Building three-dimensional modeling method based on single recent satellite image
Through a 3D building modeling method based on a single recent satellite image, using parameterized adaptive matching and semantically driven Gaussian element representation technology, the problems of high data acquisition cost and low modeling accuracy in existing technologies are solved, and efficient and accurate building 3D model updates are achieved, which is suitable for smart cities and disaster emergency response.
Patent Information
- Application Number
- CN202510803234.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-16
- Publication Date
- 2025-10-03
AI Technical Summary
Existing technologies rely on multi-perspective images and multi-source data fusion for 3D building modeling, which has high data acquisition costs, complex processing procedures, and poor update timeliness. It is difficult to achieve high-precision modeling through a single recent satellite image, and there is a lack of effective integration of building semantic attributes and temporal evolution relationships, resulting in delayed model updates and insufficient accuracy, making it difficult to meet the dynamic update needs of smart cities and disaster emergency response.
A 3D building modeling method based on a single recent satellite image is adopted. By obtaining a single recent satellite image of the building change area and comparing it with the historical 3D model, the type of change is determined. A parameterized adaptive matching mechanism and semantic-driven Gaussian element representation technology are used, combined with a satellite imaging model and a Levenberg-Marquardt algorithm with adaptive step-size control, to initialize and iteratively update the model. The semantic information and geometric constraints are integrated to achieve the diffusion and deviation reshaping of the Gaussian element set. Finally, the updated model is generated through isosurface extraction.
It enables rapid iterative updating of building three-dimensional information with only a single recent satellite image, reduces data acquisition costs, significantly improves modeling accuracy and robustness, and is suitable for emergency response and high-frequency update scenarios. It effectively integrates geometric, semantic, and temporal information to meet the dynamic update needs of smart cities and disaster emergency response.
Smart Images
Figure CN120747348A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of three-dimensional modeling and remote sensing information processing, and in particular relates to a three-dimensional building modeling method based on a single recent satellite image. Background Art
[0002] With the rapid development of satellite remote sensing technology, high-resolution satellite imagery has become a crucial data resource for urban 3D modeling and dynamic monitoring. However, traditional 3D building modeling methods rely on the fusion of multiple data sources, such as multi-view imagery, LiDAR, or oblique photography. This leads to high data acquisition costs, complex processing, and poor update timeliness. Given the rapid pace of urban change, with frequent building construction, demolition, and renovation, existing methods struggle to achieve high-precision modeling using only a single, recent satellite image. Furthermore, they lack effective integration of building semantic attributes and temporal evolution, resulting in delayed model updates, missing geometric details, and semantic mismatches. Furthermore, existing modeling techniques based on single satellite images often employ depth estimation methods. Due to variations in satellite imaging conditions, these methods struggle to capture building details, resulting in limited accuracy. This makes them unable to meet the urgent demands for dynamic, precise, and automated 3D model updates in smart cities, disaster response, and other fields. Therefore, a 3D building modeling method that can dynamically iterate from a single satellite image by integrating semantic information is urgently needed to improve the efficiency and reliability of digital urban management. Summary of the Invention
[0003] The present invention aims to solve the problems of low accuracy, low efficiency and difficulty in fusing multi-dimensional information when using satellite images to model buildings in the existing technology, and proposes a three-dimensional building modeling method based on a single recent satellite image.
[0004] A method for three-dimensional building modeling based on a single recent satellite image comprises the following steps: Step 1: Obtain a single recent satellite image of the area where the building has changed. Compare the single recent satellite image with the historical 3D model of the building to determine the type of change to the building. Global changes include new construction, demolition, and entire replacement. Local changes include partial reconstruction, facade repair, and construction.
[0005] New construction: new structures with a closed outline area ≥ 90% of the standard building base outline area and a building height ≥ 3 meters; Demolition: The outline area attenuation rate of the building unit is ≥ 95%, or the building height is ≤ 5% of the original height; Overall replacement: The change rate of the building base outline area is ≤10%, the change of the building height is ≥30%, and the modification of the load-bearing structure is involved; Partial reconstruction: 5% < building base outline area change rate ≤ 20%, and 5% < building height change ≤ 15%; External facade renovation: building base outline area change rate ≤ 5%, building height change ≤ 5%, surface texture or structure update; Construction: There are temporary structures or unclosed areas with an outline closure rate of ≤70% within the building base outline, and the building height change is ≤10%.
