A method for identifying and / or implementing the potential for supplementing streetside micro-green spaces based on deep learning

Abstract

The present invention provides a method for identifying and / or implementing the potential for supplementation of streetside micro-green spaces based on deep learning, which relates to the fields of urban and rural planning, landscape architecture, and artificial intelligence. The method for identifying the potential for supplementation of streetside micro-green spaces is based on big data of streetscape images, and proposes a standard for identifying the potential for supplementation of micro-green spaces based on the analytical theories in the fields of urban planning and landscape architecture. It quantifies the potential identification system through deep learning technology, and integrates semantic segmentation technology to specifically analyze the objective environmental factors of potential supplementation points, providing an accurate and effective basis for supplementation and implementation strategies for urban micro-green space planning and layout. The present invention makes full use of the advantages of deep learning and basic machine learning semantic segmentation technology, and performs identification and analysis in space and elements respectively, which can not only obtain the potential size of potential supplementation points, but also obtain specific supplementation implementation strategies for supplementation points with optimization potential.

Description

Technical Field

[0001] The present invention relates to the fields of urban and rural planning, landscape architecture and artificial intelligence technology, and specifically to a method for identifying and / or implementing the potential for supplementing streetside micro-green spaces based on deep learning. Background Art

[0002] In recent years, big data and AI deep learning technologies have rapidly matured and are being applied in the field of human settlement development, providing opportunities for addressing this issue. The comprehensive nature of big data allows planners to rapidly access large-scale data related to urban development, moving beyond abstract two-dimensional geographic information. The construction and improvement of street view image databases based on intelligent location-based service platforms can provide a detailed three-dimensional overview of the potential for micro-green space development along streets. The maturity of semantic segmentation technology has provided a path for quantitative analysis of street view image information. Deep learning, a key AI technology, has also been widely applied in fields such as statistical analysis and decision-making. Its application can significantly improve the accuracy, breadth, and efficiency of data analysis. By leveraging street view image data and AI deep learning to supplement and optimize micro-green spaces, the scale, accuracy, and efficiency of analysis can be significantly improved by overcoming the subjectivity and spatiotemporal limitations of traditional planning and analysis methods.

[0003] However, most of the currently published research on micro-green spaces focuses on geographical location research or visual semantic discrimination feature research, and most of the research focuses on performance evaluation, lacking in-depth research on supplementary discrimination and implementation strategies in three-dimensional space. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides a method for identifying and / or implementing the potential for supplementation of street micro-green spaces based on deep learning, which solves the problem that most of the currently published research on micro-green spaces focuses on geographical location research or visual semantic discrimination feature research, and the research mostly focuses on performance evaluation, and lacks in-depth research on supplementation identification and implementation strategies in three-dimensional space.

[0006] (2) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] First, a method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning is provided, including:

[0009] Collecting the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points;

[0010] The data set is screened to obtain a training set; and a three-level potential identification system is constructed, which includes categories, projects and semantics; the comparison indicators for the supplementary potential identification of micro-green spaces based on depth perception are defined as spatial condition category, demand condition category, facility condition category and perception condition category; among them, the spatial condition category is used to describe the necessary spatial conditions for the placement of micro-green spaces, the demand condition category is used to describe the demand intensity of recreational services of micro-green spaces in existing sites, the facility condition category is used to describe the internal and external facility resources closely related to the service status of micro-green spaces, and the perception condition category is used to describe the spatial perception factors that affect the degree of recreational satisfaction; each category is divided into multiple specific projects according to user needs, and each project is described by multiple evaluation points; and the specific semantic feature data used for each suitability identification project is extracted; according to the project description, the training atlas is sequentially compared two by two and marked as good or bad, to obtain the supplementary potential identification score of each project, and to construct the image potential label set of the training set;

[0011] A fully connected image convolutional neural network built based on a deep learning framework tool was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. This was used to establish a stable inference relationship between the complete visual scene information of micro-green spaces and the quantitative identification system for the replenishment potential of micro-green spaces.

[0012] The dataset is used to train a quantitative identification model for the deep replenishment potential of micro-green spaces using a fully connected image convolutional neural network built based on a deep learning framework tool as the underlying model. The scores of each project at each potential replenishment point are obtained, and the influence priorities of the projects are compared. The project weights are calculated by constructing a weight-limited relationship, and the potential identification project scores of each project are weighted and superimposed to obtain the comprehensive potential score of the replenishment point.

[0013] Preferably, the steps of collecting the coordinates of all potential micro-green space supplementary points and the street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and the street view images of all potential micro-green space supplementary points are as follows:

[0014] Intercept the urban navigation trajectory map of the planned area in the intelligent location service platform software;

[0015] The captured urban navigation trajectory map of the planned area was processed using a geographic information system platform, and the navigation trajectory was sampled to obtain a "number-coordinate table" of all potential micro-green space addition points. The sampling interval of the addition points was determined based on ensuring that the street view images of two adjacent potential micro-green space addition points had no duplicate information and no missing information.

[0016] Determine the crawling parameters for street view images of potential micro-green space addition points, calculate or select all crawling parameters, write a crawler program, crawl street view images along the street side of each potential micro-green space addition point at three angles of 75°, 90°, and 105° to the road, and construct a street view image dataset of all potential micro-green space addition points;

[0017] Define the street view image dataset Data of all potential supplementary points of micro-green spaces, where the longitude of the potential supplementary points is X and the latitude is Y, and the complete visual scene information, i.e., the street view image, is Pic. Then, we have:

[0018] Data = {X, Y, Pic}.

[0019] Preferably, the data set is screened to obtain a training set, specifically:

[0020] Based on the principle of overall sample coverage and spatial type uniformity, we randomly select training set sample points. The three-angle street view image set of each training sample supplement point is used as the original feature map group Gx_k of the training sample supplement point. Then, we obtain the complete visual scene information deep learning training set of the selected samples, named GTrain_X. The sample size of the training sample supplement point set is set to N, then:

[0021] GTrain_X={Gx_1,...,Gx_N}.

[0022] Preferably, the construction of a three-level potential screening system specifically includes:

[0023] Under each category, multiple specific projects are divided according to user needs, and each specific project is described by multiple semantic points. For each micro-green space supplementary potential screening project, there are several objective local spatial influencing factors that affect its micro-green space supplementary potential score based on depth perception. Assuming that there are L supplementary potential screening projects, the objective local spatial influencing factor set of project x is defined as Imp_x, and there are J objective local spatial influencing factors in total. The objective local spatial influencing factor of the xth item is defined as I_x. Then:

[0024] Imp_x={I_x1,...,I_x J}

[0025] Within the GTrain_X training sample addition point set, a web-based pairwise image selection system was constructed based on the visual information of street view images of potential micro-green space addition points. A number of field professionals were divided into two groups to conduct web page comparisons, and multiple rounds of comparisons were conducted based on each micro-green space addition potential screening project. The algorithm background recorded that the number of pairwise comparisons of all samples in the training set was greater than or equal to 15 times, and the ranking result was considered valid. After both groups of scores were judged valid, the scores of the two groups were comprehensively compared to determine consistency.

[0026] After the comparison and optimization is completed, the algorithm background obtains the ranking results and corresponding potential values ​​of multiple training samples for L micro-green space supplementation potential screening items. The image potential label set of the training set with a sample size of N is defined as Train_Score, where N is a positive integer, and the potential score set of each micro-green space supplementation potential screening item is Train_s, then:

[0027] Train_Score={Train_s1,...,Train_s L}

[0028] In the mth suitability screening project, the potential score of the kth sample point is Score k , then:

[0029] Train_s m ={Score1,...,Score N}

[0030] The obtained image potential score set Train_s of each micro-greenland supplementary potential identification project training set is matched with the feature map set GTrain_X of the deep learning training set to form the deep learning training set Train_DL.

[0031] Each score Score in the image potential score of each project training set is k One-to-one correspondence is established between each feature graph group Gx_k in the feature graph set of the deep learning training set, and a deep learning input data is constructed, which is recorded as Gt_k;

[0032] Construct a complete deep learning training set, named Train_DL, and set the sample size of the deep learning training set to N, then:

[0033] Train_DL=(GTrain_X, Train_s)={Gt_1,...,Gt_N}.

[0034] Preferably, after both groups of scores are determined to be valid, the scoring results of the two groups are comprehensively compared to determine consistency, specifically including:

[0035] Experts in the field were divided into two groups, A and B, and after multiple rounds of comparison and selection, the image potential score ranking results of the two groups of micro-green space supplementation potential identification project training sets were obtained. K-Means clustering was performed on multiple micro-green space potential supplementation point training samples, and the cluster levels were ranked according to the size of the cluster center score. In each micro-green space supplementation potential identification project, the supplementation priority of each potential supplementation point in the cluster was obtained. Assuming that the clustering level is divided into 1-M levels, the clustering level result set of all projects is defined as Pro, then:

[0036] Pro={Pro1,...,Pro L}

[0037] For the clustering classification result set Pro of the kth item k , define the potential score classification result set of group A of this project as Pro k _A; define the potential score classification result set of group B of this project as Pro k _B, then:

[0038] Pro k ={Pro k _A,Pro k _B}

[0039] For the potential score grading results of N sample sizes in the kth project, we have:

[0040] Pro k _A={Pro k _A1,…,Pro k _A N}

[0041] Pro k _B={Pro k _B1,…,Pro k _B N}

[0042] If the pth training sample supplement point is in the kth item, Pro k _A p =Pro k _B p , then the comparison results are determined to be consistent, and the consistency rate of the comparison results of the kth item is defined as η k , then:

[0043]

[0044]

[0045] Let ρ represent the proportion threshold of the overall consistency rate, then when η kWhen >ρ, the kth selection result is consistent;

[0046] When η1>ρ&&……&&η L >ρ, the overall comparison results are consistent. If the requirements are not met, the projects whose comparison results do not meet the consistency standards will be identified for discussion, the comparison rules and standards will be discussed, the comparison rules will be unified, and a new round of comparison of potential supplementary points will be carried out based on the updated comparison rules.

