A method for identifying a greenway network based on machine learning

Abstract

The application provides a greenway network identification method based on machine learning, relates to the field of greenway network identification and planning, and collects resident movement trajectories, mobile phone signaling, street views and the like to construct a status database, extracts all road-related indexes, adopts a neural network algorithm to train different attribute change rules of the movement trajectories, simulates the movement trajectories, and compares and verifies the simulated movement trajectories with the real movement trajectories, further extracts starting points and ending points of greenway route selection, inputs the starting points and the ending points into a trained artificial intelligence framework, obtains simulated resident travel routes, counts high-frequency travel routes, identifies a greenway network, and forms a greenway planning and design scheme. The application collects resident activity data, comprehensively analyzes real travel tendencies of residents by means of an artificial intelligence analysis framework, analyzes road-related index change rules of real trajectories of the residents, simulates resident travel behaviors, and identifies the greenway network, thereby supporting greenway route selection and planning and design of a city.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of greenway network identification and planning, and particularly relates to a greenway network identification method based on machine learning. BACKGROUND

[0002] Greenways can provide leisure and recreation space, protect landscape continuity, provide habitats for plants and animals, and have functions such as recreation, education, and historical resource protection. Greenway construction can meet the increasing demand of residents for beautiful ecological environment, and promote the high-quality development of cities. In the construction of greenway system, through the optimization of greenway system and the creation of slow space, the integration of urban space and ecological space is realized, and the vitality of the city is stimulated, which is an important problem faced by urban greenway system planning. With the change of information technology, new technologies such as big data and machine learning provide new data sources and analysis methods for greenway research. Resident slow activity data can accurately quantify the actual use of space by people and is widely used in greenway research.

[0003] At present, GIS path analysis method, greenway construction suitability analysis and big data method are gradually applied in greenway network identification. However, the current greenway network identification mostly adopts a top-down analysis perspective, uses static indicators of physical space for comprehensive evaluation, focuses on spatial path models or greenway carrier evaluation, and less considers the bottom-up perspective of resident travel in the analysis, lacks research on the activity behavior pattern of residents on greenways, and the application of machine learning in the method is still in the exploratory stage. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the shortcomings of the existing method by collecting and organizing all resident real activity trajectory data as a spatial background, and proposing a greenway network identification method based on machine learning.

[0005] To solve the above technical problems, the following technical means is adopted:

[0006] The present application proposes a greenway network identification method based on machine learning, which comprises:

[0007] Step 1, collect mobile signaling data, motion trajectory data, street view images, point of interest data, land use and road data, and construct a greenway network identification database.

[0008] Step 2, calculate the road attribute index, which comprises the following sub-steps:

[0009] (201) According to the road space characteristics, the key indicators are selected to measure the road attributes, including: population density, point of interest facility density, point of interest facility mixing degree, green view rate, sky openness, road width, sidewalk width, global integration, intersection density, and distance to park.

[0010] (202), the population density within a certain distance range from the road is calculated as the road population density;

[0011] (203), the density of interest point facilities within a certain range is calculated as the road interest point facility density;

[0012] (204), the entropy value method is used to measure the mixed degree of road interest point facilities;

[0013] (205), the convolutional neural network tool Segnet based on machine learning algorithm is used for semantic segmentation operation on street view pictures to obtain the vegetation proportion and sky proportion in the pictures, and the average value of the four pictures around the street view point is obtained to obtain the green view rate and sky openness of the street view point;

[0014] (206), machine learning is used to identify the road width, and the model for calculating the width of the slow lane is output;

[0015] (207), the road width, intersection density, and distance to the park are calculated by GIS spatial analysis;

[0016] (208), the global integration degree of urban space is calculated.

[0017] Step 3, construct the road topology network; specifically:

[0018] S301, according to the principle of topology, all roads are abstracted as nodes, the relationship between nodes is abstracted as connection, and the data attributes are matched with the road network to form an attribute road network, and the attribute road network is made into a topological road network by using topology;

[0019] S302, after the road is abstracted as a node, the corresponding interconnection relationship of each road is converted into the format of the networkx library that can be accessed by the artificial intelligence framework, and the dictionary generation table is used to query all roads and interconnection relationships efficiently;

[0020] S303, the starting point and the ending point of the road in the road network are one-to-one corresponding, and the interconnection relationship between the roads is converted into a topological relationship, and after the dictionary list is established, it is drawn in networkx;

[0021] S304, after the topological road network is built, the road, order rule, and road attribute of each trajectory are used as input variables.