[0006] If the change type of the building unit is an overall change, an initialization building unit model is constructed and the initialization building unit model is used as a sampling model; If the change type of the building unit is local change, the historical three-dimensional model is used as the sampling model.
[0007] The process of constructing and initializing the building monomer model is as follows: Aiming at the overall change area of the building unit, a three-dimensional model of the building unit is constructed based on the satellite imaging model through a parameterized adaptive matching mechanism. The loss between the image projected onto the plane and the recent satellite image is used as the objective function. The iteration is stopped when the objective function reaches the minimum, and the three-dimensional model of the building unit is used as the initial building unit model.
[0008] Among them, the objective function is: ; is the building model parameter, To parameterize the initial model, is the rendering function that projects the parameterized model onto the image plane, For historical moments, For the current moment, is the semantic change mask image, is the optical image at the current moment, the regularization term Physical rationality of constraint parameters, is the balance coefficient.
[0009] The satellite imaging model is based on a rational polynomial model and combines the satellite orbit altitude and imaging angle.
[0010] The adaptive matching mechanism adopts the improved Levenberg-Marquardt algorithm and introduces adaptive step size control to achieve optimal initialization fitting of the model, which is expressed as: ; in, is the Jacobian matrix, through calculate, The rendering function according to claim 3, r is the residual vector, which represents the difference between the model prediction value and the actual observation value, Dynamically adjust according to the residual decrease rate to avoid local minima.
[0011] Step 2: Sample the sampling model and perform semantically driven Gaussian meta-representation on the sampling points , forming a Gaussian element set .
[0012] The semantically driven Gaussian meta-representation is: ; in, For any point in space, is the position of the Gaussian element in the three-dimensional model, Describe the shape and direction of the Gaussian element, For color, band is the number of current satellite image bands, is the opacity, which controls the visibility of the Gaussian element. It is a semantic driving function, which includes building attribute layer, time evolution layer and Gaussian activation layer.
[0013] The Gaussian element set is: ; in, is the position of the Gaussian element in the three-dimensional model, is a highly noisy term, i.e., characterizing the deviation, is a Gaussian noise distribution with a standard deviation Related to image resolution, Describe the shape and direction of the Gaussian element, For color, band is the number of current satellite image bands, is the opacity, which controls the visibility of the Gaussian element. It is a semantically driven function.
[0014] Step 3: Based on the historical 3D model height information of the building and recent satellite images, use the depthanything or marigold algorithm to estimate the height of the building's changed area, and inject Gaussian elements based on the estimated height to characterize the height position orientation deviation. .
[0015] Step 4: Set the maximum height change according to the building scene h , combined with semantic constraints, the height position orientation representation deviation in step 3 is subjected to Gaussian diffusion in the Z-axis direction to achieve height-semantic collaborative reasoning, and the height point set after Gaussian diffusion is obtained : ; in, The height estimated based on the historical 3D model height information and satellite images in step 3, is the sensitivity coefficient, is a highly noisy term, i.e., characterizing the deviation, is a Gaussian noise distribution with a standard deviation Related to image resolution, Constrain the height variation range.
[0016] Step 5: The height point set described in step 4 As a Gaussian element representation The z coordinate of the sampling point x in the middle is reshaped using the space-time diffusion equation until the geometric constraint converges to the set value and the iteration is completed to obtain the Gaussian element set after the deviation reshaping .
[0017] The iterative process is: ; in, To characterize bias injection, k Indicates the k iterations, Reshape for deviation, is the geometrically degenerate tensor, and represents the image plane diffusion intensity, Indicates the diffusion intensity in the vertical direction of elevation, is a Gaussian diffusion process.
[0018] The geometric constraints are: ; in, is the three-dimensional Gaussian element rendering function, For recent satellite images, For rendering consistency, For semantic alignment, is the weight coefficient of semantic alignment.
[0019] Step 6: Set the Gaussian element set described in step 5 Perform isosurface extraction to obtain the updated building monomer model.
[0020] Step 7: Replace the historical 3D model of the corresponding area with the updated single building model described in step 6, and output the spatiotemporal dynamic description through the decoding network.