[0047] Preferably, the fully connected image convolutional neural network built based on the deep learning framework tool is used as the underlying model to train the micro-green space deep supplementation potential quantitative identification model, which is used to establish a stable derivation relationship between the complete visual scene information of the micro-green space and the micro-green space supplementation potential quantitative identification system, specifically including:

[0048] Based on the deep learning framework tool, a fully connected image convolutional neural network was built; the feature atlas GTrain_X of the deep learning training set and the potential score label set Train_s of each micro-green space supplement potential identification project were input, and deep learning training was carried out for each project. 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 fully connected image convolutional neural network multi-classification model, and a quantitative identification model of the deep supplement potential of micro-green space for each project based on deep feature perception was obtained.

[0049] Preferably, the dataset is trained with a fully connected image convolutional neural network built based on a deep learning framework tool as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces, obtain the scores of each project at each potential replenishment point, compare the influence priorities between the projects, calculate the project weights by constructing a weight-limited relationship, and perform weighted superposition on the potential identification project scores of each project to obtain a comprehensive potential score of the replenishment point, specifically including:

[0050] The quantitative identification model of micro-green space deep replenishment potential is applied to the Pic elements in the street view image dataset of potential replenishment points to score the potential replenishment points and obtain the scores of each item Sc_x. Then, the deep scores of each evaluation item DeepScore of all potential replenishment points are constructed; and a deep score dataset DeepScore_set of a total of n micro-green space potential replenishment points is constructed, where n is a positive integer:

[0051] DeepScorek={Sc_x1 k ,Sc_x2 k ,...,Sc_xL k}, where k is a positive integer, and x1, x2..., xL represent different items;

[0052] DeepScore_set={DeepScore1,DeepScore2,...,DeepScore n}, where n is a positive integer;

[0053] Based on the specific projects constructed in the three-level potential identification system, the priorities of different projects are ranked according to their impact on the potential points of micro-green space placement. The weight ranges of different projects are determined objectively through relationship construction. Then, the specific weights are adjusted according to actual needs to obtain the weight μ of each project. The specific operations are as follows:

[0054] Prioritize the projects based on their impact; first, in the four categories, 空间条件类 >x 需求条件类 >x 设施条件类 >x 感知条件类 , assuming that the weights of the four categories are μ1, μ2, μ3, μ4 respectively, then according to the above description, there are:

[0055]

[0056] FinalScore is defined as the comprehensive potential score of each potential supplement point, which is obtained by multiplying the depth score dataset and the weight of each project. Then, a comprehensive potential score set FinalScore_set of all potential supplement points in the micro-green space with a total number of n is constructed, and then:

[0057] k is a positive integer

[0058] FinalScore_set={FinalScore1,FinalScore2,...,FinalScore n}, n is a positive integer

[0059] The K-Means cluster analysis method is used to obtain the comprehensive potential level of each potential supplement point, which is divided into L levels. {1, 2, ..., L}, (L = 2n-1, n is a positive integer) is used to represent the first to the Lth level, and Rank k The comprehensive potential level classification result of the kth potential supplement point is represented by Rank. k The domain of is:

[0060] Rank k ∈{1,2,..,L}

[0061] Construct a comprehensive potential ranking dataset Rank_set of all potential supplementary points of micro-green spaces with a total number of n

[0062] Rank_set={Rank1,Rank2,...,Rank n}, n is a positive integer.

[0063] Secondly, a method for implementing the supplementation of streetside micro-green spaces based on deep learning is provided. Based on the comprehensive potential scores of the supplementation points, potential supplementation points of micro-green spaces with high potential are screened out, and relevant visual semantic indicators are calculated for the corresponding street view image sets to establish a potential point feature description dataset. K-means clustering and grading are performed based on the scores of the four potential identification items of the potential points to obtain a comprehensive potential level dataset. A micro-green space supplementation strategy is formulated based on the potential point street view image dataset, feature description dataset and comprehensive potential level dataset.

[0064] Preferably, the screening of potential microgreenland supplementary sites with high potential specifically includes:

[0065] For a potential supplementary point i to meet the requirements, it should meet the following requirements:

[0066] Rank i ≥(L+1) / 2

[0067] Where L represents the number of levels of comprehensive establishment level of each potential supplementary point;

[0068] After screening, m potential micro-green space supplementary points with high potential that meet the requirements are left, and the comprehensive potential level data set Rank_set1 of potential points is constructed:

[0069] Rank_set1={Rank1,Rank2,...,Rank m}

[0070] Based on the comprehensive potential level dataset Rank_set1, the corresponding elements in the constructed deep score dataset DeepScore_set are extracted. The K-Means cluster analysis method is applied to obtain the characteristic level X of each project of each micro-green space potential point. Each project is divided into L levels, and {1, 2, ..., L} is used to represent the first to the Lth level. For the kth point, it satisfies:

[0071] X k ∈{1,2,...,L}

[0072] Construct the characteristic description FD of each micro green space potential point, where the characteristic description of the kth potential point can be expressed as FD k ; Construct the feature description dataset FD_set of the total number of micro green space potential points m, then:

[0073] FD k ={X1 k ,X2k ,...,XL k},

[0074] Where k is a positive integer, X1, X2, ..., XL represent different items

[0075] FD_set={FD1,FD2,...FD m}

[0076] Where m is the number of potential points;

[0077] Based on the constructed potential point feature description dataset FD_set and the potential point comprehensive potential level dataset Rank_set1, the street view image dataset Data of the potential supplementary points is extracted according to the sequence number of the potential points. Each element in it is matched one by one according to the sequence number to construct the complete potential description PD of the potential point; for potential point k, its complete potential description PD k satisfy:

[0078] PD k ={X k ,Y k ,Pic k ,FD k ,Rank k}, k is a positive integer

[0079] Among them, the X and Y elements describe the latitude and longitude coordinates of the potential point respectively, and the two together represent the geographic spatial information of the point; the Pic element represents the complete visual information of the potential point, that is, the street view image; FD represents the feature description of the potential point; Rank represents the potential level of the potential point;

[0080] Construct a complete potential description dataset PD_set of the total number of micro-greenland supplementary potential points m, then:

[0081] PD_set={PD1,PD2,...,PD m}, m is a positive integer

[0082] Based on the constructed complete potential description dataset PD_set, for the complete potential description PD of the potential point k present in it k Feature description element FD in k Reclassify and filter out the elements whose feature description level is lower than the median value, so as to construct the low-value feature dataset XL of potential points for each evaluation project. For point i, we have:

[0083] XL={X i |X i <(L+1) / 2}

[0084] Based on the constructed low-value feature dataset of potential points of each project, and referring to the constructed potential supplementary point identification system, the objective local spatial impact factor dataset Imp_x of each project is extracted;

[0085] Based on the MXnet deep learning framework, the PSnet image semantic segmentation model is pre-trained on the CityScapes dataset. The deep learning image semantic segmentation model pre-trained on the open source city street view dataset extracts or calculates the relevant semantic elements I_x from the CityScapes dataset based on the objective local spatial influencing factors involved in each project.

[0086] Using the semantic segmentation model, the three street view images of the potential points in the three angles of each potential low-value feature dataset are semantically segmented based on their objective local spatial influence factor datasets. The proportion of the image pixels of the visual semantic elements of the three street view images of each potential point to the total number of pixels of the entire image is obtained in turn. It represents the ratio of the number of image pixels of the i-th visual semantic element of a potential micro-green space site in the direction of the relative angle θ to the road to the total number of pixels in the entire street view image. Based on the weighted average method, ω is used to represent the importance weight of the semantic elements of the street view image in the direction of the relative angle θ, thereby defining the proportion of the i-th semantic element of a potential site:

[0087]

[0088] Based on the above semantic element ratio calculation formula, semantic segmentation analysis is performed to obtain the relevant semantic element ratio of each potential point in the potential point low-value feature dataset, and each semantic element feature vector VSF_XL is constructed. After all semantic element feature vectors are extracted, the semantic element feature vector dataset VSF_XL_set is finally constructed based on the potential point low-value feature dataset of each project.

[0089] VSF_XL=(vsf I_x1 ,vsf I_x2 ,...,vsf I_xJ )

[0090] VSF_XL_set={VSF_XL i |X i ∈XL}, i is a positive integer

[0091] Based on the constructed low-value feature dataset of potential points and the constructed semantic element feature vector, the K-Means cluster analysis method is applied to obtain the level of each local spatial influence factor vsf in the semantic element feature vector VSF of each potential point. Each local spatial influence factor is divided into L levels, and {1, 2, ..., L}, (L = 2n-1, n is a positive integer) is used to represent the first level to the Lth level, and R_vsf is used to represent the level of each local spatial influence factor. i To express the level of the i-th objective local spatial influencing factor, we have:

[0092] R_vsf i ∈{1,2,...,L}

[0093] Construct the semantic feature vector level description R_VSF_XL of each potential point in the low-value feature dataset of each potential point:

[0094] R_VSF_XL={R_vsf I_x1 ,R_vsf I_x2 ,...,R_vsf I_xJ}

[0095] Screen out potential points where the semantic element feature vector level description is lower than the median: For the i-th objective local spatial influencing factor, if R_vsf is satisfied i <(L+1) / 2 (L is the number of levels), the point is retained in the original set, otherwise it is removed; construct the low-value feature vector LR_VSF_XL of the semantic element level:

[0096] LR_VSF_XL={R_vsf i |R_vsf i ∈R_VSF_XL,R_vsf i <(L+1) / 2}

[0097] For each obtained semantic factor level low-value feature vector, construct the potential point optimization factor classification dataset OE_XL for each potential point; construct the potential point optimization project dataset OP_XL for each potential point in the potential point low-value feature dataset of each project:

[0098] OE_XL={x|R_vsf x ∈LR_VSF_XL}

[0099] OP_XL={OE_XL i |X i ∈XL}

[0100] Construct the optimized feature dataset OE of the potential point. If the potential point is located in different low-value feature datasets at the same time, the optimized feature dataset OE of the point is the union of the corresponding features in the datasets.

[0101] The obtained potential point optimization factor dataset OE is combined with the constructed complete potential description dataset PD_set to construct the complete feature description and strategy dataset FINAL of each potential point. The complete feature description and strategy dataset of the i-th potential point can be expressed as:

[0102] FINAL i =[X i ,Y i ,Pic i ,FD i ,Rank i ,OE i ]

[0103] Among them, X i , Y i Represents the latitude and longitude of potential point i, and the two together describe the geographic spatial information of the potential point. i Represents the complete visual scene information of potential point i, namely the street view, FD i Represents the characteristic description of potential point i, Rank i Represents the placement potential level of potential point i, OE i Represents the spatial elements to be optimized at potential point i;

[0104] Construct a complete feature description and strategy dataset FINAL_TOTAL for a total of m potential sites, then:

[0105] FINAL_TOTAL={FINAL i |Rank i ∈Rank_set1}.