[0022] Step 4, the LSTM neural network is used to learn the index change rule of the resident motion trajectory, and the resident motion rule in the green belt is predicted; the specific process is:

[0023] S401, input the attribute change table generated according to the motion trajectory road sequence into the LSTM neural network for training; export the analyzed rule model through the neural network as the operation rule of the subsequent artificial intelligence agent, randomly select a part of the motion trajectory as the training set, and the rest as the test set, and perform neural network model training;

[0024] S402, compare the root mean square error of each attribute in the overall range according to the model test, select the factor with higher fitting degree, and eliminate the index with larger error;

[0025] S403, verify the robustness of the model, fit the index change trend of each real trajectory and simulated trajectory, analyze the similarity between the two, and verify the similarity between the real line and the simulated line.

[0026] Step 5, greenway path simulation:

[0027] (501) input the attribute change rule model and the topological road network into the artificial intelligence framework, and verify the model;

[0028] (502) calculate the real heat value and the predicted heat value of each trajectory, analyze the correlation between the real value and the predicted value, and confirm the credibility of the overall model;

[0029] (503) input the attribute change rule model and the topological road network into the artificial intelligence framework, input the starting point and the ending point of the greenway, and finally output the motion trajectory after the route simulation.

[0030] Step 6, identify the greenway network: perform heat statistical processing on the simulated trajectory, extract the road network with high resident travel frequency as the greenway network identification result.

[0031] The above technical scheme is adopted in the present application, and the beneficial effects compared with the prior art are as follows:

[0032] The method proposed in the present application integrates multiple source data, measures the influence factors of greenway alignment, constructs an artificial intelligence analysis framework for greenway alignment, and proposes a greenway network identification method. The analysis process only needs to collect relevant data and input them into the artificial intelligence framework. The analysis result in the unsupervised environment is highly aggregated, the learning cost is reduced, the intelligentization and precision of greenway alignment are improved, the greenway network is effectively identified, and guidance is provided for greenway alignment planning and construction. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The method flowchart of the present application is implemented.

[0034] Figure 2A topological road network example graph in an embodiment of the present application.

[0035] Figure 3 A simulation of resident travel heat analysis result in an embodiment of the present application. DETAILED DESCRIPTION

[0036] The technical solutions of the present application will be further described in detail below in combination with the drawings:

[0037] The present application analyzes the real travel tendency of residents by collecting resident activity data and using an artificial intelligence analysis framework, analyzes the change rule of road-related indicators of real trajectories of residents, simulates resident travel behavior and identifies greenway networks, and supports city greenway alignment and planning design. The present example is a greenway network identification method based on machine learning, and the flowchart is shown in Figure 1 , which mainly includes the following steps:

[0038] Step 1: Construct a greenway network identification database. Collect mobile signaling data, motion trajectory data, street view images, point of interest data, land use and road data, etc., and organize them into a geographic spatial database.

[0039] Step 2, calculate the road attribute index. The specific method is as follows:

[0040] (1) According to the spatial characteristics of the road, select 10 indexes of population density, point of interest facility density, point of interest facility mixed degree, green view rate, sky openness, road width, sidewalk width, global integration, intersection density, and distance to park to measure the road attribute.

[0041] (2) Calculate the population density within 100m of the road as the road population density. A double-sided 100m buffer zone is established along the road, and the number of residents contained in the buffer zone is calculated by combining mobile signaling data and assigning each buffer zone with a population attribute. The calculation formula of population density is as follows:

[0042] PopDensity i =Pop i / (L i ×100)

[0043] Where, POIDensity i is the road population density value, Pop i is the number of residents within 100m of the i-th road, and L i is the road length.

[0044] (3) The density of POI facilities within a 50m range of the road is calculated as the road POI facility density. A bilateral 50m buffer zone is established along the road network, and the number of POI facilities of different types contained in the buffer zone is counted, and the POI facility density of each buffer zone is calculated. The calculation formula of the POI facility density is as follows:

[0045] POIDensity i =Poi i / (L i ×100÷10000)

[0046] wherein Poi i is the number of POI of the type within a 50m range of the ith road segment.

[0047] (4) The entropy value method is used to calculate the mixing degree of road POI facilities, and the calculation formula of the POI facility mixing degree is as follows:

[0048]

[0049] wherein H(X) represents the entropy of random variable X; P i is the probability of X taking X i .

[0050] (5) The convolutional neural network tool (Segnet) based on the machine learning algorithm is used to perform semantic segmentation on the street view pictures to obtain the vegetation proportion and the sky proportion in the pictures, and the average value of the four pictures around the street view point is obtained to obtain the green view rate and the sky openness of the street view point. The green view rate and the sky openness of the road are the average values of the green view rate and the sky openness of all street view points in the road segment.

