Machine learning-based small greenfield augmentation siting method

Abstract

The application discloses a small and micro green land supplement site selection method based on machine learning, belongs to the technical field of urban and rural planning, and collects data through experiments; a small and micro green land site selection potential feature index system is established through the collected data, and potential discrimination features of potential site selection points of the small and micro green land in different directions are comprehensively discriminated according to characteristics of the established potential feature index system at different levels, so as to construct a small and micro green land site selection potential quantification discrimination system; a small and micro green land site selection potential discrimination model is trained through conventional machine learning and deep learning; the small and micro green land site selection potential discrimination model is applied to score the small and micro green land site selection potential, so that a final small and micro green land site selection potential quantification discrimination result is obtained, ranking and selection are performed, the supplement site selection priority of the small and micro green land site selection point is determined, and the priority supplement site selection point is selected.

Description

Technical Field

[0001] This invention belongs to the field of urban and rural planning technology, specifically relating to a method for selecting sites for small-scale green space additions based on machine learning. Background Technology

[0002] Compared with ordinary parks, small green spaces are characterized by their small scale, flexible layout, proximity to the public, low cost, and convenient maintenance. They are usually located and developed based on public spaces along urban roads and can be placed in high-density blocks that are inaccessible to traditional urban green space service networks. Their own attributes and service models are more adaptable to the urban built environment.

[0003] Existing green space site selection technologies rely heavily on manual screening, resulting in low efficiency and limited scope, while the site selection results are overly dependent on the planner's ability.

[0004] Existing quantitative methods are not very targeted at the selection of small green spaces. They rely heavily on remote sensing images and geographic analysis system platforms for quantitative analysis, but cannot effectively incorporate the visible physical spatial environment of potential sites into the analysis and identification system. Therefore, the information analysis of small green space selection is not comprehensive enough. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a machine learning-based method for selecting locations for small green spaces, thereby solving the technical problem of insufficient comprehensive information analysis for small green space location selection in existing technologies.

[0006] The objective of this invention can be achieved through the following technical solution: a machine learning-based method for selecting sites for supplementing small green spaces, the method comprising the following steps:

[0007] Data is collected, including the coordinates of all potential site selection points for small green spaces and a street view image dataset of all potential site selection points for small green spaces. The street view image dataset includes all crawling parameters of the street view images of potential site selection points for small green spaces.

[0008] A site selection potential characteristic index system for small green spaces is established using the collected data. The site selection potential characteristic index system for small green spaces includes geographical location environmental characteristics, visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view images.

[0009] Based on the different characteristics of the established small green space site selection potential characteristic index system, a quantitative identification system for small green space site selection potential is constructed.

[0010] Under the small green space site selection potential feature index system, a support vector machine classification model is used as the underlying model to train the small green space placement basic suitability identification model. A fully connected image convolutional neural network built based on deep learning framework tools is used as the underlying model to train the small green space placement deep suitability identification model. The trained small green space placement basic suitability identification model and small green space placement deep suitability identification model are used to establish a stable derivation relationship between the small green space site selection potential feature index system and the small green space site selection potential quantitative identification system.

[0011] The potential identification model for small green space sites assesses the potential of small green space sites and selects the site with the highest potential from all the coordinates of the supplementary site selection points based on the assessment results.

[0012] Preferably, the process of establishing the characteristic index system for the site selection potential of small green spaces includes the following steps:

[0013] The site selection potential characteristic index is divided into geographical location environmental characteristics, visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view.

[0014] The characteristics of geographical location environment, visual semantic elements, advanced visual semantic features of site conditions, and street view images are quantitatively measured to establish a site selection potential characteristic index system.

[0015] Establish a geographic location environmental characteristic index system for potential sites of micro-green spaces to measure the environmental factors that affect the site selection and subsequent services of micro-green spaces, such as the level of recreational demand, facility conditions, and green space construction conditions in the surrounding environment of potential sites. The geographic location environmental range of potential sites of micro-green spaces is also set within 300 meters around the potential sites of micro-green spaces.

[0016] Based on data on population, service facilities, and urban land use, a geographic location environmental characteristic index (GIF) is generated for each potential site, representing the surrounding recreational demand, facility conditions, and green space conditions. A geographic location environmental characteristic vector is constructed for each potential micro-green space site. Let g be the total defined geographic location environmental characteristic indices, denoted as GIF = {gif1,...,gif...}. g Let} represent the geographic location environmental feature vector of the site selection point. This vector represents the geographic location environmental features related to the site selection potential of a small green space. (GIF is used as an example.) k Let GIF_set represent the geographic location environmental feature vector of the k-th potential site for small green spaces among all potential site selection points. Construct a set of geographic location environmental feature vectors for all potential site selection points of small green spaces, with a total number of n, denoted by GIF_set.

[0017] GIF_set = {GIF1, ..., GIF n}

[0018] Semantic segmentation is performed on the street view images of each potential site of small green space from three angles to obtain the proportion of visual semantic elements in the three images. Each visual semantic element of each potential site of small green space is used as a feature dimension to construct the visual semantic element feature vector (VSF) of each potential site of small green space, and the visual semantic element feature vector set (VSF_set) of all potential sites of small green spaces is obtained.

[0019] Using a deep learning image semantic segmentation model pre-trained on the Cityscapes dataset, visual semantic segmentation analysis was performed on three street view images from three angles for each potential site location of a small green space in the obtained street view image dataset. The Cityscapes dataset provides 34 categories of street view visual semantic elements. Fourteen semantic elements directly related to the street-side environment were selected as visual semantic elements related to the site selection potential of potential small green space sites: vehicular roads, pedestrian walkways, buildings, walls, medians, pillars or streetlights, traffic lights, traffic signs, greenery, terrain undulations, sky, pedestrians, riders, and vehicles. The visual semantic segmentation analysis sequentially calculated the proportion of the number of pixels of each of the 14 visual semantic elements in the three street view images to the total number of pixels in the entire image. ω represents the proportion of the number of image pixels of the i-th visual semantic element in the direction with a relative angle θ to the road at a potential site for a small green space, relative to the total number of pixels in the entire street view image. θ The weights of semantic elements in the street view image at a relative angle θ represent the importance of semantic elements in potential sites for small green spaces. The proportion of the i-th semantic element in a potential site for a small green space is defined as follows:

[0020]

[0021] Based on the defined formula for the proportion of the i-th semantic element of potential site selection points for small green spaces, the vsf of each potential site selection point for small green spaces obtained from visual semantic segmentation analysis is input. θ,i The proportion of 14 semantic elements for each potential site of a small green space was calculated. A feature vector of visual semantic elements was constructed for each potential site of a small green space, using VSF = (vsf1, vsf2, ..., vsf...). 14 )express;

[0022] Use VSF kLet VSF_set represent the visual semantic feature vector of the k-th potential site of all potential sites for small green spaces. Construct a set of visual semantic feature vectors of all potential sites for small green spaces with a total of n, denoted by VSF_set:

[0023] VSF_set = {VSF1,...,VSF n}

[0024] Based on the fundamental principles of small-scale green space planning and design, and the available data types and precision, a high-level visual semantic feature index system for site conditions is established. This system measures the interrelationships between the proportions of various visual semantic elements within a site, describing the advanced characteristics of the site's spatial environment. The type of high-level visual semantic feature index for site conditions is defined according to the site characteristic description requirements for small-scale green space selection. Using the obtained visual semantic element feature vectors of all potential small-scale green space site selection points, the high-level visual semantic feature index value for each potential site selection point is calculated, constructing a high-level visual semantic feature index dataset AVSF_set. Each high-level visual semantic feature element of each potential site selection point is treated as a feature dimension, constructing a high-level visual semantic feature vector AVSF for each potential site selection point, resulting in the high-level visual semantic feature vector set AVSF_set for all potential small-scale green space site selection points.

[0025] Based on the urban needs of the research subjects, m advanced visual semantic feature indices for site conditions are defined.

[0026] Using the obtained visual semantic feature vectors of all potential site selection points for small green spaces, the high-level visual semantic feature index of each site condition for each potential site selection point is calculated, and AVSF is used to perform the calculation. i This represents the high-level visual semantic feature index of the i-th site condition;

[0027] Using the calculated advanced visual semantic feature indices for each site condition, an advanced visual semantic feature vector for each potential site location of a small green space is constructed. If a total of m advanced visual semantic feature indices for site conditions are defined, then AVSF = (avsf1, avsf2, ..., avsf...) is used. m )express;

[0028] Use AVSF k Let AVSF_set represent the high-level visual semantic feature vector of site conditions for the k-th potential site of all potential small green spaces. A dataset of high-level visual semantic feature vectors of site conditions for all potential small green spaces, with a total of n nodes, is constructed.

[0029] AVSF_set = {AVSF1, ..., AVSFn}

[0030] By combining the geographic location environmental feature vectors, visual semantic element feature vectors, and high-level visual semantic feature vectors of site conditions of all potential site selection points for small green spaces along the street, a comprehensive potential feature vector (CPF) related to the site selection potential of all potential site selection points for small green spaces is obtained, and a comprehensive potential feature vector dataset CPF_set of potential site selection points for small green spaces is constructed.

[0031] The geographic location environmental feature vector (GIF), the visual semantic element feature vector (VSF), and the site condition high-level visual semantic feature vector (AVSF) of each potential small green space site are directly concatenated to obtain the comprehensive potential feature vector, and cpf is defined. i Let represent the i-th comprehensive potential feature of a potential site for a small green space. Define the operator [a, b, c] to represent the concatenation operation of vectors a, b, and c. Let CPF represent the comprehensive potential feature vector, i.e.:

[0032] CPF=[GIF,VSF,AVSF]=(gif1,...,gif g ,vsf1,...,vsf 14 ,avsf1,...,avsf m )

[0033] =(cpf1,...,cpf 14+g+m )

[0034] CPF k Let CPF_set represent the comprehensive potential feature vector of the k-th potential site among all potential site selection points for small and micro green spaces. Construct a dataset of comprehensive potential feature vectors of all potential site selection points for small and micro green spaces with a total of n, denoted as CPF_set:

[0035] CPF_set = {CPF1, ..., CPF} n}

[0036] The obtained street view image dataset of all potential site selection points for small green spaces, Graph_set, is constructed as part of the feature index system for the site selection potential of small green spaces. k,θ Graph represents the street view image of the k-th potential site for a small green space, where the street view shooting direction is at an angle θ relative to the road. k Graph represents the complete visual scene depicted by three street view images of the k-th potential site for a small green space. k ={G k,75° G k,90° G k,105°Construct a complete visual scene dataset Graph_set containing n potential site selection points for all small green spaces:

[0037] Graph_set={Graph1,Graph2,...,Graph n}

[0038] Preferably, the area of ​​the error surface is less than two hectares.

[0039] Preferably, the process of constructing a quantitative identification system for the site selection potential of small green spaces includes the following steps:

[0040] By utilizing the comprehensive potential feature vector (CPF) of all potential site selection points for small green spaces, the comprehensive potential information of small green space site selection points is quantitatively represented. Considering the comprehensive nature of the comprehensive potential information, a method for identifying the suitability of basic infrastructure for small green space placement is proposed.

[0041] Using the complete visual scene dataset of all potential sites for small green spaces, we mine the depth perception features of the street scene of the site for small green spaces and propose a method for identifying the depth suitability of small green space placement based on depth feature perception.

[0042] A multi-class model for the basic suitability of small green spaces was trained to target the comprehensive potential feature vector (CPF).

[0043] A multi-classification model for the suitability of basic infrastructure for small green spaces was used to preliminarily determine the suitability of basic infrastructure for all potential small green space sites. K-Means cluster analysis was applied to obtain the suitability level of each potential site, classifying them into L levels, denoted as {1,2,...,L}, representing levels 1 to L. The y-value is used as the unit of measurement. k Let y represent the multi-classification model judgment result of the basic suitability of the k-th potential site for small green space, and the basic suitability level of the k-th potential site for small green space. k The domain is:

[0044] y k ∈{1,2,...,L}, where L is a positive integer

[0045] The suitability level of potential sites for small green spaces is divided into L levels, denoted by {1,2,...L}, representing levels 1 to L.