[0021] The beneficial effects of the present invention are: The present invention is based on a historical three-dimensional model or an adaptive parameterized building unit initialization model. It estimates the height of the building unit's changed area based on a single recent satellite image, injects deviations based on the estimated height of the changed area, and iterates by fusing semantic drive and geometric constraints through spatiotemporal diffusion technology to complete the deviation reshaping. The Gaussian element set after the deviation reshaping is subjected to isosurface extraction to complete the building unit modeling. The present invention only requires a single recent satellite image combined with a historical three-dimensional model to complete the modeling, significantly reducing data acquisition costs. It can quickly iterate and update the three-dimensional building information, making it suitable for emergency response or high-frequency update scenarios. It can also effectively integrate the geometric, semantic, and temporal information of satellite images, significantly improving modeling accuracy. The adaptive parameter matching mechanism and iterative process enhance the robustness of the model, and the multi-objective collaborative constraints enhance the reliability of the model. This solves the problems of low accuracy, low efficiency, and difficulty in integrating multiple information when using satellite images for building modeling in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 is a flow chart of the present invention; Figure 2 It is a design schematic diagram of the present invention. DETAILED DESCRIPTION Specific implementation method one: This embodiment is a method for 3D building modeling based on a single recent satellite image, comprising the following steps: Step 1: Obtain a single recent satellite image of the area where the building has changed. Use GIS tools to compare the single recent satellite image with the historical 3D model of the building to determine the type of change in the building. If the change type of the building unit is an overall change, including new construction, demolition and overall replacement, then an initial building unit model is constructed and the initial building unit model is used as the sampling model; If the type of change in a building unit is a local change, including local reconstruction, facade repair and construction, the historical three-dimensional model is used as the sampling model.
[0024] The specific criteria for new construction, demolition, and overall replacement are as follows: New construction: new structures with a closed outline area ≥ 90% of the standard building base outline area and a building height ≥ 3 meters; Demolition: The outline area attenuation rate of the building unit is ≥ 95%, or the building height is ≤ 5% of the original height; Overall replacement: The change rate of the building base outline area is ≤10%, the change of the building height is ≥30%, and the modification of the load-bearing structure is involved; Partial reconstruction: 5% < building base outline area change rate ≤ 20%, and 5% < building height change ≤ 15%; External facade renovation: building base outline area change rate ≤ 5%, building height change ≤ 5%, surface texture or structure update; Construction: There are temporary structures or unclosed areas with an outline closure rate of ≤70% within the building base outline, and the building height change is ≤10%.
[0025] The process of constructing the initial building monomer model is as follows: The satellite imaging model is based on the rational polynomial model, which is obtained by combining the satellite orbit altitude and imaging angle. The satellite imaging model determines the rendering function. The orbit altitude is 631km, the imaging angle is 97.9°, and the regularization term weight is =0.1.
[0026] For the overall change area of the building unit, a parameterized adaptive matching mechanism is used to construct a 3D model of the building unit based on the satellite imaging model. The loss between the image projected onto the plane and the recent satellite image is used as the objective function. The iteration is stopped when the objective function reaches the minimum. The 3D model of the building unit is used as the initial building unit model. Among them, the objective function is: ; is the building model parameter, To parameterize the initial model, is the rendering function that projects the parameterized model onto the image plane, For historical moments, For the current moment, is the semantic change mask image, is the optical image at the current moment, the regularization term Physical rationality of constraint parameters, is the balance coefficient.
[0027] The adaptive matching mechanism adopts the improved Levenberg-Marquardt algorithm and introduces adaptive step size control to achieve optimal initialization fitting of the model, which is expressed as: ; in, is the Jacobian matrix, through calculate, The rendering function of claim 3, r is a residual vector representing the difference between the model prediction value and the actual observation value, Dynamically adjust according to the residual decrease rate to avoid local minima.
[0028] Step 2: Sample the initialized building model or historical 3D model, uniformly sample 1000 points, and represent the sampled point x as a semantically driven Gaussian element representation: ; in, For any point in space, is the position of the Gaussian element in the three-dimensional model, Describe the shape and direction of the Gaussian element, For color, band is the number of current satellite image bands, is the opacity, which controls the visibility of the Gaussian element. It is a semantically driven function, consisting of a building attribute layer, a time evolution layer, and a Gaussian activation layer. The building attribute layer constrains the rationality of building reasoning; the time evolution layer models the causal relationship of time series; and the Gaussian activation layer controls the activity of Gaussian elements.
