the batch size is set to 4, and 5 convolutional layers and early stopping techniques are used for 25 iterations to optimize the training process and prevent overfitting. In addition, remote sensing indexes such as normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and soil-adjusted vegetation index (SAVI) are introduced to enhance the extraction accuracy of vegetation features. The entire process is completed on a Windows 10 operating system equipped with an NVIDIA Ge Force RTX 2070 GPU, ensuring the computational efficiency of the training and the generalization ability of the model. 5.The method of claim 1, wherein the classification of the protective green space is characterized by, The processed OSM road network is classified, and the extracted vegetation distribution map is overlaid with the buffer zone based on the OSM road network to obtain the green space function area of the protective green space, including: The processed OSM road network data is classified and buffer analysis is performed, with the following settings: using the fclass field to describe the OSM road network type, according to the standards of the Ministry of Housing and Urban-Rural Development of China, the roads are reclassified into expressways, main roads, secondary roads, and branch roads, and according to the actual road conditions, the roads are reclassified into first-class, second-class, and third-class roads. The buffer tool is used to set the buffer distance to 40m, 20m, and 10m, respectively, and the end type is set to FULL and the fusion type is set to ALL. Finally, they are merged into one whole. The extracted vegetation distribution map is overlaid with the buffer zone of the first-class road, the buffer zone of the second-class road, and the buffer zone of the third-class road, respectively, to obtain the green space function area of the protective green space. 6.The method according to claim 1, wherein the square green space, the park green space, and the subsidiary green space are classified. The extracted vegetation distribution map is overlaid with the Baidu Map boundary and road network buffer zone to calculate the UGS social function area index, and the lookup table is constructed to divide the square green space, park green space, and auxiliary green space, including: The obtained vegetation distribution map is spatially overlapped with the boundary data provided by Baidu Map to obtain the boundary of the UGS social function area. Meanwhile, according to the Urban Green Space Classification Standard CJJ / T 85-2017, four UGS social function area indexes of each UGS social function area are calculated, wherein the area index is divided into three categories, i.e., greater than 10 mu, greater than 1 mu and less than or equal to 1 mu; wherein the shape index is divided into two categories, i.e., strip and non-strip; wherein the width index is divided into two categories, i.e., greater than 12 meters and less than or equal to 12 meters; and wherein the vegetation proportion index is divided into three categories, i.e., greater than or equal to 65%, greater than or equal to 35% and less than 35%. Based on the Urban Green Space Classification Standard CJJ / T 85-2017, in combination with the above social function area indexes and the classification of POI data, a lookup table is constructed to classify square green space, park green space and accessory green space in turn. The classification order is as follows: first, based on the lookup table, square green space is extracted, then comprehensive parks, community parks, gardens, theme parks and residential green space are extracted, the above types of green space and protective green space are removed, and the remaining green space is divided into other accessory green space.

Images