[0106] In a third aspect, a method for identifying and implementing the potential for supplementing streetside micro-green spaces based on deep learning is provided, which is characterized by comprising:

[0107] Collecting the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points;

[0108] The data set is screened to obtain a training set; and a three-level potential identification system is constructed, which includes categories, projects and semantics; the comparison indicators for the supplementary potential identification of micro-green spaces based on depth perception are defined as spatial condition category, demand condition category, facility condition category and perception condition category; among them, the spatial condition category is used to describe the necessary spatial conditions for the placement of micro-green spaces, the demand condition category is used to describe the demand intensity of recreational services of micro-green spaces in existing sites, the facility condition category is used to describe the internal and external facility resources closely related to the service status of micro-green spaces, and the perception condition category is used to describe the spatial perception factors that affect the degree of recreational satisfaction; each category is divided into multiple specific projects according to user needs, and each project is described by multiple evaluation points; and the specific semantic feature data used for each suitability identification project is extracted; according to the project description, the training atlas is sequentially compared two by two and marked as good or bad, to obtain the supplementary potential identification score of each project, and to construct the image potential label set of the training set;

[0109] A fully connected image convolutional neural network built based on a deep learning framework tool was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. This was used to establish a stable inference relationship between the complete visual scene information of micro-green spaces and the quantitative identification system for the replenishment potential of micro-green spaces.

[0110] A fully connected image convolutional neural network based on a deep learning framework was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. The scores of each project at each potential replenishment point were obtained, and the impact priority of each project was compared. The project weights were calculated by constructing a weight-limited relationship, and the scores of each project potential identification project were weighted and superimposed to obtain the comprehensive potential score of the replenishment point.

[0111] Based on the comprehensive potential scores of the supplementary points, potential supplementary points of micro-green spaces with high potential were screened out, and relevant visual semantic indicators were calculated for the corresponding street view image sets to establish a potential point feature description dataset. K-means clustering was performed based on the scores of the four potential identification items of the potential points to obtain a comprehensive potential grade dataset. Based on the potential point street view image dataset, feature description dataset and comprehensive potential grade dataset, a micro-green space supplementation strategy was formulated.

[0112] (3) Beneficial effects

[0113] (1) The present invention is based on a deep learning-based method for identifying and / or implementing the potential for supplementing street micro-green spaces. The three-level potential identification system constructed can be used as a basic framework for identifying potential points and proposing supplementary implementation strategies, thus killing two birds with one stone.

[0114] (2) The present invention is based on a method for identifying and / or implementing the potential for supplementation of street micro-green spaces. Through this technical route, the entire process of urban micro-green space construction, namely, "basic database construction - potential screening of potential supplementation points - proposal of implementation strategies for potential point supplementation", can be realized, which is highly targeted and practical.

[0115] (3) The present invention is based on a deep learning method for identifying and / or implementing the potential for supplementation of streetside micro-green spaces. It fully utilizes the advantages of deep learning and basic machine learning semantic segmentation technology to identify and analyze them in space and elements respectively. It can not only obtain the potential size of potential supplementation points, but also obtain specific supplementation implementation strategies for supplementation points with optimization potential.

[0116] (4) The present invention is based on a deep learning method for identifying and / or implementing the potential for supplementing street micro-green spaces, and applies the trueskill principle to perform full sample comparison. The selection of this technical means makes the comparison results more accurate and reasonable, and effectively reduces the subsequent model training error, which can greatly improve the accuracy of the final result. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] Figure 1 Schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0118] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0119] Example

[0120] like Figure 1 As shown, one embodiment of the present invention provides a method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning, including:

[0121] Collecting the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points;

[0122] The dataset was screened to obtain a training set, and a three-level potential identification system was constructed, comprising categories, projects, and semantics. The depth perception-based selection criteria for identifying micro-green space supplementary potential were defined as spatial conditions, demand conditions, facility conditions, and perception conditions. The spatial conditions category was used to describe the necessary spatial conditions for micro-green space placement, the demand conditions category was used to describe the intensity of demand for recreational services provided by micro-green spaces at existing sites, the facility conditions category was used to describe internal and external facility resources closely related to the service status of micro-green spaces, and the perception conditions category was used to describe spatial perception factors that influence recreational satisfaction. Each category was divided into multiple specific projects based on user needs, and each project was described by multiple evaluation points. Specific semantic feature data was extracted for each suitability identification project. Based on the project descriptions, the training atlas was sequentially compared and labeled for each project, yielding a supplementary potential identification score for each project and constructing an image potential label set for the training set.

[0123] A fully connected image convolutional neural network built based on a deep learning framework tool was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. This was used to establish a stable inference relationship between the complete visual scene information of micro-green spaces and the quantitative identification system for the replenishment potential of micro-green spaces.

[0124] The dataset is used to train a quantitative identification model for the deep replenishment potential of micro-green spaces using a fully connected image convolutional neural network built based on a deep learning framework tool as the underlying model. The scores of each project at each potential replenishment point are obtained, and the influence priorities of the projects are compared. The project weights are calculated by constructing a weight-limited relationship, and the potential identification project scores of each project are weighted and superimposed to obtain the comprehensive potential score of the replenishment point.

[0125] As a further solution of an embodiment of the present invention, the steps of collecting the coordinates of all potential micro-green space supplementary points and the street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and the street view images of all potential micro-green space supplementary points are as follows:

[0126] Use electronic map interception software to intercept the city navigation trajectory map of the planned area in the intelligent location service platform software;

[0127] The captured urban navigation trajectory map of the planned area was processed using a geographic information system platform, and the navigation trajectory was sampled to obtain a "number-coordinate table" of all potential micro-green space addition points. The sampling interval of the addition points was determined based on ensuring that the street view images of two adjacent potential micro-green space addition points had no duplicate information and no missing information.

[0128] Determine the crawling parameters for street view images of potential micro-green space addition points, calculate or select all crawling parameters, write a crawler program, crawl street view images along the street side of each potential micro-green space addition point at three angles of 75°, 90°, and 105° to the road, and construct a street view image dataset of all potential micro-green space addition points;

[0129] Define the street view image dataset Data of all potential supplementary points of micro-green spaces, where the longitude of the potential supplementary points is X and the latitude is Y, and the complete visual scene information, i.e., the street view image, is Pic. Then, we have:

[0130] Data = {X, Y, Pic}.

[0131] As a further solution of an embodiment of the present invention, the data set is screened to obtain a training set, specifically:

[0132] Using a street view sampling interval of no less than 50 meters, 10,000 potential micro-green space supplementary points are uniformly selected from all potential micro-green space supplementary points as preliminary training sample supplementary points. 500 comparison training samples are further randomly selected from the 10,000 preliminary training sample supplementary points, and further verification and adjustment are performed based on spatial uniformity and spatial coverage of various types. The spatial distribution and spatial type of the samples are controlled to be uniform and representative. Based on the results of subsequent model training, it is decided whether to change the training sample type or increase the number of samples; based on the principles of overall sample coverage and spatial type uniformity, the training set sample points are randomly selected, and the three-angle (75°, 90°, 105°) street view image group along the street of each training sample supplementary point is used as the original feature map group Gx_k of the training sample supplementary point, thereby obtaining a complete visual scene information deep learning training set for the selected samples, named GTrain_X. The sample capacity of the training sample supplementary point set is set to N, then:

[0133] GTrain_X={Gx_1,...,Gx_N};

[0134] This plan is:

[0135] GTrain_X={Gx_1,...,Gx_500}.

[0136] As a further solution of an embodiment of the present invention, the construction of a three-level potential identification system specifically includes:

[0137] The selection indicators for identifying the supplementary potential of micro green spaces based on depth perception are defined as four categories: spatial conditions, demand conditions, facility conditions, and perception conditions. The spatial conditions are used to describe the necessary spatial conditions for the placement of micro green spaces; the demand conditions are used to describe the demand intensity for recreational services of micro green spaces in existing sites; the facility conditions are used to describe the internal and external facility resources that are closely related to the service status of micro green spaces; and the perception conditions are used to describe the spatial perception factors that affect the degree of recreational satisfaction.

[0138] Under each category, multiple specific items are divided according to user needs, and each specific item is described by multiple semantic points;

[0139] Specifically, regarding spatial conditions: Specific items include the area and form of available space. Specific evaluation points: 1. Area Scale: The size of available space (including various open spaces along the street). 2. Spatial Form: The regularity and usability of the available space (square is best). 3. Boundary Definition: Whether the spatial boundaries are effectively defined (through buildings, vegetation, enclosures, height differences, material materials, etc.).

[0140] Requirements: Specific project: Location activity. Specific evaluation points: 1. Current activity: The site has a large number of pedestrians, people moving around, or cyclists. 2. Potential activity: The site is surrounded by crowd points such as subway stations, bus stops, residential buildings, office buildings, and large shopping malls.

[0141] Facilities Condition Item: Specific Item: The richness of existing facilities. Specific evaluation points: 1. Internal facilities: The venue has landscape tables and chairs, pergolas, fitness and entertainment facilities, retail and dining facilities, public toilets, bicycle parking facilities, etc. 2. Surrounding facilities: The venue is surrounded by a variety of service facilities such as dining, retail, and entertainment (this can be determined by nameplates, billboards, and signs).

[0142] Perceptual Condition: Specific Item: Visual Pleasure. Specific Evaluation Points: 1. Current Greenery: A high amount of greenery within the field of view, with a rich variety and diversity of plant species. 2. Building Facades: The building facades within the field of view are aesthetically pleasing. 3. Landmark Landscape: The presence of iconic buildings, structures (sculptures, landscape elements, etc.), and natural elements (mountains, water systems, etc.) within the field of view.

