A method and system for constructing a three-dimensional model of a farmland based on remote sensing images

Abstract

The present application relates to remote sensing mapping modeling technical field, specifically to a kind of construction method and system of farmland three-dimensional model based on remote sensing image, comprising: obtaining the multi-temporal multi-angle remote sensing image data set containing stereo image pair image, multispectral image and imaging parameter. The stereo image pair image is processed using the multi-view stereo matching algorithm of optimization matching cost calculation aggregation, and the farmland digital surface model is generated. The multispectral image is subjected to semantic segmentation and spectral feature extraction, and the ground cover classification map and crop characteristic parameter atlas are generated. The three types of data are fused to input three-dimensional structure generation model, and the refined farmland three-dimensional structure grid model is generated. Through the attribute mapping module, each surface unit is given crop physiological attribute parameter, and the target farmland three-dimensional model conforming to the visualization standard is output after detail level simplification and format conversion.

Description

Technical Field

[0001] This invention relates to the field of remote sensing mapping and modeling technology, and in particular to a method and system for constructing a three-dimensional model of farmland based on remote sensing images. Background Technology

[0002] The construction of 3D models of farmland generally relies on conventional stereo matching methods of remote sensing images for terrain and surface reconstruction. These methods often employ standard multi-view stereo matching algorithms with fixed matching cost calculation logic. The algorithm parameters are adapted to general surface scenes without considering the spatial distribution characteristics of farmland crops. Remote sensing images are used only for elevation calculation based on stereo pairs, without integrating multispectral imagery for semantic analysis and feature mining, resulting in low utilization of image data resources.

[0003] Conventional modeling processes only generate simple terrain surface models, lacking the integration of crop land cover classification and spectral feature information. Terrain data and crop spectral data are independent of each other, making it impossible to incorporate the crop's own structural features for 3D inversion. The surface model only possesses spatial geometry and cannot be linked to crop physiological attributes. The model only retains the terrain framework and lacks the attribute relationships between actual farmland surface elements.

[0004] Traditionally generated 3D models lack layered detail adaptation and have a standardized mesh structure, making them unsuitable for visualization specifications across different application scenarios. The industry needs a 3D matching processing method adapted to the distribution characteristics of farmland crops, establishing a modeling process that integrates multi-source remote sensing image fusion modeling with the unified assignment of geometric structures and crop attributes. This will address the shortcomings of existing modeling methods, such as weak scenario adaptability and limited information dimensions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for constructing a three-dimensional farmland model based on remote sensing images.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing a three-dimensional model of farmland based on remote sensing imagery, comprising: Acquire a multi-temporal, multi-angle remote sensing image data set of the target farmland area, the remote sensing image data set including stereo image pairs, multispectral images and corresponding remote sensing sensor imaging parameters; An improved multi-view stereo matching algorithm is applied to the stereo image pairs to generate an initial digital surface model of farmland. The improved multi-view stereo matching algorithm optimizes the calculation and aggregation process of matching costs based on the spatial distribution pattern of farmland crops. Perform crop semantic segmentation and spectral feature extraction processing on the multispectral images to generate farmland surface cover classification maps and crop feature parameter maps; The farmland digital surface model, land cover classification map and crop feature parameter map are input into the three-dimensional structure generation model to perform terrain correction, crop three-dimensional structure inversion and detail enhancement, and generate a refined farmland three-dimensional structure mesh model. Based on the refined three-dimensional structure mesh model of farmland and the crop feature parameter map, the three-dimensional attribute mapping module is called to assign corresponding crop physiological attribute parameters to each surface unit of the three-dimensional structure mesh model of farmland, thereby constructing a three-dimensional attributed model of farmland with attribute information. The farmland 3D attribute model is simplified and format converted at multiple levels of detail to generate a target farmland 3D model that conforms to the preset 3D visualization standard.

[0007] As a further aspect of the present invention, an improved multi-view stereo matching algorithm is applied to the stereo image pair to generate an initial digital surface model of farmland, including: The stereo image pair image is preprocessed, including radiometric calibration, atmospheric correction and geometric correction, to obtain a corrected standard stereo image pair image. On the standard stereo image pair, the improved multi-view stereo matching algorithm is used to calculate the matching cost of each pixel in multiple images within a preset disparity search range. When calculating the matching cost, the improved multi-view stereo matching algorithm combines farmland edge and texture information extracted from multispectral images to enhance the distinction between farmland and non-farmland areas. The improved multi-view stereo matching algorithm employs a cost aggregation strategy, which performs adaptive weight aggregation on the matching cost of each pixel within the spatial neighborhood. The adaptive weight is dynamically adjusted based on the color similarity and spatial distance between pixels. In the aggregated cost space, the optimal cost selection strategy is used to select the disparity value corresponding to the minimum aggregate cost for each pixel to form an initial dense disparity map. Based on the initial dense parallax map and the imaging parameters of the remote sensing sensor, the three-dimensional spatial coordinates corresponding to each pixel are calculated using the forward intersection principle to generate the initial digital surface model of farmland.

[0008] As a further aspect of the present invention, the improved multi-view stereo matching algorithm optimizes the calculation and aggregation process of matching costs based on the spatial distribution patterns of farmland crops, including: In the matching cost calculation stage, a smoothness constraint based on the growth pattern of farmland area is introduced, and a penalty term is added to the cost calculation function. The weight of the penalty term decreases when crossing the boundary of different crop types and increases when in the uniform crop area, thereby promoting smooth matching within the area while maintaining clear boundaries. During the cost aggregation stage, an adaptively shaped support window is used. The shape and size of the support window are dynamically adjusted according to the texture richness of the local image and the crop row direction indicated by the farmland surface cover classification map. An expanded-size aggregation window is used in texture-rich areas or along the crop row direction to increase reliability, while a shrunken-size aggregation window is used in texture-sparse areas or perpendicular to the crop row direction to preserve details. By using the smoothness constraints and adaptive shape support windows, the matching and aggregation process becomes more adaptable to the characteristics of orderly crop distribution and obvious row and ridge structure in farmland scenarios, thereby improving the consistency between the initial farmland digital surface model and the actual landform in the crop area.

