Sponge city sunken green land infiltration and storage effect prediction method

Abstract

The application provides a kind of spongy city concave greenbelt infiltration effect prediction method, specifically includes the following steps: input step is respectively obtained soil steady infiltration rate monitoring data, precipitation monitoring data, and obtains concave greenbelt image;Training step is to obtain the soil steady infiltration rate and precipitation information condition of concave greenbelt, and obtain the area ratio of concave greenbelt and the depth of concave greenbelt;Output step is to establish the spongy city concave greenbelt infiltration effect prediction model by particle filtering algorithm.The application is based on soil steady infiltration rate monitoring data, precipitation monitoring data and concave greenbelt image, and the infiltration effect of spongy city concave greenbelt is predicted by particle filtering algorithm, which can reflect the infiltration effect of spongy city concave greenbelt in real time as the amount of data increases.

Description

Technical Field

[0001] This invention relates to the field of computer machine learning technology, and in particular to a method for predicting the infiltration and storage effect of sunken green spaces in sponge cities. Background Technology

[0002] Sponge cities integrate the concepts of sustainable urban drainage system management, blue-green cities, water-sensitive urban design, and low-impact urban design, among other stormwater management principles, resulting in integrated innovation in urban stormwater management and drainage systems. Urban green spaces are an important component of the sponge city ecosystem. Sunken green spaces, as an effective practical application of rainwater utilization and ecological stormwater drainage system management measures, play a significant role in regulating the regional ecological environment.

[0003] Traditional methods for predicting the infiltration and storage effects of sunken green spaces in cities involve analyzing and calculating the water balance relationship of sunken green spaces based on differences in various green space structures. These methods use regression algorithms to fit multivariate models and analyze relevant parameters. However, the modeling is often limited by complex urban morphology and precipitation conditions, making it impossible to update the prediction model in real time as the amount of data increases. Consequently, these methods cannot reflect the infiltration and storage effects of sunken green spaces in sponge cities in real time.

[0004] Therefore, the aforementioned technical issues still need to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the infiltration and storage effect of sunken green spaces in sponge cities, so as to solve the technical problem that the prediction model cannot be updated in real time as the amount of data increases, and cannot reflect the infiltration and storage effect of sunken green spaces in sponge cities in real time.

[0006] To address the aforementioned technical problems, the embodiments of the present invention provide the following technical solutions:

[0007] The first aspect of this invention provides a method for predicting the infiltration and storage effect of sunken green spaces in sponge cities, the prediction method comprising the following steps:

[0008] Input steps to obtain soil steady infiltration rate monitoring data, precipitation monitoring data, and images of sunken green spaces;

[0009] The training steps involve normalizing the soil stable infiltration rate monitoring data and precipitation monitoring data to obtain the soil stable infiltration rate and precipitation information conditions, and training the prediction model function based on the image of the sunken green space to obtain the area ratio and depth of the sunken green space.

[0010] The output step involves establishing a predictive model for the infiltration and storage effect of sunken green spaces in sponge cities based on the stable soil infiltration rate, precipitation information, the area ratio and depth of sunken green spaces, and using a particle filtering algorithm.

[0011] Optional, in the input step:

[0012] By using data from sunken green spaces distributed throughout sponge cities, soil stable infiltration rate monitoring data and precipitation monitoring data are saved in different community lists according to time series.

[0013] Images of sunken green spaces are collected by an online image monitoring device.

[0014] Optionally, the training steps include data training steps and image training steps;

[0015] The data training steps include:

[0016] The soil steady infiltration rate monitoring data and precipitation monitoring data were normalized and saved as soil steady infiltration rate data sample set and precipitation information condition data sample set, respectively. The uniformly extracted features were then transmitted to the computer to provide soil steady infiltration rate and precipitation information conditions.

[0017] Image training steps include:

[0018] S1: Divide the input concave green space image to be detected into multiple region boxes. Utilize the change information of the concave green space image to select potential regions containing objects in the concave green space image. Then, scale the size of the concave green space image and use a convolutional neural network algorithm to extract feature data from the candidate boxes of potential regions containing objects. Establish a set of digital labels with feature data and match the feature data with the digital labels.