[0143] For each micro-green space supplementation potential screening project, there are several objective local spatial influencing factors that affect its micro-green space supplementation potential score based on depth perception. Assuming there are L supplementation potential screening projects, the objective local spatial influencing factor set of project x is defined as Imp_x, and there are J objective local spatial influencing factors in total. The objective local spatial influencing factor of the xth item is defined as I_x. Then:

[0144] Imp_x={I_x1,...,I_xJ}

[0145] The establishment of the objective local spatial influencing factor set is based on the basic principles of micro-green space planning and design and the types of available data. This patent uses visual semantic elements to quantify them. The quantitative data includes basic semantic feature data that only reflects the proportion of a single semantic element and advanced semantic feature data that reflects the proportion relationship of each semantic element. The basic semantic feature data is obtained by pre-training a deep learning image semantic segmentation model on the Cityscapes dataset and performing visual semantic segmentation on the obtained supplementary point street view image dataset; the advanced semantic feature data is obtained by calculating the proportional relationship between the basic semantic feature data. The Cityscapes dataset provides 34 types of street view visual semantic element classifications. The selection of basic semantic feature categories and the definition of advanced semantic features are based on the specific content of the supplementary potential screening project.

[0146] use It represents the ratio of the number of image pixels of visual semantic elements of category i in the direction of the relative angle θ with the road where the potential supplementary point of a micro green space is located to the total number of pixels of the entire street view image, that is, the basic semantic feature data of category i. θ The importance weight of the semantic elements of the street view image in the direction of the relative angle θ to the semantic elements of the potential supplementary points of the micro-green space is defined as the proportion of the semantic elements of category i of the potential supplementary points of the micro-green space:

[0147]

[0148] Using VSF m Represents high-level semantic feature data of category m.

[0149] The specific semantic feature data used in each suitability screening project of this plan are as follows:

[0150] Project 1: Area and Form of Available Space

[0151] Basic semantic feature data: vsf sidewalk , vsf parking , vsf ground

[0152] High-level semantic feature data:

[0153] vsf 水平同行要素 ≠0

[0154] The objective local space influencing factor set of the “usable space area and form” project is defined as Imp_space, then:

[0155] Imp_space={vsf sidewalk , vsf parking, vsf gtound , VSF enclosure}

[0156] Project 2: Location Activity

[0157] Basic semantic feature data: vsf building ,

[0158] High-level semantic feature data: VSF crowd =vsf person +vsf rider

[0159] The objective local spatial influencing factor set of the “location liveliness” project is defined as Imp_liveness, which is:

[0160] Imp_liveness={vsf building , VSF crowd}

[0161] Project 3: Richness of existing facilities

[0162] Basic semantic feature data: vsf bicycle , vsf building , vsf pole

[0163] The objective local spatial impact factor set of the "richness of existing facilities" project is defined as Imp_facility, then:

[0164] Imp_facility = {vsf building , vsf bicycle , vsf pole}

[0165] Project 4: Visual Pleasure

[0166] Basic semantic feature data: vsf vegetation , vsf sky ,

[0167] The objective local spatial influencing factor set of the "visual pleasure" project is defined as Imp_pleasure, then:

[0168] Imp_pleasure={vsf vegetation , vsf sky}.

[0169] Within the GTrain_X training sample addition point set, a web-based pairwise image selection system was constructed based on the visual information of street view images of potential micro-green space addition points. A number of field professionals were divided into two groups to conduct web page comparisons, and multiple rounds of comparisons were conducted based on each micro-green space addition potential screening project. The algorithm background recorded that the number of pairwise comparisons of all samples in the training set was greater than or equal to 15 times, and the ranking result was considered valid. After both groups of scores were judged valid, the scores of the two groups were comprehensively compared to determine consistency.

[0170] After the comparison and optimization is completed, the algorithm background obtains the ranking results and corresponding potential values ​​of multiple training samples for L micro-green space supplementation potential screening items. The image potential label set of the training set with a sample size of N is defined as Train_Score, where N is a positive integer, and the potential score set of each micro-green space supplementation potential screening item is Train_s, then:

[0171] Train_Score={Train_s1,...,Train_s L}

[0172] In the mth suitability screening project, the potential score of the kth sample point is Score k , then:

[0173] Train_s m ={Score1,...,Score N}

[0174] The obtained image potential score set Train_s of each micro-greenland supplementary potential identification project training set is matched with the feature map set GTrain_X of the deep learning training set to form the deep learning training set Train_DL.

[0175] Each score Score in the image potential score of each project training set is k One-to-one correspondence is established between each feature graph group Gx_k in the feature graph set of the deep learning training set, and a deep learning input data is constructed, which is recorded as Gt_k;

[0176] Construct a complete deep learning training set, named Train_DL, and set the sample size of the deep learning training set to N, then:

[0177] Train_DL=(GTrain_X, Train_s)={Gt_1,...,Gt_N}.

[0178] The technology for building a web-based image pairwise selection and comparison system based on the trueskill algorithm uses the trueskill algorithm module based on the Bayesian principle. In the Python environment, an algorithm system is built that can randomly match sample point atlases in pairs, record multiple rounds of comparison results in the background, and generate ranking sequences and comparison scores; the feature atlas GTrain_X of the deep learning training set is input, and the Python background algorithm ensures that the original feature map group of the sample points is mapped in angular order on the web page. Domain experts perform pairwise selection of sample points through the front-end operation of the web page to obtain the insertion depth suitability potential score ranking of each project.

[0179] (1) TrueSkill algorithm steps: The training sample generates N set data sets based on the original feature map group, and the result score of each data set is μ = μ value +μ expected-value , μ value Represents the initial score of the dataset, μ expected-value represents the score impact value based on each comparison, δ represents the uncertainty of the score of the data set, Φ(x) represents the cumulative distribution function of the result score, N(x|μ, δ 2 ) represents the distribution function of the result score. Before the selection, the initial value μ is given. value =x,μ expected-value =y, δ=u The specific values ​​are determined through extensive adjustments and testing. The algorithm performs random matching in the background. After each pairwise comparison, the score of the dataset is updated based on the win / loss, and the μ and δ values ​​are updated:

[0180]

[0181]

[0182]

[0183]

[0184]

[0185] in,

[0186]

[0187]

[0188] After calculating μ winner , μ loser , δ winner ,δ loser , further subtract μ value part, and μ expected-valueTo update, the formula is as follows:

[0189]

[0190]

[0191] Based on the updated μ and δ, apply the formula Estimate = μ - K * δ, where K is configurable and is usually set to 3. Finally, sort by Estimate value to obtain the depth supplementation potential score ranking result.

[0192] (2) Use HTML language to build the front-end web page, and make multiple rounds of adjustments to improve the readability and operability of the web page, so as to facilitate the extensive collection of data from the public.

[0193] As a further solution of an embodiment of the present invention, after both groups of scores are determined to be valid, a comprehensive comparison is made between the scoring results of the two groups of people to determine consistency, specifically including:

[0194] Experts in the field were divided into two groups, A and B, and after multiple rounds of comparison and selection, the image potential score ranking results of the two sets of micro-green space supplementation potential identification project training sets were obtained. K-Means clustering was performed on multiple micro-green space potential supplementation point training samples, and the cluster levels were ranked according to the size of the cluster center score. The supplementation priority of each potential supplementation point cluster in the four micro-green space supplementation potential identification projects was obtained respectively; assuming that the cluster classification levels are divided into 1-M levels, and the cluster classification result set of all projects is defined as Pro, then:

[0195] Pro={Pro1,...,Pro4}

[0196] For the clustering classification result set Pro of the kth item k , define the potential score classification result set of group A of this project as Pro k _A; define the potential score classification result set of group B of this project as Pro k _B, then:

[0197] Pro k ={Pro k _A,Pro k _B}

[0198] For the potential score grading results of N sample sizes in the kth project, we have:

[0199] Pro k _A={Pro k _A1,…,Pro k _A N}

[0200] Prok _B={Pro k _B1,…,Pro k _B N}

[0201] If the pth training sample supplement point is in the kth item, Pro k _A p =Pro k _B p , then the comparison results are determined to be consistent, and the consistency rate of the comparison results of the kth item is defined as η k , then:

[0202]

[0203]

[0204] Let ρ represent the proportion threshold of the overall consistency rate, then when η k When >ρ, the kth selection result is consistent;

[0205] When η1>ρ&&……&&η4>ρ, the overall comparison result consistency is passed. If it does not meet the requirements, the projects whose comparison results do not meet the consistency standards will be found for discussion, the comparison rules and standards will be discussed, the comparison rules will be unified, and a new round of comparison of potential supplementary points will be carried out based on the updated comparison rules.

[0206] As a further solution of the embodiment of the present invention, the fully connected image convolutional neural network built based on the deep learning framework tool is used as the underlying model to train the micro-green space deep supplementation potential quantitative identification model, which is used to establish a stable derivation relationship between the complete visual scene information of the micro-green space and the micro-green space supplementation potential quantitative identification system, specifically including:

[0207] Based on the deep learning framework tool, a fully connected image convolutional neural network was built; the feature atlas GTrain_X of the deep learning training set and the potential score label set Train_s of each micro-green space supplement potential identification project were input, and deep learning training was carried out for each project. 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 fully connected image convolutional neural network multi-classification model, and a quantitative identification model of the deep supplement potential of micro-green space for each project based on deep feature perception was obtained.

[0208] Build a fully connected image convolutional neural network based on deep learning framework tools.

[0209] The feature atlas in the deep learning training set Train_DL and the f-score label set Train_s of each micro-green space supplement potential identification project are input for deep learning training. Based on the back-propagation principle of deep neural networks, the hyperparameters of the fully connected image convolutional neural network are adjusted, and the binary classification ability of the fully connected image convolutional neural network for feature images is trained, so that the fully connected image convolutional neural network can deeply mine the visual implicit information in street view images, and finally a better fully connected image convolutional neural network binary classification model is obtained, thereby obtaining a micro-green space supplement potential identification model based on deep feature perception for each evaluation project.

[0210] As a further solution of the embodiment of the present invention, the data set is trained with a fully connected image convolutional neural network built based on a deep learning framework tool as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces, obtain the scores of each project at each potential replenishment point, compare the influence priorities between the projects, calculate the project weights by constructing a weight-limited relationship, and perform weighted superposition on the potential identification project scores of each project to obtain a comprehensive potential score of the replenishment point, specifically including:

[0211] The quantitative identification model of micro-green space deep replenishment potential is applied to the Pic elements in the street view image dataset of potential replenishment points to score the potential replenishment points and obtain the scores of each item Sc_x. Then, the deep scores of each evaluation item DeepScore of all potential replenishment points are constructed; and a deep score dataset DeepScore_set of a total of n micro-green space potential replenishment points is constructed, where n is a positive integer:

[0212] DeepScore k ={Sc_x1 k ,Sc_x2 k ,...,Sc_xL k}, where k is a positive integer, and x1, x2..., xL represent different items;

[0213] DeepScore_set={DeepScore1,DeepScore2,...,DeepScore n}, where n is a positive integer;

[0214] In this case, the number of items is 4, so the depth score of the kth potential supplement point is DeepScore k It can be expressed as:

[0215] DeepScore k ={Sc_s k ,Sc_l k ,Sc_f k ,Sc_pk}, k is a positive integer

[0216] Among them Sc_s k Represents the depth potential score of the “usable space area form” project, Sc_l k Represents the depth potential score of the project "location activity", Sc_f k Represents the depth potential score of the project "richness of existing facilities", Sc_p k Represents the depth potential score of the "visual pleasure" project.

