A deep learning-based urban block land fine division method and system

Abstract

The application discloses a kind of urban neighborhood land fine division method and system based on deep learning, belong to geographic information processing field.The application first collects the remote sensing image data of case city area, building space form data, industry POI data and city land data, extracts various information conversion into two-dimensional plane image format data, and carries out gridding and data enhancement processing;Then through the grid translation window method and grid land function determination rule, the land division convolution neural network deep learning model is trained, and then the grid function determination of target area city land is carried out, and then the grid boundary is divided by the watershed method, finally the result is checked and corrected, and the final result of target area land division is obtained.The application can finely divide the land inside the city land neighborhood, realize the automatic division and function determination of city land, and promote the digitalization and intelligent application transformation of city planning work.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of geographic information processing, and particularly relates to a city block land fine division method and system based on deep learning. BACKGROUND

[0002] On the one hand, with the development of deep learning, the related research of using machines to automatically identify city land has emerged. There are a lot of researches on the automatic identification of land function in current research, which uses different identification technologies to study the automatic identification of land function from different perspectives, but there are problems of inaccurate function identification, unclear boundary division and weak interaction of identification results. On the other hand, the traditional large-scale city land identification through manual checking consumes a lot of manpower and material resources. In the traditional method, land information collection often relies on field investigation by manpower, which is time-consuming and labor-intensive and has a long update cycle. Therefore, it is necessary to combine existing remote sensing images, building space form, industry POI and other multi-source data to use deep learning technology to research an automatic and accurate method for determining city block scale land. SUMMARY

[0003] The purpose of the present application is to provide a multi-source data and deep learning city block land fine division method and system to improve the speed of identifying the current land function and the accuracy of dividing the land function boundary, shorten the cycle of updating the land function data, and reduce the consumption of manpower.

[0004] The technical scheme of the present application can be realized by the following technical scheme:

[0005] A city block land fine division method based on deep learning, comprising the following steps:

[0006] Step S1: Collect remote sensing image data, building space form data, industry POI data and city land data of a case city area, and construct a case city database;

[0007] Step S2: Extract remote sensing image information, building space form information and industry POI information of each block in the case city database, and convert various types of information into two-dimensional plane image format data, rasterize the two-dimensional information image and perform data enhancement processing, and construct a case city grid information database;

[0008] Step S3: Through the translation window method and the grid land function determination rule, the remote sensing image information, the building space form information, the industry POI information and the corresponding city land function information contained in each grid are subjected to deep learning training, and the training accuracy of the land division convolutional neural network deep learning model is recorded; the model is trained for multiple rounds with different window sizes, the results of multiple training are superimposed, and the highest value of the model training accuracy is selected;

[0009] Step S4: The remote sensing image data, building space form data, and industry POI data of the target plot are two-dimensionally converted and data enhanced, the land use division convolutional neural network deep learning model constructed in step S3 is used to determine the grid function of the target plot city land, the grid boundary is divided by the watershed method, the land use division and function determination result of the target plot is obtained, and the final result of the land use division of the target plot is obtained after checking and correcting.

[0010] Further, in step S2, the remote sensing image information, building space form information, and industry POI information of each neighborhood in the case city database are extracted, and each type of information is converted into two-dimensional planar image format data, which specifically includes the following contents: the range mask, boundary mask, and distance mask of the remote sensing image data are calculated to form a two-dimensional planar image data set of the remote sensing image information; the spatial coordinate data, land area data, total building area data, and height data of each building monomer of the building space form data are calculated, the distribution is calculated by kernel density analysis, and the distribution is divided by natural breakpoint classification method to form a two-dimensional planar image data set of the building space form information; the total number of POIs, density, mixed degree, and nearest neighbor index of each neighborhood POI data are calculated, the distribution is calculated by kernel density analysis, and the distribution is divided by natural breakpoint classification method to form a two-dimensional planar image data set of the industry POI information.

