Metaverse space and granular ball driven remote sensing change detection method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2025-12-17
- Publication Date
- 2026-07-21
AI Technical Summary
Existing deep learning-based remote sensing change detection methods have high computational complexity and strong dependence on labeled data. Furthermore, traditional methods are affected by resolution differences, noise interference, illumination changes, and terrain complexity, making it difficult to achieve efficient and accurate image registration and change detection.
By constructing digital twin terrain in the metaverse space, the remote sensing images are preprocessed and feature extracted using a particle-sphere driven method, and diffusion processing is performed using the DDPM model to generate multi-scale feature representations, and change detection is carried out in the virtual space.
It achieves low computational cost for multi-scale remote sensing image change detection, improves spatial consistency and robustness, effectively suppresses noise interference, provides three-dimensional positioning capabilities, and is suitable for applications such as geographic monitoring and urban management.
Smart Images
Figure CN121746950B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing image processing, specifically relating to a remote sensing change detection method and system based on metaverse space and particle-driven spheres. Background Technology
[0002] Remote sensing change detection involves identifying changes on the Earth's surface by analyzing images captured at different time points. It plays an increasingly important role in various applications, including climate change monitoring, natural disaster assessment, urban planning, agricultural surveys, land cover observation, and assisting military reconnaissance. Many scholars have explored various approaches to apply remote sensing change detection to solve practical problems. While initial traditional methods were simple and easy to use, their effectiveness was compromised by factors such as resolution differences, noise interference, registration errors, lighting variations between images, and landscape complexity. Deep learning-based remote sensing change detection methods were developed to improve accuracy, but they still suffer from high computational costs and heavy reliance on supervised learning models, making them both expensive and time-consuming. Recently, diffusion models have begun to be introduced into change detection tasks. Some researchers have integrated contrastive learning and time / position invariance into remote sensing pre-training to enhance representations for downstream tasks. Others have leveraged the advantages of mask image modeling to learn representations from remote sensing images that can be used in various downstream applications. However, diffusion models still have limitations, such as requiring pixel-level data as input and incurring high computational costs. Particle-sphere computation is an emerging, efficient, and robust feature representation method with multi-granularity, which can effectively address the problems existing in diffusion models. Furthermore, traditional remote sensing change detection techniques generally rely on real satellite imagery, which is limited by differences in sensor platforms, inconsistencies in imaging geometry, the influence of terrain undulations, and changes in acquisition conditions at multiple times, making it difficult to achieve strict spatial alignment of images from different times. Summary of the Invention
[0003] This invention aims to address the shortcomings of existing deep learning-based remote sensing change detection methods, such as high computational complexity and strong dependence on labeled data. It provides a remote sensing change detection method and system based on metaverse space and particle-driven approaches to achieve multi-scale, low-computational-cost remote sensing image change detection.
[0004] The objective of this invention is achieved through the following technical solution:
[0005] A remote sensing change detection method based on metaverse space and particle-driven model includes:
[0006] Step 1: Metaverse Modeling
[0007] By mapping the coordinates of the actual geographic space to the metaverse virtual space through mapping functions, the overall construction of the metaverse scene is realized. Furthermore, by interpolating the radial basis function based on the set of elevation sampling points, a continuous terrain surface is obtained, thus realizing digital twin terrain modeling.
[0008] Step 2: Data Preprocessing
[0009] Two remote sensing images of the same target area at two different times are input into the metaverse. The inverse transformation of the mapping function is used to load the remote sensing images at the two times onto the virtual surface, thereby generating the images at the two times in the metaverse.
[0010] The images at the two moments in the metaverse are processed by image registration, image normalization and other methods to generate the processed images at the two moments in the metaverse.
[0011] Step 3: Pellet Formation
[0012] The processed images at two moments in the metaverse are divided into and stitched together to obtain the set of particles corresponding to each of the two moments.