[0217] Based on the specific projects constructed in the three-level potential identification system, the priorities of different projects are ranked according to their impact on the potential points of micro-green space placement. The weight ranges of different projects are determined objectively through relationship construction. Then, the specific weights are adjusted according to actual needs to obtain the weight μ of each project. The specific operations are as follows:

[0218] Prioritize the projects based on their impact; first, in the four categories, 空间条件类 >x 需求条件类 >x 设施条件类 >x 感知条件类 , assuming that the weights of the four categories are μ1, μ2, μ3, μ4 respectively, then according to the above description, there are:

[0219]

[0220] In this case, the priority of the projects S (space), L (liveness), F (facility), and P (pleasure) can be ranked as S>L>F>P based on their impact on the potential for micro-green space placement. The specific reason is: Since S belongs to the spatial condition category and is a factor that must be met for the placement of micro-green space, the product of the minimum score of the project to reach the threshold (here refers to the median) and its weight must be greater than the maximum value of the three projects with the following priority multiplied by the sum of their weights; L belongs to the demand condition item and is the second most important existence in determining the potential for supplementation. Therefore, similar to the above principle, the product of the minimum score of the project to reach the threshold (here refers to the median) and its weight must be greater than the maximum value of the two projects with the following priority multiplied by the sum of their weights; F and P belong to the facility condition item and the perception condition item respectively. The priority between the two is F greater than P, but F is not completely dominant over P, so the weight distribution only needs to be the same as their priority ranking. Assuming that the calculation weights of S, L, F, and P are α, β, γ, and δ respectively, based on the above statement, the four need to satisfy the following relationship:

[0221]

[0222] Among them SC_S midRepresents the median value of SC_S among all potential supplementation points, SC_L max Represents the maximum value of SC_L among all potential supplementary points, SC_L mid Represents the median value of SC_L among all potential supplementation points, SC_F max represents the maximum value of SC_F among all potential supplementary points, SC_P max Represents the maximum value of SC_P among all potential supplementation points.

[0223] Finally, the final value range of α, β, γ, and δ is obtained through calculation, and the values ​​are adjusted according to actual needs to obtain the weight ratio relationship of the four projects.

[0224] FinalScore is defined as the comprehensive potential score of each potential supplement point, which is obtained by multiplying the depth score dataset and the weight of each project. Then, a comprehensive potential score set FinalScore_set of all potential supplement points in the micro-green space with a total number of n is constructed, and then:

[0225] k is a positive integer

[0226] In this case, for the kth potential supplementary point, there exists:

[0227] FinalScore k =SC_S k *α+SC_L k *β+SC_F k *γ+SC_P k *δ, k is a positive integer

[0228] Construct the comprehensive potential score set FinalScore_set of all potential supplementary points of micro-green space with a total number of n, then:

[0229] FinalScore_set={FinalScore1,FinalScore2,...,FinalScore n}, n is a positive integer

[0230] The K-Means cluster analysis method is used to obtain the comprehensive potential level of each potential supplement point, which is divided into L levels. {1, 2, ..., L}, (L = 2n-1, n is a positive integer) is used to represent the first to the Lth level, and Rank k The comprehensive potential level classification result of the kth potential supplement point is represented by Rank. k The domain of is:

[0231] Rank k ∈{1,2,..,L}

[0232] In this case, the comprehensive potential level of the potential supplementary points of the micro green space is divided into three levels, which are used to represent level 1 to level 3. The three levels represent that the potential value of the potential supplementary point is low, average or high. k ∈{1,2,3}.

[0233] Construct a comprehensive potential ranking dataset Rank_set of all potential supplementary points of micro-green spaces with a total number of n

[0234] Rank_set={Rank1,Rank2,...,Rank n}, n is a positive integer.

[0235] Another embodiment of the present invention provides a method for implementing streetside micro-green space supplementation based on deep learning. Based on the comprehensive potential scores of the supplementation points, potential micro-green space supplementation points with higher potential are screened out, relevant visual semantic indicators are calculated for the corresponding street view image set, and a potential point feature description dataset is established; K-means clustering and grading are performed based on the scores of the four potential identification items of the potential points to obtain a comprehensive potential level dataset, and a micro-green space space supplementation strategy is formulated based on the potential point street view image dataset, feature description dataset and comprehensive potential level dataset.

[0236] As a further solution of an embodiment of the present invention, screening out potential supplementary sites of micro-green land with high potential specifically includes:

[0237] For a potential supplementary point i to meet the requirements, it should meet the following requirements:

[0238] Rank i ≥(L+1) / 2

[0239] In this case, Rank i ≥2.

[0240] After screening, m potential micro-green space supplementary points with high potential that meet the requirements are left, and the comprehensive potential level data set Rank_set1 of potential points is constructed:

[0241] Rank_set1={Rank1,Rank2,...,Rank m}

[0242] Based on the comprehensive potential level dataset Rank_set1, the corresponding elements in the constructed deep score dataset DeepScore_set are extracted. The K-Means cluster analysis method is applied to obtain the characteristic level X of each project of each micro-green space potential point. Each project is divided into L levels, and {1, 2, ..., L} is used to represent the first to the Lth level. For the kth point, it satisfies:

[0243] X k ∈{1,2,...,L}

[0244] In this case, L=3 is used to represent level 1 to level 3. The three levels represent whether the level of each project characteristic of the potential site is low, average or high. k ,L k ,F k ,P k To represent the four project levels of potential point k, its domain is:

[0245] S k ∈{1,2,3}

[0246] L k ∈{1,2,3}

[0247] F k ∈{1,2,3}

[0248] P k ∈{1,2,3}

[0249] Construct the characteristic description FD of each micro green space potential point, where the characteristic description of the kth potential point can be expressed as FD k ; Construct the feature description dataset FD_set of the total number of micro green space potential points m, then:

[0250] FD k ={X1 k ,X2 k ,...,XL k},

[0251] Where k is a positive integer, X1, X2, ..., XL represent different items respectively;

[0252] FD_set={FD1,FD2,...FD m}

[0253] Where m is the number of potential points;

[0254] Taking the four projects in this case as an example, there are:

[0255] FD k ={S k ,L k ,F k ,P k}, k is a positive integer

[0256] Taking this case as an example, assuming that the "area and form of available space" of potential site k is high, the "activity of the site" is average, the "richness of existing facilities" is low, and the "visual pleasure" is high, then the characteristic description FD of this potential site is: k ={3,2,1,3}.

[0257] Based on the constructed potential point feature description dataset FD_set and the potential point comprehensive potential level dataset Rank_set1, the street view image dataset Data of the potential supplementary points is extracted according to the sequence number of the potential points. Each element in it is matched one by one according to the sequence number to construct the complete potential description PD of the potential point; for potential point k, its complete potential description PD k satisfy:

[0258] PD k ={X k ,Y k ,Pic k ,FD k ,Rank k}, k is a positive integer

[0259] Among them, the X and Y elements describe the latitude and longitude coordinates of the potential point respectively, and the two together represent the geographic spatial information of the point; the Pic element represents the complete visual information of the potential point, that is, the street view image; FD represents the feature description of the potential point; Rank represents the potential level of the potential point;

[0260] Construct a complete potential description dataset PD_set of the total number of micro-greenland supplementary potential points m, then:

[0261] PD_set={PD1,PD2,...,PD m}, m is a positive integer

[0262] Based on the constructed complete potential description dataset PD_set, for the complete potential description PD of the potential point k present in it k Feature description element FD in k Reclassify and filter out the elements whose feature description level is lower than the median value, so as to construct the low-value feature dataset XL of potential points for each evaluation project. For point i, we have:

[0263] XL={X i |X i <(L+1) / 2}

[0264] Taking this case as an example, we construct a spatial low-value feature dataset SL, an active low-value feature dataset LL, a facility low-value feature dataset FL, and a pleasure low-value feature dataset PL. For a potential point i, if it satisfies: Si <(L+1) / 2, then point i exists in SL. Similarly, if L i <(L+1) / 2, then point i exists in LL; if F i <(L+1) / 2, then point i exists in FL; if P i <(L+1) / 2, then point i exists in PL.

[0265] Through the above steps, we complete the spatial low-value feature dataset SL, the active low-value feature dataset LL, the facility low-value feature dataset FL, and the pleasure low-value feature dataset PL. When L=3, the following conditions are met:

[0266] SL={S i |S i <2}

[0267] LL={L i |L i <2}

[0268] FL={F i |F i <2}

[0269] PL={P i |P i <2}

[0270] Since the characteristic description of point k is FD k ={3,2,1,3}, where the feature level of the item “richness of existing facilities” is 1, which is lower than the median value of 2, and point k is located in the set FL.

[0271] Based on the constructed low-value feature dataset of potential points of each project, and referring to the constructed potential supplementary point identification system, the objective local spatial impact factor dataset Imp_x of each project is extracted;

[0272] Based on the MXnet deep learning framework, the PSnet image semantic segmentation model is pre-trained on the CityScapes dataset. The deep learning image semantic segmentation model pre-trained on the open source city street view dataset extracts or calculates the relevant semantic elements I_x from the CityScapes dataset based on the objective local spatial influencing factors involved in each project.

[0273] Taking this case as an example, the semantic elements that need to be extracted from potential points in the SL set are "sidewalk", "parking", "ground", and "enclosure"; the semantic elements that need to be extracted from potential points in the LL set are "building" and "crowd"; the semantic elements that need to be extracted from potential points in the FL set are "building", "bicycle", and "pole"; and the semantic elements that need to be extracted from potential points in the PL set are "vegetation" and "sky".