[0046] The basic potential score of a potential site for a small green space is calculated based on its suitability level. If the suitability level is 1, the basic potential score is 0 points. If the suitability level is L, the basic potential score is 100 points. If the suitability level is l, where l is between 1 and L, the score is... Using BaseScore k Let y represent the basic potential score of the kth potential site for a small green space. k , where y k Let ∈{1,2,...,L}, where L is a positive integer, representing the basic suitability level for the placement of the k-th potential site for a small green space. Then, the basic potential score of the k-th potential site for a small green space is defined as:

[0047]

[0048] Then, a multi-classification model for the appropriate insertion depth of street view images of all potential sites for small green spaces was trained.

[0049] A depth-appropriate multi-classification model was used to deeply mine the street view image features of all potential sites for small green spaces. Simultaneously, the depth-appropriateness of these potential sites was assessed, resulting in depth-appropriateness levels. Based on the site selection needs of the study city, the depth-appropriateness levels of potential sites were autonomously divided into L levels, denoted as {1,2,...,L}, where L is a positive integer, representing levels 1 to L. The z-axis represents the depth-appropriateness of each level. k This represents the result of the multi-class classification model for the suitability of the placement depth of the k-th potential site for small green spaces, i.e., the suitability level of the placement depth of the k-th potential site for small green spaces, z. k The domain is:

[0050] z k ∈{1,2,...,L}

[0051] The suitability level of potential site selection locations for small green spaces is divided into L levels, denoted by {1,2,...,L}. A depth potential score is calculated based on the suitability level of each potential site selection location. If the suitability level is 1, the depth potential score is 0. If the suitability level is L, the depth potential score is a maximum of 100. If the suitability level is l (where l is between 1 and L), the score is... Using DeepScore kThe depth potential score of the k-th potential site for a small green space is represented by z. k , z k Let ∈{1,2,...,L}, where L is a positive integer representing the suitability level of the placement depth of the k-th potential site for a small green space. Then, the depth potential score of the k-th potential site for a small green space is defined as:

[0052]

[0053] By combining the defined basic potential score system with the defined deep potential score system, a weighted average is calculated for the basic potential score and the deep potential score of each potential site for small green space. The weighting weights are determined based on the urban environmental characteristics of the research object and the functional development focus of the small green space, and the final potential score of each potential site for small green space is obtained.

[0054] Using BaseScore k The DeepScore represents the basic potential score of the k-th potential site for a small green space. k Let represent the deep potential score of the k-th potential site for a small green space, and let w represent the weight of the basic potential score. Then the weight of the deep potential score is (1-w). (FinalScore) k Let represent the final potential score of the k-th potential site for a small green space. Then, FinalScore k Defined as:

[0055] FinalScore k =w*BaseScore k +(1-w)*DeepScore k

[0056] By summing the final potential scores of all potential sites for small green spaces, and constructing a final potential identification result set Score of n for all potential sites for small green spaces, we have:

[0057] Score={FinalScore1,...,FinalScore n}

[0058] Preferably, the small green space site selection potential identification model uses the potential characteristic index of the potential site selection point of the small green space as the basic input information for potential identification, and uses the final potential score of the potential site selection point of the small green space as the output information for potential identification.

[0059] Preferably, the training process of the small-scale green space site selection potential identification model includes the following steps:

[0060] Construct the feature vector set Train_X for the machine learning training set, construct the score label set Train_y for the training set, construct the feature map set GTrain_X for the deep learning training set, map Train_y and Train_X together to form the training set Train_ML for conventional machine learning, map Train_y and GTrain_X together to form the training set Train_DL for deep learning.

[0061] The basic suitability assessment model for small green spaces includes the following:

[0062] Dimensionality reduction is performed on the feature vectors in the feature vector set Train_X of the machine learning training set.

[0063] Divide Train_ML into a training set and a test set;

[0064] We choose a support vector machine (SVM) classification model as the underlying model for the suitability identification model for small and micro green space placement. We define a multi-class SVM model with parameters to be trained, and call the hyperparameter optimization method and random sampling cross-validation provided by sklearn to select the hyperparameters of the model.

[0065] Input the feature vector set Train_X and the rating label set Train_y from the machine learning training set Train_ML, train the model, solve the optimization problem of the loss function, determine a set of optimal training parameters for a multi-class support vector machine, and obtain the basic suitability identification model for the placement of small green spaces.

[0066] The resulting model for discriminating the suitability of small green spaces based on depth feature perception includes the following:

[0067] A fully connected image convolutional neural network was built based on deep learning framework tools.

[0068] Input the feature map set GTrain_X and the rating label set Train_y from the deep learning training set Train_DL, perform deep learning training, adjust the hyperparameters of the fully connected image convolutional neural network, train the fully connected image convolutional neural network to classify feature images, obtain the fully connected image convolutional neural network multi-classification model, and obtain the depth suitability discrimination model for the placement of small green spaces based on deep feature perception.

[0069] Preferably, the process of constructing the feature vector set Train_X for the machine learning training set and the feature map set GTrain_X for the deep learning training set includes the following steps:

[0070] Using street view sampling intervals of no less than 50 meters, a portion of potential site selection points for small green spaces were selected from all potential site selection points for small green spaces as training sample site selection points. Simultaneously, the initially selected potential site selection points should meet the basic training quantity requirements; the initial sample size was initially set to 10,000 points per 1,000 km. 3 Based on the initial model training results, it is determined whether the number of samples needs to be increased, thus forming a set of training sample site selection points. Within this set, the comprehensive potential feature vector of each training sample site selection point is named x. k Construct the feature vector set of the machine learning training set, named Train_X, and set the sample size of the training sample location set to N. Then:

[0071] Train_X = {x1,...,x} N}

[0072] Based on the street view image dataset Graph_set of all potential site selection points for small green spaces, within the training sample site selection point set, the three-angle street view image set along the street of each training sample site selection point is used as the original feature map set Gx of the training sample site selection points. k This yields the feature map set of the deep learning training set, named GTrain_X. With the sample size N of the training sample location set set, we have:

[0073] GTrain_X={Gx1,...,Gx N}

[0074] Preferably, the process of constructing the training set's score label set Train_y, the conventional machine learning training set Train_ML, and the deep learning training set Train_DL includes the following steps:

[0075] Using the defined suitability assessment scheme for small-scale green space placement as the reference criterion for manually scoring the site selection potential of training samples, a manual scoring standard for suitability potential assessment is defined. Within the training sample site selection point set with sample sizes of GTrain_X and Train_X, based on the street view visual information of potential site selection points for small-scale green spaces, N... P Each professional observer makes an assessment, manually scoring each potential site selection point in the training samples multiple times. After each round of manual scoring, the scores from all professionals are compared to determine consistency. Simultaneously, it is necessary to ensure sufficient sample size for each score level. If the sample size for a certain score level is insufficient, new training samples are promptly added, and multiple rounds of scoring and result quality assessment are conducted until N... PThe manual scoring results of each professional meet the quality requirements; thus, the manual identification results of the basic suitability of small green space placement for each training sample potential site are determined, and these results are used as identification labels for the basic suitability of small green space placement for training representative potential sites, thereby constructing a basic suitability identification label set for small green space placement.

[0076] Multiple rounds of manual scoring were conducted. In each round, based on the visual information of the street view images of potential micro-green space sites, professionals observed and judged each potential site in the training set, performing manual visual identification and scoring. N P Several professionals sequentially visually observe street view images of potential training site locations, identifying the placement suitability level of each location. The placement suitability levels are divided into L levels, denoted by {1,2,...,L}, where L is a positive integer. If a professional manually observes the street view image of a potential training site and determines that the placement suitability is higher, the level of that potential small green space location is assigned a higher score, denoted by y. i,k Let y represent the suitability assessment result of the i-th staff member for the potential location of the k-th training sample. i,k =L, where the suitability level is determined manually to be L, L∈{1,2,...,n}, i∈{1,2,...,N}. P}

[0077] In the multi-round manual scoring process, after each round of scoring, N P Each professional staff member conducts consistency and sample size checks on the identification results of potential site selection points across the entire training sample, up to N. P The following is a method for verifying the consistency and sample size of the manual scoring results from individual professionals to meet quality requirements and for iteratively updating the scoring rules:

[0078] N P A professional manually scores N training samples, using Mp k In this round of manual scoring, N represents N. P Among the results of manual suitability assessment of potential site selection points for the k-th training sample by professionals, the most prevalent manual scoring result p k The number of values, denoted by ρ, can be used to determine the threshold for consistency. If the most frequent human-scored result has a high percentage (p), then... k Number Mp k N, accounting for the total number of professional human scoring results P If the proportion of a given score is greater than the threshold ρ for consistency, then the human scoring results for the potential location points of the k-th training sample are considered consistent. Therefore, the consistency rate η of the human scoring results is defined and calculated as follows:

[0079]

[0080] Determine whether the consistency rate meets the requirements. If it does not meet the requirements, find the site selection points with inconsistent scoring results for discussion, discuss the scoring rules and standards, further unify the manual scoring rules, and re-extract potential site selection points from the training samples for scoring.

[0081] Statistics N P The average number of samples for each level score in the professional scoring results is calculated. To ensure that the sample size for each level score meets the requirements, if the average sample size for a certain level score is significantly insufficient, new training samples are added in a timely manner.

[0082] After multiple rounds of scoring and result quality assessment, the scoring rules and standards are iteratively updated once the consistency rate meets the requirements and the sample size for each score level reaches a certain scale.

[0083] After the final iterative updates to the scoring rules and standards are completed, N P The average of the scores given by professionals to all training samples is rounded down to determine the final human suitability assessment result for potential site selection points in the training samples. This result is represented by y. i,k Let y represent the final human score given by the i-th professional to the k-th training sample. k Let represent the final human suitability assessment result of the potential location point of the k-th training sample. Then:

[0084]

[0085] The final human suitability assessment results of all potential site selection points in the training samples are used to construct a score label set for the training set, named Train_y. The sample size of the potential site selection point set is set to N. Then:

[0086] Train_y = {y1,...,y} N}

[0087] The score label set Train_y of the obtained training set is matched with the feature vector set Train_X of the obtained machine learning training set to form the training set Train_ML of conventional machine learning.

[0088] Set each label y in the rating label set k Each feature vector x in the feature vector set Train_X of the machine learning training set k A one-to-one correspondence, as a complete training data point, is denoted as t. k If the operator [a, b] is defined to represent the concatenation operation of vector a and scalar b, then:

[0089] t k =[x k ,y k ]

[0090] Construct a complete machine learning training set, named Train_ML, and set the sample size of the machine learning training set to N. Then:

[0091] Train_ML=(Train_X,Train_y)={t1,...,t N}

[0092] The obtained training set score label set Train_y is matched with the obtained deep learning training set feature map set GTrain_X to form the deep learning training set Train_DL.

[0093] Set each label y in the rating label set k Each feature map group Gx in the feature map set of the deep learning training set k A one-to-one correspondence is used to construct a deep learning input data set, denoted as Gt. k

[0094] Construct a complete deep learning training set, named Train_DL, and set the sample size of the deep learning training set to N. Then:

[0095] Train_DL=(GTrain_X,Train_y)={Gt1,...,Gt N}

[0096] Preferably, the process of using the small-scale green space site selection potential identification model to assess the potential of small-scale green space sites includes the following steps:

[0097] On the comprehensive potential feature vector dataset CPF_set of all potential sites for small green spaces, the basic suitability identification model for the placement of small green spaces is used to perform a preliminary suitability classification for each potential site. Based on the preliminary suitability classification results of each potential site for small green spaces and the basic suitability level of each potential site for placement, the basic potential score BaseScore of each potential site for small green spaces is calculated.

[0098] On the street view image dataset of all potential sites for small green spaces, a depth-based model for identifying the suitability of small green space placement is used to classify the suitability of each potential site for small green space placement. Based on the classification results of the suitability of placement of each potential site for small green space placement and the suitability level of placement of each potential site for small green space placement, the depth potential score (DeepScore) of each potential site for small green space placement is calculated.

[0099] By selecting weights, the DeepScore and BaseScore of each potential site for small green spaces are weighted and averaged to obtain the FinalScore of each potential site for small green spaces, and the FinalScore is constructed as the final potential identification result set Score of all potential sites for small green spaces.

