Modeling method and system based on semantic segmentation model and storage medium

Abstract

The invention discloses a modeling method and system based on a semantic segmentation model and a storage medium, and belongs to the technical field of three-dimensional modeling. The method comprises the steps of S1, a remote sensing data preprocessing stage; s2, a semantic segmentation model training and tuning stage; s3, a forest land area semantic segmentation recognition stage: performing pixel-level classification on the multispectral radar data of the target area by using the trained semantic segmentation model to obtain a semantic segmentation result, and performing vectorization conversion to obtain a Thiessen polygon area of an effective intensive forest land area; s4, a modeling scene construction stage; s5, a scene asset generation and output stage: generating an intensive forest land environment model based on the digital elevation model and the vegetation space coordinate point set; and performing geometric random transformation on the vegetation monomer model based on a CGA rule to obtain digital twinborn scene assets and outputting the digital twinborn scene assets. The tree space distribution characteristics based on satellite observation are accurately restored, so that the precision of environment modeling is improved within a low cost range.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional modeling, and in particular to a modeling method, system and storage medium based on a semantic segmentation model. Background Art

[0002] As the top-level enabling technology system of digital new infrastructure, digital twin technology is widely used in infrastructure construction and digital operation and maintenance, digital industrial production, smart cities, smart transportation and other scenarios. As a part of the scene construction process, vegetation modeling generally uses drone aerial photography to collect oblique photography or laser point cloud and other surveying and mapping technologies to reconstruct vegetation real-scene models. Its advantages are mature technology and can guarantee the authenticity and accuracy of model restoration to a certain extent. However, its disadvantages such as high computing and storage costs, poor data compatibility, poor close-range rendering effects, and poor linkage are all reasons that restrict the further development and application of such environmental modeling technologies in digital twin scenarios. The vegetation model constructed by 3D modeling software such as 3DSMAX and MAYA is more intuitive and realistic, but most 3D modeling software lacks data communication interfaces for geographic information system (GIS) software, including domestic super map platforms and foreign ArcGIS and other platforms. Ultimately, it is difficult to quickly deploy the created real-scene model to a local or cloud-based 3D geographic information system scene (3DGIS) for basic calculations or rendering. At the same time, its modeling under large-scale scene conditions is time-consuming and labor-intensive, and the cost becomes uncontrollable, so it is only suitable for small-scale digital twin scene vegetation modeling. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a modeling method, system and storage medium based on a semantic segmentation model.

[0004] The objective of the present invention is achieved through the following technical solutions: The first aspect of the present invention provides: a modeling method based on a semantic segmentation model, comprising the following steps: S1: remote sensing data preprocessing stage, acquiring remote sensing data, then preprocessing the remote sensing data, and finally exporting the preprocessed remote sensing data as training data; the remote sensing data includes multispectral radar data, orthophoto panchromatic image data, and digital elevation model data; the data preprocessing includes data mosaicking, radiation calibration, atmospheric correction, orthophoto correction, and enhanced vegetation index (EVI) quantitative calculation; S2: Semantic segmentation model training and tuning stage, building a semantic segmentation model and performing model training and model tuning to obtain a trained semantic segmentation model; S3: In the forest area semantic segmentation and recognition stage, the trained semantic segmentation model is used to perform pixel-level classification on the multispectral radar data of the target area to obtain the semantic segmentation results, and the semantic segmentation results are vectorized to obtain the Thiessen polygon area of ​​the effective dense forest area; S4: In the modeling scene construction phase, the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area are imported into the modeling software to construct the digital elevation model; the statistical numerical method in the geographic information system software is used to generate vegetation distribution points under the limitation of the Thiessen polygon area of ​​the effective dense forest area and import them into the modeling software; finally, the vegetation distribution points are interpolated by the elevation operation to give the vegetation distribution points three-dimensional height parameters to obtain the vegetation spatial coordinate point set; S5: In the scene asset generation and output stage, a dense forest environment model is generated based on the digital elevation model and the vegetation spatial coordinate point set. The vegetation monomer model in the dense forest environment model uses the model resources built into the modeling software or a self-built model. Based on the CGA rules, the vegetation monomer model is geometrically randomly transformed to obtain the digital twin scene asset and output it.

[0005] Preferably, the remote sensing data preprocessing stage S1 further includes the following steps: S11: remote sensing data is obtained from open source data sources, multispectral radar data is obtained from Sentinel-2 data sources, digital elevation model data is obtained from ALOS platform 12.5m resolution data sources, and orthochromatic image data is obtained from satellite image data sources; S12: Import the remote sensing data into the remote sensing image processing software and use the seamless mosaic tool to perform mosaic operation on the scattered remote sensing data to obtain the first remote sensing data; perform radiation calibration on the first remote sensing data to obtain the second remote sensing data; perform atmospheric correction on the second remote sensing data to obtain the third remote sensing data; perform orthorectification on the third remote sensing data, and determine the atmospheric model, aerosol model, and ground elevation parameters of the third remote sensing data according to the acquisition time, satellite attitude, orbital position, and data set range of the third remote sensing data to obtain the fourth remote sensing data; calculate the enhanced vegetation index EVI of the fourth remote sensing data to obtain the remote sensing data after data preprocessing; the enhanced vegetation index EVI is calculated by the following formula: Wherein, NIR is the near infrared band reflectivity, which is the 8th band data of Sentinel-2 radar; R is the red band reflectivity, which is the 4th band data of Sentinel-2 radar; B is the blue band reflectivity, which is the 2nd band data of Sentinel-2 radar; G is the gain factor, which is 2.5; C1 and C2 are the first and second atmospheric correction coefficients, which are 6 and 7.5 respectively; L is the soil adjustment parameter, which is 1; S13: Load the pre-processed multispectral radar data and orthophoto panchromatic image data into the geographic information system software, and perform mask mapping on the target data set to be processed; perform mask mapping on the contour of the pre-processed orthophoto panchromatic image data; perform mask mapping on the pre-processed multispectral radar data with reference to the pre-processed orthophoto panchromatic image data and use continuous and closed polygons; finally, export the masked data as training data.

[0006] Preferably, the S2: semantic segmentation model training and tuning stage further includes the following steps: S21: Use PyTorch framework, U-Net architecture, and DenseNet backbone network to build a semantic segmentation model; S22: Perform model training on the semantic segmentation model and adjust the hyperparameters of the semantic segmentation model to achieve a first preset quantitative index; test the semantic segmentation model, and tune the model according to the test results until the semantic segmentation model reaches a second preset quantitative index to obtain a trained semantic segmentation model.