[0009] As a further aspect of the present invention, crop semantic segmentation and spectral feature extraction processing are performed on the multispectral image to generate a farmland surface cover classification map and a crop feature parameter map, including: The multispectral image is input into a pre-trained deep semantic segmentation network, which outputs a probability distribution map of each pixel belonging to different crop categories or non-crop backgrounds. Threshold segmentation and connected component analysis are applied to the probability distribution map to obtain the farmland surface cover classification map. The farmland surface cover classification map identifies farmland boundaries, spatial distribution areas of different crop types, and bare land and road areas. For each crop type region identified in the farmland surface cover classification map, multi-band spectral reflectance data within the crop type region are extracted from the multispectral image, and a series of standardized vegetation indices and spectral characteristic parameters are calculated. The calculated vegetation indices and spectral feature parameters are organized according to their corresponding crop type regions and spatial locations to generate a crop feature parameter map that associates spatial location with feature parameters.

[0010] As a further aspect of the present invention, the farmland digital surface model, land cover classification map, and crop characteristic parameter map are input into a three-dimensional structure generation model for terrain correction, crop three-dimensional structure inversion, and detail enhancement to generate a refined farmland three-dimensional structure mesh model, including: The three-dimensional structure generation model includes a terrain correction unit, a structure inversion unit, and a detail synthesis unit; The terrain correction unit receives the farmland digital surface model and uses a preset bare land digital elevation model to correct the non-crop elevation part of the farmland area to obtain a terrain-corrected crop canopy height model. The structure inversion unit receives the topographically corrected crop canopy height model and the crop characteristic parameter map. Based on prior knowledge of the growth structure of different crop types, it inverts the three-dimensional skeleton structure and canopy volume density distribution of individual crops from canopy height and spectral characteristics. The detail synthesis unit receives the three-dimensional skeleton structure and canopy volume density distribution, combines them with texture details extracted from high-resolution images, and generates a refined three-dimensional farmland structure mesh model containing the fine geometric structure of crop stems, leaves, and ears through voxel sculpting and surface reconstruction techniques.

[0011] As a further aspect of the present invention, based on prior knowledge of the growth structure of different crop types, the three-dimensional skeletal structure and canopy volume density distribution of individual crops are inverted from canopy height and spectral characteristics, including: A mapping database of crop types and typical three-dimensional growth structures is established. The mapping database stores the typical plant morphology, branching pattern and leaf area density distribution of different crops at key growth stages. Based on the farmland surface cover classification map and the crop characteristic parameter map, determine the crop type and growth stage corresponding to each pixel in the target farmland area; Retrieve typical three-dimensional growth structures corresponding to crop types and growth stages from the mapping database as initial templates; Using the topographically corrected crop canopy height model as a spatial constraint and the vegetation index in the crop characteristic parameter map as a growth state adjustment parameter, the initial template is subjected to three-dimensional spatial deformation and structural parameter scaling. Through iterative optimization, the deformed three-dimensional growth structure achieves the optimal statistical match with the canopy height and spectral characteristics, and finally outputs the optimized three-dimensional skeleton structure and the canopy volume density distribution.

[0012] As a further aspect of the present invention, based on the refined three-dimensional farmland structure mesh model and the crop feature parameter map, a three-dimensional attribute mapping module is invoked to assign corresponding crop physiological attribute parameters to each surface unit of the three-dimensional farmland structure mesh model, thereby constructing a three-dimensional attributed farmland model with attribute information, including: The three-dimensional attribute mapping module establishes the mapping relationship between the spatial position of each surface unit of the refined three-dimensional structure mesh model of farmland and the spatial position in the crop feature parameter map; Based on the mapping relationship, the vegetation index and spectral characteristic parameters corresponding to each surface unit are queried from the crop characteristic parameter map; Using a pre-trained physiological parameter inversion model, the queried vegetation index and spectral feature parameters are used as input to predict the physiological attribute parameters of chlorophyll content, water content, and nitrogen content of the crop part represented by each surface unit. The predicted physiological attribute parameters are used as attribute values ​​and associated with and stored on the surface units corresponding to the refined three-dimensional structure mesh model of farmland, thus forming the three-dimensional attribute model of farmland.

[0013] As a further aspect of the present invention, a pre-trained physiological parameter inversion model is used, with the queried vegetation index and spectral feature parameters as input, to predict the physiological attribute parameters of chlorophyll content, water content, and nitrogen content of the crop part represented by each surface unit, including: Collect a large amount of field measurement data, including measured values ​​of chlorophyll content, water content, and nitrogen content of crop leaves or canopy acquired synchronously with remote sensing images, as well as multispectral data of the corresponding locations; Based on the field measurement data, a machine learning model is trained as the pre-trained physiological parameter inversion model. The input of the machine learning model is vegetation index and spectral feature parameters, and the output is the predicted physiological attribute parameters. During training, the hyperparameters of the machine learning model are optimized using cross-validation, and the accuracy of its prediction results is evaluated using independent validation sets. The pre-trained physiological parameter inversion model, once trained, can output predicted values ​​for chlorophyll content, water content, and nitrogen content based on the input vegetation index and spectral feature parameters.

[0014] As a further aspect of the present invention, the farmland 3D attributed model is simplified and format converted at multiple levels of detail to generate a target farmland 3D model that conforms to a preset 3D visualization standard, including: The geometric mesh of the three-dimensional attributed model of farmland is simplified at multiple levels of detail. Based on the requirements of viewpoint distance or model display clarity, multiple versions of geometric mesh with different complexities are generated. During the simplification process, the physiological attribute parameters associated with the three-dimensional attributed model of farmland are aggregated and resampled at the corresponding scale to ensure that each geometric patch after simplification still has representative attribute values. The simplified geometric mesh, along with its associated attribute parameters, is converted into a universal 3D file format supported by a preset 3D visualization standard. This universal 3D file format can store geometric, texture, and custom attribute information. During the format conversion process, the attribute parameters are standardized and encoded to ensure that they can be correctly parsed and rendered in the 3D visualization environment.