[0051] (6) For the recognition of the road width, machine learning is used to enable the machine to first recognize a large number of slow-moving roads. First, the image is processed in grayscale, and the specific formula is as follows:

[0052] Gray = 0.299·R + 0.587·G + 0.114·B

[0053] wherein Gray is the grayscale value, and R, G and B represent different values in RGB. According to the above formula, the corresponding gray photo is formed, and the image is processed by smoothing filter to avoid the occurrence of noise at multiple places. Further, the Gaussian filter method is selected for processing, and the principle is to recalculate the value of each point in the image, and the point and the surrounding points are weighted and averaged during the calculation, and the weight conforms to the Gaussian distribution. Generally, the weight is formed in the form of a matrix, which is called a Gaussian kernel. A Gaussian kernel with a size of 5 is used after the image of fisheye photography is converted into a street view. The Gaussian filter formula for a two-dimensional image is as follows:

[0054]

[0055] The parameter σ is a scale parameter. The larger σ is, the more uniform the weight distribution and the better the filtering effect, but the more blurred the image becomes. Conversely, the smaller σ is, the more the weight distribution is biased towards the center point of the window, the worse the filtering effect, but the more the image retains its original clarity.

[0056] A Gaussian kernel of size 5 is:

[0057]

[0058] After smoothing the grayscale image, the edges of the pedestrian lane and the roadway are extracted by setting a reasonable threshold, taking advantage of their relatively distinct characteristics. After manually verifying and removing obviously abnormal edges, the distance between the two lines is calculated. Through repeated calculations using machine learning, a model for calculating the width of the pedestrian lane is output.

[0059] (7) Road width, intersection density, and distance to the park were calculated using GIS spatial analysis.

[0060] (8) Global integration degree is the reciprocal of the actual relative asymmetry value. The calculation formula is:

[0061]

[0062] In the formula, I i It is the overall integration of urban space, R (n) It is the actual relative asymmetry value.

[0063]

[0064] In the formula MD i is the average depth value, which refers to the minimum number of connections required for a certain space in the system to reach other spaces; n is the sum of the number of urban unit spaces.

[0065] The third step is to construct the road topology network. Following topological principles, all roads are abstracted as nodes, and the relationships between nodes are abstracted as connections. Data attributes are mapped to the road network to form an attributed road network. This attributed road network is then transformed into a topological road network using topology principles. (Refer to...) Figure 2 As shown. After abstracting roads into nodes, the corresponding road interconnections are converted into a NetworkX library format that can be accessed by the AI ​​framework. All roads and interconnections are efficiently queried using a dictionary-generated table. The starting and ending points of roads in the road network are mapped one-to-one, and the interconnections between roads are converted into topological relationships. After establishing the dictionary list, the networkX graph is drawn. Once the topological road network is built, the roads traversed by each trajectory, their order, and road attributes can all be used as input variables.

[0066] Fourth step, machine learning simulation and verification. The present application adopts LSTM (Long Short Term Mermory network) neural network to learn the index change law of the resident motion trajectory, and predict the resident travel law in the greenway. The LSTM neural network calculation formula is as follows:

[0067] ft = σ (Wt xt + Uf ht-1 + bf)

[0068] ct = tan h (Wc xt + Ucht-1 + bc)

[0069] it = σ (Wtxt + Uiht-1 + bi)

[0070] ct = ft ⊙ ct-1 + it ⊙ ct

[0071] ot = σ (Woxt + Uoht-1 + bo)

[0072] ht = tan h (ct) ⊙ ot

[0073] y = Wd hn + bd

[0074] Where, h t is the hidden state, x t is the input at t time, f t , I t , o t , c t are the forgetting gate, input gate, output gate and cell state at t time respectively. c t is the memory update vector; σ (·) represents the Sigmoid function, tan h (·) is the hyperbolic cosine function; W, U represent the weight matrix; b represents the bias vector; ⊙ represents the vector scalar product, y is the output at t time.

[0075] (1) The attribute change table generated according to the motion trajectory road sequence is input into the LSTM neural network for training. The analyzed rule model is exported through the neural network as the running rule of the subsequent artificial intelligence agent. Further, a part of the motion trajectory is randomly selected as the training set, and the rest is selected as the test set, and the neural network model is trained.

[0076] (2) According to the model test, compare the root mean square error of each attribute in the overall range, select the factor with higher fitting degree, and eliminate the index with larger error.