[0019] S2: Based on the obtained feature data, a classification model is established. The category of the concave green space image is obtained by comparing the extracted feature data with the classification rules contained in the classification model. The classification prediction model function of the concave green space image is trained in the sample space. The weight parameters in the network are adjusted layer by layer by using a multi-task loss function. The weight of a single task is dynamically adjusted throughout the training process. The convergence is accelerated by using a deep model optimization algorithm. The accuracy of the network is improved by frequent iterative training. The nonlinear classification of the concave green space image is completed by fitting the constructed dataset with a support vector machine classification model.

[0020] S3: Read in the image of the concave green space feature area, set a mask to focus the concave green space image onto the green space, highlight the foreground through grayscale histogram statistics to enhance the features of the concave green space image, optimize the green space image through median filtering, extract the edge contour through edge detection to generate the concave green space contour image, fill the internal area of ​​the concave green space contour image through the seed filling algorithm, sum the area of ​​all concave green space contour images, divide the area of ​​the concave green space contour image by the total area of ​​the green space to calculate the area ratio of concave green spaces, and establish the area ratio of concave green spaces distributed in the sponge city.

[0021] S4: By reading the feature region image of the sunken green space through the feature region bounding box, the image of the sunken green space collected by the online monitoring device is segmented into superpixels using a simple linear iterative clustering superpixel segmentation algorithm. The feature region image of the sunken green space is mapped to the superpixel segmentation result and the corresponding region is marked. Then, the spatial relationship features between the sunken green space and the road surface or curbstone are used to extract the features of the sunken green space and calculate the depth of the sunken green space in the sponge city.

[0022] Optionally, in the data training step:

[0023] By using data from sunken green spaces distributed throughout sponge cities, soil stable infiltration rate monitoring data are stored in different community lists according to time series. This represents the information set corresponding to the first to t time nodes of N different communities, which constitutes the soil stable infiltration rate information training sample set. The soil stable infiltration rate data is uniformly normalized to obtain the soil stable infiltration rate data sample set.

[0024] The precipitation monitoring data of the sponge city project will be stored in different lists of communities according to time series, forming a training sample set of precipitation information data. This represents the set of precipitation information from the 1st to the tth time point for N different communities in the corresponding sponge city. The precipitation information data is uniformly normalized and stored as a sample set of precipitation information data.

[0025] The collected and stored soil stable infiltration rate data sample set and precipitation information data sample set are used to extract features and transmit them to the computer.

[0026] Optional, in S1:

[0027] S11: The image of the sunken green space is acquired by the online image monitoring device. The sunken green space image to be detected is divided into multiple potential region boxes. The potential regions containing green space are selected in the image using the image change information. Candidate boxes are extracted from the potential regions and directly applied to feature building. The position labels of the boxes are stored in an array.

[0028] S12: After scaling the image size, use convolutional layers and sampling layers alternately to extract features, search for potential regions where objects exist, extract features from the potential regions where objects exist, establish digital labels with feature data, and match the data features with the digital labels.

[0029] Optional, in S2:

[0030] S21: Use image training step S1 to complete the feature extraction of the sunken green space image, establish a classification model based on the obtained feature data, complete the classifier training of the sunken green space image, and obtain the category of the sunken green space image by comparing the extracted features with the classification rules contained in the classification model.

[0031] S22: Establish a set of digital labels with feature data, correspond the features representing the category of concave green space images with the digital labels, normalize the digital labels, and pass both the category features and the digital labels to the computer for training.

[0032] S23: Train a classification and prediction model function for concave green space images in the sample space. By employing a multi-task loss function to adjust the weight parameters of the network layer by layer inversely, and by dynamically adjusting the weights of individual tasks throughout the training process, a deep model optimization algorithm is used to accelerate convergence.

[0033]

[0034] in L represents the multi-task loss function. T This represents the three-part graph loss function. Indicates that the output triangulation is adaptive, L α This represents the loss function estimated by α. Let α represent the estimation, σ1 and σ2 represent the dynamically adjusted task weights, and T represent the α estimation. gt Let α represent the cross-entropy loss of the trie graph. gt α represents the estimated loss.

[0035] Optional, in S3:

[0036] S31: Obtain the feature region bounding box of the concave green space through the concave green space image classifier trained in step S2, read in the feature region image of the concave green space, set a mask to focus the feature region image of the concave green space onto the concave green space, highlight the foreground through grayscale histogram statistics to enhance the features of the concave green space image, optimize the image through median filtering, extract the edge contour through edge detection, and generate the contour image of the concave green space.