[0015] As a further aspect of the present invention, the present invention also includes a system for constructing a three-dimensional farmland model based on remote sensing imagery. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the method for constructing a three-dimensional farmland model based on remote sensing imagery as described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Based on the spatial distribution patterns of farmland crops, the matching cost calculation and aggregation process of the multi-view stereo matching algorithm is optimized. Matching processing is carried out on stereo image pairs, which breaks away from the constraints of fixed matching operation logic in general scenarios. The image matching operation is completed in accordance with the inherent characteristics of farmland surface vegetation distribution, which weakens the matching deviation caused by general algorithms in densely vegetated areas of farmland and produces a farmland digital surface model that conforms to the real undulation of the land surface.

[0017] Multispectral imagery was processed for crop semantic segmentation and spectral feature extraction, resulting in pixel-level land surface type classification and spectral dimension information analysis, forming a structured farmland land cover classification map and crop feature parameter map. The farmland digital surface model, land cover classification map, and crop feature parameter map were then integrated into a unified 3D structure generation model. Simultaneously, terrain correction, crop 3D structure inversion, and mesh detail enhancement were performed, achieving integrated fusion modeling of terrain elevation data, land cover classification data, and spectral feature data, and improving the geometric refinement of the 3D mesh structure.

[0018] The 3D attribute mapping module matches and binds corresponding crop physiological attribute parameters to each surface unit of the farmland 3D structural mesh model, enabling the 3D mesh geometry to carry the corresponding crop attribute information, forming a farmland 3D attributed model that integrates geometry and attribute information. The model undergoes multi-level detail simplification processing, adjusting the mesh density and structure, and simultaneously completing standardized format conversion to adapt to preset 3D visualization access specifications for different application scenarios and to meet the requirements of different platforms for 3D model layer precision and file format. Attached Figure Description

[0019] Figure 1 This is a state diagram of a method for constructing a three-dimensional farmland model based on remote sensing imagery, as described in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a method for constructing a 3D model of farmland based on remote sensing imagery. (See attached document.) Figure 1 The specific implementation method is as follows: A multi-temporal, multi-angle remote sensing image dataset of the target farmland area is acquired. This dataset includes stereo image pairs, multispectral images, and corresponding remote sensing sensor imaging parameters. An improved multi-view stereo matching algorithm is applied to the stereo image pairs to generate an initial farmland digital surface model. This improved algorithm optimizes the calculation and aggregation process of matching costs based on the spatial distribution patterns of farmland crops. Crop semantic segmentation and spectral feature extraction are performed on the multispectral images to generate a farmland land cover classification map and a crop feature parameter map. The farmland digital surface model, land cover classification map, and crop feature parameter map are input into a 3D structure generation model for terrain correction, crop 3D structure inversion, and detail enhancement, generating a refined 3D farmland structure mesh model. Based on the refined 3D farmland structure mesh model and crop feature parameter map, a 3D attribute mapping module is invoked to assign corresponding crop physiological attribute parameters to each surface unit of the 3D farmland structure mesh model, constructing a 3D attributed farmland model with attribute information. The farmland 3D attribute model is simplified and format converted at multiple levels to generate a target farmland 3D model that conforms to the preset 3D visualization standard.

[0022] Example 1: The stereo image pairs are preprocessed, including radiometric calibration, atmospheric correction, and geometric correction, to obtain corrected standard stereo image pairs. On these standard stereo image pairs, an improved multi-view stereo matching algorithm is used to calculate the matching cost of each pixel across multiple images within a preset disparity search range. This improved algorithm incorporates farmland edge and texture information extracted from multispectral images to enhance the distinction between farmland and non-farmland areas when calculating the matching cost. A cost aggregation strategy based on the improved multi-view stereo matching algorithm is employed to adaptively weight the matching cost of each pixel within its spatial neighborhood. This adaptive weight is dynamically adjusted based on color similarity and spatial distance between pixels. In the aggregated cost space, an optimal cost selection strategy is used to select the disparity value corresponding to the minimum aggregated cost for each pixel, forming an initial dense disparity map. Based on the initial dense disparity map and remote sensing sensor imaging parameters, the three-dimensional spatial coordinates of each pixel are calculated using the forward intersection principle to generate an initial digital surface model of the farmland.

[0023] In the matching cost calculation stage, a smoothness constraint based on the growth pattern of farmland areas is introduced. A penalty term is added to the cost calculation function, with its weight decreasing when crossing the boundaries of different crop types and increasing its weight within uniform crop areas. This promotes smooth matching within regions while maintaining clear boundaries. In the cost aggregation stage, an adaptively shaped support window is used. The shape and size of this support window are dynamically adjusted based on the texture richness of the local image and the crop row direction indicated by the farmland surface cover classification map. An expanded aggregation window is used in texture-rich areas or along the crop row direction to increase reliability, while a contracted aggregation window is used in sparsely textured areas or perpendicular to the crop row direction to preserve detail. Through this smoothness constraint and the adaptively shaped support window, the matching and aggregation process is better adapted to the orderly distribution of crops and the obvious row and ridge structure in farmland scenes, improving the consistency between the initial farmland digital surface model and the actual terrain in crop areas.