[0013] Step 4: Feature Extraction
[0014] The encoder in the DDPM model is used to extract features of each sphere in the two sphere sets;
[0015] Step 5: Denoising and Reconstruction
[0016] Using the denoising and reconstruction module in the DDPM model, the noise component in each particle is gradually removed through the reverse denoising process to obtain the particle representation corresponding to each particle.
[0017] Step 6: Generation of change detection results
[0018] Based on the particle representation of each particle in the particle set corresponding to the two time points, a change detection network is used to generate the change detection results of the target region in the metaverse virtual space.
[0019] Furthermore, the method also includes:
[0020] The change detection results are back-projected onto the metaverse virtual 3D surface model to achieve a visual overlay display of the change detection results in virtual space;
[0021] Based on the imaging parameters of the virtual camera model, the changing state in the three-dimensional space of the metaverse is converted into a planar image according to the pixel mapping rules. The planar image is in a binarized form, where one value in the binarization represents that there is a change on the ground surface, and the other value in the binarization represents that there is no change on the ground surface.
[0022] Furthermore, the particle-sphere division in step 3 includes:
[0023] For each image in the processed images at two moments in the metaverse, perform the following operations:
[0024] Using the center of the image as a reference, the image is divided equally along the horizontal and vertical directions, so that the image is divided into four equal-sized sub-regions.
[0025] Calculate the purity of each sub-region image;
[0026] Determine whether the purity of each sub-region image is greater than or equal to the purity threshold. If it is, the sub-region image is no longer subdivided. Otherwise, determine whether the sub-region image has been refined to the minimum limit. If so, no further subdivision is required. Otherwise, the sub-region image is divided equally along the horizontal and vertical directions to generate four new sub-region images of equal size. The purity of each new sub-region image is recalculated. The above process is repeated until the image division termination condition is met. The termination condition is that the purity of the new sub-region images is greater than the purity threshold or the number of pixels in the new sub-region images has reached the minimum limit.
[0027] After the image is segmented, each sub-image region corresponds to a sphere unit, thus obtaining the sphere segmentation result for that image.
[0028] Furthermore, the calculation of the purity of each sub-region image includes:
[0029] Determine the difference between the pixel value of each pixel in the sub-region image and the pixel value of the center pixel in the sub-region image;
[0030] Count the number of pixels in the sub-region image whose differences satisfy a preset pixel threshold;
[0031] The ratio of the number of pixels that meet a preset pixel threshold to the total number of pixels in the sub-region image is determined, and the ratio is the purity of the sub-region image.
[0032] Furthermore, the granule splicing in step 3 includes:
[0033] For each image in the processed images at two moments in the metaverse, perform the following operations:
[0034] For the sphere units in the sphere segmentation result of the image, when multiple sphere units satisfy the spatial adjacency relationship and spatial scale conditions, the multiple sphere units are spliced together. In the splicing process, boundary smoothing constraints and feature similarity weighting strategies need to be combined to ensure that the fused spheres are consistent in spectral features, texture structure and spatial layout, thereby generating a sphere set corresponding to the image.
[0035] Furthermore, step 4 specifically includes:
[0036] The spheres in the sphere set corresponding to each of the two time points are input into the DDPM model. The DDPM model is used to perform multi-time-step noise addition processing on each sphere, and multi-scale features of multiple time steps corresponding to each sphere are extracted.
[0037] For each sphere, at each spatial scale, the features from multiple time steps are stitched together along the channel dimension to obtain the stitched features at each spatial scale, which are the features of each sphere.
[0038] Furthermore, step 6 specifically includes:
[0039] The spheres in the sphere sets corresponding to the two time points are denoted as bi-temporal sphere characteristics;
[0040] The feature fusion is achieved by spatial alignment, splicing or attention weighting strategies for the dual-temporal particle-sphere features, generating particle-sphere fused features;
[0041] Contextual modeling is performed on the particle-sphere fusion features to obtain particle-sphere enhancement features;
[0042] The enhanced features of the particles are mapped to a binary classification decision space, and the probability of the changed state of each particle is output, thus obtaining the change detection result based on the particles.