[0037] S32: Fill the internal area of ​​the concave green space outline in S31 using the seed filling algorithm. Use stack to fill the polygons in the internal area of ​​the outline. Sum the areas of all concave green space outlines. Divide the area of ​​the concave green space outline by the total area to calculate the area ratio of the concave green space.

[0038] Optional, in S4:

[0039] S41: Using the concave green space feature region bounding box, read in the image of the concave green space feature region. Utilize a simple linear iterative clustering superpixel segmentation algorithm to segment the image of the concave green space acquired by the online monitoring device to obtain superpixels. Initialize seed points and perform similarity measurement, calculating the similarity L between each pixel i and its nearest seed point. i The label of the most similar seed point is assigned to pixel i, and this process is iterated until convergence. Using superpixels as the basic objects, a support vector machine classification algorithm is used to segment the vertical boundary of the concave green space based on the features of the superpixels, with similarity L. i for:

[0040]

[0041] Among them, l c For the color difference between pixels, l xy L represents the spatial distance between pixels. i For similarity, k is a balancing parameter to measure the weight of color value and spatial distance in the similarity metric, and d is the spacing between seed points;

[0042] S42: Map the green space feature area image to the superpixel segmentation result and label the corresponding area. Since the size of the concave green space feature area image in the binary image is the same as the size in the superpixel segmentation image, the coordinates (x, y) of the concave green space feature area image in the binary image are the same as the coordinates (x', y') in the superpixel segmentation image. Therefore, the mapping relationship is: (x', y') = (x, y). Calculate the difference between the average brightness of all concave green space superpixels and the average brightness of the superpixels of the road surface and curbstone adjacent to the target area. If the difference is greater than the threshold ε, then the area is the location of the concave green space. Utilize the spatial relationship features between the concave green space and the road surface or curbstone to extract the vertical concave green space features and calculate the depth of the concave green space in the sponge city.

[0043] Optionally, in the output step:

[0044] With the increase in monitoring data on soil stable infiltration rate and precipitation in sponge cities, the proportion of sunken green space area and the depth of sunken green space can be obtained by collecting green space images from online image monitoring devices.

[0045] Optionally, in the output step:

[0046] The state-space model of a nonlinear dynamic system is Where γ represents the state transition function. This represents the predicted infiltration rate of the concave green space in community k at time t. This represents the predicted rainwater infiltration and storage rate of the sunken green space at time t-1 using station k. Where λ represents the measurement function. This represents the observation results of the sunken green space area ratio, green space depression depth, soil stable infiltration rate, and precipitation information conditions at time t in cell k. Let the observation results of cell k at time t-1 be represented, and construct the posterior probability density function. Let the set of N particles represent the posterior probability density function of cell k at time t for each parameter of the sunken green space. The i-th sample particle is drawn from the set of N particles (i = 1, 2, ..., N), and its weight is... It means, and Introducing the sampling density function according to The weights of the particles are recursively calculated and normalized to train a prediction model for the infiltration rate of sunken green spaces in sponge cities. A particle filtering algorithm is used to establish a model function relating the infiltration rate of sunken green spaces to the area ratio of sunken green spaces, the depth of the sunken green spaces, the stable soil infiltration rate, and precipitation information conditions. This model can then predict the infiltration effect of sunken green spaces in sponge cities in real time.

[0047] Compared to existing technologies, this invention provides a method for predicting the infiltration and storage effect of sunken green spaces in sponge cities. The method includes: an input step that acquires soil stable infiltration rate monitoring data, precipitation monitoring data, and images of sunken green spaces; a training step that normalizes the soil stable infiltration rate monitoring data and precipitation monitoring data, and trains a classifier for the sunken green space images to obtain the area ratio and depth of the sunken green spaces; and an output step that establishes a predictive model for the infiltration and storage effect of sunken green spaces in sponge cities using a particle filtering algorithm based on the area ratio and depth of the sunken green spaces. In other words, by using soil stable infiltration rate monitoring data, precipitation monitoring data, and images of sunken green spaces, and employing a particle filtering algorithm to predict the infiltration and storage effect of sunken green spaces in sponge cities, this method can reflect the infiltration and storage effect of sunken green spaces in sponge cities in real time as the amount of data increases. Attached Figure Description

[0048] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0049] Figure 1 This is a flowchart of the sponge city sunken green space infiltration and storage effect prediction system of the present invention;

[0050] Figure 2 This is a flowchart of the process for extracting the location features of sunken green spaces according to the present invention;

[0051] Figure 3 This is a flowchart of the recessed green space outline recognition process of the present invention;

[0052] Figure 4 This is a flowchart illustrating the extraction process of the recessed depth of concave green spaces based on superpixel spatial relationship features according to the present invention. Detailed Implementation

[0053] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. Unless otherwise specified, the techniques used in the embodiments are conventional means well known to those skilled in the art.