[0024] In specific implementations, after acquiring multi-temporal and multi-angle stereo image pairs and multispectral images of the target farmland area, along with corresponding remote sensing sensor imaging parameters, radiometric calibration, atmospheric correction, and geometric correction are performed on the stereo image pairs to obtain corrected standard stereo image pairs. In specific implementations, on the standard stereo image pairs, an improved multi-view stereo matching algorithm is used to calculate the matching cost of each pixel across multiple images within a preset disparity search range. When calculating the matching cost, the improved multi-view stereo matching algorithm incorporates farmland edge and texture information extracted from the multispectral images to enhance the distinction between farmland and non-farmland areas. In some embodiments, during the matching cost calculation stage, a smoothness constraint based on the farmland area's growth pattern is introduced. A penalty term is added to the cost calculation function. The penalty term has a reduced weight when crossing boundaries of different crop types and a increased weight within uniform crop areas, thereby promoting smooth matching within the area while maintaining clear boundaries. The matching cost calculation function is expressed as: in: Represents pixels In disparity value The total matching cost is as follows. This represents the cost of matching data items based on pixel grayscale or color differences. The adjustment coefficient representing the smoothness constraint. This represents a penalty term based on the growth pattern of farmland areas, when the pixel... When located in the boundary area of ​​crop type Take the smaller value, when the pixel point When located within a uniform crop region Take the larger value. In specific implementation, an improved cost aggregation strategy of the multi-view stereo matching algorithm is adopted. The matching cost of each pixel is adaptively weighted and aggregated within the spatial neighborhood. The adaptive weights are dynamically adjusted based on the color similarity and spatial distance between pixels. Optionally, in the cost aggregation stage, an adaptively shaped support window is used. The shape and size of the support window are dynamically adjusted based on the texture richness of the local image and the crop row direction indicated by the farmland cover classification map. An expanded-size aggregation window is used in texture-rich areas or along the crop row direction to increase reliability, while a sparse-size aggregation window is used in texture-sparse areas or perpendicular to the crop row direction to preserve detail.

[0025] In specific implementation, within the aggregated cost space, an optimal cost selection strategy is employed to select the disparity value corresponding to the minimum aggregation cost for each pixel, forming an initial dense disparity map. Based on the initial dense disparity map and remote sensing sensor imaging parameters, the three-dimensional spatial coordinates corresponding to each pixel are calculated using the forward intersection principle, generating an initial farmland digital surface model. It can be understood that through smoothness constraints and adaptive shape support windows, the matching and aggregation process becomes more adaptable to the characteristics of orderly crop distribution and distinct row and ridge structures in farmland scenarios, improving the consistency between the initial farmland digital surface model and the actual terrain in the crop area. In some embodiments, the improved multi-view stereo matching algorithm optimizes the calculation and aggregation process of the matching cost based on the spatial distribution pattern of farmland crops, specifically including the penalty term in the aforementioned matching cost calculation function. The settings and cost aggregation phase allow for adaptive shape support window adjustments.

[0026] Optionally, during adaptive weight aggregation, the weight coefficients... Based on pixels and neighboring pixels Color similarity between and spatial distance Dynamic calculations are performed, and color similarity is measured by the sum of the absolute differences between two pixels across different bands of the multispectral image. It can be understood that when calculating 3D spatial coordinates using the forward intersection principle, the image coordinates of the same ground point on at least two standard stereo image pairs, along with the interior and exterior orientation elements from the remote sensing sensor's imaging parameters, are used to solve for the 3D coordinates of that ground point through collinearity equations.

[0027] Example 2: Multispectral imagery is input into a pre-trained deep semantic segmentation network, which outputs a probability distribution map of each pixel belonging to different crop categories or non-crop backgrounds. Thresholding segmentation and connected component analysis are applied to the probability distribution map to obtain a farmland cover classification map, which identifies farmland boundaries, spatial distribution areas of different crop types, and bare land and road areas. For each crop type region identified in the farmland cover classification map, multi-band spectral reflectance data within the crop type region are extracted from the multispectral imagery, and a series of standardized vegetation indices and spectral feature parameters are calculated. The calculated vegetation indices and spectral feature parameters are organized according to their corresponding crop type regions and spatial locations to generate a crop feature parameter map that associates spatial location with feature parameters.

[0028] In specific implementations, after acquiring multi-temporal and multi-angle multispectral images of the target farmland area, the multispectral images are input into a pre-trained deep semantic segmentation network. The pre-trained deep semantic segmentation network outputs a probability distribution map of each pixel belonging to different crop categories or non-crop backgrounds. In some embodiments, the pre-trained deep semantic segmentation network adopts an encoder-decoder structure. The encoder part uses a residual network to extract multi-level features from the multispectral images, and the decoder part restores the original image resolution through upsampling operations. A flexible maximum function is then used in the output layer to calculate the probability value of each pixel belonging to each category in a preset category set. In specific implementations, threshold segmentation and connected component analysis are applied to the probability distribution map to obtain a farmland surface cover classification map. This farmland surface cover classification map identifies farmland boundaries, spatial distribution areas of different crop types, and bare land and road areas.

[0029] Optionally, during threshold segmentation, for each pixel, if the maximum probability value of belonging to a certain crop category is greater than a preset probability threshold, the pixel is classified as belonging to that crop category; otherwise, it is classified as non-crop background. During connected component analysis, adjacent pixels classified as belonging to the same category are marked as eight-connected regions, and noise regions with an area smaller than a preset area threshold are removed to form a continuous and complete spatial distribution region boundary. In specific implementation, for each crop type region identified in the farmland surface cover classification map, multi-band spectral reflectance data within the crop type region are extracted from multispectral images, and a series of standardized vegetation indices and spectral feature parameters are calculated. One formula for calculating the vegetation index can be understood as follows: in: Represents the normalized difference vegetation index. This represents the reflectance value in the near-infrared band of a multispectral image. This represents the reflectance value of the red band in the multispectral image. In specific implementations, in addition to the normalized difference vegetation index, the ratio vegetation index, difference vegetation index, enhanced vegetation index, and chlorophyll absorption ratio index are also calculated as standardized vegetation indices. Simultaneously, the mean, standard deviation, and first derivative characteristics of the reflectance of each band are extracted from the multispectral reflectance data as spectral feature parameters. Optionally, for areas of the same crop type identified in the farmland cover classification map, spatial division is performed according to a regular grid or crop row direction. Within each block, statistical values ​​of the vegetation index and spectral feature parameters are calculated to reflect the spatial heterogeneity of crop growth. In some embodiments, the calculated vegetation index and spectral feature parameters are organized according to their corresponding crop type areas and spatial locations to generate a crop feature parameter map that associates spatial location with feature parameters. It can be understood that the crop feature parameter map adopts a raster data structure with the same spatial resolution as the multispectral image, where each raster cell stores the crop type identifier of that location and the corresponding vegetation index and spectral feature parameter values. In practice, vegetation index and spectral characteristic parameters are not calculated for bare land and road areas in the crop characteristic parameter map. The characteristic parameters corresponding to these areas are stored as null values.