[0043] Furthermore, the optimization of the classifier in the change detection network adopts the binary cross-entropy loss function as the optimization objective.
[0044] The present invention also provides a remote sensing change detection system based on metaverse space and particle-driven system, which includes a memory and a processor;
[0045] The memory is used to store computer-executable instructions;
[0046] The processor is used to execute the computer-executable instructions to implement the method described above.
[0047] The beneficial effects of this invention are:
[0048] First, the addition of the metaverse effectively solves the key problems of previous technologies, such as difficulty in image registration, serious interference from terrain factors, and difficulty in ensuring spatial consistency, and improves the performance of change detection in terms of spatial consistency, feature stability and system robustness.
[0049] Second, by compressing the original pixel-level image into structured granular units, the dimensionality of the input data can be significantly reduced, thus lowering computational complexity and data dependency.
[0050] Third, introducing a Gaussian noise training mechanism during the forward diffusion process and restoring structural information through the reverse denoising process can effectively suppress errors caused by illumination differences and sensor noise in remote sensing images, ensuring the stability and reliability of detection results.
[0051] Fourth, by employing a multi-timestep, multi-scale diffusion feature extraction strategy, it is possible to capture change information at different spatial levels;
[0052] Fifth, the introduction of the metaverse not only retains the intuitiveness of traditional change detection results, but also realizes the three-dimensional positioning of change information through the virtual space of the metaverse, so that the results have both two-dimensional display and three-dimensional spatial expression capabilities, making them more suitable for subsequent applications such as geographic monitoring, urban management and spatiotemporal analysis.
[0053] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:
[0055] Figure 1 This is a schematic flowchart of a remote sensing change detection method based on metaverse space and particle-driven spheres;
[0056] Figure 2 This is a schematic diagram of the diffusion model;
[0057] Figure 3 This is a flowchart of particle generation;
[0058] Figure 4 This is a schematic block diagram of a remote sensing change detection method based on metaverse space and particle-driven spheres. Detailed Implementation
[0059] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.
[0060] This invention proposes a remote sensing change detection method based on metaverse space and particle-driven approach. This method maps two remote sensing images of the same geographic region (one representing the image before change and the other representing the image after change) onto a metaverse for processing. After image preprocessing, the images are then divided into particles, and subsequent image processing is performed on a particle-sphere basis. Specifically, this includes using Denoising Diffusion Probabilistic Models (DDPM) to perform forward diffusion (i.e., forward noise addition) and backward diffusion (backward denoising) on the particle-sphere images. A subsequent change detection module generates change detection results, which are then rendered as a planar change detection result map. This allows for the reconstruction of the true spatial location of the changed area within a 3D digital twin scene, providing more intuitive and spatially logical 3D visualization support for subsequent applications such as geographic monitoring, disaster analysis, and urban supervision.
[0061] Combination Figure 1 and Figure 4 Remote sensing change detection methods based on metaverse space and particle-driven approaches can include:
[0062] Step 1: Metaverse Modeling:
[0063] By mapping the coordinates of the actual geographic space to the metaverse virtual space through mapping functions, the overall construction of the metaverse scene is realized. Furthermore, by interpolating the radial basis function based on the set of elevation sampling points, a continuous terrain surface is obtained, thus realizing digital twin terrain modeling.
[0064] Step 2: Data Preprocessing
[0065] Two remote sensing images of the same target area at two different times are input into the metaverse. The inverse transformation of the mapping function is used to load the remote sensing images at the two times onto the virtual surface, thereby generating the images at the two times in the metaverse.
[0066] The images at the two moments in the metaverse are processed by image registration, image normalization, noise suppression, etc., to generate the processed images at the two moments in the metaverse.