[0054] It should be noted that, unless otherwise stated, the technical or scientific terms used in this invention should be understood in their ordinary sense by those skilled in the art. The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0055] like Figure 1 As shown in the figure, this invention provides a method for predicting the infiltration and storage effect of sunken green spaces in sponge cities. The prediction method includes the following steps:

[0056] Input steps to obtain soil steady infiltration rate monitoring data, precipitation monitoring data, and images of sunken green spaces;

[0057] In the input steps:

[0058] By using data from sunken green spaces distributed throughout sponge cities, soil stable infiltration rate monitoring data and precipitation monitoring data are saved in different community lists according to time series.

[0059] Images of sunken green spaces are collected by an online image monitoring device.

[0060] The training steps involve normalizing the soil stable infiltration rate monitoring data and precipitation monitoring data to obtain the soil stable infiltration rate (K) and precipitation information conditions (Q), and training the prediction model function based on the image of the sunken green space to obtain the area ratio (f) and depth (Δh) of the sunken green space.

[0061] like Figure 1 and Figure 2 As shown, the training steps in this embodiment include a data training step and an image training step;

[0062] The data training steps in this embodiment include:

[0063] The soil steady infiltration rate monitoring data and precipitation monitoring data were normalized and saved as soil steady infiltration rate data sample set and precipitation information data sample set, respectively. The uniformly extracted features were then transmitted to the computer to provide soil steady infiltration rate (K) and precipitation information conditions (Q).

[0064] In the data training steps of this embodiment:

[0065] By using data from sunken green spaces distributed throughout sponge cities, soil stable infiltration rate monitoring data are stored in different community lists according to time series. This represents the information set corresponding to the first to t time nodes of N different communities, which constitutes the soil stable infiltration rate information training sample set. The soil stable infiltration rate data is uniformly normalized and stored as a soil stable infiltration rate data sample set.

[0066] The precipitation monitoring data of the sponge city project will be stored in different lists of communities according to time series, forming a training sample set of precipitation information data. This represents the set of precipitation information from the 1st to the tth time point for N different communities in the corresponding sponge city. The precipitation information data is uniformly normalized and stored as a sample set of precipitation information data.

[0067] The collected and stored soil stable infiltration rate data sample set and precipitation information data sample set are used to extract features and transmit them to the computer.

[0068] The image training steps in this embodiment include:

[0069] like Figure 2As shown, S1: Divide the input concave green space image to be detected into multiple region boxes. Utilize the change information of the concave green space image to select potential regions containing objects in the concave green space image. Then, scale the size of the concave green space image and use a convolutional neural network algorithm to extract feature data from the candidate boxes of potential regions containing objects. Establish a set of digital labels with feature data and match the feature data with the digital labels.

[0070] In S1 of this embodiment:

[0071] S11: The image of the concave green space is acquired by the online image monitoring device. The concave image to be detected is divided into multiple potential region boxes. The potential regions containing green space are selected in the image using the image change information. Candidate boxes are extracted from the potential regions and directly applied to feature building. The position labels of the boxes are stored in an array.

[0072] In this embodiment, the image change information includes one or more of the following: edge, texture, color, and hue.

[0073] S12: After scaling the image size, use convolutional layers and sampling layers alternately to extract features, search for potential regions where objects exist, extract features from the potential regions where objects exist, establish a set of digital labels with feature data, and match the data features with the digital labels.

[0074] S2: Based on the obtained feature data, a classification model is established. The category of the concave green space image is obtained by comparing the extracted feature data with the classification rules contained in the classification model. The classification prediction model function of the concave green space image is trained in the sample space. The weight parameters in the network are adjusted layer by layer by using a multi-task loss function. The weight of a single task is dynamically adjusted throughout the training process. The convergence is accelerated by using a deep model optimization algorithm. The accuracy of the network is improved by frequent iterative training. The nonlinear classification of the concave green space image is completed by fitting the constructed dataset with a support vector machine classification model.