[0030] Example 3: The 3D structure generation model comprises a terrain correction unit, a structure inversion unit, and a detail synthesis unit. The terrain correction unit receives a farmland digital surface model and uses a pre-defined bare land digital elevation model to correct the non-crop elevation portions of the farmland area, resulting in a terrain-corrected crop canopy height model. The structure inversion unit receives the terrain-corrected crop canopy height model and crop characteristic parameter maps. Based on prior knowledge of the growth structure of different crop types, it inverts the 3D skeleton structure and canopy volume density distribution of individual crops from canopy height and spectral characteristics. The detail synthesis unit receives the 3D skeleton structure and canopy volume density distribution, and combines texture details extracted from high-resolution imagery. Through voxel sculpting and surface reconstruction techniques, it generates a refined 3D farmland structure mesh model containing the fine geometric structure of crop stems, leaves, and ears.

[0031] A mapping database of crop types and typical three-dimensional growth structures was established, storing typical plant morphology, branching patterns, and leaf area density distributions of different crops at key growth stages. Based on farmland cover classification maps and crop characteristic parameter maps, the crop type and growth stage corresponding to each pixel within the target farmland area were determined. Typical three-dimensional growth structures for the corresponding crop type and growth stage were retrieved from the mapping database as initial templates. Using a topographically corrected crop canopy height model as a spatial constraint and vegetation indices from the crop characteristic parameter maps as growth state adjustment parameters, the initial templates were subjected to three-dimensional spatial deformation and structural parameter scaling. Through iterative optimization, the deformed three-dimensional growth structure achieved a statistically optimal match with canopy height and spectral characteristics, ultimately outputting the optimized three-dimensional skeleton structure and canopy volume density distribution.

[0032] In specific implementation, the 3D structure generation model includes a terrain correction unit, a structure inversion unit, and a detail synthesis unit. The terrain correction unit receives a farmland digital surface model and uses a pre-defined bare land digital elevation model to correct the non-crop elevation portion of the farmland area, obtaining a terrain-corrected crop canopy height model. In some embodiments, the pre-defined bare land digital elevation model is generated by acquiring remote sensing stereo image pairs during periods without crop cover, or by extracting bare land elevation information from existing regional digital elevation model data. In specific implementation, the structure inversion unit receives the terrain-corrected crop canopy height model and crop characteristic parameter maps. Based on prior knowledge of the growth structure of different crop types, it inverts the 3D skeleton structure and canopy volume density distribution of individual crops from canopy height and spectral characteristics. Optionally, a mapping relationship library between crop types and typical 3D growth structures is established. This library stores typical plant morphology, branching patterns, and leaf area density distributions of different crops at key growth stages. See Table 1 for the mapping relationship library.

[0033] Table 1: Mapping Relationship between Crop Type and Typical Three-Dimensional Growth Structure In practice, based on farmland surface cover classification maps and crop characteristic parameter maps, the crop type and growth stage corresponding to each pixel within the target farmland area are determined. Typical three-dimensional growth structures corresponding to the crop type and growth stage are retrieved from the mapping database as initial templates. These initial templates can be understood as containing the three-dimensional skeletal topology of the plant and the relative spatial relationships of its organs. In practice, using the topographically corrected crop canopy height model as a spatial constraint and the vegetation index from the crop characteristic parameter map as a growth state adjustment parameter, the initial template undergoes three-dimensional spatial deformation and structural parameter scaling. The structural parameter scaling process adjusts the plant height and canopy radius using the following formula: in: This indicates the adjusted plant height. This represents the standard plant height in the initial template. This represents the canopy height value at the corresponding pixel location in the crop canopy height model. This represents the reference canopy height value corresponding to the initial template. This indicates the adjusted canopy radius. This represents the standard canopy radius in the initial template. This represents the vegetation index value in the crop characteristic parameter map. This represents the reference vegetation index value corresponding to the initial template. This represents the growth state regulation index. In some embodiments, through iterative optimization, the deformed 3D growth structure is statistically optimally matched with canopy height and spectral characteristics, outputting the optimized 3D skeleton structure and canopy volume density distribution. Optionally, the iterative optimization employs a gradient descent method, with the objective function simultaneously measuring the difference between the simulated canopy height generated by the projection of the 3D skeleton structure and the actual crop canopy height model, as well as the difference between the simulated spectral characteristics and the vegetation indices in the actual crop characteristic parameter map. It can be understood that the canopy volume density distribution is stored in a voxel grid format, with each voxel storing the leaf area density value at that location. In a specific implementation, the detail synthesis unit receives the 3D skeleton structure and canopy volume density distribution, combines it with texture details extracted from high-resolution images, and generates a refined 3D farmland structure mesh model containing the fine geometric structure of crop stems, leaves, and ears through voxel sculpting and surface reconstruction techniques. Optionally, during the voxel sculpting process, the voxels are sculpted in three-dimensional space with the canopy volume density distribution as a constraint, and voxel regions with density values ​​higher than the threshold are retained; the surface reconstruction technology uses the moving cube algorithm to extract isosurface meshes from the voxel data and maps the texture of the corresponding region in the high-resolution image onto the mesh surface.

[0034] Example 4: The 3D attribute mapping module establishes a mapping relationship between the spatial location of each surface unit in the refined 3D farmland structure mesh model and its spatial location in the crop feature parameter map. Based on this mapping relationship, the vegetation index and spectral feature parameters corresponding to each surface unit are retrieved from the crop feature parameter map. Using a pre-trained physiological parameter inversion model, the retrieved vegetation index and spectral feature parameters are used as input to predict the chlorophyll content, water content, and nitrogen content physiological attribute parameters of the crop part represented by each surface unit. The predicted physiological attribute parameters are used as attribute values ​​and associatedly stored on the corresponding surface unit of the refined 3D farmland structure mesh model, forming a 3D attributed farmland model.