[0274] Using the semantic segmentation model, the three street view images of the potential points in the three angles of each potential low-value feature dataset are semantically segmented based on their objective local spatial influence factor datasets. The proportion of the image pixels of the visual semantic elements of the three street view images of each potential point to the total number of pixels of the entire image is obtained in turn. It represents the ratio of the number of image pixels of the i-th visual semantic element of a potential micro-green space site in the direction of the relative angle θ to the road to the total number of pixels in the entire street view image. Based on the weighted average method, ω is used to represent the importance weight of the semantic elements of the street view image in the direction of the relative angle θ, thereby defining the proportion of the i-th semantic element of a potential site:

[0275]

[0276] Based on the above semantic element ratio calculation formula, semantic segmentation analysis is performed to obtain the relevant semantic element ratio of each potential point in the potential point low-value feature dataset, and each semantic element feature vector VSF_XL is constructed. After all semantic element feature vectors are extracted, the semantic element feature vector dataset VSF_XL_set is finally constructed based on the potential point low-value feature dataset of each project.

[0277] VSF_XL=(vsf I_x1 ,vsf I_x2 ,...,vsf I_xJ )

[0278] Taking this case as an example, construct the semantic element feature vectors VSF_SL, VSF_LL, VSF_FL, and VSF_PL. For each element in SL, LL, FL, and PL, a semantic element feature vector is constructed that satisfies:

[0279] VSF_SL=(vsf sidewalk ,vsf parking ,vsf ground ,vsf enclosure )

[0280] VSF_LL=(vsf building ,vsfcrowd )

[0281] VSF_FL=(vsf building ,vsf bicycle ,vsf pole )

[0282] VSF_PL=(vsf vegetation ,vsf sky )

[0283] Construct the semantic feature vector dataset VSF_XJ_set based on the low-value feature dataset of potential points of each project:

[0284] VSF_XL_set={VSF_XL i |X i ∈XL}, i is a positive integer

[0285] Taking this case as an example, we construct the semantic element feature vector datasets VSF_SL_set, VSF_LL_set, VSF_FL_set, and VSF_PL_set for each potential point in SL, LL, FL, and PL. Each element in the dataset represents the proportion of all objective local spatial influencing factors involved in the low-value feature description of a potential point.

[0286] VSF_SL_set={VSF_SL i |S i ∈SL}

[0287] VSF_LL_set={VSF_LL i |L i ∈LL}

[0288] VSF_FL_set={VSF_FL i |F i ∈FL}

[0289] VSF_PL_set={VSF_PL i |P∈PL}

[0290] Taking the second set as an example, the elements in the VSF_FL_set set represent the proportion of "objective local spatial influencing factors" such as "building," "bicycle," and "pole" that influence the "richness of existing facilities" within each potential site with a low rating. The remaining three sets are constructed in the same way.

[0291] Based on the constructed low-value feature dataset of potential points and the constructed semantic element feature vector, the K-Means cluster analysis method is applied to obtain the level of each local spatial influence factor vsf in the semantic element feature vector VSF of each potential point. Each local spatial influence factor is divided into L levels, and {1, 2, ..., L}, (L = 2n-1, n is a positive integer) is used to represent the first level to the Lth level, and R_vsf is used to represent the level of each local spatial influence factor. i To express the level of the i-th objective local spatial influencing factor, we have:

[0292] R_vsf i ∈{1,2,...,L}

[0293] Taking this case as an example, L = 3, then the objective local spatial influencing factors of the micro green space supplement potential points are divided into 3 levels, and {1, 2.3} is used to represent the first to third levels, indicating that the levels of the objective local spatial influencing factors are low, average and high, respectively. i ∈{1,2,3}.

[0294] Construct the semantic feature vector level description R_VSF_XL of each potential point in the low-value feature dataset of each potential point:

[0295] R_VSF_XL={R_vsf I_x1 ,R_vsf I_x2 ,...,R_vsf I_xJ}

[0296] Taking this case as an example, the semantic element feature vector level of each point in the four projects is described as follows:

[0297] R_VSF_SL={R_vsf sidewalk ,R_vsf parking ,R_vsf ground ,R_vsf enclosure}

[0298] R_VSF_LL={R_vsf building ,R_vsf crowd}

[0299] R_VSF_FL={R_vsf building ,R_vsf bicycle ,R_vsf pole}

[0300] R_VSF_PL={R_vsf vegetation ,R_vsf sky}

[0301] Assume that in the potential point k in the low-value facility feature dataset FL, the level of the objective local spatial influencing factor "building" is low, the level of "bicycle" is high, and the level of "pole" is average. Then the semantic feature vector level of point k is described as:

[0302] R_VSF_FL k ={1,3,2}

[0303] The same is true for the description of the feature vector levels of the remaining semantic elements.

[0304] Screen out potential points where the semantic element feature vector level description is lower than the median: For the i-th objective local spatial influencing factor, if R_vsf is satisfied i <(L+1) / 2 (L is the number of levels), the point is retained in the original set, otherwise it is removed; construct the low-value feature vector LR_VSF_XL of the semantic element level:

[0305] LR_VSF_XL={R_vsf i |R_vsf i ∈R_VSF_XL,R_vsf i <(L+1) / 2}

[0306] Taking this case as an example, L = 3. For each potential point in the SL, LL, FL, PL set, a low-value feature vector of the semantic element level is constructed to meet the following requirements:

[0307] LR_VSF_SL={R_vsf i |R_vsf i ∈R_VSF_SL,R_vsf i <2}

[0308] LR_VSF_LL={R_vsf i |R_vsf i ∈R_VSF_LL,R_vsf i <2}

[0309] LR_VSF_FL={R_vsf i |R_vsf i ∈R_VSF_FL,R_vsf i <2}

[0310] LR_VSF_PL = {R_vsf i |R_vsf i ∈R_VSF_PL,R_vsf i <2}

[0311] Filter out the elements with semantic feature vector description level lower than 2 and keep them in the set. At the same time, based on the hypothetical description R_VSF_FL for point k k ={1,3,2}, after this screening step, the semantic element level low value feature vector of point k becomes: LR_VSF_FL k ={1}, and the construction of the low-value feature vectors of the remaining three semantic elements is similar.

[0312] For each obtained low-value feature vector of the semantic element level, the potential point optimization element classification dataset OE_XL of each potential point is constructed respectively;

[0313] OE_XL={x|R_vsf x ∈LR_VSF_XL}

[0314] Taking this case as an example, construct OE_SL, OE_LL, OE_FL, OE_PL to meet the following requirements:

[0315] OE_SL={s|R_vsf s ∈LR_VSF_SL}

[0316] OE_LL={l|R_vsf l ∈LR_VSF_LL}

[0317] OE_FL={f|R_vsf f ∈LR_VSF_FL}

[0318] OE_PL={p|R_vsf p ∈LR_VSF_PL}

[0319] Among them, s, l, f, and p represent the types of factors that need to be optimized at each potential point.

[0320] Taking this case as an example, based on the hypothetical result (LR_VSF_FL k ={1}), the potential point k optimization factor classification data set obtained through this step is: OE_SL={"bicycle"}.

[0321] Construct the potential point optimization project dataset OP_XL for each potential point in the low-value feature dataset of each project:

[0322] OP_XL={OE_XL i |X i ∈XL}

[0323] Taking this case as an example, construct OP_SL, OP_LL, OP_FL, OP_PL, and there are:

[0324] OP_SL={OE_SLi |S i ∈SL}

[0325] OP_LL={OE_LL i |L i ∈LL}

[0326] OP_FL={OE_FL i |F i ∈FL}

[0327] OP_PL = {OE_PL i |P i ∈PL}.

[0328] Construct the optimized feature dataset OE of the potential point. If the potential point is located in different low-value feature datasets at the same time, the optimized feature dataset OE of the point is the union of the corresponding features in the datasets.

[0329] Taking this case as an example, assuming that potential point i is in both FL and PL, the optimized factor dataset OE of potential point k is i ={OE_FL i , OE_PL i After this step, for any potential point i in this case, there must be an OE i ∈{OP_SL, OP_LL, OP_FL, OP_PL}, thereby obtaining the factors to be optimized for this potential point.

[0330] Taking this case as an example, if the potential point q is in both the potential point low-value feature datasets FL and PL, and OE_FL q ={"bicycle"}, OE_PL q = {"vegetation"}, then the optimized feature dataset OE of point q q ={"bicycle", "vegetation"}.

[0331] The obtained potential point optimization factor dataset OE is combined with the constructed complete potential description dataset PD_set to construct the complete feature description and strategy dataset FINAL of each potential point. The complete feature description and strategy dataset of the i-th potential point can be expressed as:

[0332] FINAL i =[X i ,Y i ,Pic i ,FD i ,Rank i ,OE i ]

[0333] Among them, Xi , Y i Represents the latitude and longitude of potential point i, and the two together describe the geographic spatial information of the potential point. i Represents the complete visual scene information of potential point i, namely the street view, FD i Represents the characteristic description of potential point i, Rank i Represents the placement potential level of potential point i, OE i The spatial elements representing potential point i to be optimized;

[0334] Construct a complete feature description and strategy dataset FINAL_TOTAL for a total of m potential sites, then:

[0335] FINAL_TOTAL={FINAL i |Rank i ∈Rank_set1}.

[0336] Another embodiment of the present invention provides a method for identifying and implementing the potential for supplementing streetside micro-green spaces based on deep learning, which is characterized by comprising:

[0337] Collecting the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points;

[0338] The data set is screened to obtain a training set; a three-level potential identification system is constructed, which includes categories, projects and semantics; the comparison indicators for the supplementary potential identification of micro-green spaces based on depth perception are defined as spatial condition class, demand condition class, facility condition class and perception condition class; among them, the spatial condition class is used to describe the necessary spatial conditions for the placement of micro-green spaces, the demand condition class is used to describe the demand intensity of recreational services of micro-green spaces in existing sites, the facility condition class is used to describe the internal and external facility resources closely related to the service status of micro-green spaces, and the perception condition class is used to describe the spatial perception factors that affect the degree of recreational satisfaction; each category is divided into multiple specific projects according to user needs, and each project is described by multiple evaluation points; and the specific semantic feature data used for each suitability identification project is extracted to construct the image potential label set of the training set;

[0339] A fully connected image convolutional neural network built based on a deep learning framework tool was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. This was used to establish a stable inference relationship between the complete visual scene information of micro-green spaces and the quantitative identification system for the replenishment potential of micro-green spaces.