[0007] Preferably, the S3: forest area semantic segmentation and recognition stage further comprises the following steps: S31: Use the trained semantic segmentation model to segment the forest area of ​​the multispectral radar data after data preprocessing into multiple sub-areas and perform data prediction on each sub-area to obtain local segmentation maps. Then, perform data mosaicking on the local segmentation maps to merge them to obtain a complete segmentation map. The input multispectral radar data is segmented using the parameters in the following formula: In the formula, C x is the cutting step length for the input multi-spectrum radar data in the horizontal direction; I x is the number of samples cut out in the horizontal direction from the input multi-spectrum radar data; T x is the number of pixels in the horizontal direction of the input multi-spectrum radar data, in pixels; S x is the number of pixels in the horizontal direction of a single sample obtained by cutting, in pixels; G x is the number of pixels in the overlapping area of ​​two adjacent samples in the horizontal direction, in pixels; C y The cutting step size for the input multispectral radar data in the vertical direction; I y is the number of samples cut out in the vertical direction from the input multi-spectrum radar data; T yis the number of pixels in the vertical direction of the input multispectral radar data, in pixels; S y is the number of pixels in the vertical direction of a single sample obtained by cutting, in pixels; G y is the number of pixels in the overlapping area of ​​two adjacent samples in the vertical direction, in pixels; “⌈...⌉” means rounding up; C x , I x , T x , S x , G x , C y , I y , T y , S y , G y are all positive integers; S32: Load the complete segmentation map, set the judgment condition to clear the background to obtain the first segmentation map; use the raster to vector surface conversion function to convert the effective dense woodland area in the raster of the first segmentation map into a vector surface to obtain the first raster data; use the first raster data as a mask and clear the non-forest area of ​​the multispectral radar data after the data preprocessing initially input in S31 to obtain the second raster data; convert the second raster data into a raster point set, and the data field inherited by the raster point set is the value of each pixel of the input raster; input the raster point set into the create Thiessen polygon function to obtain multiple Thiessen polygon areas centered on the points in the raster point set and with adjacent sides; use the geometric intersection calculation function to calculate the intersection of the first raster data and the Thiessen polygon area to obtain the Thiessen polygon area of ​​the effective dense woodland area.

[0008] Preferably, the S4: modeling scene construction phase further includes the following steps: S41: Import the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area into the modeling software, use the terrain import mode to input the terrain data of the same range as the multispectral radar data of the target area, and perform seamless mosaicking and clipping operations on the terrain data to obtain the digital elevation model; S42: Taking any rectangular unit in the Thiessen polygon area of ​​the effective dense forest area as the benchmark, calculate the number of vegetation per unit area TD: Where A is the area of ​​the current measurement area; is the average canopy area of ​​vegetation in the current measurement area; EVI is the EVI value of the current rectangular unit in the Thiessen polygon area of ​​the effective dense forest area; “⌈...⌉” is a rounding operation; After obtaining the number of vegetation per unit area TD, calculate the vegetation distribution points: in, is the easting distance of the current t-th random coordinate point, and satisfies t≤TD; is the northing distance of the current t-th random coordinate point, and satisfies t≤TD; D east is the westernmost east distance of the current rectangular unit; D west is the easternmost east distance of the current rectangular unit; D north is the northernmost north distance of the current rectangular unit; D south is the southernmost north distance of the current rectangular unit; N r is a random number that follows a Gaussian distribution; n The number of samples generated to satisfy the uniform distribution condition; X i is the i-th sample value that satisfies uniform distribution; S43: Importing vegetation distribution points into modeling software, using the digital elevation model as the elevation data source, matching the vegetation distribution points with the elevation data source to obtain a vegetation spatial coordinate point set.

[0009] Preferably, the S5: scene asset generation and output stage further includes the following steps: S51: Generate a vegetation monomer model and initialize the attributes of the vegetation monomer model to obtain a rule file, select all coordinate points in the vegetation space coordinate point set and bind them to the rule file, and then start the model generation instruction to obtain a dense forest environment model; S52: Perform geometric random transformation on the vegetation monomer model based on CGA rules, and then export the vegetation monomer model and dense forest environment model as digital twin scene assets and output them.

[0010] Preferably, the semantic segmentation model is a U-Net architecture.

[0011] Preferably, the modeling software is CityEngine, and the geographic information system software is ArcGIS Pro.

[0012] The second aspect of the present invention provides: a modeling system based on a semantic segmentation model, used to implement any of the above-mentioned modeling methods based on a semantic segmentation model, comprising: A remote sensing data preprocessing module is used to obtain remote sensing data, then preprocess the remote sensing data, and finally export the preprocessed remote sensing data as training data; the remote sensing data includes multispectral radar data, orthophoto panchromatic image data, and digital elevation model data; the data preprocessing includes data mosaicking, radiation calibration, atmospheric correction, orthophoto correction, and enhanced vegetation index (EVI) quantitative calculation; The semantic segmentation model training and tuning module is used to build a semantic segmentation model and perform model training and model tuning to obtain a trained semantic segmentation model; The forest area semantic segmentation recognition module is used to use the trained semantic segmentation model to perform pixel-level classification on the multispectral radar data of the target area to obtain semantic segmentation results, and to vectorize the semantic segmentation results to obtain the Thiessen polygon area of ​​the effective dense forest area; The modeling scene construction module is used to import the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area into the modeling software to construct the digital elevation model; use the statistical numerical method in the geographic information system software to generate vegetation distribution points under the limitation of the Thiessen polygon area of ​​the effective dense forest area and import them into the modeling software; finally, perform elevation interpolation operation on the vegetation distribution points to give the vegetation distribution points three-dimensional height parameters to obtain the vegetation space coordinate point set; The scene asset generation and output module is used to generate a dense forest environment model based on the digital elevation model and the vegetation spatial coordinate point set. The vegetation monomer model in the dense forest environment model uses the model resources built into the modeling software or a self-built model; based on the CGA rules, the vegetation monomer model is geometrically randomly transformed to obtain the digital twin scene asset and output it.

[0013] The third aspect of the present invention provides: a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by a processor, any of the above-mentioned modeling methods based on the semantic segmentation model is implemented.

[0014] The beneficial effects of the present invention are: 1) The training data annotation process based on orthophoto panchromatic image data adopts the idea of ​​visualizing mask plotting based on orthophoto panchromatic image data and applying it to multispectral radar data for output, which circumvents the problem that multispectral radar data cannot be recognized by the human eye and indirectly improves the quality of semantic segmentation training data.

[0015] 2) By using the EVI index, with the support of Thiessen polygons and raster conversion workflows, we can obtain fine-grained vegetation distribution feature vector data, which plays a key role in the subsequent workflow of describing the spatial distribution characteristics of vegetation distribution and performing random vector distribution of dense woodland vegetation. The spatial distribution characteristics of trees based on satellite observations are restored more accurately, so that the accuracy of environmental modeling under the direction of UAV aerial survey technology can be achieved as much as possible within a low cost range.