[0100] Based on the final potential identification result set (Score) of all potential site selection points for small and micro green spaces, K-Means clustering is performed on all potential site selection points according to the final potential score (FinalScore). The cluster levels are sorted according to the cluster center scores to obtain the supplementary site selection priority of each potential site selection point's cluster. Finally, the coordinates and site selection priorities of all supplementary site selection points are used as the decision basis for the site selection of small and micro green spaces.

[0101] The beneficial effects of this invention are as follows: During its use, this invention involves data collection experiments; establishing a potential characteristic index system for small-scale green space site selection based on the collected data; considering the characteristics of different levels of the established potential characteristic index system, and comprehensively considering the potential identification features of potential site selection points for small-scale green spaces in different directions implied by the potential characteristic results at each level, constructing a quantitative identification system for small-scale green space site selection potential; training the small-scale green space site selection potential identification model through conventional machine learning and deep learning; applying the small-scale green space site selection potential identification model to score the site selection potential of small-scale green spaces, obtaining the final quantitative identification result of small-scale green space site selection potential, ranking and selecting, and determining priority supplementary site selection points for small-scale green spaces; the constructed automatic identification system can efficiently and comprehensively make preliminary judgments on site selection potential, improving efficiency and reducing labor costs; the model establishes a stable derivation relationship between the small-scale green space site selection potential characteristic index system and the potential quantitative identification system, providing efficient and accurate site selection technology for the construction of small-scale green spaces in the urban built environment, and has application value. Attached Figure Description

[0102] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0103] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0104] 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, and 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.

[0105] like Figure 1 As shown, the method for selecting sites for small-scale green space additions based on machine learning includes the following steps:

[0106] Conduct data collection experiments;

[0107] A characteristic index system for the site selection potential of small green spaces was established based on the collected data.

[0108] Based on the characteristics of the different levels of potential characteristic index system, and by integrating the potential identification characteristics of potential site selection points of micro green spaces in different directions implied by the potential characteristic results of each level, a quantitative identification system for site selection potential of micro green spaces is constructed.

[0109] A model for identifying the site selection potential of small green spaces was trained using conventional machine learning and deep learning.

[0110] A small-scale green space site selection potential identification model is applied to score the site selection potential of small-scale green spaces, and the final quantitative identification results of the site selection potential of small-scale green spaces are obtained. K-Means clustering is performed, and the cluster levels are sorted according to the cluster center scores to obtain the supplementary site selection priority of each potential site selection point in the cluster. Finally, the coordinates and site selection priorities of all supplementary site selection points are used as the decision basis for the site selection of small-scale green spaces.

[0111] 1. Data collection and verification

[0112] By crawling and cleaning street view images through an intelligent location service platform, the main data foundation for the site potential characteristic index system for potential site selection points of small and micro green spaces can be obtained.

[0113] 1.1. Use electronic map capture software to capture the urban navigation trajectory map of the planned area in the intelligent location service platform software:

[0114] Based on the actual planning or analysis needs of small-scale green spaces, the scope of the planning area is determined, thereby determining the cut-off range for the urban navigation trajectory map.

[0115] Select an appropriate map cropping scaling scale based on the required cropping range and planning accuracy.

[0116] Use a map coordinate picking website to query and analyze the latitude and longitude coordinates of the center of the region.

[0117] Using electronic map cropping software, given the latitude and longitude coordinates of the area center and a zoom level, capture the corresponding urban navigation trajectory map of the study area.

[0118] 1.2. Using a Geographic Information System (GIS) platform, process the urban navigation trajectory map of the divided area, sample the navigation trajectories, and obtain a "number-coordinate table" of all potential site selection points for small green spaces.

[0119] Using a geographic information system platform, the city navigation trajectory map is vectorized to obtain the vectorized city navigation trajectory map.

[0120] Using a geographic information system platform, georegistration is performed on the vectorized urban navigation trajectory map to obtain a georegistered vectorized urban navigation trajectory map.

[0121] Based on the street view image capture results in the intelligent location service platform software, the sampling interval for potential site selection points of small green spaces with the highest actual coverage efficiency of street view images is determined to ensure that the street view images of two adjacent potential site selection points of small green spaces have exactly no duplicate information and no missing information.

[0122] Using a geographic information system platform, potential site selection points for small green spaces are sampled from the georegistered vectorized urban navigation trajectory map to obtain the latitude and longitude coordinates of all potential site selection points for small green spaces. Each sampled potential site selection point is assigned a number, and each potential site selection point number and its corresponding latitude and longitude coordinates are combined into a table entry to obtain a "number-coordinate table" for all potential site selection points for small green spaces.

[0123] 1.3. Analyze the API structure of the intelligent location service platform software, determine the crawling parameters for street view images of potential small green space sites, calculate or select all crawling parameters, write a crawler program to crawl street view images from three angles along the street for each potential small green space site, and construct a street view image dataset for all potential small green space sites:

[0124] The analysis of the intelligent location service platform software API composition primarily determines the following crawling parameters for street view images of potential small green space locations: a) the absolute angle between the road where the potential small green space location is located and due north; b) the relative angle between the street view shooting direction along the street and the road where the potential small green space location is located; c) the longitude of the potential small green space location; d) the latitude of the potential small green space location; e) the viewing angle of the street view shooting along the street.

[0125] Calculate or select all crawling parameters: a) Using a Geographic Information System (GIS) platform, analyze the urban navigation trajectory map of the study area to calculate the absolute angle between the road where all potential small green space sites are located and due north. b) Select three angles, 75°, 90°, and 105°, as the three relative angles between the street view shooting direction and the road where the potential small green space sites are located. c) Obtain the longitude of the potential small green space sites based on the "Longitude" attribute of the "Number-Coordinate Table". d) Obtain the latitude of the potential small green space sites based on the "Latitude" attribute of the "Number-Coordinate Table". e) Select a 90° angle as the viewing angle range for street view shooting along the street of the potential small green space sites. Combine the number of each potential small green space site with the above five crawling parameters as an entry to obtain the "Number-Crawling Parameter Table" for all potential small green space sites.

[0126] Based on the API structure of the intelligent location service platform software, a crawling program was written. Inputting a "number-crawling parameter table", it crawls three-angle street view images along the street for each potential site of a small green space. The three-angle street view images of each potential site are saved to a folder named "number + longitude + latitude" for that site. This process is repeated to obtain the three-angle street view images of all potential sites of small green spaces, constructing a street view image dataset for potential sites of small green spaces, Graph_set. Using G... k,θ The street view image representing the k-th potential site for a small green space, with a relative angle θ between the street view shooting direction and the road it is located on, is represented by a Graph. k Graph represents the complete visual scene depicted by three street view images of the k-th potential site for a small green space. k =z{G k,75° G k,90° G k,105° Then we can construct a complete visual scene dataset Graph_set containing n potential site locations for all small green spaces:

[0127] Graph_set={Graph1,Graph2,...,Graph n}

[0128] 2. Small Green Space Site Selection Potential Characteristic Index System

[0129] Based on the data obtained through crawling, and by using and processing this data, we can now apply the methods of data analysis from the geographic information system platform and semantic segmentation analysis of street view images to quantify potential site selection points for small green spaces.

[0130] 2.1 To replace the survey technology solution of "conducting manual on-site surveys and observations of potential sites for small green spaces", a two-level quantitative potential feature index system of "comprehensive potential feature vector + complete visual scene" is constructed to quantitatively represent the site selection potential-related feature information of potential sites for small green spaces. The details will be elaborated in subsequent steps.

[0131] 2.2 Establish a geographic location environmental characteristic index system for potential micro-green space sites to measure the environmental factors affecting micro-green space site selection and subsequent services, such as recreational demand levels (e.g., resident size and density), facility conditions (scale and density of recreational facilities), and green space construction conditions (current green space grade and scale). Based on the 300-meter service radius for micro-green spaces set in the "Urban Green Space Planning Standard (GB / T 51346-2019)," the geographic location environmental scope of potential micro-green space sites is also set within a 300-meter radius of the potential site.

[0132] Based on population, service facilities, and urban land use data, a geographic location environmental characteristic index (GIF) is generated on a geographic information system platform, representing the recreational needs, facility conditions, and green space conditions around each potential site. A geographic location environmental characteristic vector is constructed for each potential site for a small green space. If a total of g geographic location environmental characteristic indices are defined, then GIF = {gif1,...,gif...} g Let} represent the geographic location environmental feature vector of the selected site. This feature vector is used to represent the geographic location environmental features related to the site selection potential of a small green space potential site.

[0133] Therefore, the site selection potential of small green spaces is jointly determined by both the surrounding environment and site conditions. In this regard, the site selection potential characteristic index is divided into geographical location environmental characteristics, visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view. The geographical location environmental characteristics index mainly reflects the positive and negative factors related to the development of small green spaces in the surrounding environment of potential sites, representing characteristics of the surrounding environment. The visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view mainly reflect the suitability characteristics of small green space development on the site of potential sites, representing characteristics of site conditions.

[0134] Based on the fundamental principles and mechanisms of public green space services, positive factors in the geographical location environment for the development of micro-green spaces include resident demand and existing recreational facilities (such as restaurants, retail, and public toilets), while negative factors include existing public green spaces. Site suitability characteristics include lighting conditions, spatial enclosure, pedestrian friendliness, green space coverage, and safety. By quantifying and measuring geographical location environmental characteristics, visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view images, a site selection potential characteristic index system can be established. This index system can be added to, subtracted from, and replaced according to the characteristics of the planning area, the types of data available, and the required precision. The geographical location environmental characteristic index, as a feature of the surrounding environment, provides information on the spatial environment characteristics of the site at a macro scale and is highly systematic. Visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view images, as features of site conditions, provide visual material environmental characteristics of the site from a human perspective and are highly accurate. The two types of characteristics corresponding to the surrounding environment and site conditions complement each other, enabling a comprehensive and multifaceted measurement of the site selection potential characteristics of micro-green spaces.

[0135] Taking Nanjing's main urban area as an example, the geographical location environment characteristic index system includes: a) the surrounding recreation demand intensity index; b) the surrounding retail facility density distribution index; c) the surrounding catering facility density distribution index; d) the surrounding public toilet density distribution index; and e) the surrounding existing park competition effect index. Among these, index a reflects the recreation demand level of the surrounding environment of potential site selection points; indices b, c, and d reflect the facility conditions of the surrounding environment of potential resettlement sites and their supporting effect on small-scale green space services; and index e reflects the green space construction conditions of the surrounding environment of potential resettlement sites and their competitive effect on small-scale green space services.

[0136] A) Geographic location and environmental characteristics index in the case of Nanjing's main urban area:

[0137] 1) Demand for recreation in the surrounding area: The core function of micro-green spaces is to provide daily recreational services for urban residents within a five-minute living circle (approximately 300 meters). Therefore, the larger the population within the service area, the higher the demand for green spaces, and the greater the potential for incorporating micro-green spaces.

[0138] 2) Supporting effect of surrounding recreational facilities: Small green spaces are small in scale and cannot use internal service facilities like large parks. Therefore, they need the support of existing recreational service facilities in the surrounding built environment. The denser the distribution of existing recreational service facilities, the stronger the supporting effect on small green spaces, and the greater the potential for incorporating small green spaces.

[0139] 3) Competition Effect of Surrounding Existing Green Spaces: As a component of the urban green space system, small green spaces are subject to competition from surrounding green spaces during their recreational services. Existing research shows that the larger the area of ​​surrounding green spaces and the closer they are, the stronger the competitive effect of those green spaces on the recreational services of the small green spaces; conversely, the smaller the area of ​​surrounding green spaces and the farther they are, the weaker the competitive effect of those green spaces on the recreational services of the small green spaces.

[0140] B) Advanced visual semantic feature index of site conditions in the Nanjing main urban area case:

[0141] Sky openness: The sky openness index reflects the lighting conditions of a site. Generally speaking, the higher the sky openness, the more sunlight the area receives, and the more suitable it is for the introduction of small green spaces.

[0142] Enclosure: Generally speaking, street spaces with a high degree of enclosure provide a sense of comfort and shelter. Studies have shown that streets with a strong sense of enclosure offer a greater sense of security and are more suitable for the integration of small green spaces.