[0067] Step 3: Pellet formation:
[0068] The processed images at two different moments in the metaverse are divided into and stitched together to obtain the set of particles corresponding to each of the two moments. For details on the particle generation method, please refer to [link to relevant documentation]. Figure 3And its related descriptions.
[0069] Step 4: Feature Extraction
[0070] The encoder in the DDPM model is used to extract features of each sphere in the two sphere sets;
[0071] Step 5: Denoising and Reconstruction
[0072] Using the denoising and reconstruction module in the DDPM model, the noise component in each particle is gradually removed through the reverse denoising process to obtain the particle representation corresponding to each particle.
[0073] Step 6: Generation of change detection results:
[0074] Based on the particle representation of each particle in the particle set corresponding to the two time points, a change detection network is used to generate the change detection results of the target region in the metaverse virtual space.
[0075] In step 1, the actual geographic coordinate system can be denoted as... :
[0076]
[0077] in, Latitude refers to the angular coordinates of a geographical location in the north-south direction. Longitude refers to the angular coordinates of a geographical location in the east-west direction. Geodetic height refers to the vertical height of a feature point relative to the Earth's reference ellipsoid, and is used for three-dimensional terrain construction.
[0078] The unified virtual geographic coordinate system (a three-dimensional virtual coordinate system) established in the metaverse space is denoted as... :
[0079]
[0080] in, For virtual space Axis coordinates; For virtual space Axis coordinates; For virtual space Axis coordinates, i.e., height values.
[0081] A unified mapping function is used to establish the coordinate correspondence between virtual space and actual geographic space. The specific mapping formula can be expressed as follows:
[0082]
[0083] in, They refer to virtual space. The scaling factor of the axis; The Earth's average radius; These refer to the reference points in the virtual scene.
[0084] For ease of use later, this mapping function will be denoted as :
[0085]
[0086] This mapping function can convert real coordinate values into virtual coordinates in the metaverse space.
[0087] To accurately construct the digital twin terrain in the metaverse, the following terrain interpolation model is used, given a set of discrete elevation sampling points:
[0088]
[0089] After obtaining the elevation sampling points, the actual geographic coordinates can be converted into coordinates in the metaverse virtual space according to the mapping function above. Then, the continuous terrain surface can be obtained by interpolation using radial basis functions (RBF). (i.e., the height value of the m-th point):
[0090]
[0091] in, This represents the total number of sampling points used to construct the digital twin terrain or the virtual surface of the metaverse; From 1 to , indicating a total of A number of real geographic sampling points were used for digital twin terrain reconstruction (i.e., RBF interpolation). The coordinates of any point on the Earth's surface in the metaverse whose height is to be calculated; Known terrain sampling points used to construct digital twin terrain (the first) The planar position coordinates of (number of sampling points); Refers to the RBF kernel function; This represents the interpolation weights obtained by solving the RBF system. This enables smooth, continuous, and renderable 3D reconstruction of realistic terrain.
[0092] Next, in step 2, remote sensing images of the same target area at two different times are input into the metaverse, denoted as... ,in, They refer to the first Remote sensing images at two different times; These represent the image's height, width, and number of channels, respectively. For these two remote sensing images from different times, it is necessary to map the real remote sensing images onto the metaverse virtual geographic landing. This can be achieved using the inverse transform of the mapping function (denoted as...). The remote sensing images from the two specified moments are loaded onto the virtual surface to generate the images from the metaverse at those two moments. This is because, to calculate the pixel value of a point on the virtual surface, it is necessary to first find the coordinates of the remote sensing image corresponding to that point. This requires an inverse transformation of the mapping function. Then, the pixel value corresponding to the remote sensing image coordinates is filled into that position on the virtual surface. This process can be expressed by the formula:
[0093]
[0094] in, Represents the coordinates of a point on the virtual surface of the metaverse, through The location of the point on the remote sensing image can be determined, and then the pixel value corresponding to that location can be obtained. This pixel value is filled into the virtual surface. At the point, thus obtaining Corresponding pixel value .