[0035] Extensive field data was collected, including measured values ​​of chlorophyll content, moisture content, and nitrogen content in crop leaves or canopies acquired synchronously with remote sensing imagery, along with corresponding multispectral data. Based on this field data, a machine learning model was trained as a pre-trained physiological parameter inversion model. The input to this machine learning model was vegetation indices and spectral feature parameters, and the output was the predicted physiological attribute parameters. During training, cross-validation was used to optimize the hyperparameters of the machine learning model, and independent validation sets were used to evaluate the accuracy of its predictions. The completed pre-trained physiological parameter inversion model can output predicted values ​​of chlorophyll content, moisture content, and nitrogen content based on the input vegetation indices and spectral feature parameters.

[0036] In specific implementations, the 3D attribute mapping module establishes a mapping relationship between the spatial location of each surface unit in the refined 3D farmland structure mesh model and its spatial location in the crop characteristic parameter map. In some embodiments, spatial location mapping involves projecting the vertex coordinates of the surface units in the refined 3D farmland structure mesh model onto the geographic reference plane coordinate system where the crop characteristic parameter map is located, and locating the corresponding raster unit index in the crop characteristic parameter map based on the projected 2D coordinates. In specific implementations, based on the mapping relationship, the vegetation index and spectral characteristic parameters corresponding to each surface unit are queried from the crop characteristic parameter map. Optionally, when a surface unit, after projection, covers multiple raster units in the crop characteristic parameter map, the vegetation index and spectral characteristic parameter values ​​corresponding to that surface unit are calculated using a bilinear interpolation method. In specific implementations, using a pre-trained physiological parameter inversion model, the queried vegetation index and spectral characteristic parameters are used as input to predict the chlorophyll content, water content, and nitrogen content physiological attribute parameters of the crop part represented by each surface unit. It is understandable that the physiological parameter inversion model adopts a gradient boosting regression tree structure. The input feature vector includes red band reflectance, near-infrared band reflectance, normalized difference vegetation index, difference vegetation index, ratio vegetation index, and spectral first derivative features. The pre-trained physiological parameter inversion model is shown in Table 2.

[0037] Table 2: Correspondence between input features and output parameters of the pre-trained physiological parameter inversion model In practice, a large amount of field measurement data is collected. This data includes measured values ​​of chlorophyll content, moisture content, and nitrogen content in crop leaves or canopies, acquired synchronously with remote sensing images, as well as corresponding multispectral data. In some embodiments, field measurement data collection employs a grid sampling method, with sampling points evenly distributed across the target farmland area. At each sampling point, a handheld chlorophyll meter is used to measure leaf chlorophyll content, a drying and weighing method is used to measure leaf moisture content, and the Kjeldahl method is used to measure leaf nitrogen content. Simultaneously, the coordinates of the sampling points are recorded to match the reflectance values ​​at corresponding locations in the multispectral image. Optionally, the field measurement data is divided into a training set and a validation set in a 7:3 ratio. The training set is used to train the machine learning model, and the validation set is used to evaluate the model's predictive performance. In practice, based on the field measurement data, a machine learning model is trained as a pre-trained physiological parameter inversion model. The input to the machine learning model is vegetation index and spectral feature parameters, and the output is the predicted physiological attribute parameters. It can be understood that for the prediction of chlorophyll content, the pre-trained physiological parameter inversion model internally adopts the following linear combination form: in: This indicates the predicted chlorophyll content, expressed in micrograms per square centimeter. This represents the normalized difference vegetation index, which is a dimensionless value. This represents the reflectivity in the red light band and is a dimensionless value. This represents the reflectance in the near-infrared band and is a dimensionless value. , , The weight coefficients obtained from the model learning are represented in micrograms per square centimeter. The term represents the bias term, also expressed in micrograms per square centimeter. In practice, during training, cross-validation is used to optimize the hyperparameters of the machine learning model, and an independent validation set is used to evaluate the accuracy of its predictions. Optionally, a five-fold cross-validation strategy is used, dividing the training set into five equal parts, alternating between four parts as training subsets and one as a validation subset, and selecting the hyperparameter combination with the smallest average validation error after iterating through all combinations as the optimal hyperparameters. In practice, the pre-trained physiological parameter inversion model can output predicted values ​​for chlorophyll content, water content, and nitrogen content based on the input vegetation index and spectral feature parameters. In practice, the predicted physiological attribute parameters are used as attribute values ​​and associated with and stored on the corresponding surface units of the refined three-dimensional farmland structure mesh model, forming a three-dimensional attributed farmland model. In some embodiments, each surface unit of the refined three-dimensional farmland structure mesh model contains a vertex array and a patch index array. The association of attribute values ​​is achieved by adding attribute fields to the patch data structure, with each patch field storing chlorophyll content, water content, and nitrogen content values ​​respectively. It is understandable that the attribute values ​​are stored in floating-point form, retaining two decimal places of precision.

[0038] Example 5: The geometric mesh of the 3D attributed model of farmland is simplified at multiple levels of detail. Multiple versions of the geometric mesh with varying complexity are generated based on viewpoint distance or model display clarity requirements. During the simplification process, the physiological attribute parameters associated with the 3D attributed model of farmland are aggregated and resampled at appropriate scales to ensure that each simplified geometric patch still has representative attribute values. The simplified geometric mesh, along with its associated attribute parameters, is converted into a universal 3D file format supported by a preset 3D visualization standard. This universal 3D file format can store geometric, texture, and custom attribute information. During the format conversion process, the attribute parameters are standardized and encoded to ensure they can be correctly parsed and rendered in the 3D visualization environment.