[0016] 3) In the process of creating the model using CGA rules, the present invention fully considers the randomness of dense forest vegetation growth, adopts the idea of ​​random vegetation point distribution plus random model size scaling, so that the dense forest vegetation characteristics produced by the present invention can simulate the random distribution of vegetation in space and morphology to a certain extent. And the use of code generation modeling rules greatly reduces the modeling workload and saves the time and energy of technical implementation personnel.

[0017] 4) Except for the data labeling work required for the semantic segmentation task, which is a one-time human investment in the early stage, the rest of the processes can be converted into code programs with low or no human intervention for automatic execution, so that the environmental modeling of dense woodland areas can truly automatically produce dense woodland environment scenes that conform to the laws of remote sensing observations, and output scene assets that meet the relevant digital twin standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flowchart of the modeling method based on the semantic segmentation model. DETAILED DESCRIPTION

[0019] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0020] See also Figure 1 The first aspect of the present invention provides: a modeling method based on a semantic segmentation model, comprising the following steps: S1: remote sensing data preprocessing stage, acquiring remote sensing data, then preprocessing the remote sensing data, and finally exporting the preprocessed remote sensing data as training data; the remote sensing data includes multispectral radar data, orthophoto panchromatic image data, and digital elevation model data; the data preprocessing includes data mosaicking, radiation calibration, atmospheric correction, orthophoto correction, and enhanced vegetation index (EVI) quantitative calculation; S2: Semantic segmentation model training and tuning stage, building a semantic segmentation model and performing model training and model tuning to obtain a trained semantic segmentation model; S3: In the forest area semantic segmentation and recognition stage, the trained semantic segmentation model is used to perform pixel-level classification on the multispectral radar data of the target area to obtain the semantic segmentation results, and the semantic segmentation results are vectorized to obtain the Thiessen polygon area of ​​the effective dense forest area; S4: In the modeling scene construction phase, the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area are imported into the modeling software to construct the digital elevation model; the statistical numerical method in the geographic information system software is used to generate vegetation distribution points under the limitation of the Thiessen polygon area of ​​the effective dense forest area and import them into the modeling software; finally, the vegetation distribution points are interpolated by the elevation operation to give the vegetation distribution points three-dimensional height parameters to obtain the vegetation spatial coordinate point set; S5: In the scene asset generation and output stage, a dense forest environment model is generated based on the digital elevation model and the vegetation spatial coordinate point set. The vegetation monomer model in the dense forest environment model uses the model resources built into the modeling software or a self-built model. Based on the CGA rules, the vegetation monomer model is geometrically randomly transformed to obtain the digital twin scene asset and output it.

[0021] In this embodiment, S1 first obtains open source high-resolution remote sensing data, including multispectral radar, orthophoto panchromatic image, and digital elevation model (DEM); then performs data preprocessing, including data mosaic, radiation calibration, atmospheric correction, orthophoto correction, and quantitative calculation of vegetation index EVI; finally, training data is exported, including target detection mask calibration, semantic segmentation mask calibration, and training data export. In S2, the segmentation model adopts the U-Net architecture in the deep neural network to obtain the pixel classification results of the dense woodland area when the vegetation index (NDVI) is greater than 0. This model is suitable for the recognition task of red light and infrared band data in multispectral radar, and is mainly used to convert remote sensing images into vegetation density-related parameters. This step mainly includes model code writing, model training and testing. The Thiessen polygon area of ​​the effective dense woodland area obtained by vectorization conversion in S3 can be used as the basis for the modeling point range in the case of dense dense woodland. S4 is the process of establishing a digital terrain that matches the real world and the modeling basic spatial coordinate input modeling software. The two-dimensional raster data layer cannot truly reflect the terrain undulation of the data, nor does it have the coordinate point information required for the creation of the vegetation model. Therefore, it is necessary to first use the modeling software system CityEngine of the present invention to import the ALOS open source terrain data and the Thiessen polygon area of ​​the effective dense woodland area; and build a digital elevation model (DEM) based on the terrain data; then use the statistical numerical method that conforms to the rules in ArcGIS and generate vegetation distribution points under the limitation of "Thyssen polygon area of ​​the effective dense woodland area", and import them into CityEngine; finally, perform elevation interpolation operation on this vegetation distribution point based on the DEM model to give it a three-dimensional height parameter. Therefore, this process mainly includes DEM modeling, vegetation distribution point generation, and elevation interpolation. In S5, based on the elevation modeling and elevation interpolation results, the CGA code of CityEngine is written to generate a dense woodland environment model that conforms to certain rules. The vegetation monomer model in the dense woodland model uses the model resources in the built-in obj format of CityEngine, and the implementer can also build the model by himself according to needs. Based on the geometric random transformation of CGA rules, the single vegetation model is appropriately adjusted within the allowable range to avoid the regular repetition of the visual model and improve the naturalness of the visual perception. The model scene is obtained, and the digital twin scene assets that meet the Omniverse-USD and Unreal-DataSmith data standards are output to obtain a vegetation environment close to the real density distribution. Therefore, this step includes CGA code writing and calling, and scene asset saving.

[0022] In some embodiments, the remote sensing data preprocessing stage S1 further includes the following steps: S11: remote sensing data is obtained from open source data sources, multispectral radar data is obtained from Sentinel-2 data sources, digital elevation model data is obtained from ALOS platform 12.5m resolution data sources, and orthochromatic image data is obtained from satellite image data sources; S12: Import the remote sensing data into the remote sensing image processing software and use the seamless mosaic tool to perform mosaic operation on the scattered remote sensing data to obtain the first remote sensing data; perform radiation calibration on the first remote sensing data to obtain the second remote sensing data; perform atmospheric correction on the second remote sensing data to obtain the third remote sensing data; perform orthorectification on the third remote sensing data, and determine the atmospheric model, aerosol model, and ground elevation parameters of the third remote sensing data according to the acquisition time, satellite attitude, orbital position, and data set range of the third remote sensing data to obtain the fourth remote sensing data; calculate the enhanced vegetation index EVI of the fourth remote sensing data to obtain the remote sensing data after data preprocessing; the enhanced vegetation index EVI is calculated by the following formula: Wherein, NIR is the near infrared band reflectivity, which is the 8th band data of Sentinel-2 radar; R is the red band reflectivity, which is the 4th band data of Sentinel-2 radar; B is the blue band reflectivity, which is the 2nd band data of Sentinel-2 radar; G is the gain factor, which is 2.5; C1 and C2 are the first and second atmospheric correction coefficients, which are 6 and 7.5 respectively; L is the soil adjustment parameter, which is 1; S13: Load the pre-processed multispectral radar data and orthophoto panchromatic image data into the geographic information system software, and perform mask mapping on the target data set to be processed; perform mask mapping on the contour of the pre-processed orthophoto panchromatic image data; perform mask mapping on the pre-processed multispectral radar data with reference to the pre-processed orthophoto panchromatic image data and use continuous and closed polygons; finally, export the masked data as training data.