[0143] Pedestrian feasibility index: This reflects how pedestrian-friendly a street is and how attractive it is to walk-in visitors. It can be assessed using indicators such as the number of motor vehicles and the proportion of motor vehicle lanes. Areas with higher feasibility tend to have a greater need for small green spaces.

[0144] Crowd Gathering Index: The higher the crowd gathering density, the greater the traffic flow to the venue, and the greater the potential demand for small green spaces that provide recreational services.

[0145] Green space ratio: The more complete the existing green space construction of the site, the stronger the competitive effect and the smaller the potential for incorporating small green spaces.

[0146] 2.3 Based on urban land use vector files, urban road network vector files, and information from the SouFun website, geographic information system platforms, intelligent location service platform software, crawling software, and geographic conversion software are used to calculate the geographic location environment characteristic index.

[0147] 1) Calculation of Green Space Demand Intensity Index: In calculating the green space demand intensity, the names, total number of houses, and addresses of each community within the planning area are crawled from the SouFun website using web scraping software. Geographic conversion software is then used to convert the addresses into latitude and longitude coordinates. The population size index is obtained by multiplying the total number of houses by the average number of households per capita in the area. Finally, the total population of different blocks is obtained using the distribution of community point data and population size in the geographic information system platform, with population size reflecting the green space demand intensity.

[0148] The population size of all residential areas within the planning scope was selected as the research object based on the potential characteristic index system.

[0149] Using a data collector, we crawled a dataset from the SouFun website, collecting the names, information links, geographic data, total number of houses, and housing types for all residences within the research area.

[0150] The data was exported as an Excel spreadsheet, and problematic data was manually deleted and modified based on the scraped links.

[0151] The average number of households in each area within the planning scope was obtained based on the data from the seventh national census. The total number of houses in each community was multiplied by the average number of households in each area to estimate the relatively accurate population size of each residential area.

[0152] Geographic data of each residential location is converted into latitude and longitude coordinates using geographic conversion software.

[0153] Import the city road network vector file and population data Excel file into the geographic information system platform, define the projection coordinates, and perform topological relationship checks and processing.

[0154] Add a field MJ (Area) to the layer property table. Use computational geometry tools to calculate the area (in hectares) of each facet. Select faces with obvious errors (area values ​​< 2) and remove them using the elimination tool to obtain a new facet data layer. Use the face-to-line tool to reverse the process, converting the facet data layer into individual vector line segments.

[0155] Open the attribute table of the line layer obtained by converting faces to lines, add a field XHB (line merge), use the field calculator tool, select the Python parser, set the type to numeric, and use the following formula:

[0156] XHB=str(!LEFT_FID!)+str(!RIGHT_FID!)

[0157] XHB is the ID field of each newly defined vector line, LEFT_FID is the FID field of the vector face to the left of the vector line, and RIGHT_FID is the FID field of the vector face to the right of the vector line. This operation assigns the same XHB field to two or more broken lines sandwiched between two vector faces, so that the two or more broken lines sandwiched between two faces have the same XHB value, which is convenient for subsequent broken line merging.

[0158] Using the fusion tool, input the line layer of the trunk road network, merge fields with the same FID, export a shapefile, and obtain a relatively independent and continuous road network vector line segment layer.

[0159] Python was used to calculate the population size of each block (road-enclosed area). Potential site locations along the street within a block were assigned the block's population size, while potential site locations along the main road network were assigned the average population size of the plots on both sides of the road. The final result is a population size index for each potential site (gif1). The higher the index, the greater the residents' demand for green space, and the greater the relative potential for incorporating small green spaces.

[0160] 2) Calculation of the Support Effect Index of Surrounding Recreation Facilities: Points of Interest (POI) location data for public toilets, restaurants, and retail facilities within the main urban area of ​​Nanjing were extracted using the JavaScript Application Programming Interface (JAPI) provided by Baidu Maps. Kernel density analysis tools were then used in the Geographic Information System (GIS) platform to calculate the density of various related POI points within the planned area, obtaining the density distribution index of different POI facilities. Specifically, this represents the density of point features within a 300-meter radius around each output raster pixel, with the point weight decreasing with distance from the center pixel. The resulting density distribution indices are: Retail POI density distribution index (GIF2), Restaurant POI density distribution index (GIF3), and Public Toilet POI density distribution index (GIF4).

[0161] Based on the provisions of the Park Design Code (GB 51192-2016) regarding the types of recreational service facilities in parks and green spaces, we extracted three types of recreational service facilities in the built environment: retail, catering, and public toilets, and analyzed their supporting effects on small green spaces.

[0162] Register as a software developer for the intelligent location service platform to obtain a key for POI crawling.

[0163] In the geographic information system platform, the planning area is first divided into zones, and then the polygon search POI method in the platform's web service API is used to crawl data in order to deal with the problem of the upper limit of the number of crawled data.

[0164] Convert the JSON format results returned by the webpage into an Excel spreadsheet, and then convert the spreadsheet content into point data shapefiles in the geographic information system platform.

[0165] Based on the "Urban Green Space Planning Standard (GB / T 51346-2019)" and the "Urban Residential Area Planning and Design Standard (GB50180-2018)," the service radius of micro-green spaces is set at 300 meters (a 5-minute living circle). Using a kernel density analysis tool on a geographic information system platform, the density of various relevant Points of Interest (POIs) within a 300-meter radius (5-minute living circle) of potential sites within the research area is calculated. The analysis results represent the density of point features within a certain radius around each output raster pixel. The weight of a point decreases with distance from the center pixel, resulting in retail POI density distribution index GIF2, catering POI density distribution index GIF3, and public toilet POI density distribution index GIF4. These indices represent the influence of POIs within a 300-meter radius of a given location; the higher the value, the greater the relative potential for inclusion in a micro-green space.

[0166] 3) Calculation of the competition effect index of existing green space in the surrounding area: In the analysis of green space competition effect, based on the urban land use vector file, the Euclidean distance tool is used in the geographic information system platform to linearly attenuate the competition effect of different levels of green space within its service radius according to distance, and the superposition is used to obtain the competition effect index of existing green space within the planning area.

[0167] Based on the characteristic description, the research objects were selected as urban parks and green spaces, including comprehensive parks, free specialized parks, community parks, and street parks.

[0168] The urban land use vector file is imported into the geographic information system platform for data processing, including defining projection coordinates and checking and processing topological relationships.

[0169] Based on the "Urban Green Space Classification Standard (CJJ / T 85-2017)" and the distribution of urban green space in the green space planning documents of the study area, the information was screened and corrected, leaving out comprehensive parks, free specialized parks, community parks and street parks.

[0170] Based on the Nanjing City land use map, comprehensive parks, free specialized parks, community parks, and street parks within the main urban area were extracted as existing open recreational green spaces in the city. According to the relevant provisions of the "Urban Green Space Planning Standard (GB / T 51346-2019)" regarding the service radius of different levels of green space, the service radius of comprehensive parks was set at 1000 meters, community parks and specialized parks at 500 meters, and street parks at 300 meters. Distance attenuation simulations were performed to assess the competitive effect of different park green spaces within their service radii, yielding the recreational competition effects (grif1, griff2, griff3) of each existing park green space on potential sites of small green spaces within its service radius. This effect can be defined by the following formula:

[0171]

[0172] Where S is the area of ​​park i, D is the distance between park i and study point j, and α and β are the coefficients related to the park area and distance, respectively.

[0173] Using the grid calculator tool, the GRIFF images were weighted according to their competition effect and then superimposed. The stronger the competition effect of the park green space, the higher the weight. The results were then normalized to obtain the park green space competition effect index (gif5). The larger the result value, the smaller the relative potential for placing small green spaces.

[0174] Construct a geographic location environmental feature vector for each potential site of a small green space, represented by a GIF, i.e.:

[0175] GIF={gif1,gif2,gif3,gif4,gif5}

[0176] Use GIF k Let GIF_set represent the geographic location environmental feature vector of the k-th potential site for small green spaces among all potential site selection points. Construct a set of geographic location environmental feature vectors for all potential site selection points of small green spaces, with a total number of n, denoted by GIF_set.

[0177] GIF_set = {GIF1, ..., GIF n}

[0178] 2.4. Using a pre-trained deep learning image semantic segmentation model, semantic segmentation is performed on street view images from three angles for each potential site of a small green space. The proportion of visual semantic elements in the three images is obtained and used to calculate the proportion of visual semantic elements for each potential site of a small green space. Each visual semantic element of each potential site of a small green space is used as a feature dimension to construct the visual semantic element feature vector (VSF) for each potential site of a small green space. The result is the visual semantic element feature vector set (VSF_set) for all potential sites of small green spaces. This allows for the quantitative representation of visual semantic features related to site selection potential along the street using objective street view image information.

[0179] Based on the MXNet deep learning framework, a PSPnet image semantic segmentation model was pre-trained on the Cityscapes dataset (a deep learning image semantic segmentation model pre-trained on an open-source city street view dataset). The Cityscapes dataset provides 34 categories of street view visual semantic elements. In the application scenario of small green space site selection, 14 semantic elements directly related to the measurement of the street site environment were selected as the visual semantic elements related to the site selection potential of small green space sites. These are: a) vehicular roads b) pedestrian roads c) buildings d) walls e) medians f) pillars or streetlights g) traffic lights h) traffic signs i) greenery j) terrain undulations (such as peaks, mounds, rocks, etc.) k) sky l) pedestrians m) riders n) vehicles (including cars, trucks, buses, rail transit vehicles, motorcycles, and bicycles)

[0180] Using a semantic segmentation model, visual semantic segmentation analysis was performed on three street view images from three angles for each potential site location of a small green space in the site selection point street view image dataset. The proportion of each visual semantic element's pixel count to the total pixel count of the entire image was calculated sequentially. This represents the proportion of image pixels of the i-th visual semantic element at a potential location of a small green space in a direction with a relative angle θ to the road, relative to the total number of pixels in the entire street view image. Based on a weighted average method, ω... θ The weight of the semantic elements of the street view image in the direction of relative angle θ to the semantic elements of potential sites for small green spaces is used to define the proportion of the i-th semantic element of a potential site for a small green space:

[0181]

[0182] Based on the formula for the proportion of the i-th semantic element of potential site selection points for small green spaces, the vsf of each potential site selection point for small green spaces obtained from semantic segmentation analysis is input. θ,i Calculate the proportion of 14 semantic elements for each potential site selection point of a small green space, i.e., 14 vsf. i .

[0183] Based on the semantic element type, a visual semantic element feature vector is constructed for each potential site selection point of a small green space, using VSF = (vsf1, vsf2, ..., vsf...). 14 This means that this feature vector is used to quantify the visual semantic features related to the site selection potential of a small green space.

[0184] Based on the definition of feature vectors of visual semantic elements, VSF is used. kLet VSF_set represent the visual semantic feature vector of the k-th potential site of all potential small green spaces. Then, we can construct a set of visual semantic feature vectors of all n potential small green space sites, denoted by VSF_set, i.e.:

[0185] VSF_set = {VSF1,...,VSF n}

[0186] 2.5. Based on the fundamental principles of small-scale green space planning and site selection, and considering the available data types and precision, and taking into account the "high correlation" between advanced visual semantic information of site conditions and the site selection potential of small-scale green spaces, an advanced visual semantic feature index system for site conditions is established. This system measures the interrelationships between the proportions of various visual semantic elements in a site, describing the advanced characteristics of the site's spatial environment. The type of advanced visual semantic feature index for site conditions is defined according to the site characteristic description requirements for small-scale green space selection. Using the visual semantic element feature vectors of all potential site selection points for small-scale green spaces, the advanced visual semantic feature index values ​​for each potential site selection point are calculated, constructing an advanced visual semantic feature index dataset AVSF_set. Each advanced visual semantic feature element of each potential site selection point for small-scale green spaces is treated as a feature dimension, constructing an advanced visual semantic feature vector AVSF for each potential site selection point. This yields the advanced visual semantic feature vector set AVSF_set for all potential site selection points for small-scale green spaces, thus enabling the quantitative representation of advanced visual semantic features related to site selection potential along streets using objective street view image information.