[0039] In specific implementation, the geometric mesh of the 3D attributed model of farmland undergoes multi-level detail simplification. Multiple geometric mesh versions with varying complexities are generated based on viewpoint distance or model display clarity requirements. In some embodiments, the multi-level detail simplification employs an edge-folding simplification algorithm. For the geometric mesh of the 3D attributed model of farmland, vertices are sorted according to the magnitude of their geometric errors. The vertex pairs with the smallest errors are merged into a single vertex, while simultaneously updating the connectivity of adjacent faces. This process is repeated until a simplified version with the target number of faces is generated. In specific implementation, during the simplification process, the physiological attribute parameters associated with the 3D attributed model of farmland are aggregated and resampled at appropriate scales to ensure that each simplified geometric face still has representative attribute values. Optionally, when multiple original faces are folded into a simplified face, the aggregation of physiological attribute parameters uses a weighted average method. The weights are determined based on the ratio of the area of ​​each original face to the total area of ​​the simplified face. The aggregated attribute values ​​are assigned to the simplified geometric face. It can be understood that independent aggregation calculations are performed for the three physiological attribute parameters—chlorophyll content, water content, and nitrogen content—while maintaining the physical units of each attribute parameter. In practice, the simplified geometric mesh, along with its associated attribute parameters, is converted into a universal 3D file format supported by a preset 3D visualization standard. This universal 3D file format can store geometric, texture, and custom attribute information. In some embodiments, the preset 3D visualization standard is the WebGL standard, and the universal 3D file format uses the glTF format. The glTF format includes a node hierarchy, mesh data, material definitions, and an extended attribute storage area. In practice, during the format conversion process, the attribute parameters are standardized and encoded to ensure they can be correctly parsed and rendered in the 3D visualization environment. One form of attribute parameter standardization encoding is: in: This represents the standardized encoded attribute value, with a value range between 0 and 1. Indicates the original physiological attribute parameter values; This represents the minimum value of the same physiological attribute parameter across all surface units in the entire farmland three-dimensional attributed model; This represents the maximum value of the same physiological attribute parameter across all surface units in the entire farmland 3D attributed model. Optionally, the standardized attribute values ​​are stored as four-byte floating-point numbers in an extended attribute storage area in glTF format, with each physiological attribute parameter corresponding to an independent attribute channel. It can be understood that the 3D visualization environment parses the extended attribute storage area in the glTF file, reads the standardized attribute values ​​of each vertex or facet, and maps them back to the attribute values ​​of the original physical units for rendering or analysis as needed. In practice, for different display clarity requirements, corresponding multi-level-of-detail versions of the glTF file are generated, and different versions are dynamically switched and loaded in the 3D visualization environment based on the viewpoint distance.

[0040] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for constructing a three-dimensional model of farmland based on remote sensing imagery, characterized in that, include: Acquire a multi-temporal, multi-angle remote sensing image data set of the target farmland area, the remote sensing image data set including stereo image pairs, multispectral images and corresponding remote sensing sensor imaging parameters; An improved multi-view stereo matching algorithm is applied to the stereo image pairs to generate an initial digital surface model of farmland. The improved multi-view stereo matching algorithm optimizes the calculation and aggregation process of matching costs based on the spatial distribution pattern of farmland crops. Perform crop semantic segmentation and spectral feature extraction processing on the multispectral images to generate farmland surface cover classification maps and crop feature parameter maps; The farmland digital surface model, land cover classification map and crop feature parameter map are input into the three-dimensional structure generation model to perform terrain correction, crop three-dimensional structure inversion and detail enhancement, and generate a refined farmland three-dimensional structure mesh model. Based on the refined three-dimensional structure mesh model of farmland and the crop feature parameter map, the three-dimensional attribute mapping module is called to assign corresponding crop physiological attribute parameters to each surface unit of the three-dimensional structure mesh model of farmland, thereby constructing a three-dimensional attributed model of farmland with attribute information. The farmland 3D attribute model is simplified and format converted at multiple levels of detail to generate a target farmland 3D model that conforms to the preset 3D visualization standard. An improved multi-view stereo matching algorithm is applied to the stereo image pairs to generate an initial digital surface model of farmland, including: The stereo image pair image is preprocessed, including radiometric calibration, atmospheric correction and geometric correction, to obtain a corrected standard stereo image pair image. On the standard stereo image pair, the improved multi-view stereo matching algorithm is used to calculate the matching cost of each pixel in multiple images within a preset disparity search range. When calculating the matching cost, the improved multi-view stereo matching algorithm combines farmland edge and texture information extracted from multispectral images to enhance the distinction between farmland and non-farmland areas. The improved multi-view stereo matching algorithm employs a cost aggregation strategy, which performs adaptive weight aggregation on the matching cost of each pixel within the spatial neighborhood. The adaptive weight is dynamically adjusted based on the color similarity and spatial distance between pixels. In the aggregated cost space, the optimal cost selection strategy is used to select the disparity value corresponding to the minimum aggregate cost for each pixel to form an initial dense disparity map. Based on the initial dense parallax map and the imaging parameters of the remote sensing sensor, the three-dimensional spatial coordinates corresponding to each pixel are calculated using the forward intersection principle to generate the initial digital surface model of farmland. The improved multi-view stereo matching algorithm optimizes the calculation and aggregation process of matching costs based on the spatial distribution patterns of farmland crops, including: In the matching cost calculation stage, a smoothness constraint based on the growth pattern of farmland area is introduced, and a penalty term is added to the cost calculation function. The weight of the penalty term decreases when crossing the boundary of different crop types and increases when in the uniform crop area, thereby promoting smooth matching within the area while maintaining clear boundaries. During the cost aggregation stage, an adaptively shaped support window is used. The shape and size of the support window are dynamically adjusted according to the texture richness of the local image and the crop row direction indicated by the farmland surface cover classification map. An expanded-size aggregation window is used in texture rich areas or along the crop row direction to increase reliability, while a shrunken-size aggregation window is used in texture sparse areas or perpendicular to the crop row direction to preserve details. By using the smoothness constraints and adaptive shape support windows, the matching and aggregation process becomes more adaptable to the characteristics of orderly crop distribution and obvious row and ridge structure in farmland scenarios, thereby improving the consistency between the initial farmland digital surface model and the actual landform in the crop area.

2. The method for constructing a three-dimensional farmland model based on remote sensing imagery according to claim 1, characterized in that, The multispectral image is subjected to crop semantic segmentation and spectral feature extraction processing to generate a farmland surface cover classification map and a crop feature parameter map, including: The multispectral image is input into a pre-trained deep semantic segmentation network, which outputs a probability distribution map of each pixel belonging to different crop categories or non-crop backgrounds. Threshold segmentation and connected component analysis are applied to the probability distribution map to obtain the farmland surface cover classification map. The farmland surface cover classification map identifies farmland boundaries, spatial distribution areas of different crop types, and bare land and road areas. For each crop type region identified in the farmland surface cover classification map, multi-band spectral reflectance data within the crop type region are extracted from the multispectral image, and a series of standardized vegetation indices and spectral characteristic parameters are calculated. The calculated vegetation indices and spectral feature parameters are organized according to their corresponding crop type regions and spatial locations to generate a crop feature parameter map that associates spatial location with feature parameters.