[0095] The images at the two moments in the metaverse are processed by image registration, image normalization, and image enhancement to generate the processed images at the two moments in the metaverse.
[0096] To ensure strict spatial alignment of images, image registration is performed using the following formula:
[0097]
[0098] In the above formula, The first in the metaverse Images of moments; The first in the metaverse The image after time-registration; As a spatial transformation operator, different calculation methods can be selected according to different types (such as translation, affine transformation, projection transformation, nonlinear transformation, etc.).
[0099] To enable subsequent particle generation and diffusion models to process images, the system needs to normalize each channel of the metaverse to eliminate amplitude shifts caused by differences in illumination and sensors.
[0100]
[0101] in, The image after normalization; The first image obtained after image registration A metaverse image of time; The first one is automatically calculated by the metaverse. The global brightness mean and standard deviation of the metaverse image at each time step ensure the consistency of the same-domain distribution of images across multiple time steps.
[0102] The above preprocessing steps can effectively improve the quality and spatiotemporal alignment accuracy of the input image, providing high-quality and comparable input data for subsequent processing.
[0103] Next, in step 3, particle sphere generation is performed. The preprocessed image space is divided into multiple local region units, and the "purity" of each region is calculated. Purity reflects the differences in characteristics within the region. When the purity of a region is lower than a set threshold, further subdivision is required, and this process is repeated recursively. Finally, each image is divided into several "particle sphere units," and a set of particle spheres is output. This invention uses particle spheres as the basic structural unit, which can efficiently capture local characteristics. Figure 3 This is a schematic flowchart illustrating the formation of granules. For example... Figure 3 As shown, the image can be first divided into particles and then the particles and spheres can be stitched together.
[0104] Specifically, for each image in the processed images at two moments in the metaverse, the following particle-sphere partitioning operation is performed:
[0105] Using the center of the image as a reference, the image is divided equally along the horizontal and vertical directions, so that the image is divided into four equal-sized sub-regions.
[0106] Calculate the purity of each sub-region image;
[0107] In some embodiments, calculating the purity of each sub-region image includes:
[0108] Determine the difference between the pixel value of each pixel in the sub-region image and the pixel value of the center pixel in the sub-region image;
[0109] Count the number of pixels in the sub-region image whose differences satisfy a preset pixel threshold;
[0110] The ratio of the number of pixels that meet a preset pixel threshold to the total number of pixels in the sub-region image is determined, and the ratio is the purity of the sub-region image.
[0111] The formula for calculating purity is as follows:
[0112]
[0113]
[0114] in, For image Each pixel Pixel values; For image The pixel value at the center point; For a threshold; This indicates that when the difference between the pixel value of a single pixel and the pixel value of the image center is within a set threshold... If the condition is met within the specified time, it is recorded as the value 1. This indicates the number of pixels that meet the condition. Images of the divided sub-regions The number of all pixels contained in the image. Therefore, for each region image All can obtain the corresponding purity. .
[0115] Next, it is determined whether the purity of each sub-region image is greater than or equal to a purity threshold. If it is greater than or equal to the purity threshold If the sub-region image is not subdivided, then the sub-region image is not further subdivided; otherwise, it is determined whether the sub-region image has been refined to the minimum limit. If so, it is not further subdivided. Otherwise, the sub-region image is divided into equal parts along the horizontal and vertical directions to generate four new sub-region images of equal size. The purity of each new sub-region image is recalculated. The above process is repeated until the image division termination condition is met. The termination condition for image division is: the purity of the new sub-region images is greater than the purity threshold or the number of pixels in the new sub-region images has reached the minimum limit.
[0116] This can be expressed as a formula:
[0117]
[0118] When the purity value of a certain region map is lower than a preset purity threshold If the result is negative, it indicates that there is heterogeneity within the region, which does not meet the spatial consistency requirement and needs to be further subdivided (further subdivision is possible if the region has not been refined to the minimum limit). By further subdividing, four new sub-region images of equal size are generated, and the purity of each new sub-region image is recalculated.