[0029] Gaussian element representation forms a Gaussian element set: ; in, is the position of the Gaussian element in the three-dimensional model, is a highly noisy term, i.e., characterizing the deviation, is a Gaussian noise distribution with a standard deviation Related to image resolution, Describe the shape and direction of the Gaussian element, For color, band is the number of current satellite image bands, is the opacity, which controls the visibility of the Gaussian element. It is a semantically driven function.
[0030] Step 3: Based on the historical 3D model height information of the building and recent satellite images, use the depthanything or marigold algorithm to estimate the height of the changed area of the building, and inject Gaussian element height position orientation characterization deviation based on the estimated height.
[0031] Step 4: Set the maximum height change according to the building scene h = 15m, combined with semantic constraints, the height position orientation representation deviation in step 3 is subjected to Gaussian diffusion in the Z direction to achieve height-semantic collaborative reasoning and obtain the height point set after Gaussian diffusion. : ; in, The height estimated based on the historical 3D model height information and satellite images in step 3, is the sensitivity coefficient, is a highly noisy term, i.e., characterizing the deviation, is a Gaussian noise distribution with a standard deviation Related to image resolution, Constrain the height variation range.
[0032] Step 5: The height point set described in step 4 As a Gaussian element representation The z coordinate of the sampling point x in the middle is reshaped using the space-time diffusion equation until the geometric constraint converges to the set value and the iteration is completed to obtain the Gaussian element set after the deviation reshaping .
[0033] The iterative process is: ; in, To characterize bias injection, k Indicates the k iterations, Reshape for deviation, is the geometrically degenerate tensor, and represents the image plane diffusion intensity, Indicates the diffusion intensity in the vertical direction of elevation, is a Gaussian diffusion process.
[0034] Set the diffusion intensity parameters 、 、 , by adaptively iteratively optimizing the Gaussian element set, the residual converges to below 0.05.
[0035] The geometric constraints are: ; in, is the three-dimensional Gaussian element rendering function, For recent satellite images, For rendering consistency, For semantic alignment, is the weight coefficient of semantic alignment.
[0036] Step 6: Set the Gaussian element set described in step 5 Perform isosurface extraction, use the MarchingCubes algorithm to generate a Mesh model, convert the model format to OBJ, and obtain the updated building monomer model.
[0037] Step 7: Replace the historical 3D model of the corresponding area with the updated single building model described in step 6, and output a standardized spatiotemporal dynamic description JSON format file through the decoding network, including the change timestamp, geographic coordinates, and building attributes.
Claims
1. A method for three-dimensional building modeling based on a single recent satellite image, characterized in that: The following steps are involved: Step 1: Obtain a single recent satellite image of the area where the building has changed, compare the single recent satellite image with the historical 3D model of the building, and determine the type of change in the building; If the change type of the building unit is an overall change, an initialization building unit model is constructed and the initialization building unit model is used as a sampling model; If the change type of the building unit is local change, the historical 3D model is used as the sampling model; Step 2: Sample the sampling model and perform semantically driven Gaussian meta-representation on the sampling points , forming a Gaussian element set ; Step 3: Based on the historical 3D model height information of the building and recent satellite images, use the depthanything algorithm or the marigold algorithm to estimate the height of the building's changed area, and inject Gaussian elements based on the estimated height to characterize the height position orientation deviation. ; Step 4: Set the maximum height change according to the building scene h , combined with semantic constraints, the height position orientation representation deviation in step 3 is subjected to Gaussian diffusion in the Z-axis direction to obtain the height point set after Gaussian diffusion ; Step 5: The height point set described in step 4 As a Gaussian element representation The z coordinate of the sampling point x in the middle is reshaped using the space-time diffusion equation until the geometric constraint converges to the set value and the iteration is completed to obtain the Gaussian element set after the deviation reshaping ; Step 6: Set the Gaussian element set described in step 5 Perform isosurface extraction to obtain the updated building monomer model; Step 7: Replace the historical 3D model of the corresponding area with the updated single building model described in step 6, and output the spatiotemporal dynamic description through the decoding network.