[0011] Further, in step S3, the deep learning training is performed by the translation window method, which specifically includes: using the ResUNet-a algorithm, establishing a plot plane coordinate system, recording any land image grid as S(x, y), where x is the horizontal coordinate and y is the vertical coordinate, establishing a 2x2 scale inference window W(x, y) composed of four adjacent grids, Each grid S(x, y) is contained in the inference window W(x, y), W(x-1, y), W(x, y-1), W(x-1, y-1), and becomes the intersection of the four windows, so that the inference model based on the ResUNet-a algorithm is used to perform deep learning inference on the land function determination of S(x, y).

[0012] Further, the grid land function determination rule of the grid S(x, y) in step S3 comprises two steps: the first step is to analyze the semantic features of each grid remote sensing image based on the translation window method, and the preliminary determination result of the grid attribute is obtained through image semantic feature clustering; the second step is to first connect the grid with missing spatial form and POI data to the nearest grid of the same type based on the preliminary determination result of the grid attribute, and assign the spatial form and POI data of the same grid to it; based on the grid function determination rule system of spatial form and POI data shown in Table 1, the actual land use properties corresponding to each grid are compared, the significance of each spatial form feature index of the grid is calculated, and the grid is divided according to the determination rule and the urban land classification and planning construction land standard;

[0013] Table 1 Grid function determination rule system of spatial form and POI data

[0014]

[0015]

[0016] The image semantic feature clustering specifically refers to that the ResUNet-a algorithm model determines the grid according to the multi-element features of image composition element types, shapes, area proportions, relative positions and directionality;

[0017] The grid attribute specifically refers to that the grid is only attributed to R class, B class and G class according to the feature difference, which does not strictly correspond to the three types of grids in the second step.

[0018] The significance of the spatial form feature index specifically refers to the influence degree of each spatial form feature index on the grid determination result.

[0019] Further, in step S3, the model is trained in multiple rounds with different window sizes, and the multiple training results are superimposed, and the highest value of the model training accuracy is selected, that is, the function of S(x, y) is inferred by using multiple specifications of windows such as 2x2, 3x3, 4x4, 5x5 and 6x6, until the inference window area reaches 25% of the overall plot area, and the training result with the highest accuracy is selected.

[0020] Further, the step S4 adopts the watershed method to divide the grid boundary, specifically, the non-continuous grid is determined as a preset foreground, the remaining grid is determined as a preset background, an inflation operation is performed on the preset foreground, the image is processed by using a distance transform function cv2.distanceTransform(), and the result is segmented based on the grid feature similarity, to obtain a determined foreground region and an unknown foreground region, the preset foreground and the determined foreground are taken as foreground, the unknown foreground region and other remaining grid are taken as background, and then the foreground and the background are retained or deleted by combining the inference rules shown in Table 2, to complete the division of the land use boundary.

[0021] Table 2: Grid boundary division rule system

[0022]

[0023] The inflation operation specifically refers to connecting the non-continuous grid to form a preset region with non-continuous grid features, and then judging whether the preset region has a significant difference from the background.

[0024] Further, for each land use unit in the step S4, the volume rate data of the building space form information, the building density number, and the average height data of the building are calculated, the density data and the mixed degree data of the POI information of the business format are calculated, and according to the city land classification and the planning construction land standard, a parameterized rule system is constructed to check and automatically correct the function determination result.

[0025] Further, the method further includes a step S5 of establishing a holographic projection sand table, outputting the land division result of the target area, and interactively displaying various information of the target area.

[0026] A city block land fine division system based on deep learning, comprising:

[0027] A data acquisition and database construction module is configured to acquire remote sensing image data, building space form data, business format POI data, and city land data of a case city area, and construct a case city database.

[0028] A multi-source data two-dimensional conversion and data enhancement processing module is configured to extract remote sensing image information, building space form information, and business format POI information of each block in the case city database, convert various types of information into two-dimensional image format data, rasterize the two-dimensional information image and perform data enhancement processing, and construct a case city grid information database.

[0029] The deep learning model construction and database training module is used for deep learning training of remote sensing image information, building space form information, industry POI information and corresponding urban land function information contained in each grid through the sliding window method and the grid land function determination rule, and recording the training accuracy of the land division convolutional neural network deep learning model; the model is trained for multiple rounds with different window sizes, the training results are superimposed multiple times, and the highest value of the model training accuracy is selected;

[0030] The target area land division and correction module is used for two-dimensional conversion and data enhancement processing of remote sensing image data, building space form data and industry POI data of the target area, and carries out grid function determination of the urban land of the target area based on the trained land division convolutional neural network deep learning model; the grid boundary is divided through the watershed method, the land division and function determination result of the target area is obtained, and the result is checked and corrected to obtain the final result of the land division of the target area.