[0077] (3) Verify the robustness of the model, fit the index change trend of each real trajectory and simulated trajectory, analyze the similarity between them, and verify the similarity between the real line and the simulated line.

[0078] Step 5: Greenway path simulation. The specific approach is as follows:

[0079] (1) Input the attribute change law model and the topological road network into the artificial intelligence framework to verify the model. Input the starting point and the ending point of the real movement trajectory, and let the agent in the framework run the trajectory simulation in the topological road network according to the attribute change law model. The agent walks out the most reasonable path by using the A* algorithm, and finally outputs the simulated trajectory. In this study, the improved A* algorithm is used to calculate the Pearson correlation coefficient between the predicted value and the actual value. The correlation coefficient is used to calculate the dispersion coefficient, and the minimum value of the weighted sum is the optimal path with the smallest difference from the predicted optimal value. The specific calculation formula of the A* algorithm is as follows:

[0080] f(n) = g(n) + h(n)

[0081] where f(n) is the evaluation function of node n from the initial point to the target point, g(n) is the actual cost from the initial node to node n in the state space, and h(n) is the estimated cost of the best path from n to the target node.

[0082] Corresponding conditions of the A* algorithm:

[0083] f(n) = g * (n) + h * (n)

[0084] h * (n) ≤ h'(n)

[0085] h * (n) ≤ h'(n) * , h'(n) is the shortest distance from n to the target, which ensures that the heuristic function selected in the A* algorithm is optimal, and the corresponding path is the shortest path.

[0086] (2) Calculate the real heat and predicted heat value of each trajectory, analyze the correlation between the real value and the predicted value, confirm the credibility of the overall model, and refer to the figure shown in Figure 3 .

[0087] (3) Input the attribute change law model and the topological road network into the artificial intelligence framework, input the starting point and the ending point of the greenway, and let the agent in the framework run the trajectory simulation in the topological road network according to the attribute change law. Finally, output the movement trajectory after simulating the route.

[0088] Step 6: Identify the greenway network. Perform heat statistical processing on the simulated trajectory, extract the road network with high resident travel frequency, and take it as the identification result of the greenway network.

[0089] The above describes only the possible implementation of the present application, but cannot be understood as a limitation to the patent scope of the present application. For those skilled in the art, corresponding modifications and improvements can be made without departing from the concept of the present application, which are all within the protection scope of the present application. Therefore, the patent protection scope of the present application shall be subject to the appended claims.