[0023] In this embodiment, S11 first downloads open source commercial remote sensing data through open source data sources such as the Geospatial Data Cloud or the European Space Agency Copernicus Data Platform, and it is noted that the present invention requires a large amount of high-resolution data for model training and tuning. Therefore, in addition to elevation data, the Sentinel-2 data source is selected as the main data source for multispectral radar, the ALOS platform 12.5m resolution data source is selected as the main data source for digital elevation, and mainstream satellite image products such as Google or Bing are selected as the main data source for orthophoto full-color satellite images. The above-mentioned raster data is generally stored in GeoTIFF format with a suffix of ".tif". In S12, the source data is imported through ENVI software, and the scattered data originally in multiple independent files is mosaicked using the Seamless Mosaic tool (SeamlessMosaic). It can be seen from the principle of remote sensing that the source data of each band is the original brightness or DN (Digital Number) value, so it is necessary to convert it into a quantifiable physical quantity through radiation calibration. First, use radiometric calibration, and use the calibration type corresponding to the data set to calibrate the data. The tools are all automated processing for specific satellite types, and the parameters do not need to be debugged. Secondly, use the atmospheric correction tool to reprocess the above radiometric calibration result data to eliminate the influence of factors such as the atmospheric environment on the data. It is necessary to select the corresponding parameters in the tool tab. For example, when Sentinel-2 is selected, the correction method should use the FLAASH algorithm, and determine its atmospheric model, aerosol model, ground elevation and other parameters for the acquisition time and satellite attitude, orbital position, data set range and other related parameters of the processed data set. Finally, it is necessary to calculate the Enhanced Vegetation Index (EVI) to evaluate the density of vegetation in a unit area when the vegetation density is large. In S13, the general data processing tools for data label masks in the direction of machine vision in deep learning can usually be used, such as LableAnything, LableMe, etc., and the present invention uses ArcGIS Pro's own label annotation tool for mask annotation to obtain compatibility with image data containing geographic location information. First, load the ENVI preprocessed multispectral radar data and orthophoto panchromatic image data into ArcGIS Pro software, and ensure that the target dataset to be processed is selected; secondly, use the "label objects for deep learning" tool in the classification toolbar to mask the dataset. For multispectral radar data, it is mainly used for dense vegetation, such as artificial forests, natural forests, etc., where the vegetation coverage is dense and the crown is continuously distributed and cannot be distinguished by the naked eye. Therefore, it is necessary to mask the dense forest type area data using continuous and closed polygons to depict the boundary contours of the dense forest.Because multispectral radar data is expressed in numerical values, it is usually not directly interpretable by humans, so the reference for mask drawing is the orthophoto panchromatic image. Based on the fact that the boundary of the significant dense forestland feature in the orthophoto panchromatic image is distinguishable by the human eye, its contour is marked with a mask, and this mask is applied to the multispectral radar data for export. The export method is still through the "Export Dataset for Deep Learning" tool, and ensure that the "metadata format" of such data is "panoramic segmentation" when it is exported. Finally, in the above export steps, two branches are formed, one of which is the case of single non-dense vegetation with "metadata format" of "RCNN", and the other is the case of dense and continuous vegetation with "metadata format" of "panoramic segmentation".

[0024] In some embodiments, the S2: semantic segmentation model training and tuning stage further includes the following steps: S21: Use PyTorch framework, U-Net architecture, and DenseNet backbone network to build a semantic segmentation model; S22: Perform model training on the semantic segmentation model and adjust the hyperparameters of the semantic segmentation model to achieve a first preset quantitative index; test the semantic segmentation model, and tune the model according to the test results until the semantic segmentation model reaches a second preset quantitative index to obtain a trained semantic segmentation model.

[0025] In this embodiment, the semantic segmentation model in S21 uses the PyTorch framework and the DenseNet backbone network as the core algorithm components for image processing. The main algorithm adopts the U-Net architecture, and uses the encoding-decoding idea of ​​convolution + pooling for downsampling and transposed convolution and channel splicing for upsampling to perform pixel-level segmentation of remote sensing multispectral data. S22 first involves the semantic segmentation model training and tuning process, which requires tuning and controlling the hyperparameters so that their performance meets the corresponding quantitative indicators; followed by the semantic segmentation model test, which tests the results of the trained semantic segmentation model in combination with quantitative indicators. The semantic segmentation model training and tuning process mainly focuses on whether there are gradient vanishing and gradient explosion in the loss iteration, and dynamically tunes the non-descriptive model structure hyperparameters such as learning rate, batch size, and number of iterations to improve the model until the quantitative indicators of its final performance evaluation meet the requirements, such as the Dice coefficient of the semantic segmentation model is greater than 0.75. The relevant single or multiple quantitative indicators are freely selected by the technicians, and the present invention does not limit them. The final trained data uses ONNX, a universal deep learning model data format, to store the model data to ensure its portability and versatility. The semantic segmentation model result test is a key step in using the reserved validation set data to verify the performance of the model. Generally, 10% of random data is pre-selected and separated from the original training set for the final test, and the final test index results of the semantic segmentation model of the present invention meet the same trend as the relevant performance index obtained after the model training, such as the Dice coefficient of the semantic segmentation model is greater than 0.75, otherwise it means that the overall performance of the model is poor and the model needs to be re-tuned.