2. The method for three-dimensional building modeling based on a single recent satellite image according to claim 1, characterized in that: The overall changes mentioned in step 1 include new construction, demolition and overall replacement, while the local changes include local reconstruction, facade repair and construction; New construction: new structures with a closed outline area ≥ 90% of the standard building base outline area and a building height ≥ 3 meters; Demolition: The outline area attenuation rate of the building unit is ≥ 95%, or the building height is ≤ 5% of the original height; Overall replacement: The change rate of the building base outline area is ≤10%, the change of the building height is ≥30%, and the modification of the load-bearing structure is involved; Partial reconstruction: 5% < building base outline area change rate ≤ 20%, and 5% < building height change ≤ 15%; External facade renovation: building base outline area change rate ≤ 5%, building height change ≤ 5%, surface texture or structure update; Construction: There are temporary structures or unclosed areas with an outline closure rate of ≤70% within the building base outline, and the building height change is ≤10%.
3. The method for three-dimensional building modeling based on a single recent satellite image according to claim 1, characterized in that: The process of constructing the initial building monomer model in step 1 is as follows: For the overall change area of the building unit, a parameterized adaptive matching mechanism is used to construct a 3D model of the building unit based on the satellite imaging model. The loss between the image projected onto the plane and the recent satellite image is used as the objective function. The iteration is stopped when the objective function reaches the minimum. The 3D model of the building unit is used as the initial building unit model. Among them, the objective function is: ; is the building model parameter, To parameterize the initial model, is the rendering function that projects the parameterized model onto the image plane, For historical moments, For the current moment, is the semantic change mask image, is the optical image at the current moment, the regularization term Physical rationality of constraint parameters, is the balance coefficient.
4. The method for three-dimensional building modeling based on a single recent satellite image according to claim 3, characterized in that: The satellite imaging model is based on a rational polynomial model and is obtained by combining the satellite orbit height and imaging angle.
5. The method for three-dimensional building modeling based on a single recent satellite image according to claim 3, characterized in that: The adaptive matching mechanism adopts the improved Levenberg-Marquardt algorithm and introduces adaptive step size control, which is expressed as: ； in, is the Jacobian matrix, r is the residual vector, Dynamically adjusted according to the residual decrease rate.
6. The method for three-dimensional building modeling based on a single recent satellite image according to claim 1, characterized in that: The semantically driven Gaussian element representation in step 2 is: ； in, For any point in space, is the position of the Gaussian element in the three-dimensional model, Describe the shape and direction of the Gaussian element, For color, band is the number of current satellite image bands, is opacity, It is a semantically driven function.
7. The method for three-dimensional building modeling based on a single recent satellite image according to claim 1, characterized in that: The Gaussian element set in step 2 is: ; in, is the position of the Gaussian element in the three-dimensional model, is a highly noisy term, i.e., characterizing the deviation, is a Gaussian noise distribution with a standard deviation Related to image resolution, Describe the shape and direction of the Gaussian element, For color, band is the number of current satellite image bands, is the opacity, which controls the visibility of the Gaussian element. It is a semantically driven function.
8. The method for three-dimensional building modeling based on a single recent satellite image according to claim 1, characterized in that: The height point set after Gaussian diffusion in step 4 for: ； in, The height estimated based on the historical 3D model height information and satellite images in step 3, is the sensitivity coefficient, is a highly noisy term, i.e., characterizing the deviation, is a Gaussian noise distribution with a standard deviation Depends on the image resolution.
9. The method for three-dimensional building modeling based on a single recent satellite image according to claim 1, characterized in that: The iterative process described in step 5 is: ； in, To characterize bias injection, k Indicates the k iterations, Reshape for deviation, is the geometrically degenerate tensor, and represents the image plane diffusion intensity, Indicates the diffusion intensity in the vertical direction of elevation, is a Gaussian diffusion process.
10. The method for three-dimensional building modeling based on a single recent satellite image according to claim 1, characterized in that: The geometric constraints in step 5 are: ； in, is the three-dimensional Gaussian element rendering function, For recent satellite images, For rendering consistency, For semantic alignment, is the weight coefficient of semantic alignment.