[0187] Based on the fundamental principles of small-scale green space planning and design, and the available data types and precision, an advanced visual semantic feature index system for site conditions is established. This advanced visual semantic feature index is a characteristic index calculated using the proportions of various visual semantic elements within the basic site, targeting the identification requirements for small-scale green space site selection potential. Users can independently define the advanced visual semantic feature index for site conditions to identify the site selection potential of small-scale green spaces, based on the site selection needs of the city under study. Taking Nanjing's main urban area as an example, the advanced visual semantic feature index for site conditions includes: a) sky openness; b) site enclosure; c) pedestrian feasibility index; d) crowd gathering index; and e) green view rate.

[0188] Combined with the proportion of the i-th semantic element of a potential site for a small green space (psf) iAnd 14 semantic elements directly related to the measurement of street site environment, namely: a) vehicular roads; b) pedestrian walkways; c) buildings; d) walls; e) medians; f) pillars or streetlights; g) traffic lights; h) traffic signs; i) greenery; j) terrain undulation; k) sky; l) pedestrians; m) riders; n) vehicles. A specific mathematical calculation formula is defined for the advanced visual semantic feature index of site conditions determined in the Nanjing main urban area case in section 1):

[0189] Sky openness, denoted by SVF:

[0190] SVF = vsf 天空

[0191] The enclosure degree of the site is represented by ED:

[0192]

[0193] Pedestrian Feasibility Index, denoted by PFI:

[0194] PFI = 1 + (vsf) 行人 +vsf 人行道路 )-(vsf 车行道路 +vsf 交通信号灯 +vsf 交通工具 )

[0195] Crowd Aggregation Index, denoted by CCI:

[0196] CCI = vsf 行人 +vsf 骑手

[0197] Green visibility, denoted by GLR:

[0198] GLR=vsf 绿化

[0199] Based on the advanced visual semantic feature index of site conditions, a specific mathematical calculation formula is used. Using the visual semantic element feature vectors of all potential site selection points for small green spaces, the advanced visual semantic feature index of each site condition for each potential site selection point is calculated, and AVSF is used as the formula. i Let represent the high-level visual semantic feature index of the i-th site condition. For example, for the five high-level visual semantic feature indices of site conditions determined in the case of Nanjing's main urban area, the high-level visual semantic feature index of a certain site condition is represented as follows:

[0200]

[0201] Based on the advanced visual semantic feature indices of various site conditions, an advanced visual semantic feature vector is constructed for each potential site selection point of a small green space. If a total of m advanced visual semantic feature indices of site conditions are defined, then AVSF = (avsf1, avsf2, ..., avsf...) is used. m This feature vector is used to quantify the high-level visual semantic features of site conditions related to the site selection potential of a small green space.

[0202] By using AVSF k Let represent the high-level visual semantic feature vector of site conditions for the k-th potential site of all small green spaces. Then, we can construct a dataset of high-level visual semantic feature vectors of site conditions for all potential site locations of small green spaces, with a total of n points, denoted by AVSF_set, i.e.:

[0203] AVSF_set = {AVSF1, ..., AVSF n}

[0204] 2.6. By combining the geographic location environmental feature vectors, visual semantic element feature vectors, and site condition high-level visual semantic feature vectors of all potential site selection points for street-side micro-green spaces obtained in 2.3, 2.4, and 2.5, a comprehensive potential feature vector (CPF) related to the site selection potential of all potential site selection points for micro-green spaces is obtained, and a comprehensive potential feature vector dataset CPF_set is constructed for potential site selection points of micro-green spaces. This dataset will be used for training conventional machine learning models and for identifying the suitability of micro-green spaces for placement, so as to achieve a complete and accurate representation of the characteristics of potential site selection points for micro-green spaces using quantitative methods, and to establish a quantitative relationship between the objective and analyzable characteristics of potential site selection points for micro-green spaces and the site selection potential of micro-green spaces that needs to be evaluated in urban green space construction.

[0205] For all potential sites of small green spaces, the geographic location environment feature vector (GIF) obtained in 2.3, the visual semantic element feature vector (VSF) obtained in 2.4, and the site condition high-level visual semantic feature vector (AVSF) obtained in 2.5 for each potential site of small green spaces are directly concatenated to obtain the comprehensive potential feature vector, and cpf is defined. i Let represent the i-th comprehensive potential feature of a potential site for a small green space. Define the operator [a, b, c] to represent the concatenation operation of vectors a, b, and c. Thus, the comprehensive potential feature vector is represented by CPF, i.e.:

[0206] CPF=[GIF,VSF,AVSF]=(gif1,...,gif g ,vsf1,...,vsf14 ,avsf1,...,avsf m )

[0207] =(cpf1,...,cpf 14+g+m )

[0208] Based on the definition of the comprehensive potential eigenvector, using CPF k Let CPF_set represent the comprehensive potential feature vector of the k-th potential site among all potential sites for small green spaces. Then, we can construct a dataset of comprehensive potential feature vectors for all potential sites for small green spaces, with a total of n nodes, denoted as CPF_set.

[0209] CPF_set = {CPF1, ..., CPF} n}

[0210] 2.7. Visual scene information of potential sites for small green spaces is the main basis for identifying their potential, and it needs to be represented as accurately and completely as possible. In steps 2.1 to 2.6 above, the comprehensive potential information of all potential sites for small green spaces was quantified using manual feature engineering. The basic visual semantic features and the advanced visual semantic features of site conditions are both methods of representing the potential information of potential sites for small green spaces, abstracted and generalized manually. To further explore the potential information of potential sites for small green spaces that is difficult to abstract and generalize manually but is implicitly present in the visual scene, the street scene image dataset Graph_set of all potential sites for small green spaces is constructed as part of the potential site selection feature index system for small green spaces, using G... k,θ The street view image representing the k-th potential site for a small green space, with a relative angle θ between the street view shooting direction and the road it is located on, is represented by a Graph. k Graph represents the complete visual scene depicted by three street view images of the k-th potential site for a small green space. k ={G k,75° G k,90° G k,105° Then we can construct a complete visual scene dataset Graph_set containing n potential site locations for all small green spaces:

[0211] Graph_set={Graph1,Graph2,...,Graph n}

[0212] Establishment of a quantitative identification system for the site selection potential of 3 micro-green spaces

[0213] Based on the data obtained in step 1, these data were processed and analyzed to obtain step 2, the potential characteristic index system for small and micro green space site selection. Based on the characteristics of different levels of potential characteristic index systems in the potential characteristic index system for small and micro green space site selection, the potential identification characteristics of potential site selection points of small and micro green spaces in different directions implied by the potential characteristic results of each level were combined to construct an abstract quantitative identification system for the potential of small and micro green space site selection.

[0214] 3.1 To replace the traditional screening technology of "scorching potential sites for small green spaces after conducting manual on-site surveys", a two-layer automatic screening system is constructed, consisting of "basic suitability screening for small green spaces + deep suitability screening for small green spaces based on deep feature perception", which is used to screen the comprehensive potential of potential sites for small green spaces to supplement the site selection.

[0215] 3.1.1 Using the comprehensive potential feature vector (CPF) of all potential site selection points for small and micro green spaces constructed in 2.6, the comprehensive potential information of the site selection points for small and micro green spaces is quantitatively represented. In view of the "comprehensiveness" of the comprehensive potential information, the "suitability identification of the basic site selection for small and micro green spaces" method is proposed to make a basic multi-classification judgment on the site selection suitability of potential site selection points for small and micro green spaces, and to initially distinguish the basic site selection suitability level of each potential site selection point for small and micro green spaces. This will be explained in detail in 3.2.

[0216] 3.1.2 Based on the complete visual scene dataset of all potential small green space sites obtained in 2.7, a deep learning method based on image fully convolutional neural networks is applied to mine the depth perception features of the street scene of small green space sites. This feature is used to quantify the depth perception information related to the site selection potential of small green spaces, which is difficult to extract manually. In view of the "difficult-to-interpret" nature of depth perception information, a method of "deep feature perception-based deep suitability identification for small green space placement" is proposed. This method explores in depth the impact of the perception information hidden in the street scene images of small green spaces on their placement suitability. It assists the multi-class classification of placement basic suitability in 3.1.1 and effectively prevents the omission problem caused by the incomplete feature information extracted manually. This will be explained in detail in 3.3.

[0217] 3.2 To implement the "suitability identification of small green space placement" method defined in 3.1.1, based on the sampling of representative training sample site selection points, conventional machine learning methods are used to train a multi-classification model of the suitability of small green space placement on the comprehensive potential feature vector (CPF) of the training sample site selection points. The specific training implementation scheme will be explained in section 4. The model is applied to all potential site selection points of small green spaces, a basic potential score system is defined, and the model judgment results are transformed into basic potential scores of potential site selection points of small green spaces.

[0218] 3.2.1 Based on the "suitability identification of small green space placement" method proposed in 3.1.1, a multi-classification model for the suitability of small green space placement is trained using conventional machine learning training methods to target the comprehensive potential feature vector CPF.

[0219] 3.2.2 Using the model in 3.2.1, the suitability of the basic infrastructure for all potential small green space sites is initially assessed. This yields the suitability level for each potential site. Users can independently classify the suitability levels for potential small green space sites based on the site selection needs of the city under study. A total of L levels are defined, denoted by {1,2,...,L} (where L is a positive integer). The higher the level, the greater the suitability of the basic infrastructure for the potential small green space site. This is represented by y. k Let y represent the multi-classification model judgment result of the basic suitability of the k-th potential site for small green space, that is, the basic suitability level of the k-th potential site for small green space. k The domain is:

[0220] y k ∈{1,2,...,n}, where n is a positive integer

[0221] Taking the case of Nanjing's main urban area as an example, the suitability level of potential sites for small green spaces is divided into three levels, with {1,2,3} representing levels 1 to 3. The higher the level, the greater the suitability of potential sites for small green spaces.

[0222] 3.2.3 Calculate the basic potential score of potential small green space sites based on their suitability level. If the suitability level is 1, the basic potential score is 0. If the suitability level is L, the basic potential score is full (100). If the suitability level is l, where l is between 1 and L, the score is... Using BaseScore k Let y represent the basic potential score of the kth potential site for a small green space. k (y k Let ∈{1,2,...,L}, where L is a positive integer) represent the multi-classification model judgment result of the basic suitability of the k-th potential site for small green space placement, i.e., the basic suitability level of the k-th potential site for small green space placement. Then, the basic potential score of the k-th potential site for small green space placement is defined as:

[0223]

[0224] Taking Nanjing's main urban area as an example, the basic suitability level of potential small green space sites is divided into three levels, denoted by {1,2,3}, representing levels 1 to 3. The basic potential score of the k-th potential small green space site is defined as:

[0225]

[0226] 3.3 To implement the "deep suitability identification of small green space placement based on depth feature perception" method defined in 3.1.2, based on the sampling of representative training sample site selection points, a deep learning method is used to train a multi-classification model of small green space placement depth suitability for street view images of the training sample site selection points. The specific training implementation scheme will be explained in 4. The model is applied to all potential site selection points of small green spaces, a depth potential score system is defined, and the model judgment results are converted into depth potential scores of potential site selection points of small green spaces.

[0227] 3.3.1 Based on the "deep feature perception-based identification of placement depth suitability of small green spaces" method proposed in 3.1.2, a multi-classification model of placement depth suitability for all potential site selection points of small green spaces is trained based on deep learning training method. This depth discrimination model can effectively mine the deep hidden features in the street view image.

[0228] 3.3.2 Using the model in 3.3.1, we deeply mine the street view image features of all potential sites for small green spaces, and simultaneously determine the suitability of the placement depth of these potential sites, obtaining the suitability level of the placement depth. Users can independently classify the suitability level of the placement depth of potential small green spaces according to the site selection needs of the research city. There are a total of L levels, denoted by {1,2,...,L} (L is a positive integer). The higher the level, the greater the suitability of the placement depth of the small green space. (The last part, "z", appears to be a typo and can be omitted.) k Let z represent the result of the multi-class classification model for the suitability of the placement depth of the k-th potential site for small green spaces, i.e., the suitability level of the placement depth of the k-th potential site for small green spaces. k The domain is:

[0229] z k ∈{1,2,...,L}, where L is a positive integer

[0230] Taking the case of Nanjing's main urban area as an example, the suitability level of the placement depth of potential small green space sites is divided into 3 levels, with {1,2,3} representing the first to third levels. The higher the level, the greater the suitability of the placement depth.