[0026] In some embodiments, the S3: forest area semantic segmentation and recognition stage further includes the following steps: S31: Use the trained semantic segmentation model to segment the forest area of ​​the multispectral radar data after data preprocessing into multiple sub-areas and perform data prediction on each sub-area to obtain local segmentation maps. Then, perform data mosaicking on the local segmentation maps to merge them to obtain a complete segmentation map. The input multispectral radar data is segmented using the parameters in the following formula: In the formula, C x is the cutting step length for the input multi-spectrum radar data in the horizontal direction; I x is the number of samples cut out in the horizontal direction from the input multi-spectrum radar data; T x is the number of pixels in the horizontal direction of the input multi-spectrum radar data, in pixels; S xis the number of pixels in the horizontal direction of a single sample obtained by cutting, in pixels; G x is the number of pixels in the overlapping area of ​​two adjacent samples in the horizontal direction, in pixels; C y The cutting step size for the input multispectral radar data in the vertical direction; I y is the number of samples cut out in the vertical direction from the input multi-spectrum radar data; T y is the number of pixels in the vertical direction of the input multispectral radar data, in pixels; S y is the number of pixels in the vertical direction of a single sample obtained by cutting, in pixels; G y is the number of pixels in the overlapping area of ​​two adjacent samples in the vertical direction, in pixels; “⌈...⌉” means rounding up; C x , I x , T x , S x , G x , C y , I y , T y , S y , G y are all positive integers; S32: Load the complete segmentation map, set the judgment condition to clear the background to obtain the first segmentation map; use the raster to vector surface conversion function to convert the effective dense woodland area in the raster of the first segmentation map into a vector surface to obtain the first raster data; use the first raster data as a mask and clear the non-forest area of ​​the multispectral radar data after the data preprocessing initially input in S31 to obtain the second raster data; convert the second raster data into a raster point set, and the data field inherited by the raster point set is the value of each pixel of the input raster; input the raster point set into the create Thiessen polygon function to obtain multiple Thiessen polygon areas centered on the points in the raster point set and with adjacent sides; use the geometric intersection calculation function to calculate the intersection of the first raster data and the Thiessen polygon area to obtain the Thiessen polygon area of ​​the effective dense woodland area.

[0027] In this embodiment, the semantic segmentation model in S31 is loaded using the "torch.load" method in the PyTorch library, and the affine transformation matrix and projection information of the data are obtained and stored based on the "GetGeoTransform" and "GetProjection" methods of the GDAL library. Note that in general, the multispectral remote sensing data of this type of task are large-scale data (the unit of measurement is square kilometers), so sending a single large-scale data into the model for prediction at one time will increase the computer hardware load. Therefore, the present invention reasonably cuts the data and predicts them one by one in batches before merging the prediction results to avoid such problems. S x , S y , G x , G y The settings of the following four parameters shall comply with the following standards: S x = S y =1024, G x = G y ; and meet G x Can be T x divisible, and G y Can be T yInteger divisibility. In particular, when the aforementioned integer divisibility condition cannot be met, it is necessary to reduce and adjust the last cutting step length of the sample in the horizontal or vertical direction to ensure the normal output of the cutting data. For example, if-else judgment is added to the cutting output logic based on the Python-based GDAL environment to calculate the number of remaining pixels of the last sample from the boundary of one side of the input data, and finally obtain the cutting step length of the last sample. After the cutting is completed, the data is predicted and the results are output one by one, and a separate segmentation map is formed, and its storage format is the same as the input data. After completing the entire data segmentation and prediction process, the "gdal.Warp" function in GDAL is used to mosaic the above results (the mosaic operation of the raster, that is, the operation of merging sub-grids based on spatial position) to obtain a complete segmentation map. The value of each pixel in the figure reflects the type of the corresponding area, that is, it is predicted to be dense woodland or other. In S32, the ArcGISPro geographic data processing library ArcPy based on Python is used to vectorize the segmentation results, and the complete segmentation map of the prediction results is converted into vector data as the source data to achieve the purpose of framing the scope of vegetation modeling. First, the complete segmentation map is loaded using the ArcPy raster data loading function "arcpy.Raster", and the raster pixel nulling function "SetNull" in the ArcPy library is used to set the judgment condition to set the background to a null value, that is, the pixel that is represented as "NoData" in the raster data; based on the obtained segmentation map with the nulled background, the raster to vector surface conversion function "RasterToPolygon" is used to convert the effective dense forest area in the raster into a vector surface, where the third parameter value of the function is used as the control of the output vector edge simplification, and its value is set to "false" and the remaining parameters are adjusted and filled in according to actual needs. The raster after the nulling process is used as a mask, and the "SetNull" function is used again to perform the nulling operation on the original multispectral radar data at the prediction input end, and the raster data of the non-forest area is obtained. Secondly, the "RasterToPoint" function is used to convert the raster after the nulling process into a raster point set, and its inherited data field is the value of each pixel of the input raster, that is, the EVI value. Subsequently, the “CreateThiessenPolygons” function is used to take the grid point set created above as input. In order to ensure that the output data completely retains the EVI value carried by the point set, the third parameter of the function needs to be set to “ALL”. Finally, a number of rectangular areas centered on the point set and adjacent to each other are obtained, and their number is the same as the number of input points.Finally, the geometric intersection calculation function "Intersect" is used to calculate the intersection of the aforementioned effective dense forest area vector surface and the Thiessen polygon vector set to obtain the Thiessen polygon area of ​​the effective dense forest area, which is manifested as any effective dense forest area is completely covered by the rectangular surface adjacent to each side. Each rectangle corresponds to a valid raster pixel, and the value inherited by its vector attribute is the same as the EVI value of the raster pixel at the same spatial position. The final processing results need to be stored in a file geodatabase (File GeoDatabase). Use ArcPy's "arcpy.CreateFileGDB_management" to create an empty database, and use the "ExportFeatures" function to output the processing results. The second item directory parameter of the output function needs to be set to the storage path of the file geodatabase created above.

[0028] In some embodiments, the S4: modeling scene construction phase further includes the following steps: S41: Import the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area into the modeling software, use the terrain import mode to input the terrain data of the same range as the multispectral radar data of the target area, and perform seamless mosaicking and clipping operations on the terrain data to obtain the digital elevation model; S42: Taking any rectangular unit in the Thiessen polygon area of ​​the effective dense forest area as the benchmark, calculate the number of vegetation per unit area TD: Where A is the area of ​​the current measurement area; is the average canopy area of ​​vegetation in the current measurement area; EVI is the EVI value of the current rectangular unit in the Thiessen polygon area of ​​the effective dense forest area; “⌈...⌉” is a rounding operation; After obtaining the number of vegetation per unit area TD, calculate the vegetation distribution points: in, is the easting distance of the current t-th random coordinate point, and satisfies t≤TD; is the northing distance of the current t-th random coordinate point, and satisfies t≤TD; D east is the westernmost east distance of the current rectangular unit; D west is the easternmost east distance of the current rectangular unit; D north is the northernmost north distance of the current rectangular unit; D south is the southernmost north distance of the current rectangular unit; Nr is a random number that follows a Gaussian distribution; n The number of samples generated to satisfy the uniform distribution condition; X i is the i-th sample value that satisfies uniform distribution; S43: Importing vegetation distribution points into modeling software, using the digital elevation model as the elevation data source, matching the vegetation distribution points with the elevation data source to obtain a vegetation spatial coordinate point set.