[0075] In S2 of this embodiment:

[0076] S21: Use image training step S1 to complete the feature extraction of the sunken green space image, establish a classification model based on the obtained feature data, complete the classifier training of the sunken green space image, and obtain the category of the sunken green space image by comparing the extracted features with the classification rules contained in the classification model.

[0077] S22: Establish a set of digital labels with feature data, correspond the features representing the category of concave green space images with the digital labels, normalize the digital labels, and pass both the category features and the digital labels to the computer for training.

[0078] S23: Train a classification and prediction model function for concave green space images in the sample space. By employing a multi-task loss function to adjust the weight parameters of the network layer by layer inversely, and by dynamically adjusting the weights of individual tasks throughout the training process, a deep model optimization algorithm is used to accelerate convergence.

[0079]

[0080] in, L represents the multi-task loss function. T This represents the three-part graph loss function. Indicates that the output triangulation is adaptive, L α This represents the loss function estimated by α. Let α represent the estimation, σ1 and σ2 represent the dynamically adjusted task weights, and T represent the α estimation. gt Let α represent the cross-entropy loss of the trie graph. gt This represents the α estimation loss. The α estimation loss is only used to estimate the unknown regions in the trilateration map. The accuracy of the network is improved through frequent iterative training. The constructed dataset is fitted by a support vector machine classification model to complete the nonlinear classification of the concave green space image.

[0081] like Figure 3 As shown, S3: Read in the image of the concave green space feature area, set a mask to focus the concave green space image onto the green space, highlight the foreground through grayscale histogram statistics to enhance the features of the concave green space image, optimize the green space image through median filtering, extract the edge contour through edge detection to generate the concave green space contour image, fill the internal region of the concave green space contour image through the seed filling algorithm, sum the area of ​​all concave green space contour images, divide the area of ​​the concave green space contour image by the total area of ​​the green space to calculate the concave green space area ratio, and establish the concave green space area ratio distributed in the sponge city. Here, the green space area ratio of each community in the sponge city should not be lower than the green space ratio of the residential area of ​​each community in the sponge city.

[0082] In S3 of this embodiment:

[0083] S31: Obtain the concave green space feature region bounding box through the concave green space image classifier trained in step S2 of the embodiment, read in the concave green space feature region image, set a mask to focus the concave green space feature region image onto the concave green space, highlight the foreground through grayscale histogram statistics to enhance the concave green space image features, optimize the image through median filtering, extract the edge contour through edge detection, and generate the contour image of the concave green space.

[0084] S32: The seed filling algorithm is used to fill the internal area of ​​the concave green space outline in S31. Stacking is used to fill the polygons in the internal area of ​​the outline. The area of ​​all concave green space outlines is summed. The area ratio of concave green space can be calculated by dividing the area of ​​the concave green space outline by the total area. The area ratio of concave green space distributed in the sponge city can be obtained. The area ratio of concave green space in the sponge city should not be lower than the green space ratio of the residential area.

[0085] S4: By reading the feature region image of the sunken green space through the feature region bounding box, the image of the sunken green space collected by the online monitoring device is segmented into superpixels using a simple linear iterative clustering superpixel segmentation algorithm. The feature region image of the sunken green space is mapped to the superpixel segmentation result and the corresponding region is marked. Then, the vertical features of the sunken green space are extracted by using the spatial relationship features between the sunken green space and the road surface or curbstone, and the depth of the sunken green space in the sponge city is calculated.

[0086] like Figure 4 As shown, in S4 of this embodiment:

[0087] S41: Using the concave green space feature region bounding box, read in the image of the concave green space feature region. Utilize a simple linear iterative clustering superpixel segmentation algorithm to segment the image of the concave green space acquired by the online monitoring device to obtain superpixels. Initialize seed points and perform similarity measurement, calculating the similarity L between each pixel i and its nearest seed point. i The label of the most similar seed point is assigned to pixel i, and this process is iterated until convergence. Using superpixels as the basic objects, the color and texture features of the superpixels are extracted. The vertical boundary of the concave green space is segmented using a support vector machine classification algorithm, and the similarity L is calculated. i for:

[0088]