[0231] 3.3.3 Calculate the depth potential score of potential site selection points for small green spaces based on their placement depth suitability level. If the placement depth suitability level of a potential site selection point is 1, its depth potential score is 0 points. If the placement depth suitability level of a potential site selection point is L, its depth potential score is full (100). If the placement depth suitability level of a potential site selection point is l, where l is between 1 and L, the score is... Using DeepScore k The depth potential score of the k-th potential site for a small green space is represented by z. k (z k Let ∈{1,2,...,L}, where L is a positive integer) represent the multi-classification model judgment result of the placement depth suitability of the k-th potential site for small green spaces, i.e., the placement depth suitability level of the k-th potential site for small green spaces. Then, the depth potential score of the k-th potential site for small green spaces is defined as:

[0232]

[0233] Taking Nanjing's main urban area as an example, the suitability level of potential small green space locations is divided into three levels, denoted by {1,2,3}, representing levels 1 to 3. The depth potential score of the k-th potential small green space location is defined as:

[0234]

[0235] 3.4 Combine the basic potential score system defined in 3.2 with the deep potential score system defined in 3.3, and calculate the weighted average of the basic potential score and the deep potential score of each potential site for small green space obtained in 3.2 and 3.3. Users can determine the weighting weights based on the urban environmental characteristics of the research object and the functional development focus of the small green space to obtain the final potential score of each potential site for small green space.

[0236] 3.4.1 Using BaseScore k The DeepScore represents the basic potential score of the k-th potential site for a small green space. k Let represent the deep potential score of the k-th potential site for a small green space, and let w represent the weight of the basic potential score. Then the weight of the deep potential score is (1-w). (FinalScore) k Let represent the final potential score of the k-th potential site for a small green space. The final potential score can be defined as follows:

[0237] FinalScore k =w*BaseScore k+(1-w)*DeepScore k

[0238] 3.4.2 Summarize the final potential scores of potential sites for small green spaces, and construct a final potential identification result set (Score) of n for all potential sites for small green spaces. Then:

[0239] Score={FinalScore1,...,FinalScore n}

[0240] 4. Train a model for identifying the site selection potential of small green spaces using conventional machine learning and deep learning. Use this model to establish a stable derivation relationship between the characteristic index system for site selection potential of small green spaces and the quantitative identification system for site selection potential of small green spaces.

[0241] Based on the results of the first three stages, a potential discrimination model was trained using conventional machine learning and deep learning techniques. The model takes the potential characteristic index of potential small-scale green space sites as the basic input information for potential discrimination, and outputs the final potential score of the potential small-scale green space sites as the potential discrimination result. By training the small-scale green space site selection potential discrimination model using conventional machine learning and deep learning, a stable derivation relationship can be established between the small-scale green space site selection potential characteristic index system and the small-scale green space site selection potential quantitative discrimination system.

[0242] 4.1 Training Set Construction Techniques

[0243] 4.1.1 Based on the comprehensive potential feature vector dataset CPF_set of all potential site selection points for small green spaces, the differences and representativeness of potential site selection points for small green spaces are clarified. A representative subset of potential site selection points for small green spaces (taking Nanjing's main urban area as an example, with no fewer than 10,000 points) are selected as training sample site selection points, forming a training sample site selection point set. Within the training sample site selection point set, the comprehensive potential feature vector x of each training sample site selection point is... k Construct the feature vector set Train_X as a machine learning training set.

[0244] Based on the sample size estimation theory of machine learning, the greater the difference between the selected training samples, the better the model training effect. According to the definition of the comprehensive potential feature vector of potential site selection points for small green spaces and the sampling method of street view images of small green spaces, we can propose Hypothesis 1: Among all potential site selection points for small green spaces, the differences between two potential site selection points with relatively large relative distances are greater. If potential site selection points with relatively large relative distances are selected as training samples, the differences between the selected training samples will be relatively large, and each individual training sample will be more representative.

[0245] Based on street view similarity identification of potential site selection points for small green spaces, this study analyzes the pixel distribution similarity (or frequency-weighted intersection-union ratio) of 14 visual semantic elements in the street view images of two potential site selection points L1 and L2 at different spacings. Some pixels are classified as the i-th semantic element in the street view image of L1, while they are classified as the j-th semantic element in the street view image of L2. This is represented by vsf_pn. i,j This represents the number of pixels. Let FWIoU represent the frequency weight intersection-union ratio (FWIoU) of the street view images of two potential site locations for small green spaces, L1 and L2. Then, FWIoU is defined as follows:

[0246]

[0247] Based on the definition of FWIoU, a street view similarity discrimination experiment was conducted on the visual semantic element feature vector set VSF_set for potential site selection points of small green spaces. Different street view sampling intervals were selected, and a certain number (no less than 100 samples) of potential site selection points of small green spaces were sampled as experimental objects. Corresponding visual semantic element feature vectors were extracted from the visual semantic element feature vector set. The average FWIoU between adjacent sampling points was calculated under different street view sampling intervals, and the magnitude of the average FWIoU between adjacent sampling points under different intervals was compared. The conclusion was that the larger the sampling distance, the larger the average FWIoU between adjacent sampling points, indicating that the differences between the selected training samples are relatively large, and a single training sample is more representative, thus confirming Hypothesis 1. Therefore, setting the sampling interval to a value of no less than 50 meters is a suitable sampling interval selection.

[0248] To ensure training effectiveness, a portion of potential site selection points for small green spaces will be selected from all potential site selection points using a sampling interval of no less than 50 meters as training sample site selection points. Simultaneously, the initially selected potential site selection points should meet the basic training quantity requirements (taking Nanjing's main urban area as an example, the initial number should be no less than 10,000). In practice, the initial sample size can be set to 10,000 per 1000 km. 3 Based on the initial model training results, it is determined whether the number of samples needs to be increased. This results in the formation of a training sample site selection set. Within this set, the comprehensive potential feature vector of each training sample site selection point is named x. k Construct the feature vector set of the machine learning training set, named Train_X, and set the sample size of the training sample location set to N (taking the main urban area of ​​Nanjing as an example, N is not less than 10000). Then:

[0249] Train_X = {x1,...,x} N}

[0250] 4.1.2 Based on the street view image dataset Graph_set of all potential site selection points for small green spaces, within the training sample site selection point set constructed in 4.1.1, the three-angle street view image set along the street of each training sample site selection point is used as the original feature map set Gx for that training sample site selection point. k This yields the feature map set of the deep learning training set, named GTrain_X. With the sample size N of the training sample location set set, we have:

[0251] GTrain_X={Gx1,...,Gx N}

[0252] 4.1.3 Using the "Suitability Identification of Small Green Spaces" scheme defined in 3.1 as the reference criterion for manually scoring the site selection potential of the training samples, a manual scoring standard for suitability potential identification is defined. Within the training sample site selection point set of size N (N not less than 10000) obtained in 4.1.1, based on the street view visual information of the potential site selection points of the small green spaces, N... P Each professional observer makes an assessment, manually scoring each potential site selection point in the training samples multiple times. After each round of manual scoring, the scores from all professionals are compared to determine consistency. Simultaneously, the number of samples for each score level must be guaranteed; if the number of samples for a certain score level is insufficient, new training samples are added promptly. Multiple rounds of scoring and result quality assessment are performed until N... P The manual scoring results from individual professionals meet quality requirements (e.g., consistency rate above 90%, and a certain sample size for each score level (e.g., no less than 1000 samples)). This determines the manual assessment results of the suitability of small green space placement for each training sample potential site, which serves as the suitability assessment label for small green space placement for representative potential sites, thus constructing a suitability assessment label set for small green space placement.

[0253] Multiple rounds of manual scoring were conducted. In each round, based on the street view visual information of potential micro-green space sites, professionals observed and judged each potential site in the training set, performing manual visual perception identification and scoring. N P (N P(where L is a positive integer) professionals sequentially perform visual perception observations on street view images of potential training site locations, identifying the placement suitability level of each potential training site location. The placement suitability levels of the potential training site locations are divided into L levels, denoted by {1, 2, ..., L} (where L is a positive integer). If, after manually observing the street view image of a potential training site location, a professional judges the basic suitability of the location to be higher, then the level of that potential small green space location location is assigned a higher score. Let y i,k Let represent the suitability assessment result of the i-th staff member for the potential location of the k-th training sample.

[0254] y i,k =l, the suitability level of the implantation is determined manually to be l

[0255] l∈{1,2,...,L}

[0256] i∈{1,2,...,N P}

[0257] In the multi-round manual scoring process, after each round of scoring, N P Each professional staff member conducts consistency and sample size checks on the identification results of potential site selection points across the entire training sample, up to N. P The manual scoring results from individual professionals must meet quality requirements (e.g., consistency rate above 90%, sample size for each score level must reach a certain scale (e.g., no less than 1000 samples each)). The specific methods for consistency and sample size verification at each level, along with the iterative updates to the scoring rules, are as follows:

[0258] N P A professional manually scores N training samples, using Mp k In this round of manual scoring, N represents N. P Among the results of human suitability assessment (i.e., human scoring) of potential site selection points for the k-th training sample by professionals, the most prevalent human scoring result p is... k The number of [number] scores. Let ρ represent the threshold percentage for determining consistency (e.g., ρ = 90%). If the most prevalent human-scored results have a p-value... k Number Mp k N, accounting for the total number of professional human scoring results P If the proportion of a given score is greater than the threshold ρ for consistency, then the human scoring results for the potential location points of the k-th training sample are considered consistent. Therefore, the consistency rate η of the human scoring results is defined and calculated as follows:

[0259]

[0260] Determine whether the consistency rate meets the requirements (e.g., consistency rate reaches 90% or more, i.e., η≥90%). If the requirements are not met, identify the site selection points with inconsistent scoring results for discussion, discuss the scoring rules and standards, further unify the manual scoring rules, and re-extract potential site selection points from the training samples for scoring.

[0261] Statistics N P The average number of samples for each level score in the professional scoring results is calculated. To ensure that the number of samples for each level score meets the requirements (e.g., the number of training samples for each level score is not less than 1000), if the average number of samples for a certain level score is significantly insufficient (e.g., less than 1000), new training samples are added in a timely manner (e.g., one-tenth of the current sample size is added).

[0262] After multiple rounds of scoring and result quality assessment, the scoring rules and standards are iteratively updated once the consistency rate meets the requirements (e.g., consistency rate reaches 90% or more) and the sample size for each level score reaches a certain scale (e.g., no less than 1000 samples each).

[0263] After the final iterative updates to the scoring rules and standards are completed, N P The average of the scores given by professionals to all training samples is rounded down to determine the final human suitability assessment result for potential site selection points in the training samples. This result is represented by y. i,k Let y represent the final human score given by the i-th professional to the k-th training sample. k Let represent the final human suitability assessment result (i.e., human scoring result) of the potential location point of the k-th training sample. Then:

[0264]

[0265] The final human suitability assessment results of all potential site selection points in the training samples are used to construct a score label set for the training set, named Train_y. The sample size of the potential site selection point set is set to N. Then:

[0266] Train_y = {y1,...,y} N}

[0267] 4.1.4 The score label set Train_y obtained in 4.1.3 is matched with the feature vector set Train_X obtained in 4.1.1 to form the training set Train_ML for conventional machine learning.

[0268] Set each label y in the rating label set k Each feature vector x in the feature vector set Train_X of the machine learning training set kA one-to-one correspondence, as a complete training data point, is denoted as t. k If the operator [a, b] is defined to represent the concatenation operation of vector a and scalar b, then:

[0269] t k =[x k ,y k ]

[0270] Construct a complete machine learning training set, named Train_ML, and set the sample size of the machine learning training set to N. Then:

[0271] Train_ML=(Train_X,Train_y)={t1,...,t N}

[0272] 4.1.5 The score label set Train_y of the training set obtained in 4.1.3 is matched with the feature map set GTrain_X of the deep learning training set obtained in 4.1.2 to form the deep learning training set Train_DL.