[0029] In this embodiment, S41 imports ALOS open source terrain data and Thiessen polygon areas of effective dense woodland areas based on the modeling software system CityEngine. Use the "Terrain" terrain import mode to input terrain data of the same range as the multispectral radar data of the target area. ArcGIS Pro, QGIS and other geographic information processing tool software or Python's GDAL library can be used to seamlessly mosaic and crop the basic terrain raster data. The final input is a DEM model (digital elevation model) that includes accurate terrain undulations. The biggest feature of CityEngine in S42 is that it can use CGA rule code to generate a three-dimensional model. Therefore, it is necessary to estimate the unit space coverage of vegetation that conforms to the remote sensing observation law within the target range through a program, and then generate the coordinates of the spatial points within the calculation range through Gaussian random numbers. It is worth noting that the layout of random points is established after the TD value is determined. The generation of random points requires the generation of 5000 to 10000 random numbers that obey a uniform distribution based on Python's built-in random number generation function "random.uniform", and then calculate the random numbers that obey a Gaussian distribution. N r , based on which the random coordinates are calculated. Then, based on ArcPy, the easting and northing value pairs of the generated random coordinates are stored as a data table, and the aforementioned data table is converted into coordinate points using "XYTableToPoint", and finally, it is converted into shapefile format geographic vector data with the suffix ".shp" using "FeatureClassToShapefile", and the vegetation distribution points are obtained. In S43, the vegetation distribution points are imported into CityEngine, and the digital elevation model is used as the elevation data source. The "Align shapes to terrain" command is used to match the vegetation distribution point data with the elevation data source to obtain a set of spatial coordinate points that are completely consistent with the actual elevation, that is, the vegetation spatial coordinate point set.

[0030] In some embodiments, the S5: scene asset generation and output stage further includes the following steps: S51: Generate a vegetation monomer model and initialize the attributes of the vegetation monomer model to obtain a rule file, select all coordinate points in the vegetation space coordinate point set and bind them to the rule file, and then start the model generation instruction to obtain a dense forest environment model; S52: Perform geometric random transformation on the vegetation monomer model based on CGA rules, and then export the vegetation monomer model and dense forest environment model as digital twin scene assets and output them.

[0031] In this embodiment, the core purpose of S51 is to write CGA code for the creation of a dense forest environment model, which mainly includes two sections: attribute parameter predefinition and driving rule coding; after the coding is completed, the coding rules are called to generate the model to obtain the result. The attribute parameters are predefined in the form of "attr" keyword + attribute name to initialize the attributes of a single vegetation model, including: the horizontal and horizontal scaling ratio (zoom_x), horizontal and vertical scaling ratio (zoom_y), vertical scaling ratio (zoom_z), rotation angle around the vertical axis (rotate_y), file path array (model_path), the file path array (such as "array name ['string 1', 'string 2', 'string 3', ...]"), all model paths involved need to be initialized in the form of string array, and the other attribute parameters mentioned above can be initialized to 0; the driving rule encoding adopts the form of main rule + sub-rule, and a line of "@StartRule" needs to be inserted before the main rule as an entry mark. Its declaration is in the form of "main rule name-->", which only acts on vector elements such as points, lines, and surfaces that are assigned corresponding initialization rules, so that the program can perform arbitrary operations based on geometric elements; the main rule is followed by the model instantiation function "i(model_path)", the instance scaling function "s(zoom_x, zoom_z, The parameter assignment format of “model_path” in the instantiation function is “array name [floor (rand (max))]”, floor (…) is the floor rounding function, rand (…) is the uniform random number generation function, and max is the total number of tree model types; the instance zoom factor also uses the rand (…) function for restrictive random number generation, and its value should refer to the interval [0.5, 2.5]; the instance rotation only considers the transformation of the parameter “rotate_y”, that is, random rotation around the vertical axis y, and the angle value is generated by the rand (…) function. Based on the above encoding process, a rule file in the ".cga" format is obtained, all the vegetation spatial coordinate point sets are selected, the rule file (Rule file) of the vector point coordinates selected in the "Inspector" window is bound to the rule file in the ".cga" format, and the initial rule (StartRule) is set as the main rule in the rule file in the ".cga" format, and the model generation instruction (Generate models of selectedshapes) is started to obtain the final scene result, that is, the dense woodland environment model.In S52, the vegetation monomer model and the dense woodland environment model are exported as scene assets. The main export mode used is the Python script (scripting library) of CityEngine. By writing the code "ce.getObjectsFrom(ce.scene)", the output model is defined as all models of the current scene (that is, the model retrieval domain is "ce.scene"), and the output parameters are limited by any of the export parameter modes "USDExportModelSettings()" and "UnrealExportModelSettings()". The former corresponds to the USD data exchange standard of Ominiverse, and the latter corresponds to the DataSmith data exchange standard of Unreal. Finally, ce.export({objects},{settings}) is used to output the target model. The {objects} parameter corresponds to the selected model and DEM layer, and {setteing} corresponds to the data output settings.

[0032] In some embodiments, the semantic segmentation model is a U-Net architecture.

[0033] In some embodiments, the modeling software is CityEngine and the geographic information system software is ArcGIS Pro.

[0034] The second aspect of the present invention provides: a modeling system based on a semantic segmentation model, used to implement any of the above-mentioned modeling methods based on a semantic segmentation model, comprising: A remote sensing data preprocessing module is used to obtain remote sensing data, then preprocess the remote sensing data, and finally export the preprocessed remote sensing data as training data; the remote sensing data includes multispectral radar data, orthophoto panchromatic image data, and digital elevation model data; the data preprocessing includes data mosaicking, radiation calibration, atmospheric correction, orthophoto correction, and enhanced vegetation index (EVI) quantitative calculation; The semantic segmentation model training and tuning module is used to build a semantic segmentation model and perform model training and model tuning to obtain a trained semantic segmentation model; The forest area semantic segmentation recognition module is used to use the trained semantic segmentation model to perform pixel-level classification on the multispectral radar data of the target area to obtain semantic segmentation results, and to vectorize the semantic segmentation results to obtain the Thiessen polygon area of ​​the effective dense forest area; The modeling scene construction module is used to import the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area into the modeling software to construct the digital elevation model; use the statistical numerical method in the geographic information system software to generate vegetation distribution points under the limitation of the Thiessen polygon area of ​​the effective dense forest area and import them into the modeling software; finally, perform elevation interpolation operation on the vegetation distribution points to give the vegetation distribution points three-dimensional height parameters to obtain the vegetation space coordinate point set; The scene asset generation and output module is used to generate a dense forest environment model based on the digital elevation model and the vegetation spatial coordinate point set. The vegetation monomer model in the dense forest environment model uses the model resources built into the modeling software or a self-built model; based on the CGA rules, the vegetation monomer model is geometrically randomly transformed to obtain the digital twin scene asset and output it.

[0035] The third aspect of the present invention provides: a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by a processor, any of the above-mentioned modeling methods based on the semantic segmentation model is implemented.