[0031] A computer program product comprising computer programs / instructions which, when executed by a processor, implement the steps of the deep learning-based urban block land fine division method.

[0032] Advantages: Compared with the prior art, the present application has the following advantages:

[0033] 1、The present application carries out deep learning training of remote sensing image information, building space form information, industry POI information and corresponding urban land function information contained in each grid through the sliding window method and the grid land function determination rule, and selects the highest value of the model training accuracy by training the model for multiple rounds with different window sizes, thereby improving the accuracy of land function recognition and the accuracy of grid division.

[0034] 2、The present application divides the grid boundary by the watershed method, checks and corrects the function determination result after land division, can effectively improve the division speed and accuracy, improves the accuracy of land boundary division, improves the credibility of the division result, greatly reduces the workload of manual checking.

[0035] 3、The present application further constructs a parameterized rule system for each land unit, after obtaining each land function, increases the volume rate data, building density number and average building height data of building space form information of each land unit, calculates the density data and mixed degree data of industry POI information, and further refines the accuracy of the identification result. The method effectively improves the accuracy of the identification result compared with the traditional identification method, limits the threshold value of the land attribute data, reduces the possibility of identification error of some special land, and effectively guarantees the credibility of the result. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a method flowchart of an embodiment of the present application;

[0037] Figure 2 is a street block land grid schematic diagram in an embodiment of the present application;

[0038] Figure 3 is a street block land grid translation window method schematic diagram in an embodiment of the present application;

[0039] Figure 4 is a watershed method land grid boundary division schematic diagram in an embodiment of the present application;

[0040] Figure 5 is a visual display example in an embodiment of the present application. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0042] As shown in Figure 1 , the embodiment of the present application discloses a city street block land fine division method based on deep learning, which comprises the following steps:

[0043] Step S1, data collection and database construction, collecting remote sensing image data, building space form data, industry POI (Point of Interest) data and urban land data of a case city area, and constructing a case city database.

[0044] In some embodiments, the collected various data are also subjected to spatial calibration and error correction, specifically including the following contents: importing all remote sensing image data, building space form data, industry POI data and urban land data into the same GIS database, and unifying the coordinates of all data to projection coordinates WGS1984UTM ZONE 51N. Among them, the industry POI data is cleaned, the irrelevant and missing value data are removed, and the spatial registration is performed on the inaccurate coordinate data.

[0045] Step S2, two-dimensional conversion and data enhancement processing of multiple data, extracting remote sensing image information, building space form information and industry POI information of each street block in the case city database, and converting various information into two-dimensional plane image format data, further rasterizing the two-dimensional information image, and performing data enhancement processing on each street block data, to construct a case city grid information database.

[0046] As shown in Figure 2 In this embodiment, the grid size is 5m*5m. The remote sensing image information, building space form information and industry POI information of each neighborhood in the calculation case city database are calculated, and each type of information is converted into image format data in a two-dimensional plane. The specific contents include the following: the range mask, boundary mask and distance mask of the remote sensing image information are calculated to form a two-dimensional plane image dataset of the remote sensing image information; the spatial coordinate data, land area data, total building area data and height data of each building monomer of the building space form data are calculated, the distribution thereof is calculated by kernel density analysis in the ArcGIS platform, the distribution thereof is divided by the natural breakpoint classification method to form a two-dimensional plane image dataset of the building space form information; the total number of POIs, density, mixing degree and nearest neighbor index of each neighborhood POI data are calculated, the distribution thereof is calculated by kernel density analysis in the ArcGIS platform, the distribution thereof is divided by the natural breakpoint classification method to form a two-dimensional plane image dataset of the industry POI information.

[0047] The data enhancement processing of each neighborhood data specifically includes: geometric transformation and color space transformation are performed on the images in the case city grid information database to increase data diversity, improve model generalization ability, optimize scale adaptability and avoid overfitting. Specifically, the geometric transformation includes horizontal and vertical flipping and perspective transformation, and the color space transformation includes adjusting the brightness of the picture, increasing or decreasing the saturation of the picture, adding Gaussian noise and color offset.