[0119] After the image is segmented, each sub-image region corresponds to a sphere unit, thus obtaining the sphere segmentation result for that image.
[0120] Through the above segmentation process, images with high purity and consistent internal features can be obtained. This recursive segmentation and purity evaluation mechanism can adaptively control the spatial scale of the spheres. Furthermore, due to pixel information differences within image blocks, some spheres contain little or no information after segmentation. To reduce the number of such spheres and improve subsequent computational efficiency, they can be stitched together.
[0121] Specifically, the following particle-sphere stitching operation is performed on each image in the processed images at two moments in the metaverse:
[0122] For the sphere units in the sphere partitioning result of the image, when sphere units from different regions simultaneously satisfy both spatial adjacency and spatial scale conditions, a splicing operation is performed on the multiple sphere units. First, the sphere units to be spliced must be located in different regions and be spatially adjacent or connected. Second, since spheres of different sizes can be obtained through the above sphere partitioning, the spatial scale condition consists of multiple hierarchical side length levels. That is, only spheres of the same size (same side length) can be spliced; spheres with different side lengths cannot be spliced. This ensures the consistency of scale and spatial continuity during the splicing process.
[0123] For eligible spheres, a splicing operation is performed. During the splicing process, a boundary smoothing constraint and a feature similarity weighting strategy must be combined to ensure that the fused spheres are consistent in spectral features, texture structure, and spatial layout. Through this splicing operation, small spheres can be merged into larger sphere units.
[0124] Through the above particle-sphere segmentation and particle-sphere stitching, adaptive optimization of particle-sphere segmentation can be achieved under different image features. Each image can yield a particle-sphere set.
[0125] The set of particles corresponding to two time-phase images in the metaverse can be represented as:
[0126]
[0127] in, The resulting set of particles; arrive for Each particle is a sphere. It should be noted that the number of particles in the particle set corresponding to two moments in the metaverse can be the same or different, and the specific number depends on the actual image situation. This indicates that the number of spheres in the sphere cluster is much smaller than the height x width of the input image. This operation replaces pixels with local regions of spheres, achieving data compression and structuring.
[0128] After obtaining spatially consistent and highly pure spheres, this invention further extracts features from the remote sensing images at these two time points. A multi-scale feature representation is constructed using the diffusion model within the DDPM model to fully exploit spatiotemporal variation information and enhance the robustness and noise resistance of the features. Feature extraction employs a diffusion model as the generation framework, introducing random Gaussian noise into the image to learn the hierarchical structural features of the image at different noise levels. The diffusion model progressively "perturbs" and "recovers" the image at multiple time steps. Through this multi-scale modeling mechanism, multi-level feature representations are provided for change detection.
[0129] Figure 2 This is a schematic diagram of a diffusion model. (For example...) Figure 2 As shown, the forward diffusion process is a process of gradually adding Gaussian noise to the original image X0, with the aim of gradually degrading the original image X0 into pure noise X. T . Figure 2 middle, Let X represent the conditional probability density during the forward diffusion process, and let X represent the noisy image X at each step, which depends only on the noisy image at the previous step. t Relying solely on the noisy image X from the previous step t-1 The backdiffusion process originates from a purely noisy image X. T Starting from this point, after T-step denoising, X0 is restored to be close to the original image distribution. represents the conditional probability density in the backdiffusion process, and represents the reconstruction of the previous image from the current noisy image.
[0130] Specifically, a multi-time-step noise version can be constructed for each particle, and the noisy images obtained at time step t are as follows:
[0131]
[0132]
[0133] in, Gaussian noise that follows a standard normal distribution; and These are the particles before and after the change (i.e., any one particle from the particle set corresponding to the two moments). and They respectively added standard Gaussian noise at time step t. Before and after the change in the granules; Control the amount of noise added. The generated noisy image can be used for subsequent feature extraction, making the features noise-resistant and robust, which helps with subsequent change detection.