[0036] The present invention combines mainstream open source multispectral radar remote sensing, orthophoto panchromatic image data, and digital elevation model, with dense woodland area range image segmentation as the first-level purpose, and initializes the data within the dense woodland range to obtain the EVI index as the measurement value for the subsequent judgment of the vegetation density in the unit dense woodland area of ​​this scheme, bypassing the problem of high initial investment in mainstream oblique photography and laser point cloud for three-dimensional dense woodland scene modeling under conventional circumstances; at the same time, ENVI and ArcGIS Pro software are combined for the preliminary preparation of image segmentation task data, so that the image segmentation results within the dense woodland range can directly carry correct geographic spatial information; and this scheme respectively utilizes the characteristics of multispectral radar remote sensing and orthophoto panchromatic images, that is, the characteristics that orthophoto panchromatic images can be clearly identified by the human eye, and accordingly depicts a mask layer for multispectral radar remote sensing image segmentation, and acts on multispectral radar data for output to obtain training data, so that the quality of model training data for the dense woodland range segmentation task is improved.

[0037] The division of dense woodland areas mainly adopts the U-Net architecture to train the semantic segmentation model based on the ResNet-50 backbone network, and uses the Dice coefficient to evaluate the performance of the model, outputting a more transferable ONNX format model file, ensuring that the model has better performance in the semantic segmentation task of dense woodland areas based on the remote sensing image EVI index.

[0038] The semantic segmentation model combines GDAL and Pytorch libraries to perform miniaturized segmentation operations on the input data, which reduces the pressure of data calculations. The two are used to implement complete data segmentation, semantic segmentation, vectorized output and other processes, making the semantic segmentation task of dense woodland areas based on multispectral remote sensing data of the EVI index more standardized and more practical. The raster is transformed into a vector grid surface based on a single pixel by using the conversion of raster points and Thiessen polygons, which provides data support for the subsequent generation of refined coordinates based on EVI.

[0039] Based on the vector grid surface carrying the EVI index, the vegetation modeling base points were generated based on the Gaussian distribution, so that the distribution of the vegetation model was consistent with the results of satellite observations to a certain extent. It is not a mechanical and uniform distribution form, but a random and natural distribution form, which makes the randomness of the modeling results close to the real natural distribution form.

[0040] The model generation of dense woodland vegetation is based on CGA rules. The model creation process is described using a programming language with strong descriptiveness. The vegetation model is quickly and randomly generated in large quantities based on the alignment of initial rules and three-dimensional point element rules. The randomness of the individual vegetation model is based on the restrictive random number scaling transformation of the model that controls the insertion of the i function after the main rule. The dense woodland scene assets output by the present invention are based on Python scripts, and the obtained data exchange format complies with the USD of Omniverse or the DataSmith data standard of Unreal, and can be widely and quickly connected with mainstream digital twin platforms.

[0041] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A modeling method based on a semantic segmentation model, characterized in that: The following steps are involved: S1: remote sensing data preprocessing stage, acquiring remote sensing data, then preprocessing the remote sensing data, and finally exporting the preprocessed remote sensing data as training data; the remote sensing data includes multispectral radar data, orthophoto panchromatic image data, and digital elevation model data; the data preprocessing includes data mosaicking, radiation calibration, atmospheric correction, orthophoto correction, and enhanced vegetation index (EVI) quantitative calculation; S2: Semantic segmentation model training and tuning stage, building a semantic segmentation model and performing model training and model tuning to obtain a trained semantic segmentation model; S3: In the forest area semantic segmentation and recognition stage, the trained semantic segmentation model is used to perform pixel-level classification on the multispectral radar data of the target area to obtain the semantic segmentation results, and the semantic segmentation results are vectorized to obtain the Thiessen polygon area of ​​the effective dense forest area; S4: Modeling scenario construction phase, digital elevation model data and Thiessen polygon areas of effective dense forest areas are imported into the modeling software to construct a digital elevation model; Using the statistical numerical method in the geographic information system software, the vegetation distribution points are generated under the limitation of the Thiessen polygon area of ​​the effective dense forest area and imported into the modeling software; finally, the vegetation distribution points are interpolated by the elevation operation to give the vegetation distribution points three-dimensional height parameters to obtain the vegetation spatial coordinate point set; S5: In the scene asset generation and output stage, a dense forest environment model is generated based on the digital elevation model and the vegetation spatial coordinate point set. The vegetation monomer model in the dense forest environment model uses the model resources built into the modeling software or a self-built model. Based on the CGA rules, the vegetation monomer model is geometrically randomly transformed to obtain the digital twin scene asset and output it.

2. The modeling method based on the semantic segmentation model according to claim 1, characterized in that: The S1: remote sensing data preprocessing stage further includes the following steps: S11: remote sensing data is obtained from open source data sources, multispectral radar data is obtained from Sentinel-2 data sources, digital elevation model data is obtained from ALOS platform 12.5m resolution data sources, and orthochromatic image data is obtained from satellite image data sources; S12: Import the remote sensing data into the remote sensing image processing software and use the seamless mosaic tool to perform mosaic operation on the scattered remote sensing data to obtain the first remote sensing data; perform radiation calibration on the first remote sensing data to obtain the second remote sensing data; perform atmospheric correction on the second remote sensing data to obtain the third remote sensing data; perform orthorectification on the third remote sensing data, and determine the atmospheric model, aerosol model, and ground elevation parameters of the third remote sensing data according to the acquisition time, satellite attitude, orbital position, and data set range of the third remote sensing data to obtain the fourth remote sensing data; calculate the enhanced vegetation index EVI of the fourth remote sensing data to obtain the remote sensing data after data preprocessing; the enhanced vegetation index EVI is calculated by the following formula: Wherein, NIR is the near infrared band reflectivity, which is the 8th band data of Sentinel-2 radar; R is the red band reflectivity, which is the 4th band data of Sentinel-2 radar; B is the blue band reflectivity, which is the 2nd band data of Sentinel-2 radar; G is the gain factor, which is 2.5; C1 and C2 are the first and second atmospheric correction coefficients, which are 6 and 7.5 respectively; L is the soil adjustment parameter, which is 1; S13: Load the pre-processed multispectral radar data and orthophoto panchromatic image data into the geographic information system software, and perform mask mapping on the target data set to be processed; perform mask mapping on the contour of the pre-processed orthophoto panchromatic image data; perform mask mapping on the pre-processed multispectral radar data with reference to the pre-processed orthophoto panchromatic image data and use continuous and closed polygons; finally, export the masked data as training data.

3. The modeling method based on the semantic segmentation model according to claim 1, characterized in that: The S2: semantic segmentation model training and tuning phase also includes the following steps: S21: Use PyTorch framework, U-Net architecture, and DenseNet backbone network to build a semantic segmentation model; S22: Perform model training on the semantic segmentation model and adjust the hyperparameters of the semantic segmentation model to achieve a first preset quantitative index; test the semantic segmentation model, and tune the model according to the test results until the semantic segmentation model reaches a second preset quantitative index to obtain a trained semantic segmentation model.