Claims

1.A method for identifying a greenway network based on machine learning, characterized in that, The application relates to a greenway network identification method based on artificial intelligence. Step S1, constructing a greenway network identification database; Step S2, calculating road attribute indexes, specifically including the following substeps: S201, selecting key index to measure road attribute according to road space characteristics; S202, counting population density within a certain distance range from the road as road population density; S203, unifying interest point facility density within a certain range as road interest point facility density; S204, using an entropy value method to measure the mixed degree of road interest point facilities; S205, using a convolutional neural network tool Segnet based on a machine learning algorithm to perform semantic segmentation operation on street view pictures to obtain vegetation proportion and sky proportion in the pictures, and taking average values of four pictures around the street view point to obtain green view rate and sky openness of the street view point; S206, using machine learning to identify road width and output a model for calculating slow lane width; S207, calculating road width, intersection density and distance to a park through GIS spatial analysis; S208, calculating global integration of urban space; Step 3, constructing a road topology network; Step 4, using an LSTM neural network to learn index change rules of resident movement and predict resident movement rules in the greenway; Step 5, greenway path simulation, specifically including the following substeps: S501, inputting the attribute change rule model and the topology road network into an artificial intelligence framework to verify the model, specifically as follows: inputting a real movement trajectory starting point and an ending point, making an agent in the framework run a simulation trajectory in the topology road network according to the attribute change rule model, and using the agent to walk out the most reasonable path through a class A* algorithm, and finally outputting a simulation trajectory; the improved class A* algorithm calculates a Pearson correlation coefficient between a value predicted by using the rule model and an actual value, inversely calculates a dispersion coefficient by using the correlation coefficient, and the minimum value of weighted summation is the optimal path with the minimum difference from the best prediction value, and a specific calculation formula of the class A* algorithm is as follows: f(n) = g(n) + h(n) Wherein f(n) is an evaluation function of the node n from the initial point to the target point, g(n) is an actual cost from the initial node to the node n in the state space, and h(n) is an estimated cost of the best path from n to the target node; Corresponding conditions of the class A* algorithm: f(n) = g * (n) + h * (n) h * (n) is additionally conditioned by h * (n) ≤ h'(n), h'(n) being the shortest distance from n to the target. S502, calculating real heat value and predicted heat value of each trajectory, analyzing correlation between the real value and the predicted value, and confirming credibility of the overall model; S503, inputting the attribute change rule model and the topology road network into the artificial intelligence framework, inputting a greenway starting point and an ending point, making the agent in the framework run a simulation trajectory in the topology road network according to the attribute change rule, and finally outputting a movement trajectory after route simulation; Step 6, identifying the greenway network: performing heat statistical processing on the simulation trajectory, extracting a road network with higher resident travel frequency as a greenway network identification result. 2.The method of claim 1, wherein, Step S202, the population density within a certain distance range from the road is calculated as the road population density. A bilateral 100m buffer zone is established along the road, and the number of residents within the buffer zone is calculated based on the mobile signaling data. The population density is calculated according to the following formula: PopDensity i = Pop i / (L i x 100) wherein PopDensity i is the road population density value, Pop i is the number of population within 100m of the ith road segment, L i is the road length. 3.The method of claim 1, wherein, Step S203, the density of interest point facilities within a certain range is calculated as the road interest point facility density. A bilateral 50m buffer zone is established along the road network, and the number of different types of interest points within the buffer zone is calculated. The interest point facility density is calculated according to the following formula: POI Density i = Poi i / (L i × 100 ÷ 10000) where Poi i is the number of such POIs in the range of the i-th segment 50m. 4.The method of claim 1, wherein, Step S204, the entropy method is used to calculate the mixed degree of road interest point facilities. The calculation formula of the interest point facility mixed degree is as follows: where H(X) denotes the entropy of the random variable X, P i is the probability that X takes the value X i . 5.The method of claim 1, wherein, Step S206, machine learning is used to identify the road width. First, the image is processed in grayscale according to the following formula: Gray = 0.299 * R + 0.587 * G + 0.114 * B where Gray is the grayscale value, R, G, and B represent different values in RGB. According to the above formula, the corresponding gray photo is formed, and the image is smoothed. Secondly, Gaussian filtering method is selected for processing. The value of each point in the image is recalculated by weighted average of the point and its surrounding points. The weight is arranged in matrix form, called Gaussian kernel. The Gaussian filtering formula for two-dimensional image is as follows: where parameter σ is the scale parameter, The size of the Gaussian kernel is 5, as follows: After smoothing the grayscale image, the obvious characteristics of the slow lane and the carriageway edge are used to set a reasonable threshold to extract them. Finally, after manual correction to remove the obviously abnormal edges, the distance between the two straight lines is calculated. After repeated calculation by machine learning, the model for calculating the slow lane width is output. Step S208, the global integration degree of urban space is calculated as follows: 6.The method of claim 1, wherein, The global integration degree is the inverse of the actual relative asymmetry value, and the calculation formula is as follows: n is the number of urban unit spaces. where I i is the global integration of urban space, R (n) is the actual relative asymmetry value, MD i is the average depth value, indicating the minimum number of connections required for a space to reach other spaces; Step S3, the road topology network is constructed as follows: 7.The method of claim 1, wherein, S301, according to the principles of topology, all roads are abstracted as nodes, and the relationship between nodes is abstracted as connection. The data attributes are matched with the road network to form an attribute road network. The attribute road network is converted into a topology road network using topology; S302, after abstracting the road as a node, the corresponding intercommunication relationship of each road is converted into the format of networkx library that can be accessed by artificial intelligence framework. The dictionary generation table is used to query all roads and intercommunication relationships efficiently; S303, the starting point and the ending point of the road in the road network are corresponded one by one, and the intercommunication relationship between roads is converted into a topological relationship. After establishing the dictionary list, the networkx is drawn; S304, after the topology road network is built, the road, order rule, and road attribute of each trajectory are taken as input variables. Step S4, the LSTM neural network is used to learn the index change rule of the resident movement trajectory. The specific process is as follows: 8.The method of claim 1, wherein, ​ S401, input the attribute change table generated according to the motion trajectory road sequence into the LSTM neural network for training; export the analyzed rule model through the neural network as the operation rule of the subsequent artificial intelligence agent, randomly select a part of the motion trajectory as the training set, and the rest as the test set, and train the neural network model; S402, compare the root mean square error of each attribute in the overall range according to the model test, select the factor with higher fitting degree, and eliminate the indicators with error greater than the preset value; S403, verify the model robustness, fit the index change trend of each real trajectory and simulated trajectory, analyze the similarity between them, and verify the similarity between the real line and the simulated line.

Images