[0048] Step S3, land division convolutional neural network deep learning model construction and database training, based on the two-dimensional information database of the case city established in step S2, an initial grid of the case city neighborhood information is established in the Arcgis platform, the remote sensing image information, building space form information and industry POI information contained in each grid are subjected to deep learning training with the corresponding urban land function information by the translation window method and the grid land function determination rule, and the training accuracy of the land division convolutional neural network deep learning model is recorded. The model is trained in different window sizes for multiple rounds, the results of multiple training are superimposed, and the highest value of the model training accuracy is selected.

[0049] In this embodiment, the deep learning training by the translation window method specifically includes: a ResUNet-a algorithm is adopted, a district plane coordinate system is established, an image grid of any land block is recorded as S(x, y), wherein x is the horizontal coordinate and y is the vertical coordinate, a 2*2 scale inference window W(x, y) composed of four adjacent grids is established, Each grid S(x, y) is contained by inference windows W(x, y), W(x-1, y), W(x, y-1), W(x-1, y-1), becomes the intersection of the four windows, and thus the inference model based on the ResUNet-a algorithm [(W(x, y)), …, (W(x-1, y-1))] is established to perform deep learning inference on the land use function determination of S(x, y).

[0050] The grid land function determination rule for S(x, y) contains two steps: the first step is to analyze the semantic features of each grid remote sensing image based on the translation window method, and the preliminary determination result of the grid attribute is obtained through image semantic feature clustering. The second step is to first connect the grid with missing spatial form and POI data to the nearest grid of the same type based on the preliminary determination result of the grid attribute, and assign it the spatial form and POI data of the same grid. Further based on the grid function determination rule system based on spatial form and POI data shown in Table 1, the significance of each spatial form feature index of the grid is calculated by comparing the actual land use properties corresponding to each grid, and the grid function is divided in combination with the determination rule and the urban land classification and planning construction land standard (GB50137-2011), wherein the R-class grid represents the residential function, the B-class grid represents the commercial function, and the G-class grid represents the green land function.

[0051] Table 1 Grid function determination rule system based on spatial form and POI data

[0052]

[0053] The image semantic feature clustering specifically refers to the ResUNet-a algorithm model determining the grid according to the multiple features of image composition element types, shapes, area proportions, relative positions, and directionality.

[0054] The grid attribute specifically refers to classifying the grid into R-class, B-class, and G-class only according to feature difference, which does not strictly correspond to the three types of grids in the second step.

[0055] The significance of the spatial form feature index specifically refers to the influence degree of each spatial form feature index on the grid determination result.

[0056] By comparing the actual land use properties corresponding to each grid, the significance of each spatial form feature index of the grid is calculated, and the grid is further determined according to the rules in the table.

[0057] The model is trained in multiple rounds with different window sizes, the results of multiple training are superimposed, and the highest correct rate of model training is selected. The function of S(x, y) is inferred by using multiple window sizes such as 2x2, 3x3, 4x4, 5x5, 6x6, and the like, until the inference window area reaches 25% of the overall plot area, and the training result with the highest correct rate is selected. The neighborhood land grid translation window method with different window sizes is as shown in Figure 3

[0058] Step S4, target plot land refinement and parameterization rule correction, the remote sensing image data, building space form data, and industry POI data of the target plot are processed by the method of step S2 for two-dimensional conversion and data enhancement, and the land division convolutional neural network deep learning model constructed in step S3 is used to determine the grid function of the target plot city land. Further, the watershed method is used to divide the grid boundary to obtain the land division and function determination result of the target plot and to check and correct it to obtain the final result of the target plot land division.

[0059] In this embodiment, the watershed method is used to divide the grid boundary, which specifically refers to determining the non-continuous grid as a preset foreground, determining the remaining grid as a preset background, performing an inflation operation on the preset foreground, processing the image using the distance transform function cv2.distanceTransform(), and segmenting the result based on the grid feature similarity to obtain the determined foreground area and unknown foreground area. The preset foreground and the determined foreground are regarded as foreground, and the unknown foreground area and other remaining grids are regarded as background, and then the foreground and the background are retained or deleted according to the inference rules shown in Table 2 to complete the division of the land boundary. The watershed method land grid boundary division schematic diagram is as shown in Figure 4 (a) is an example of non-continuous grid before correction, and (b) is the grid boundary after correction.