[0089] Among them l c For the color difference between pixels, l xy L represents the spatial distance between pixels. i For similarity, k is a balancing parameter to measure the weight of color value and spatial distance in the similarity metric, and d is the spacing between seed points;

[0090] S42: Map the green space feature area image to the superpixel segmentation result and label the corresponding area. Since the size of the concave green space feature area image in the binary image is the same as its size in the superpixel segmentation image, the coordinates (x, y) of the concave green space feature area image in the binary image are the same as the coordinates (x', y') in the superpixel segmentation image. Therefore, the mapping relationship is: (x', y') = (x, y). Calculate the difference between the average brightness of all concave green space superpixels and the average brightness of the superpixels of the road surface and curbstone adjacent to the target area. If the difference is greater than the threshold ε, then the area is the location of the concave green space. Utilize the spatial relationship features between the concave green space and the road surface or curbstone to extract the vertical concave green space features, calculate the depth of the concave green space in the sponge city, and use the completeness and accuracy of target extraction as evaluation indicators for the accuracy of concave green space depth calculation.

[0091] The output step involves establishing a predictive model for the infiltration and storage effect of sunken green spaces in sponge cities based on the stable soil infiltration rate (K), precipitation information (Q), the area ratio (f) of sunken green spaces, and the depth (Δh) of sunken green spaces using a particle filtering algorithm.

[0092] In the output steps of this embodiment:

[0093] With the increase of monitoring data on soil stable infiltration rate and precipitation in sponge cities, the area ratio (f) and depth (Δh) of sunken green space can be obtained by collecting green space images from online image monitoring devices.

[0094] In the output steps of this embodiment:

[0095] The state-space model of a nonlinear dynamic system is Where γ represents the state transition function. This represents the predicted infiltration rate of the concave green space in community k at time t. This represents the predicted infiltration rate of the sunken green space at time t-1 using station k. Where λ represents the measurement function. The data represent the observed results of the sunken green space area ratio (f), green space depression depth (Δh), soil stable infiltration rate (K), and precipitation information conditions (Q) at time t in cell k. Let the observation results of cell k at time t-1 be represented, and construct the posterior probability density function. Let the set of N particles represent the posterior probability density function of cell k at time t for each parameter of the sunken green space. The i-th sample particle is drawn from the set of N particles (i = 1, 2, ..., N), and its weight is... It means, and Introducing the sampling density function according to The weights of particles are recursively calculated and normalized to train a prediction model for the infiltration rate of sunken green spaces in sponge cities. A particle filtering algorithm is used to establish a relationship model function between the infiltration rate of sunken green spaces and the area ratio (f), the depth of the sunken area (Δh), the stable soil infiltration rate (K), and the precipitation information conditions (Q). This model is used to predict the infiltration effect of sunken green spaces in sponge cities in real time.