[0273] Set each label y in the rating label set k Each feature map group Gx in the feature map set of the deep learning training set k A one-to-one correspondence is used to construct a deep learning input data set, denoted as Gt. k

[0274] Construct a complete deep learning training set, named Train_DL, and set the sample size of the deep learning training set to N. Then:

[0275] Train_DL=(GTrain_X,Train_y)={Gt1,...,Gt N}

[0276] 4.2 Model Training Techniques:

[0277] 4.2.1 Based on the basic suitability identification requirements for small-scale green space placement defined in section 3, conventional machine learning methods are selected to model the problem. A classification model for the basic suitability of small-scale green space placement based on multi-class support vector machine is constructed. The small-scale green space placement basic suitability classification model based on support vector machine is trained on the machine learning training set Train_ML constructed in section 4.1.4. The model performance is optimized to obtain the basic suitability identification model for small-scale green space placement.

[0278] Principal component analysis is used to reduce the dimensionality of feature vectors in the conventional machine learning feature vector set Train_X, thereby optimizing the features.

[0279] Choosing the cross-validation partitioning method will result in Train_ML being alternately partitioned into training and test sets during training.

[0280] Support vector machines (SVMs) and classification models are chosen as the underlying models for the suitability assessment of small and micro green space placement. A multi-class SVM model containing parameters to be trained is defined.

[0281] At the same time, the general hyperparameter optimization methods provided by sklearn (such as GridSearchCV and RandomizedSearchCV) are called to select the hyperparameters of the model.

[0282] Input the feature vector set Train_X and the rating label set Train_y from the machine learning training set Train_ML, perform conventional machine learning training, construct the soft-margin loss function of the multi-class support vector machine based on the principle of convex optimization problem, then construct the dual problem, solve the optimization problem of the loss function, determine a set of optimal training parameters for the multi-class support vector machine, and construct the basic suitability identification model for the placement of small green spaces.

[0283] Observe the results of cross-validation to observe the effect of the basic suitability identification model for small green spaces.

[0284] 4.2.2. Based on the depth-feature-aware small green space placement depth suitability identification requirement defined in section 3, a deep learning method is selected to model the problem. A small green space placement depth suitability classification model based on a fully connected image convolutional neural network is constructed. The small green space placement depth suitability classification model based on a fully connected image convolutional neural network is trained on the deep learning training set Train_DL constructed in section 4.1.5. The model performance is optimized to obtain a small green space placement depth suitability identification model based on depth feature awareness.

[0285] A fully connected image convolutional neural network was built using deep learning framework tools.

[0286] Input the feature map set GTrain_X and the rating label set Train_y from the deep learning training set Train_DL, and perform deep learning training. Based on the backpropagation principle of deep neural networks, adjust the hyperparameters of the fully connected image convolutional neural network to train the fully connected image convolutional neural network to classify feature images in multiple ways. This allows the fully connected image convolutional neural network to deeply mine the visual implicit information in street scene images, and finally obtain a better fully connected image convolutional neural network multi-classification model, thereby obtaining a deep feature perception-based model for the deep suitability of small green space placement.

[0287] 5. The application model is used to score, rank, and select potential sites for all small green spaces.

[0288] In step 2, a characteristic index system for the site selection potential of all small green spaces has been obtained. Based on the potential characteristic index results of all potential site selection points of small green spaces, the basic suitability identification model and deep suitability identification model trained in step 4 are applied to implement the quantitative identification system for the site selection potential of small green spaces in step 3. This ultimately enables the scoring, ranking, and selection of all potential site selection points of small green spaces.

[0289] On the CPF_set dataset of comprehensive potential feature vectors of all potential sites for small green spaces, a preliminary suitability classification of each potential site is performed using the basic suitability identification model for small green space placement. Based on the basic suitability identification method for small green space placement defined in section 3, the basic potential score BaseScore of each potential site for small green space is calculated according to the preliminary suitability classification results (i.e., the basic suitability level of each potential site for placement).

[0290] On the street view image dataset of all potential sites for small green spaces, a depth suitability discrimination model based on depth feature perception is used to classify the site suitability depth of each potential site for small green spaces. Based on the depth suitability discrimination method for small green spaces defined in section 3, the depth potential score (DeepScore) of each potential site for small green spaces is calculated according to the depth suitability classification result of each potential site for small green spaces (i.e., the depth suitability level of each potential site for small green spaces).

[0291] By selecting appropriate weights, the DeepScore and BaseScore of each potential site for small green spaces are weighted and averaged to obtain the FinalScore of each potential site for small green spaces. This FinalScore is then used to construct the final potential identification result set Score for all potential sites for small green spaces.

[0292] Based on the final potential identification result set (Score) of all potential site selection points for small and micro green spaces, K-Means clustering is performed on all potential site selection points according to the final potential score (FinalScore). The cluster levels are sorted according to the cluster center scores to obtain the supplementary site selection priority of each potential site selection point's cluster. Finally, the coordinates and site selection priorities of all supplementary site selection points are used as the decision basis for the site selection of small and micro green spaces.

[0293] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0294] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for selecting sites for small-scale green space additions based on machine learning, characterized in that: The method includes the following steps: Data is collected, including the coordinates of all potential site selection points for small green spaces and a street view image dataset of all potential site selection points for small green spaces. The street view image dataset includes all crawling parameters of the street view images of potential site selection points for small green spaces. A site selection potential characteristic index system for small green spaces was established using the collected data. The site selection potential characteristic index system for small green spaces includes geographical location environmental characteristics, visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view images. Based on the different characteristics of the established small green space site selection potential characteristic index system, a quantitative identification system for small green space site selection potential is constructed. Under the small green space site selection potential feature index system, a support vector machine classification model is used as the underlying model to train the basic suitability identification model for small green space placement, and a fully connected image convolutional neural network built based on deep learning framework tools is used as the underlying model to train the deep suitability identification model for small green space placement. The trained basic suitability identification model and deep suitability identification model for small green space placement are used to establish a stable derivation relationship between the small green space site selection potential feature index system and the small green space site selection potential quantitative identification system. The potential identification model for small green space site selection assesses the potential of small green space site selection. From the coordinates of all supplementary site selection points, the model selects the one with the highest potential as the basis for decision-making on small green space site selection based on the results of the potential assessment. The training process of the small-scale green space site selection potential identification model includes the following steps: Constructing the feature vector set of the machine learning training set Construct a score label set for the training set. Construct a feature map set for the deep learning training set. ,Will and Correspondingly, the training set that makes up conventional machine learning ,Will and Correspondingly, they together form the training set for deep learning. ; The resulting suitability assessment model for small-scale green space placement includes the following: Feature vector set of machine learning training set Dimensionality reduction of the feature vectors in the model; Will Divided into training set and test set; We choose a support vector machine (SVM) classification model as the underlying model for the suitability identification model for small green space placement. We define a multi-class SVM model with parameters to be trained, and call the hyperparameter optimization method and random sampling cross-validation provided by sklearn to select the hyperparameters of the model. Input machine learning training set Feature vector set in and rating tag set Training is performed, the loss function is optimized, a set of optimal training parameters for a multi-class support vector machine is determined, and a basic suitability identification model for the placement of small green spaces is obtained. The resulting model for discriminating the suitability of small green spaces based on depth feature perception includes the following: A fully connected image convolutional neural network was built based on deep learning framework tools. Input deep learning training set Feature atlas in and rating tag set Deep learning training was conducted, the hyperparameters of the fully connected image convolutional neural network were adjusted, and the multi-classification ability of the fully connected image convolutional neural network for feature images was trained to obtain a multi-classification model of the fully connected image convolutional neural network, and a depth suitability discrimination model for the placement of small green spaces based on deep feature perception was obtained.