[0134] The particles in the particle set corresponding to each of the two time points are input into the DDPM model, and the DDPM model is used to perform multi-time-step noise addition processing on each particle (to... and As input to DDPM, multi-scale features corresponding to multiple time steps for each sphere are extracted. The multi-scale features can be represented as:
[0135]
[0136]
[0137] in, and This represents the multi-scale features extracted from the images before and after the change. The scale representing the feature. Indicates the highest resolution ( ), Indicates the minimum resolution ( It should be noted that the features presented here are at five scales, which are merely examples and are not intended to limit the invention.
[0138] To fully integrate the complementary information contained in the features at multiple time points, the images before and after the change were analyzed at different time steps at each spatial scale. The features below are then stitched together. This stitching process is performed along the channel dimension, specifically as follows:
[0139]
[0140]
[0141] Where t=t0 to t=t n Indicates the selected time step. Refers to the splicing operation; The splicing operation is performed along the channel dimension; and The concatenated features (i ranges from 0 to 4) pass through two 2D convolutional layers to retain semantic information across multiple time scales while controlling model complexity, providing more discriminative feature inputs for subsequent change detection tasks. The concatenated features at each spatial scale constitute the features of each sphere.
[0142] Next, using the denoising and reconstruction module in the DDPM model, the noise component in each particle is gradually removed through the reverse denoising process to obtain the particle representation corresponding to each particle.
[0143] The denoising and reconstruction module is based on a diffusion model. It gradually removes noise components through a reverse denoising process to restore high-fidelity structural features. The denoising process can be formally represented as:
[0144]
[0145]
[0146] In the above formula, Indicates at time step Noise characteristics; For denoising networks; The image after denoising; To predict noise; A coefficient used to control noise intensity; This represents a random disturbance that follows a standard normal distribution. This indicates the proportion of "preserving the content of the previous image" during the current reverse denoising process; Control the magnitude of "noise removal"; Used to calibrate the cumulative noise ratio. Through multi-time-step reverse denoising inference, the spatial and semantic structure information of the features can be gradually recovered, resulting in a clear representation after noise suppression.
[0147] The feature representation processed by the denoising and reconstruction module is input into the change detection network, where it works in conjunction with the sphere structure features of the images at previous and subsequent time points to generate a denoised change detection result. This change detection result is at the sphere level, meaning each sphere corresponds to a "changed" or "unchanged" label.
[0148] The spheres in the sphere sets corresponding to the two time points are denoted as biphase sphere features.
[0149] The change detection network performs the following process: spatial alignment, splicing, or attention weighting strategies are used to fuse the two-phase particle-sphere features to generate particle-sphere fused features;
[0150] Contextual modeling is performed on the particle-sphere fusion features to obtain particle-sphere enhancement features;
[0151] The enhanced features of the particles are mapped to a binary classification decision space, and the probability of the changed state of each particle is output, thus obtaining the change detection result based on the particles.
[0152] Specifically, firstly, using pre-trained and parameter-frozen... The model acts as a fixed feature extractor, extracting features from the target scene respectively. Multi-layer feature extraction is performed on each sphere at two time points, where shallow features capture low-level visual information and deep features capture high-level abstract information; subsequently, ... Spatial alignment is performed on the features of the corresponding particles at two time points to ensure the correspondence of the features in the spatial dimension (i.e., particles in the same geographical area before and after the change should correspond to each other). Strategies such as difference feature calculation, direct feature concatenation, or attention-weighted fusion (e.g., channel attention mechanism can be used to assign weight coefficients to features in each dimension, with higher weight coefficients assigned to features in change-sensitive dimensions) are used to complete the fusion of the features of the particles in the two time points. Finally, the fused particle features are mapped to the decision space, and spatial context modeling is introduced to consider the spatial relationships between the neighborhoods of the particles. The probability of the changed state corresponding to each particle is output through the binary classification decision module, thereby obtaining the particle-based change detection results in this scenario.