[0096] The present invention provides a method for predicting the infiltration and storage effect of sunken green spaces in sponge cities. This method can achieve real-time prediction of the infiltration and storage effect of sunken green spaces in different communities throughout the sponge city and can correct and update the prediction model based on the increase of data. It is not affected by the amount of observation data or information. Even for complex green space monitoring image information, as long as it is within the monitoring range of the community, real-time prediction can be performed. The prediction model automatically extracts data features without the need for a large amount of specific measured data or estimation of green space coverage. It can train and update the image recognition model in real time using existing monitoring information and perform real-time analysis of accuracy and computational efficiency. This enables timely updates of the current infiltration and storage effect of sunken green spaces based on changing parameter information, as well as prediction of the infiltration and storage effect of sunken green spaces in the sponge city at future moments. This reflects the rainwater drainage regulation capacity of sponge city communities in a timely manner, thereby better serving the management of urban rainwater infiltration capacity.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the infiltration and storage effect of sunken green spaces in sponge cities, characterized in that, The prediction method includes the following steps: Input steps to obtain soil steady infiltration rate monitoring data, precipitation monitoring data, and images of sunken green spaces; The training steps involve normalizing the soil stable infiltration rate monitoring data and precipitation monitoring data to obtain the soil stable infiltration rate and precipitation information conditions, and training the prediction model function based on the image of the sunken green space to obtain the area ratio and depth of the sunken green space. The output step involves establishing a predictive model for the infiltration and storage effect of sunken green spaces in sponge cities based on the stable soil infiltration rate, precipitation information, the area ratio and depth of sunken green spaces, and using a particle filtering algorithm. The training steps include data training steps and image training steps; The data training steps include: The soil steady infiltration rate monitoring data and precipitation monitoring data were normalized and saved as soil steady infiltration rate data sample set and precipitation information condition data sample set, respectively. The uniformly extracted features were then transmitted to the computer to provide soil steady infiltration rate and precipitation information conditions. The image training steps include: S1: Divide the input concave green space image to be detected into multiple region boxes. Utilize the change information of the concave green space image to select potential regions containing objects in the concave green space image. Then, scale the size of the concave green space image and use a convolutional neural network algorithm to extract feature data from the candidate boxes of potential regions containing objects. Establish a set of digital labels with feature data and match the feature data with the digital labels. S2: Based on the obtained feature data, a classification model is established. The category of the concave green space image is obtained by comparing the extracted feature data with the classification rules contained in the classification model. The classification prediction model function of the concave green space image is trained in the sample space. The weight parameters in the network are adjusted layer by layer by using a multi-task loss function. The weight of a single task is dynamically adjusted throughout the training process. The convergence is accelerated by using a deep model optimization algorithm. The accuracy of the network is improved by frequent iterative training. The nonlinear classification of the concave green space image is completed by fitting the constructed dataset with a support vector machine classification model. S3: Read in the image of the concave green space feature area, set a mask to focus the concave green space image onto the green space, highlight the foreground through grayscale histogram statistics to enhance the features of the concave green space image, optimize the green space image through median filtering, extract the edge contour through edge detection to generate the concave green space contour image, fill the internal region of the concave green space contour image through the seed filling algorithm, sum the area of ​​all concave green space contour images, divide the area of ​​the concave green space contour image by the total area of ​​the green space to calculate the area ratio of concave green spaces, and establish the area ratio of concave green spaces distributed in the sponge city; S4: By reading the feature region image of the sunken green space through the feature region bounding box, the image of the sunken green space collected by the online monitoring device is segmented into superpixels using a simple linear iterative clustering superpixel segmentation algorithm. The feature region image of the sunken green space is mapped to the superpixel segmentation result and the corresponding region is marked. Then, the spatial relationship features between the sunken green space and the road surface or curbstone are used to extract the features of the sunken green space and calculate the depth of the sunken green space in the sponge city.

2. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 1, characterized in that, In the input step: By using data from sunken green spaces distributed throughout sponge cities, soil stable infiltration rate monitoring data and precipitation monitoring data are saved in different community lists according to time series. Images of sunken green spaces are collected by an online image monitoring device.

3. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 2, characterized in that, In the data training step: By using data from sunken green spaces distributed throughout sponge cities, soil stable infiltration rate monitoring data are stored in different community lists according to time series. This represents the information set corresponding to the first to t time nodes of N different communities, which constitutes the soil stable infiltration rate information training sample set. The soil stable infiltration rate data is uniformly normalized to obtain the soil stable infiltration rate data sample set. The precipitation monitoring data of the sponge city project will be stored in different lists of communities according to time series, forming a training sample set of precipitation information data. This represents the set of precipitation information from the 1st to the tth time point for N different communities in the corresponding sponge city. The precipitation information data is uniformly normalized and stored as a sample set of precipitation information data. The collected and stored soil stable infiltration rate data sample set and precipitation information data sample set are used to extract features and transmit them to the computer.

4. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 1, characterized in that, In S1: S11: The image of the concave green space is acquired by the online image monitoring device. The concave image to be detected is divided into multiple potential region boxes. The potential regions containing green space are selected in the image using the image change information. Candidate boxes are extracted from the potential regions and directly applied to feature building. The position labels of the boxes are stored in an array. S12: After scaling the image size, use convolutional layers and sampling layers alternately to extract features, search for potential regions where objects exist, extract features from the potential regions where objects exist, establish digital labels with feature data, and match the data features with the digital labels.

5. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 3, characterized in that, In S2: S21: Use image training step S1 to complete the feature extraction of the sunken green space image, establish a classification model based on the obtained feature data, complete the classifier training of the sunken green space image, and obtain the category of the sunken green space image by comparing the extracted features with the classification rules contained in the classification model. S22: Establish a set of digital labels with feature data, correspond the features representing the category of concave green space images with the digital labels, normalize the digital labels, and pass both the category features and the digital labels to the computer for training. S23: Train a classification and prediction model function for concave green space images in the sample space. By employing a multi-task loss function to adjust the weight parameters of the network layer by layer inversely, and by dynamically adjusting the weights of individual tasks throughout the training process, a deep model optimization algorithm is used to accelerate convergence. ， in, Represents the multi-task loss function. This represents the three-part graph loss function. This indicates that the output triangulation is adaptive. This represents the loss function estimated by α. Indicates α estimation, and This indicates dynamically adjusted task weights. This represents the cross-entropy loss of the trie graph. α represents the estimated loss.

6. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 3, characterized in that, In S3: S31: Obtain the feature region bounding box of the concave green space through the concave green space image classifier trained in step S2, read in the feature region image of the concave green space, set a mask to focus the feature region image of the concave green space onto the concave green space, highlight the foreground through grayscale histogram statistics to enhance the features of the concave green space image, optimize the image through median filtering, extract the edge contour through edge detection, and generate the contour image of the concave green space. S32: Fill the internal area of ​​the concave green space outline in S31 using the seed filling algorithm. Use stack to fill the polygons in the internal area of ​​the outline. Sum the areas of all concave green space outlines. Divide the area of ​​the concave green space outline by the total area to calculate the area ratio of the concave green space.

7. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 3, characterized in that, In S4: S41: Using the concave green space feature region bounding box, read in the image of the concave green space feature region. Utilize a simple linear iterative clustering superpixel segmentation algorithm to segment the image of the concave green space acquired by the online monitoring device to obtain superpixels. Initialize seed points for similarity measurement and calculate the superpixel value for each pixel. i Similarity to its nearest seed point L i Assign the label of the most similar seed point to the pixel. i The process is iterated until convergence. Using superpixels as the basic objects, a support vector machine classification algorithm is used to segment the vertical boundary of the concave green space based on the features of the superpixels, and the similarity is calculated. L i for: ， in, l c For the color difference between pixels, l xy The spatial distance between pixels. L i For similarity, k To balance parameters to measure the weight of color value and spatial distance in similarity metrics, d The spacing between seed points; S42: Map the green space feature region image to the superpixel segmentation result and label the corresponding region. Since the size of the concave green space feature region image in the binary image is the same as its size in the superpixel segmentation image, the coordinates of the concave green space feature region image in the binary image are ( x , y ) and coordinates in the superpixel segmentation map ( x ', y Since ') are the same, the mapping relationship is: ( x ', y ')=( x , y The difference between the average brightness of all superpixels of sunken green spaces and the average brightness of superpixels of adjacent road surfaces and curbs in the target area is calculated. If the difference is greater than the threshold ε, then the area is the location of the sunken green space. The spatial relationship between the sunken green space and the road surface or curb is used to extract the vertical features of the sunken green space and calculate the depth of the sunken green space in the sponge city.

8. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 1, characterized in that, In the output step: With the increase in monitoring data on soil stable infiltration rate and precipitation in sponge cities, the proportion of sunken green space area and the depth of sunken green space can be obtained by collecting green space images from online image monitoring devices.

9. The method for predicting the infiltration and storage effect of sunken green spaces in sponge cities according to claim 1, characterized in that, In the output step: The state-space model of a nonlinear dynamic system is ,in γ Represents the state transition function. express k residential area t Predicted infiltration rate of sunken green space at any given time express k Taiwan Station t Predicted rainwater infiltration rate of sunken green space at time -1 in λ Represents the measurement function. express k residential area t The observation results of the proportion of sunken green space area, the depth of the green space depression, the stable soil infiltration rate, and precipitation information at any given time. express k residential area t Construct the posterior probability density function based on the observations at time -1. ,make N A set consisting of particles t time k The posterior probability density function of the community regarding various parameters of the sunken green space, from... The first particle extracted from the particle i The sample particle, the first i The weights of each sample particle are... It means, and Introducing the sampling density function ,according to The weights of the particles are recursively calculated and normalized to train a prediction model for the infiltration rate of sunken green spaces in sponge cities. A particle filtering algorithm is used to establish a model function relating the infiltration rate of sunken green spaces to the area ratio of sunken green spaces, the depth of the sunken green spaces, the stable soil infiltration rate, and precipitation information conditions. This model can then predict the infiltration effect of sunken green spaces in sponge cities in real time.

Images