2. The method for selecting sites for small-scale green space additions based on machine learning according to claim 1, characterized in that, The process of establishing the characteristic index system for the site selection potential of small green spaces includes the following steps: The site selection potential characteristic index is divided into geographical location environmental characteristics, visual semantic element characteristics, site condition advanced visual semantic characteristics, and street view; Quantitatively measure geographical location environmental characteristics, visual semantic element characteristics, advanced visual semantic characteristics of site conditions, and street view images to establish a site selection potential characteristic index system. Establish a geographic location environmental characteristic index system for potential sites of micro-green spaces to measure the environmental factors that affect the selection of micro-green spaces and subsequent services in the surrounding environment of potential sites, such as recreational demand, facility conditions, and green space construction conditions. The geographic location environmental range of potential sites of micro-green spaces is also set within 300 meters of the potential sites. Based on data on population, service facilities, and urban land use, a geographical location environmental characteristic index is generated for each selected site, including its surrounding recreational demand, facility conditions, and green space conditions. For each potential site selection location of a small green space, a geographic location environmental feature vector is constructed. Let g geographic location environmental feature indices be defined in total. To represent the geographic location environmental feature vector of a site selection point, the geographic location environmental feature vector is used to represent the geographic location environmental features related to the site selection potential of a small green space potential site. Let the geographic location environmental feature vector of the k-th potential site for small green spaces be represented. Construct a set of geographic location environmental feature vectors for all potential sites for small green spaces with a total of n. This means, that is: Semantic segmentation was performed on the street view images from three angles for each potential site of a small green space to obtain the proportion of visual semantic elements in the three images. Each visual semantic element of each potential site of a small green space was used as a feature dimension to construct a visual semantic element feature vector for each potential site of a small green space. The visual semantic feature vector set of all potential site selection points for small green spaces is obtained. ; Using a deep learning image semantic segmentation model pre-trained on the Cityscapes dataset, visual semantic segmentation analysis was performed on three street view images from three angles for each potential site location of a small green space in the obtained street view image dataset. The Cityscapes dataset provides 34 categories of street view visual semantic elements. Fourteen semantic elements directly related to the street-side environment were selected as visual semantic elements related to the site selection potential of potential small green space sites: vehicular roads, pedestrian walkways, buildings, walls, medians, pillars or streetlights, traffic lights, traffic signs, greenery, terrain undulations, sky, pedestrians, riders, and vehicles. The visual semantic segmentation analysis sequentially calculated the proportion of the number of pixels of each of the 14 visual semantic elements in the three street view images to the total number of pixels in the entire image. This indicates the angle between a potential site for a small green space and the road it is located at. The proportion of the number of image pixels of the i-th visual semantic element in the direction to the total number of pixels in the entire street view image is expressed as... Indicates the relative angle as The importance weight of semantic elements of the street view image in the direction of the potential site of small green space to the semantic elements of the potential site of small green space is defined as: Based on the defined formula for the proportion of the i-th semantic element of potential site selection points for small green spaces, the v of each potential site selection point for small green spaces obtained from visual semantic segmentation analysis is input. The proportion of 14 semantic elements for each potential site of a small green space was calculated, and a feature vector of visual semantic elements was constructed for each potential site of a small green space. express; use This represents the visual semantic feature vector of the k-th potential site for small green spaces among all potential site selection points. A set of visual semantic feature vectors for all potential site selection points of small green spaces with a total size of n is constructed. express: Based on the fundamental principles of small-scale green space planning and design, and the available data types and precision, a high-level visual semantic feature index system for site conditions is established to measure the interrelationships between the proportions of various visual semantic elements in a site and to describe the advanced characteristics of the site's spatial environment. The types of high-level visual semantic feature indices for site conditions are defined according to the site characteristic description requirements for small-scale green space selection. Using the obtained visual semantic element feature vectors of all potential site selection points for small-scale green spaces, the high-level visual semantic feature index values ​​for each potential site selection point are calculated, thus constructing a dataset of high-level visual semantic feature indices for site conditions. Each high-level visual semantic feature element of the site conditions for each potential site of a small green space is used as a feature dimension to construct a feature vector of high-level visual semantic features of the site conditions for each potential site of a small green space. This yields the high-level visual semantic feature vector set of site conditions for all potential small green space sites. , Based on the urban needs of the research subjects, m advanced visual semantic feature indices for site conditions are defined. Using the obtained visual semantic feature vectors of all potential site selection points for small green spaces, the high-level visual semantic feature indices of each site condition are calculated. A high-level visual semantic feature vector of site conditions is constructed for each potential site selection point for small green spaces. If a total of m high-level visual semantic feature indices of site conditions are defined, then... This feature vector represents the high-level visual semantic features of site conditions related to the site selection potential of a small green space. use This represents the high-level visual semantic feature vector of site conditions for the k-th potential site of all small green spaces. A dataset of high-level visual semantic feature vectors of site conditions for all potential site locations of small green spaces, with a total of n, is constructed. express: By combining the geographic location environmental feature vector, visual semantic element feature vector, and high-level visual semantic feature vector of site conditions of all potential site selection points for small green spaces along the street, a comprehensive potential feature vector related to the site selection potential of all potential site selection points for small green spaces is obtained. Construct a comprehensive potential feature vector dataset for potential site selection points of small green spaces. , For all potential sites of small green spaces, the geographic location environmental feature vector of each potential site of small green space will be obtained. The visual semantic feature vectors of each potential site for a small green space were obtained. The high-level visual semantic feature vector of site conditions for each potential site of a small green space was obtained. By directly concatenating the features, we obtain the comprehensive potential feature vector, which is defined as follows: Let the i-th comprehensive potential feature of a potential site for a small green space be represented, and define the operators. Representing vectors ,vector sum vector The concatenation operation, using This represents the comprehensive potential feature vector, namely: This represents the comprehensive potential feature vector of the k-th potential site among all potential site selection points for small and micro green spaces. A dataset of comprehensive potential feature vectors for all potential site selection points for small and micro green spaces with a total size of n is constructed. express: The obtained street view image dataset of all potential site selection points for small green spaces This is part of a system of indexes describing the potential for site selection of small green spaces. The street view shooting direction of the k-th potential site for a small green space is relative to the road it is located at. Street view images, This represents the complete visual scene depicted by three street view images of the k-th potential site for a small green space. Construct a complete visual scene dataset of n potential site selection points for all small green spaces. : 。 3. The method for selecting sites for small-scale green space additions based on machine learning according to claim 1, characterized in that, The process of constructing a quantitative identification system for the site selection potential of small green spaces includes the following steps: Using the constructed comprehensive potential feature vector of all potential site selection points for small green spaces This study quantifies the comprehensive potential information of small green space sites and proposes a method for identifying the suitability of basic infrastructure for small green spaces, taking into account the comprehensive nature of this information. Using the complete visual scene dataset of all potential sites for small green spaces, we mine the depth perception features of the street scene of the site for small green spaces and propose a method for identifying the depth suitability of small green space placement based on depth feature perception. Training yields feature vectors targeting comprehensive potential. A multi-classification model for the suitability of small green spaces; Using feature vectors targeting comprehensive potential A multi-class classification model for the suitability of basic infrastructure for small green spaces was used to preliminarily assess the suitability of basic infrastructure for all potential small green space sites. K-Means cluster analysis was applied to obtain the suitability level of basic infrastructure for each potential small green space site, classifying the potential sites into L levels. To represent levels 1 to L, use Let represent the multi-classification model judgment result of the basic suitability of the k-th potential site for small green space, and let represent the basic suitability level of the k-th potential site for small green space. The domain is: The suitability level of potential sites for small green spaces is classified into Level L, using... To represent levels 1 to L, The basic potential score of a potential site for a small green space is calculated based on its suitability level. If the suitability level is 1, the basic potential score is 0 points. If the suitability level is L, the basic potential score is 100 points. If the suitability level is l, where l is between 1 and L, the score is... ,use The basic potential score of the k-th potential site for a small green space is represented by... Let represent the basic suitability level of the kth potential site for a small green space. Then, the basic potential score of the kth potential site for a small green space is defined as: Then, a multi-classification model for the appropriate insertion depth of street view images of all potential sites for small green spaces was trained. A depth-appropriate multi-classification model was used to deeply mine the street view image features of all potential sites for small green spaces. Simultaneously, the depth-appropriateness of these potential sites was assessed, resulting in depth-appropriateness levels. Based on the site selection needs of the study city, these levels were autonomously divided into L levels. To represent levels 1 to L, use This represents the multi-class classification model's judgment result on the suitability of the placement depth of the k-th potential site for small green spaces, i.e., the suitability level of the placement depth of the k-th potential site for small green spaces. The domain is: The suitability level of the placement depth of potential sites for small green spaces is divided into Level L, using... Let L represent levels 1 to L. The depth potential score of a potential small green space site is calculated based on its suitability level. If the suitability level is 1, the depth potential score is 0. If the suitability level is L, the depth potential score is 100. If the suitability level is l (where l is between 1 and L), the score is... ,use The deep potential score of the k-th potential site for a small green space is represented by... Let represent the suitability level of the placement depth of the k-th potential site for a small green space. Then, the depth potential score of the k-th potential site for a small green space is defined as follows: By combining the defined basic potential score system with the defined deep potential score system, a weighted average is calculated for the basic potential score and the deep potential score of each potential site for small green space. The weighting weights are determined based on the urban environmental characteristics of the research object and the functional development focus of the small green space, and the final potential score of each potential site for small green space is obtained. use The basic potential score of the k-th potential site for a small green space is represented by... The deep potential score of the k-th potential site for a small green space is represented by... If the base potential score represents the weight, then the deep potential score has the following weight: ,use Let represent the final potential score of the k-th potential site for a small green space. Defined as: By summarizing the final potential scores of potential sites for small green spaces, a final potential identification result set of n potential sites for small green spaces is constructed. Then we have: 。 4. The method for selecting sites for small-scale green space additions based on machine learning according to claim 1, characterized in that, The small green space site selection potential identification model uses the potential characteristic index of the potential site selection point of the small green space as the basic input information for potential identification, and uses the final potential score of the potential site selection point of the small green space as the output information for potential identification.

5. The method for selecting sites for supplementing small green spaces based on machine learning according to claim 1, characterized in that, Constructing the feature vector set of the machine learning training set Feature maps of deep learning training sets The process includes the following steps: Using a street view sampling interval of no less than 50 meters, a portion of potential site selection points for small green spaces were selected from all potential site selection points for small green spaces as training sample site selection points. Simultaneously, the initially selected potential site selection points should meet the basic training quantity requirements. The initial sample size was initially set to... Based on the initial model training results, it is determined whether the number of samples needs to be increased, thus forming a training sample site selection set. Within this set, the comprehensive potential feature vector of each training sample site selection point is named... The feature vector set of the machine learning training set is constructed and named Set the sample size of the training sample location set to be [value]. Then we have: Street view image dataset based on all potential site selection points of small green spaces Within the training sample site selection set, the three-angle street view images of each training sample site selection point are used as the original feature image set of the training sample site selection point. This leads to the feature map set of the deep learning training set, named Set the sample size of the training sample location set to be [value]. Then we have: 。 6. The method for selecting sites for supplementing small green spaces based on machine learning according to claim 5, characterized in that, Constructing the score label set of the training set and deep learning training set The process includes the following steps: Using the defined suitability assessment scheme for small-scale green space placement as a training sample, a manual scoring reference criterion for site selection potential is defined, and a manual scoring standard for suitability potential assessment is defined. Sample size and Within the training sample site selection set of a sample size, based on the street view visual information of potential site selection points for micro-green spaces, the selection is made by... Each professional observer makes an assessment, conducting multiple rounds of manual scoring on each potential site selection point in the training samples. After each round of manual scoring, the scores from all professionals are compared to determine consistency. Simultaneously, it is necessary to ensure sufficient sample size for each score level. If the sample size for a certain score level is insufficient, new training samples are promptly added, and multiple rounds of scoring and result quality assessment are performed until... The manual scoring results of each professional meet the quality requirements; thus, the manual identification results of the basic suitability of small green space placement for each training sample potential site are determined, and these results are used as identification labels for the basic suitability of small green space placement for training representative potential sites, thereby constructing a basic suitability identification label set for small green space placement. Multiple rounds of manual scoring were conducted. In each round, based on the visual information of the street view images of potential micro-green space sites, professionals observed and judged each potential site in the training set through manual visual perception identification and scoring. Several professionals sequentially performed visual perception observations on street view images of potential site selection points in the training samples, identifying the placement suitability level of each potential site selection point. The placement suitability levels of the potential site selection points in the training samples were divided into a total of L levels, using... To represent levels 1 to L, If professionals manually observe street view images of potential site selection locations in a training sample and determine that the greater the suitability of the site selection location's foundation, the higher the score will be given to that potential site selection location for the small green space. Indicates the first The staff member on the first The suitability assessment results of potential site selection points for each training sample are then , , In the multi-round manual scoring process, after each round of scoring, the results are... A team of professionals conducted consistency and sample size checks on the identification results of potential site selection points across the entire training sample, until... The following is a method for verifying the consistency and sample size of the manual scoring results from individual professionals to meet quality requirements and for iteratively updating the scoring rules: A professional manually scores N training samples, using... This indicates that in this round of manual scoring, The professional on the first Among the manual suitability assessment results for potential site selection points in the training samples, the most prevalent result is the manual scoring. The number of, using This indicates the threshold for determining the proportion of consistent scores; if the highest percentage of the manually assigned scores is... Number This accounts for a significant portion of the total number of scores given by professionals. The proportion is greater than the threshold for determining consistency. Then it is considered that for the first The human scoring results for all training sample potential site selection points are consistent. Therefore, the consistency rate of human scoring results is defined and calculated. : Determine whether the consistency rate meets the requirements. If it does not meet the requirements, find the site selection points with inconsistent scoring results for discussion, discuss the scoring rules and standards, further unify the manual scoring rules, and re-extract potential site selection points from the training samples for scoring. statistics The average number of samples for each level score in the professional scoring results is calculated. To ensure that the sample size for each level score meets the requirements, if the average sample size for a certain level score is significantly insufficient, new training samples are added in a timely manner. After multiple rounds of scoring and result quality assessment, the scoring rules and standards are iteratively updated once the consistency rate meets the requirements and the sample size for each score level reaches a certain scale. After the final iteration and update of the scoring rules and standards are completed The average of the scores given by professionals to all training samples, rounded down, is used to determine the final human suitability assessment result for potential site selection points in the training samples. This represents the final human score given by the i-th professional to the k-th training sample, expressed in terms of... Indicates the first The final manual suitability assessment result for each training sample potential location point is as follows: The final manual suitability assessment results of all potential site selection points in the training samples are used to construct a score label set for the training set, named as follows: Set the sample size of the training sample potential location point set to be [value missing]. Then we have: The obtained training set score label set The feature vector set obtained from the training set of the learning machine Correspondingly, they together form the training set for conventional machine learning. ; Set each tag in the rating tag set Feature vector set of machine learning training set Each feature vector A one-to-one correspondence, as a complete training data point, is denoted as... Define operators Representing vectors and scalar The concatenation operation is as follows: Construct a complete machine learning training set, named Set the sample size of the machine learning training set to . Then we have: The obtained training set score label set , and the feature map set obtained from the deep learning training set Correspondingly, they together form the training set for deep learning. ; Set each tag in the rating tag set Each feature map group in the feature map set of the deep learning training set A one-to-one correspondence is used to construct a deep learning input data set, denoted as... Construct a complete deep learning training set, named Set the sample size of the deep learning training set to be... Then we have: 。 7. The method for selecting sites for small-scale green space additions based on machine learning according to claim 6, characterized in that, The potential assessment process for small-scale green space site selection using the small-scale green space site selection potential identification model includes the following steps: Comprehensive potential feature vector dataset of all potential site selection points for small and micro green spaces The suitability assessment model for small-scale green space placement was used to perform a preliminary suitability classification for each potential site. Based on the preliminary suitability classification results and the suitability level of each potential site, the basic potential score for each potential site was calculated. ; On a street view image dataset of all potential sites for small green spaces, a depth-feature-based model for assessing the suitability of small green space placement is used to classify the depth suitability of each potential site. Based on the classification results and suitability level of each site, a depth potential score is calculated for each potential site. ; Select weights and assign depth potential scores to each potential site for a small green space. and the basic potential score of each potential site for a small green space Calculate the weighted average to obtain the final potential score for each potential site for a small green space. This will form the final potential identification result set for all potential site selection locations of small and micro green spaces. ; The final potential identification results set of all potential sites for small and micro green spaces Above, based on the final potential score Based on the potential identification results, K-Means clustering was performed on all potential site selection points of small green spaces. The cluster levels were sorted according to the cluster center scores to obtain the supplementary site selection priority of each potential site selection point's cluster. Finally, the coordinates and site selection priorities of all supplementary site selection points were used as the decision-making basis for the site selection of small green spaces.

8. A computer device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the machine learning-based small green space supplementation site selection method as described in any one of claims 1-7.

9. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the machine learning-based small-scale green space supplementation site selection method as described in any one of claims 1-7.

Images