[0340] A fully connected image convolutional neural network based on a deep learning framework was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. The scores of each project at each potential replenishment point were obtained, and the impact priority of each project was compared. The project weights were calculated by constructing a weight-limited relationship, and the scores of each project potential identification project were weighted and superimposed to obtain the comprehensive potential score of the replenishment point.

[0341] Based on the comprehensive potential scores of the supplementary points, potential supplementary points of micro-green spaces with high potential were screened out, and relevant visual semantic indicators were calculated for the corresponding street view image sets to establish a potential point feature description dataset. K-means clustering was performed based on the scores of the four potential identification items of the potential points to obtain a comprehensive potential grade dataset. Based on the potential point street view image dataset, feature description dataset and comprehensive potential grade dataset, a micro-green space supplementation strategy was formulated.

[0342] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

Claims

1. A method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning, characterized in that: include: Collecting the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points; Filter the data set to obtain the training set; A three-level potential identification system is constructed, which includes categories, projects and semantics; the comparison index categories for the supplementary potential identification of micro-green spaces based on depth perception are defined as spatial condition category, demand condition category, facility condition category and perception condition category; among them, the spatial condition category is used to describe the necessary spatial conditions for the placement of micro-green spaces, the demand condition category is used to describe the demand intensity of recreational services of micro-green spaces in existing sites, the facility condition category is used to describe the internal and external facility resources closely related to the service status of micro-green spaces, and the perception condition category is used to describe the spatial perception factors that affect the degree of recreational satisfaction; each category is divided into multiple specific projects according to user needs, and each project is described by multiple evaluation points; and the specific semantic feature data used for each suitability identification project is extracted; according to the project description, the training atlas is sequentially compared two by two and marked with pros and cons, and the supplementary potential identification score of each project is obtained, and the image potential label set of the training set is constructed; A fully connected image convolutional neural network built based on a deep learning framework tool was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. This was used to establish a stable inference relationship between the complete visual scene information of micro-green spaces and the quantitative identification system for the replenishment potential of micro-green spaces. The dataset is used to train a quantitative identification model for the deep replenishment potential of micro-green spaces using a fully connected image convolutional neural network built based on a deep learning framework tool as the underlying model. The scores of each project at each potential replenishment point are obtained, and the influence priorities of the projects are compared. The project weights are calculated by constructing a weight-limited relationship, and the potential identification project scores of each project are weighted and superimposed to obtain the comprehensive potential score of the replenishment point.

2. The method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning according to claim 1, characterized in that: The steps of collecting the coordinates of all potential micro-greenland supplementary points and the street view images of all potential micro-greenland supplementary points, and constructing a data set including the coordinates of all potential micro-greenland supplementary points and the street view images of all potential micro-greenland supplementary points are as follows: Intercept the urban navigation trajectory map of the planned area in the intelligent location service platform software; Using a geographic information system platform to process the captured urban navigation trajectory map of the planned area, the navigation trajectories were sampled to obtain a "number-coordinate table" for all potential micro-green space addition points. The sampling interval for the addition points was determined by ensuring that the street view images of two adjacent potential micro-green space addition points had no duplicate information and no missing information. Determine the crawling parameters for street view images of potential micro-green space addition points, calculate or select all crawling parameters, write a crawler program, crawl street view images along the street side of each potential micro-green space addition point at three angles of 75°, 90°, and 105° to the road, and construct a street view image dataset of all potential micro-green space addition points; Define the street view image dataset Data of all potential supplementary points of micro-green spaces, where the longitude of the potential supplementary points is X and the latitude is Y, and the complete visual scene information, i.e., the street view image, is Pic. Then, we have: Data = {X, Y, Pic}.

3. The method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning according to claim 1, characterized in that: The training set is obtained by screening the data set, specifically: Based on the principle of overall sample coverage and spatial type uniformity, we randomly select training set sample points. The three-angle street view image set of each training sample supplement point is used as the original feature map group Gx_k of the training sample supplement point. Then, we obtain the complete visual scene information deep learning training set of the selected samples, named GTrain_X. The sample size of the training sample supplement point set is set to N, then: GTrain_X={Gx_1,...,Gx_N}.

4. The method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning according to claim 1, characterized in that: The construction of the three-level potential identification system specifically includes: Under each category, multiple specific projects are divided according to user needs, and each specific project is described by multiple semantic points. For each micro-green space supplementary potential screening project, there are several objective local spatial influencing factors that affect its micro-green space supplementary potential score based on depth perception. Assuming that there are L supplementary potential screening projects, the objective local spatial influencing factor set of project x is defined as Imp_x, and there are J objective local spatial influencing factors in total. The objective local spatial influencing factor of the xth item is defined as I_x. Then: Imp_x={I_x1,...,I_x J } Within the GTrain_X training sample addition point set, a web-based pairwise image selection system was constructed based on the visual information of street view images of potential micro-green space addition points. A number of field professionals were divided into two groups to conduct web-based comparisons, and multiple rounds of comparisons were conducted based on each micro-green space addition potential screening item. The algorithm background recorded that the number of pairwise comparisons of all samples in the training set was greater than or equal to 15, and the ranking result was considered valid. After both groups of scores were judged valid, the scores of the two groups were comprehensively compared to determine consistency. After the comparison and optimization is completed, the algorithm background obtains the ranking results and corresponding potential values ​​of multiple training samples for L micro-green space supplementation potential screening items. The image potential label set of the training set with a sample size of N is defined as Train_Score, where N is a positive integer, and the potential score set of each micro-green space supplementation potential screening item is Train_s, then: Train_Score={Train_s1,...,Train_s L } In the mth supplementary potential screening project, the potential score of the kth sample point is Score k , then: Train_s m ={Score1,...,Score N } The obtained image potential score set Train_s of each micro-greenland supplementary potential identification project training set is matched with the feature map set GTrain_X of the deep learning training set to form the deep learning training set Train_DL. Each score Score in the image potential score of each project training set is k One-to-one correspondence is established between each feature graph group Gx_k in the feature graph set of the deep learning training set, and a deep learning input data is constructed, which is recorded as Gt_k; Construct a complete deep learning training set, named Train_DL, and set the sample size of the deep learning training set to N, then: Train_DL=(GTrain_X, Train_s)={Gt_1,...,Gt_N}.

5. The method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning according to claim 4, characterized in that: After both groups of scores are determined to be valid, the scores of the two groups are comprehensively compared to determine their consistency, specifically including: Experts in the field were divided into two groups, A and B, and after multiple rounds of comparison and selection, the image potential score ranking results of the two sets of micro-green space supplementation potential identification project training sets were obtained. K-Means clustering was performed on multiple micro-green space potential supplementation point training samples, and the cluster levels were ranked according to the size of the cluster center scores. In each micro-green space supplementation potential identification project, the supplementation potential priority of each potential supplementation point cluster was obtained. Assuming that the cluster classification levels are divided into 1-M levels, the cluster classification result set of all projects is defined as Pro, then: For={For1,...,For L } For the clustering classification result set Pro of the kth item k , define the potential score classification result set of group A of this project as Pro k _A; define the potential score classification result set of group B of this project as Pro k _B, then: For k ={For k _A,For k _B} For the potential score grading results of N sample sizes in the kth project, we have: For k _A={For k _A1,…,For k _AND N } For k _B={For k _B1,…,For k _B N } If the pth training sample supplement point is in the kth item, Pro k _A p =Pro k _B p , then the comparison results are determined to be consistent, and the consistency rate of the comparison results of the kth item is defined as η k , then: Let ρ represent the proportion threshold of the overall consistency rate, then when η k When >ρ, the kth selection result is consistent; When η1>ρ&&……&&η L >ρ, the overall comparison results are consistent. If the requirements are not met, the projects whose comparison results do not meet the consistency standards will be identified for discussion, the comparison rules and standards will be discussed, the comparison rules will be unified, and a new round of comparison of potential supplementary points will be carried out based on the updated comparison rules.

6. The method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning according to claim 1, characterized in that: The fully connected image convolutional neural network built based on the deep learning framework tool is used as the underlying model to train the micro-green space deep supplementation potential quantitative identification model, which is used to establish a stable derivation relationship between the complete visual scene information of the micro-green space and the micro-green space supplementation potential quantitative identification system, specifically including: Based on the deep learning framework tool, a fully connected image convolutional neural network was built; the feature atlas GTrain_X of the deep learning training set and the potential score label set Train_s of each micro-green space supplement potential identification project were input, and deep learning training was carried out for each project. 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 fully connected image convolutional neural network multi-classification model, and a quantitative identification model of the deep supplement potential of micro-green space for each project based on deep feature perception was obtained.

7. The method for identifying the potential for supplementation of streetside micro-green spaces based on deep learning according to claim 1, characterized in that: The dataset is trained using a fully connected image convolutional neural network built based on a deep learning framework tool as the underlying model to quantify the potential for deep replenishment of micro-green spaces. The scores of each project at each potential replenishment point are obtained, and the impact priority between the projects is compared. The project weights are calculated by constructing a weight-limited relationship, and the potential identification project scores of each project are weighted and superimposed to obtain the comprehensive potential score of the replenishment point, which specifically includes: The quantitative identification model of micro-green space deep replenishment potential is applied to the Pic elements in the street view image dataset of potential replenishment points to score the potential replenishment points and obtain the scores of each item Sc_x. Then, the deep scores of each evaluation item DeepScore of all potential replenishment points are constructed; and a deep score dataset DeepScore_set of a total of n micro-green space potential replenishment points is constructed, where n is a positive integer: DeepScore k ={Sc_x1 k ,Sc_x2 k ,...,Sc_xL k }, where k is a positive integer, and x1, x2..., xL represent different items; DeepScore_set={DeepScore1,DeepScore2,...,DeepScore n }, where n is a positive integer; Based on the specific projects constructed in the three-level potential identification system, the priorities of different projects are ranked according to their impact on the potential points of micro-green space placement. The weight ranges of different projects are determined objectively through relationship construction. Then, the specific weights are adjusted according to actual needs to obtain the weight μ of each project. The specific operations are as follows: Prioritize the projects based on their impact; first, in the four categories, 空间条件类 >x 需求条件类 >x 设施条件类 >x 感知条件类 , assuming that the weights of the four categories are μ1, μ2, μ3, μ4 respectively, then according to the above description, there are: FinalScore is defined as the comprehensive potential score of each potential supplementary point. This score is obtained by multiplying and summing the depth score dataset and the weight of each project. Then, a comprehensive potential score set FinalScore_set of all potential supplementary points in the micro-green space with a total number of n is constructed, which is: k is a positive integer FinalScore_set={FinalScore1,FinalScore2,...,FinalScore n }, n is a positive integer The K-Means cluster analysis method is used to obtain the comprehensive potential level of each potential supplement point, which is divided into L levels. {1, 2, ..., L}, L = 2n-1, n is a positive integer, to represent the first to the Lth level, and Rank k The comprehensive potential level classification result of the kth potential supplement point is represented by Rank. k The domain of is: Rank k ∈{1,2,..,L} Construct a comprehensive potential ranking dataset Rank_set of all potential supplementary points of micro-green spaces with a total number of n Rank_set={Rank1,Rank2,...,Rank n }, n is a positive integer.