3. The method for constructing a three-dimensional farmland model based on remote sensing imagery according to claim 1, characterized in that, The farmland digital surface model, land cover classification map, and crop feature parameter map are input into the 3D structure generation model. Terrain correction, crop 3D structure inversion, and detail enhancement are performed to generate a refined 3D farmland structure mesh model, including: The three-dimensional structure generation model includes a terrain correction unit, a structure inversion unit, and a detail synthesis unit; The terrain correction unit receives the farmland digital surface model and uses a preset bare land digital elevation model to correct the non-crop elevation part of the farmland area to obtain a terrain-corrected crop canopy height model. The structure inversion unit receives the topographically corrected crop canopy height model and the crop characteristic parameter map. Based on prior knowledge of the growth structure of different crop types, it inverts the three-dimensional skeleton structure and canopy volume density distribution of individual crops from canopy height and spectral characteristics. The detail synthesis unit receives the three-dimensional skeleton structure and canopy volume density distribution, combines them with texture details extracted from high-resolution images, and generates a refined three-dimensional farmland structure mesh model containing the fine geometric structure of crop stems, leaves, and ears through voxel sculpting and surface reconstruction techniques.

4. The method for constructing a three-dimensional farmland model based on remote sensing imagery according to claim 3, characterized in that, Based on prior knowledge of the growth structure of different crop types, the three-dimensional skeletal structure and canopy volume density distribution of individual crops are inverted from canopy height and spectral characteristics, including: A mapping database of crop types and typical three-dimensional growth structures is established. The mapping database stores the typical plant morphology, branching pattern and leaf area density distribution of different crops at key growth stages. Based on the farmland surface cover classification map and the crop characteristic parameter map, determine the crop type and growth stage corresponding to each pixel in the target farmland area; Retrieve typical three-dimensional growth structures corresponding to crop types and growth stages from the mapping database as initial templates; Using the topographically corrected crop canopy height model as a spatial constraint and the vegetation index in the crop characteristic parameter map as a growth state adjustment parameter, the initial template is subjected to three-dimensional spatial deformation and structural parameter scaling. Through iterative optimization, the deformed three-dimensional growth structure achieves the optimal statistical match with the canopy height and spectral characteristics, and finally outputs the optimized three-dimensional skeleton structure and the canopy volume density distribution.

5. The method for constructing a three-dimensional farmland model based on remote sensing imagery according to claim 1, characterized in that, Based on the refined three-dimensional farmland structure mesh model and the crop feature parameter map, the three-dimensional attribute mapping module is invoked to assign corresponding crop physiological attribute parameters to each surface unit of the three-dimensional farmland structure mesh model, thereby constructing a three-dimensional attributed farmland model with attribute information, including: The three-dimensional attribute mapping module establishes the mapping relationship between the spatial position of each surface unit of the refined three-dimensional structure mesh model of farmland and the spatial position in the crop feature parameter map; Based on the mapping relationship, the vegetation index and spectral characteristic parameters corresponding to each surface unit are queried from the crop characteristic parameter map; Using a pre-trained physiological parameter inversion model, the queried vegetation index and spectral feature parameters are used as input to predict the chlorophyll content, water content, and nitrogen content physiological attribute parameters of the crop part represented by each surface unit. The predicted physiological attribute parameters are used as attribute values ​​and associated with and stored on the surface units corresponding to the refined three-dimensional structure mesh model of farmland, thus forming the three-dimensional attribute model of farmland.

6. The method for constructing a three-dimensional farmland model based on remote sensing imagery according to claim 5, characterized in that, Using a pre-trained physiological parameter inversion model, and taking the queried vegetation index and spectral feature parameters as input, the model predicts the physiological attribute parameters of chlorophyll content, water content, and nitrogen content of the crop part represented by each surface unit, including: A large amount of field measurement data was collected, including measured values ​​of chlorophyll content, water content, and nitrogen content of crop leaves or canopy acquired synchronously with remote sensing images, as well as multispectral data of the corresponding locations. Based on the field measurement data, a machine learning model is trained as the pre-trained physiological parameter inversion model. The input of the machine learning model is vegetation index and spectral feature parameters, and the output is the predicted physiological attribute parameters. During training, the hyperparameters of the machine learning model are optimized using cross-validation, and the accuracy of its prediction results is evaluated using independent validation sets. The pre-trained physiological parameter inversion model, once trained, can output predicted values ​​for chlorophyll content, water content, and nitrogen content based on the input vegetation index and spectral feature parameters.

7. The method for constructing a three-dimensional farmland model based on remote sensing imagery according to claim 1, characterized in that, Perform multi-level detail simplification and format conversion on the farmland 3D attributed model to generate a target farmland 3D model that conforms to a preset 3D visualization standard, including: The geometric mesh of the three-dimensional attributed model of farmland is simplified at multiple levels of detail. Based on the requirements of viewpoint distance or model display clarity, multiple versions of geometric mesh with different complexities are generated. During the simplification process, the physiological attribute parameters associated with the three-dimensional attributed model of farmland are aggregated and resampled at the corresponding scale to ensure that each geometric patch after simplification still has representative attribute values. The simplified geometric mesh, along with its associated attribute parameters, is converted into a universal 3D file format supported by a preset 3D visualization standard. This universal 3D file format can store geometric, texture, and custom attribute information. During the format conversion process, the attribute parameters are standardized and encoded to ensure that they can be correctly parsed and rendered in the 3D visualization environment.

8. A system for constructing a 3D model of farmland based on remote sensing imagery, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for constructing a three-dimensional farmland model based on remote sensing imagery as described in any one of claims 1 to 7.

Images