[0060] Table 2 Grid boundary division rule system

[0061]

[0062] The inflation operation specifically refers to connecting the non-continuous grid to form a preset area with non-continuous grid features, and then determining whether the preset area has a significant difference from the background.

[0063] ​In some embodiments, a parameterized rule system for each land unit is also constructed, specifically, for each land unit, the volume rate data of the building space form information, the building density number, the average height data of the building, the density data of the industry POI information, and the mixed degree data are calculated, and according to the urban land classification and planning construction land standard (GB50137-2011), the parameterized rule system is constructed to check and automatically correct the function determination result. Exemplarily, the constructed parameterized rule system is shown in Table 3:

[0064] Table 3 Parameterized rule system for each land unit

[0065]

[0066] In some embodiments, the method further comprises: step S5, result output and interactive display, establishing a holographic projection sand table, outputting the land division result of the target area, and interactively displaying various information of the target area, and the required equipment includes a three-dimensional physical sand table, a digital holographic projector and a gesture recognizer.

[0067] The holographic projection sand table is established, the land division result of the target area is output, and various information of the target area is interactively displayed, specifically including using a projector with a resolution of 4K or above, a depth sensor, a motion sensor, a touch sensor, a sand table to construct a three-dimensional digital sand table, and cooperating with a gesture recognizer to display a virtual scene of the land information of the target area. Figure 5 The visualization display example is shown.

[0068] The urban block land fine division method based on deep learning has great significance in quickly and accurately identifying the function of urban block land and automatically determining, solves the problems of traditional land recognition difficulty, unclear boundary division, large artificial recognition workload, difficult determination of machine recognition boundary and the like, and realizes fine division of urban block land and automatic checking of the result.

[0069] The technical solutions of the present application will be described in detail below taking the urban block land fine division of a certain area in Nanjing as an example.

[0070] (1) Taking a certain area in Nanjing as a target area, collecting remote sensing image data, building space form data, industry POI data and urban land data of the area, and constructing a database of the certain area in Nanjing, specifically including:

[0071] (1.1) Import remote sensing image data, building space form data, industry POI data and urban land data into the same GIS database, and unify the coordinates of all data as projection coordinates WGS1984UTM ZONE 51N. Among them, the industry POI data is the SHP file of each POI point; the urban land data includes the present situation closed land CAD / SHP file in the area, the present situation building CAD / SHP file in the land.

[0072] (1.2) Clean the industry POI data, eliminate irrelevant and missing value data, and perform spatial registration on the inaccurate coordinate data.

[0073] (2) Convert the remote sensing image information, building space form information and industry POI information of each neighborhood in the database of a certain area in Nanjing into two-dimensional plane image format data, which specifically includes the following contents: calculate the range mask, boundary mask and distance mask of remote sensing image information to form a two-dimensional plane image data set of remote sensing image information; calculate the spatial coordinate data, land area data, total building area data and height data of each building monomer of building space form information, use kernel density analysis to calculate its distribution in the ArcGIS platform, and use the natural breakpoint classification method to divide its distribution to form a two-dimensional plane image data set of building space form information; calculate the total number of POI, density, mixed degree and nearest neighbor index of each neighborhood POI information, use kernel density analysis to calculate its distribution in the ArcGIS platform, and use the natural breakpoint classification method to divide its distribution to form a two-dimensional plane image data set of industry POI information.

[0074] The obtained two-dimensional plane image data set is first geometrically transformed, including horizontal and vertical flipping, perspective transformation; secondly, color space transformation is performed, including adjusting the brightness of the picture, increasing or decreasing the saturation of the picture, adding Gaussian noise and color offset, in order to increase the diversity of data, improve the generalization ability of the model, optimize the scale adaptability, and avoid overfitting.