4. The modeling method based on the semantic segmentation model according to claim 1, characterized in that: The S3: forest area semantic segmentation and recognition stage further includes the following steps: S31: Use the trained semantic segmentation model to segment the forest area of ​​the multispectral radar data after data preprocessing into multiple sub-areas and perform data prediction on each sub-area to obtain local segmentation maps. Then, perform data mosaicking on the local segmentation maps to merge them to obtain a complete segmentation map. The input multispectral radar data is segmented using the parameters in the following formula: In the formula, C x is the cutting step length for the input multi-spectrum radar data in the horizontal direction; I x is the number of samples cut out in the horizontal direction from the input multi-spectrum radar data; T x is the number of pixels in the horizontal direction of the input multi-spectrum radar data, in pixels; S x is the number of pixels in the horizontal direction of a single sample obtained by cutting, in pixels; G x is the number of pixels in the overlapping area of ​​two adjacent samples in the horizontal direction, in pixels; C y The cutting step size for the input multispectral radar data in the vertical direction; I y is the number of samples cut out in the vertical direction from the input multi-spectrum radar data; T y is the number of pixels in the vertical direction of the input multispectral radar data, in pixels; S y is the number of pixels in the vertical direction of a single sample obtained by cutting, in pixels; G y is the number of pixels in the overlapping area of ​​two adjacent samples in the vertical direction, in pixels; "⌈...⌉" is a round-up operation; C x , I x , T x , S x , G x , C y , I y , T y , S y , G y are all positive integers; S32: Load the complete segmentation map, set the judgment condition to clear the background to obtain the first segmentation map; use the raster to vector surface conversion function to convert the effective dense woodland area in the raster of the first segmentation map into a vector surface to obtain the first raster data; use the first raster data as a mask and clear the non-forest area of ​​the multispectral radar data after the data preprocessing initially input in S31 to obtain the second raster data; convert the second raster data into a raster point set, and the data field inherited by the raster point set is the value of each pixel of the input raster; input the raster point set into the create Thiessen polygon function to obtain multiple Thiessen polygon areas centered on the points in the raster point set and with adjacent sides; use the geometric intersection calculation function to calculate the intersection of the first raster data and the Thiessen polygon area to obtain the Thiessen polygon area of ​​the effective dense woodland area.

5. The modeling method based on the semantic segmentation model according to claim 1, characterized in that: The S4: modeling scenario construction phase also includes the following steps: S41: Import the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area into the modeling software, use the terrain import mode to input the terrain data of the same range as the multispectral radar data of the target area, and perform seamless mosaicking and clipping operations on the terrain data to obtain the digital elevation model; S42: Taking any rectangular unit in the Thiessen polygon area of ​​the effective dense forest area as the benchmark, calculate the number of vegetation per unit area TD: Where A is the area of ​​the current measurement area; is the average canopy area of ​​vegetation in the current measurement area; EVI is the EVI value of the current rectangular unit in the Thiessen polygon area of ​​the effective dense forest area; "⌈...⌉" is the rounding operation; After obtaining the number of vegetation per unit area TD, calculate the vegetation distribution points: in, is the easting distance of the current t-th random coordinate point, and satisfies t≤TD; is the northing distance of the current t-th random coordinate point, and satisfies t≤TD; D east is the westernmost east distance of the current rectangular unit; D west is the easternmost east distance of the current rectangular unit; D north is the northernmost north distance of the current rectangular unit; D south is the southernmost north distance of the current rectangular unit; N r is a random number that follows a Gaussian distribution; n The number of samples generated to satisfy the uniform distribution condition; X i is the i-th sample value that satisfies uniform distribution; S43: Importing vegetation distribution points into modeling software, using the digital elevation model as the elevation data source, matching the vegetation distribution points with the elevation data source to obtain a vegetation spatial coordinate point set.

6. The modeling method based on the semantic segmentation model according to claim 1, characterized in that: The S5: scene asset generation and output stage further includes the following steps: S51: Generate a vegetation monomer model and initialize the attributes of the vegetation monomer model to obtain a rule file, select all coordinate points in the vegetation space coordinate point set and bind them to the rule file, and then start the model generation instruction to obtain a dense forest environment model; S52: Perform geometric random transformation on the vegetation monomer model based on CGA rules, and then export the vegetation monomer model and dense forest environment model as digital twin scene assets and output them.

7. The modeling method based on the semantic segmentation model according to any one of claims 1 to 6, characterized in that: The semantic segmentation model is a U-Net architecture.

8. The modeling method based on the semantic segmentation model according to claim 1, characterized in that: The modeling software is CityEngine, and the geographic information system software is ArcGIS Pro.

9. A modeling system based on a semantic segmentation model, characterized in that: The method for implementing a modeling method based on a semantic segmentation model as claimed in any one of claims 1 to 8 comprises: A remote sensing data preprocessing module is used to obtain remote sensing data, then preprocess the remote sensing data, and finally export the preprocessed remote sensing data as training data; the remote sensing data includes multispectral radar data, orthophoto panchromatic image data, and digital elevation model data; the data preprocessing includes data mosaicking, radiation calibration, atmospheric correction, orthophoto correction, and enhanced vegetation index (EVI) quantitative calculation; The semantic segmentation model training and tuning module is used to build a semantic segmentation model and perform model training and model tuning to obtain a trained semantic segmentation model; The forest area semantic segmentation recognition module is used to use the trained semantic segmentation model to perform pixel-level classification on the multispectral radar data of the target area to obtain semantic segmentation results, and to vectorize the semantic segmentation results to obtain the Thiessen polygon area of ​​the effective dense forest area; The modeling scene construction module is used to import the digital elevation model data and the Thiessen polygon area of ​​the effective dense forest area into the modeling software to construct the digital elevation model; use the statistical numerical method in the geographic information system software to generate vegetation distribution points under the limitation of the Thiessen polygon area of ​​the effective dense forest area and import them into the modeling software; finally, perform elevation interpolation operation on the vegetation distribution points to give the vegetation distribution points three-dimensional height parameters to obtain the vegetation space coordinate point set; The scene asset generation and output module is used to generate a dense forest environment model based on the digital elevation model and the vegetation spatial coordinate point set. The vegetation monomer model in the dense forest environment model uses the model resources built into the modeling software or a self-built model; based on the CGA rules, the vegetation monomer model is geometrically randomly transformed to obtain the digital twin scene asset and output it.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by the processor, the modeling method based on the semantic segmentation model as described in any one of claims 1 to 8 is implemented.

Images