8. A method for implementing the supplementation of streetside micro green spaces based on deep learning, characterized in that: Based on the comprehensive potential score of the supplementary points obtained by the street micro-green space supplementation potential identification method described in any one of claims 1 to 7, potential supplementary points of micro-green spaces with higher potential are screened out, relevant visual semantic indicators are calculated for the corresponding street view image set, and a potential point feature description data set is established; K-means clustering and grading are performed based on the scores of the four potential identification items of the potential points to obtain a comprehensive potential grade data set, and a micro-green space space supplementation strategy is formulated based on the potential point street view image data set, the feature description data set and the comprehensive potential grade data set.

9. The method for implementing the supplementation of streetside micro-green spaces based on deep learning according to claim 8, characterized in that: The potential supplementary sites of micro green land with high potential are screened out, specifically including: For a potential supplementary point i to meet the requirements, it should meet the following requirements: Rank i ≥(L+1) / 2 Where L represents the number of levels of comprehensive establishment level of each potential supplementary point; After screening, m potential micro-green space supplementary points with high potential that meet the requirements are left, and the comprehensive potential level data set Rank_set1 of potential points is constructed: Rank_set1={Rank1,Rank2,...,Rank m } Based on the comprehensive potential level dataset Rank_set1, the corresponding elements in the constructed deep score dataset DeepScore_set are extracted. The K-Means cluster analysis method is applied to obtain the characteristic level X of each project of each micro-green space potential point. Each project is divided into L levels, and {1, 2, ..., L} is used to represent the first to the Lth level. For the kth point, it satisfies: X k ∈{1,2,...,L} Construct the characteristic description FD of each micro green space potential point, where the characteristic description of the kth potential point can be expressed as FD k ; Construct the feature description dataset FD_set of the total number of micro green space potential points m, then: FD k ={X1 k ,X2 k ,...,XL k }, Where k is a positive integer, X1, X2, ..., XL represent different items FD_set={FD1,FD2,...FD m } Where m is the number of potential points; Based on the constructed potential point feature description dataset FD_set and the potential point comprehensive potential level dataset Rank_set1, the street view image dataset Data of the potential supplementary points is extracted according to the sequence number of the potential points. Each element in it is matched one by one according to the sequence number to construct the complete potential description PD of the potential point; for potential point k, its complete potential description PD k satisfy: PD k ={X k ,Y k ,Pic k ,FD k ,Rank k }, k is a positive integer Among them, the X and Y elements describe the latitude and longitude coordinates of the potential point respectively, and the two together represent the geographic spatial information of the point; the Pic element represents the complete visual information of the potential point, that is, the street view image; FD represents the feature description of the potential point; Rank represents the potential level of the potential point; Construct a complete potential description dataset PD_set of the total number of micro-greenland supplementary potential points m, then: PD_set={PD1,PD2,...,PD m }, m is a positive integer Based on the constructed complete potential description dataset PD_set, for the complete potential description PD of the potential point k present in it k Feature description element FD in k Reclassify and filter out the elements whose feature description level is lower than the median value, so as to construct the low-value feature dataset XL of potential points for each evaluation project. For point i, we have: XL={X i |X i <(L+1) / 2} Based on the constructed low-value feature dataset of potential points of each project, and referring to the constructed potential supplementary point identification system, the objective local spatial impact factor dataset Imp_x of each project is extracted; Based on the MXnet deep learning framework, the PSnet image semantic segmentation model is pre-trained on the CityScapes dataset. The deep learning image semantic segmentation model pre-trained on the open source city street view dataset extracts or calculates the relevant semantic elements I_x from the CityScapes dataset based on the objective local spatial influencing factors involved in each project. Using the semantic segmentation model, the three street view images of the potential points in the three angles of each potential low-value feature dataset are semantically segmented based on their objective local spatial influence factor datasets. The proportion of the image pixels of the visual semantic elements of the three street view images of each potential point to the total number of pixels of the entire image is obtained in turn. It represents the ratio of the number of image pixels of the i-th visual semantic element of a potential micro-green space site in the direction of the relative angle θ to the road to the total number of pixels in the entire street view image. Based on the weighted average method, ω is used to represent the importance weight of the semantic elements of the street view image in the direction of the relative angle θ, thereby defining the proportion of the i-th semantic element of a potential site: Based on the above semantic element ratio calculation formula, semantic segmentation analysis is performed to obtain the relevant semantic element ratio of each potential point in the potential point low-value feature dataset, and each semantic element feature vector VSF_XL is constructed. After all semantic element feature vectors are extracted, the semantic element feature vector dataset VSF_XL_set is finally constructed based on the potential point low-value feature dataset of each project. VSF_XL=(vsf I_x1 ,vsf I_x2 ,...,vsf I_xJ ) VSF_XL_set={VSF_XL i |X i ∈XL}, i is a positive integer Based on the constructed low-value feature dataset of potential points and the constructed semantic element feature vector, the K-Means cluster analysis method is applied to obtain the level of each local spatial influence factor vsf in the semantic element feature vector VSF of each potential point. Each local spatial influence factor is divided into L levels, and {1, 2, ..., L}, L = 2n-1, n is a positive integer, to represent the first level to the Lth level, and R_vsf is used. i To express the level of the i-th objective local spatial influencing factor, we have: R_vsf i ∈{1,2,...,L} Construct the semantic feature vector level description R_VSF_XL of each potential point in the low-value feature dataset of each potential point: R_VSF_XL={R_vsf I_x1 ,R_vsf I_x2 ,...,R_vsf I_xJ } Screen out potential points where the semantic element feature vector level description is lower than the median: For the i-th objective local spatial influencing factor, if R_vsf is satisfied i <(L+1) / 2, where L is the number of levels, the point is retained in the original set, otherwise it is removed; construct the low-value feature vector LR_VSF_XL of the semantic element level: LR_VSF_XL={R_vsf i |R_vsf i ∈R_VSF_XL,R_vsf i <(L+1) / 2} For each obtained semantic factor level low-value feature vector, construct the potential point optimization factor classification dataset OE_XL for each potential point; construct the potential point optimization project dataset OP_XL for each potential point in the potential point low-value feature dataset of each project: OE_XL={x|R_vsf x ∈LR_VSF_XL} OP_XL={OE_XL i |X i XL Construct the optimized feature dataset OE of the potential point. If the potential point is located in different low-value feature datasets at the same time, the optimized feature dataset OE of the point is the union of the corresponding features in the datasets. The obtained potential point optimization factor dataset OE is combined with the constructed complete potential description dataset PD_set to construct the complete feature description and strategy dataset FINAL of each potential point. The complete feature description and strategy dataset of the i-th potential point can be expressed as: FINAL i =[X i ,Y i ,Pic i ,FD i ,Rank i ,OE i ] Among them, X i , Y i Represents the latitude and longitude of potential point i, and the two together describe the geographic spatial information of the potential point. i Represents the complete visual scene information of potential point i, namely the street view, FD i Represents the characteristic description of potential point i, Rank i Represents the placement potential level of potential point i, OE i The spatial elements representing potential point i to be optimized; Construct a complete feature description and strategy dataset FINAL_TOTAL for a total of m potential sites, then: FINAL_TOTAL={FINAL i |Rank i ∈Rank_set1}。 10. A method for identifying and implementing the potential for supplementing streetside micro-green spaces based on deep learning, characterized in that: include: Collecting the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points, and constructing a data set including the coordinates of all potential micro-green space supplementary points and street view images of all potential micro-green space supplementary points; Filter the data set to obtain the training set; A three-level potential identification system is constructed, which includes categories, projects and semantics; the comparison indicators for the supplementary potential identification of micro-green spaces based on depth perception are defined as spatial condition category, demand condition category, facility condition category and perception condition category; among them, the spatial condition category is used to describe the necessary spatial conditions for the placement of micro-green spaces, the demand condition category is used to describe the demand intensity of recreational services of micro-green spaces in existing sites, the facility condition category is used to describe the internal and external facility resources closely related to the service status of micro-green spaces, and the perception condition category is used to describe the spatial perception factors that affect the degree of recreational satisfaction; each category is divided into multiple specific projects according to user needs, and each project is described by multiple evaluation points; and the specific semantic feature data used for each suitability identification project is extracted; according to the project description, the training atlas is sequentially compared two by two and marked with pros and cons, and the supplementary potential identification score of each project is obtained, and the image potential label set of the training set is constructed; A fully connected image convolutional neural network built based on a deep learning framework tool was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. This was used to establish a stable inference relationship between the complete visual scene information of micro-green spaces and the quantitative identification system for the replenishment potential of micro-green spaces. A fully connected image convolutional neural network based on a deep learning framework was used as the underlying model to train a quantitative identification model for the deep replenishment potential of micro-green spaces. The scores of each project at each potential replenishment point were obtained, and the impact priority of each project was compared. The project weights were calculated by constructing a weight-limited relationship, and the scores of each project potential identification project were weighted and superimposed to obtain the comprehensive potential score of the replenishment point. Based on the comprehensive potential scores of the supplementary points, potential supplementary points of micro-green spaces with high potential were screened out, and relevant visual semantic indicators were calculated for the corresponding street view image sets to establish a potential point feature description dataset. K-means clustering was performed based on the scores of the four potential identification items of the potential points to obtain a comprehensive potential grade dataset. Based on the potential point street view image dataset, feature description dataset and comprehensive potential grade dataset, a micro-green space supplementation strategy was formulated.

Images