[0075] (3) land division convolutional neural network deep learning model construction and database training, based on a certain area database in Nanjing city, establish an initial size of 5m*5m grid on the street information in the area in Arcgis platform, through the translation window method and the grid land function determination rule, the remote sensing image information, building space form information, industry POI information and corresponding urban land function information contained in each grid are trained by deep learning, and the training accuracy of land division convolutional neural network deep learning model is recorded. A plurality of specifications such as 2*2, 3*3, 4*4, 5*5 and 6*6 are used for multiple rounds of training of the function of S(x, y) with different window sizes, until the inference window area reaches 25% of the overall plot area, stop, superimpose the training results, and select the highest value of the model training accuracy.

[0076] (4) target area land fine division and parameterization rule correction, the watershed method is used to divide the grid boundary, specifically, the non-continuous grid is determined as the preset foreground, the remaining grid is determined as the preset background, the dilation operation is performed on the preset foreground, the image is processed by using the distance transform function cv2.distanceTransform(), and the result is segmented based on the grid feature similarity, to obtain the determined foreground area and unknown foreground area. The preset foreground and the determined foreground are regarded as the foreground, and the unknown foreground area and the other remaining grid are regarded as the background, and then the foreground and the background are retained or deleted by combining the inference rules shown in table 2, to complete the division of the land boundary.

[0077] The parameterization rule system for each land unit is constructed, specifically, for each land unit, the volume rate data, building density number and average building height data of the building space form information are calculated, the density data and mixed degree data of the industry POI information are calculated, and according to the urban land classification and planning construction land standard (GB50137-2011), the rule system is constructed as shown in table 3.

[0078] (5) result output and interactive display, a holographic projection sand table is established, the land division results of the area are output, the interactive display of various information of the target area is realized, and the required equipment includes a three-dimensional physical sand table, a digital holographic projector and a gesture recognizer.

[0079] Among them, a holographic projection sand table is established, the land division results of the target area are output, and the interactive display of various information of the target area is realized, specifically, a 4K resolution or above projector, a depth sensor, a motion sensor, a touch sensor and a three-dimensional digital sand table are used to construct a three-dimensional digital sand table, and a gesture recognizer is used to display the virtual scene of the land information of the target area.

[0080] Based on the same inventive concept, the embodiment of the present application discloses a kind of urban street land fine division system based on deep learning, comprising:

[0081] a data acquisition and database construction module for acquiring remote sensing image data, building space form data, industry POI data and urban land data of a case city district, and constructing a case city database;

[0082] a multi-source data two-dimensional conversion and data enhancement processing module for extracting remote sensing image information, building space form information and industry POI information of each neighborhood in the case city database, converting various types of information into two-dimensional planar image format data, rasterizing the two-dimensional information image and performing data enhancement processing, and constructing a case city raster information database;

[0083] a deep learning model construction and database training module for performing deep learning training on remote sensing image information, building space form information, industry POI information and corresponding urban land function information contained in each grid by using a translation window method and a grid land function determination rule, recording the training accuracy of a land division convolutional neural network deep learning model, performing multiple rounds of training on the model with different window sizes, superimposing multiple training results, and selecting the highest value of the model training accuracy;

[0084] a target district land division and correction module for performing two-dimensional conversion and data enhancement processing on remote sensing image data, building space form data and industry POI data of a target district, performing rasterized function determination on the urban land of the target district based on the trained land division convolutional neural network deep learning model, dividing the grid boundary by using a watershed method to obtain land division and function determination results of the target district, and performing checking and correction to obtain the final result of the land division of the target district.

[0085] The embodiment of the present application also discloses a computer program product comprising computer programs / instructions, which, when executed by a processor, realize the steps of the deep learning-based fine division method of urban neighborhood land.

Claims

1. A method for fine-grained urban block land use division based on deep learning, characterized in that, Includes the following steps: Step S1: Collect remote sensing image data, building spatial form data, business type POI data, and urban land use data of the case city area to construct the case city database; Step S2: Extract remote sensing image information, building spatial form information, and business POI information of each neighborhood in the case city database, and convert all types of information into two-dimensional planar image format data. Rasterize the two-dimensional information images and perform data enhancement processing to construct a case city raster information database. Step S3: Using the translation window method and grid land use function determination rules, the remote sensing image information, building spatial form information, business type POI information and corresponding urban land use function information contained in each grid are subjected to deep learning training, and the training accuracy of the land use division convolutional neural network deep learning model is recorded; the model is trained in multiple rounds with different window sizes, the training results are superimposed, and the model with the highest training accuracy is selected. Step S4: Perform two-dimensional transformation and data augmentation on the remote sensing image data, building spatial form data, and business POI data of the target area. Based on the land use division convolutional neural network deep learning model constructed in Step S3, perform rasterization function determination on the urban land use of the target area. Divide the raster boundary using the watershed method to obtain the land use division and function determination results of the target area. Then, verify and correct the results to obtain the final land use division result of the target area. Step S3 involves deep learning training using the translation window method, specifically including: employing the ResUNet-a algorithm to establish a regional planar coordinate system; denoting the image grid of any plot as S(x,y), where x is the horizontal coordinate and y is the vertical coordinate; and establishing a 2×2 inference window W(x,y) composed of four adjacent grids. Each grid S(x,y) is contained by inference windows W(x,y), W(x-1,y), W(x,y-1), and W(x-1,y-1), forming the intersection of the four windows. Thus, an inference model based on the ResUNet-a algorithm is established to perform deep learning inference on the land use function of S(x,y). Step S3, which determines the land use function of grid S(x,y), involves two steps: First, the semantic features of each grid remote sensing image are analyzed using the translation window method, and the grid affiliation is preliminarily determined through image semantic feature clustering. Second, based on the preliminary determination results of grid affiliation, grids lacking spatial morphology and POI data are connected to the nearest grid of the same type, and assigned the spatial morphology and POI data of that grid. Based on the grid function determination rule system of spatial morphology and POI data shown in Table 1, the significance of each spatial morphology feature index of the grid is calculated by comparing it with the actual land use nature corresponding to each grid, and the grids are divided in combination with the determination rules and urban land use classification and planning construction land standards. Table 1. Raster Function Determination Rules for Spatial Morphology and POI Data The image semantic feature clustering specifically refers to the ResUNet-a algorithm model determining the grid based on the multi-dimensional features of image constituent elements, such as type, shape, area ratio, relative position, and direction. The raster classification specifically refers to classifying rasters into three categories—R, B, and G—based solely on feature differences, which does not strictly correspond to the three raster categories in the second step. The significance of the spatial morphological feature indicators specifically refers to the degree of influence of each spatial morphological feature indicator on the raster determination result.

2. The method for fine-grained urban block land use based on deep learning according to claim 1, characterized in that, Step S2 extracts remote sensing image information, building spatial form information, and business POI information of each neighborhood in the case city database, and converts all types of information into two-dimensional planar image format data, specifically including the following: Calculate the range mask, boundary mask, and distance mask of the remote sensing image data to form a two-dimensional planar image dataset of the remote sensing image information; The spatial coordinates, floor area, total building area, and height of each building in the building spatial form data are calculated. Kernel density analysis is used to calculate their distribution, and the distribution is further divided using the natural discontinuity grading method, forming a two-dimensional planar image dataset of building spatial form information. The total number, density, mixing degree, and nearest neighbor index of each neighborhood POI data are calculated. Kernel density analysis is used to calculate their distribution, and the distribution is further divided using the natural discontinuity grading method, forming a two-dimensional planar image dataset of neighborhood POI information.

3. The method for fine-grained urban block land use division based on deep learning according to claim 1, characterized in that, In step S3, the model is trained in multiple rounds with different window sizes. The training results are superimposed, and the model with the highest training accuracy is selected. This means using multiple windows of various sizes, such as 2×2, 3×3, 4×4, 5×5, 6×6, etc., to infer the function of S(x,y) until the inference window area reaches 25% of the total land area. The training result with the highest accuracy is then selected.

4. The method for fine-grained urban block land use based on deep learning according to claim 1, characterized in that, In step S4, the watershed method is used to divide the grid boundary. Specifically, discontinuous grids are identified as the preset foreground, the remaining grids are identified as the preset background, the preset foreground is dilated, the image is processed using the distance transformation function cv2.distanceTransform(), and the results are segmented based on the grid feature similarity to obtain the determined foreground region and the unknown foreground region. The preset foreground and the determined foreground are used as the foreground, and the unknown foreground region and other remaining grids are used as the background. Then, the foreground and background are retained or deleted according to the inference rules shown in Table 2 to complete the division of the land boundary. Table 2. Grid Boundary Delineation Rule System The dilation operation specifically refers to connecting discontinuous grids to form a preset region with discontinuous grid characteristics, and then determining whether the preset region has a significant difference from the background.

5. The method for fine-grained urban block land use division based on deep learning according to claim 1, characterized in that, In step S4, for each land use unit, the plot ratio, building density, and average building height of its building spatial form information are calculated, and the density and mixed-use data of the business type POI information are calculated. Based on the urban land use classification and planning construction land standards, a parameterized rule system is constructed to verify and automatically correct the functional judgment results.

6. The method for fine-grained urban block land use division based on deep learning according to claim 1, characterized in that, It also includes step S5: establishing a holographic projection sand table, outputting the land use division results of the target area, and interactively displaying various information of the target area.

7. A deep learning-based system for fine-grained urban block land use division, characterized in that, include: The data acquisition and database construction module is used to collect remote sensing image data, building spatial form data, business POI data and urban land use data of the case city area, and to construct the case city database. The multi-source data 2D conversion and data enhancement processing module is used to extract remote sensing image information, building spatial form information and business POI information of each block in the case city database, and convert various types of information into 2D planar image format data. The 2D information images are rasterized and data enhancement processing is performed to construct a case city raster information database. The deep learning model construction and database training module is used to train the remote sensing image information, building spatial form information, business type POI information and corresponding urban land use function information contained in each grid through the translation window method and grid land use function determination rules. It records the training accuracy of the land use division convolutional neural network deep learning model; it performs multiple rounds of training on the model with different window sizes, superimposes the training results, and selects the model with the highest training accuracy. The deep learning training employed a translation window method, specifically including: using the ResUNet-a algorithm to establish a regional planar coordinate system; denoting the image raster of any plot as S(x,y), where x is the horizontal coordinate and y is the vertical coordinate; and establishing a 2×2 inference window W(x,y) consisting of four adjacent raster cells. Each grid S(x,y) is contained by inference windows W(x,y), W(x-1,y), W(x,y-1), and W(x-1,y-1), forming the intersection of the four windows. Thus, an inference model based on the ResUNet-a algorithm is established to perform deep learning inference on the land use function of S(x,y). The raster land use function determination rule for raster S(x,y) includes two steps: First, based on the translation window method, the semantic features of each raster remote sensing image are analyzed, and the raster attribution is preliminarily determined by image semantic feature clustering; Second, based on the preliminary raster attribution determination result, rasters lacking spatial morphology and POI data are connected to the nearest similar rasters, and the spatial morphology and POI data of that similar rasters are assigned to them; Based on the raster function determination rule system of spatial morphology and POI data shown in Table 1, the significance of each spatial morphology feature index of each raster is calculated by comparing with the actual land use nature corresponding to each raster, and the rasters are divided in combination with the determination rules and urban land use classification and planning construction land standards; Table 1. Raster Function Determination Rules for Spatial Morphology and POI Data The image semantic feature clustering specifically refers to the ResUNet-a algorithm model determining the grid based on the multi-dimensional features of image constituent elements, such as type, shape, area ratio, relative position, and direction. The raster classification specifically refers to classifying rasters into three categories—R, B, and G—based solely on feature differences, which does not strictly correspond to the three raster categories in the second step. The significance of the spatial morphological feature indicators specifically refers to the degree of influence of each spatial morphological feature indicator on the raster determination result. The target area land use delineation and correction module is used to perform two-dimensional transformation and data augmentation processing on remote sensing image data, building spatial form data, and business type POI data of the target area. Based on the trained land use delineation convolutional neural network deep learning model, it performs rasterization function determination on the urban land use of the target area; it divides the raster boundary through the watershed method to obtain the land use delineation and function determination results of the target area, and performs verification and correction to obtain the final land use delineation result of the target area.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the deep learning-based urban block land fine division method according to any one of claims 1-6.

Images