[0153] This process not only effectively suppresses Gaussian noise interference, but also significantly improves the model's sensitivity to subtle changes and detection accuracy, providing reliable input for subsequent change classification and boundary optimization.
[0154] In this invention, the model is fine-tuned using the LEVIR building Change Detection dataset (LEVIR-CD dataset), and the binary cross-entropy (BCE) loss function is used as the optimization objective.
[0155]
[0156] in, Represents the binary cross-entropy loss function; Indicates whether the actual label has changed; The probability of change predicted by the model; Based on the natural constant Logarithm to base 0; It can convert the difference in probability into a quantifiable value of loss, making the loss more sensitive to changes in probability. and Category weight coefficients ( =1), generally set This is to strengthen the learning of the "change zone".
[0157] After determining the particle-level change detection results, post-processing optimization is required. This optimization includes first mapping the particle-level change results back to the unified 3D surface of the metaverse virtual geographic space. Based on the continuous structure and neighborhood relationships of the virtual terrain, spatial consistency corrections are performed on isolated noise points, pseudo-change regions, and boundary breaks that may appear in the initial detection. Subsequently, morphological filtering, connected component analysis, and edge smoothing methods are used to refine and complete the boundaries of the change regions, making them more closely resemble the actual terrain features.
[0158] Since the classifier outputs sphere-level detection result labels, where a sphere is a structured segmentation unit of an image, while the metaverse virtual geospatial requires "pixel-level images," after generating the change detection results, these results can be back-projected onto the metaverse virtual 3D surface model (i.e., data-to-space mapping) to achieve a visual overlay of the change detection results in virtual space. Then, based on the imaging parameters of the virtual camera model, the change state in the metaverse 3D space is converted into a planar image according to pixel mapping rules. Specifically, the 3D change area can be re-rendered to maintain the same size as the virtual image. The planar change detection image is generated using a binary format (e.g., 0 and 1). One value in the binary format (e.g., "1", a white pixel) represents a change in the land surface, while the other value (e.g., "0", a black pixel) represents no change. The results generated based on the metaverse virtual geospatial model not only accurately reflect the changes in the same geographical location at different times in a two-dimensional image, but also restore the true spatial location of the changed area in a three-dimensional digital twin scene. This provides more intuitive and spatially logical three-dimensional visualization support for subsequent applications such as geographic monitoring, disaster analysis, and urban supervision.
[0159] In summary, this method first constructs a metaverse digital twin surface based on real geographic coordinates and sampling points. Multi-timeframe remote sensing images are then loaded as virtual surface textures via inverse coordinate mapping. A spatially consistent multi-timeframe virtual image is obtained through virtual model rendering, and the virtual image is preprocessed. Subsequently, a grain-sphere generation algorithm is used to adaptively partition the image space, constructing grain-sphere units. Regional grain-spheres replace pixel input, enabling parallel data input and structured feature representation. The grain-sphere units are then input into a pre-trained diffusion model to extract multi-timestep, multi-scale diffusion features, thereby obtaining deep feature representations robust to illumination changes, noise, and registration errors. Finally, the denoised features and grain-sphere structure features are input into a change detection network, outputting change detection results. Simultaneously, based on the unified 3D scene and virtual coordinate system of the metaverse digital twin space, the change detection results can be mapped back to the virtual surface, enabling visualization and interactive analysis of changed regions within the metaverse space. This invention improves change detection accuracy while reducing computational costs. Furthermore, by leveraging the unified coordinate system and three-dimensional structural representation of the metaverse digital twin space, it enhances the spatial consistency of multi-temporal images, thereby further strengthening the reliability of detection results. It exhibits stronger stability and generalization under illumination disturbances, noise interference, and complex terrain conditions. This method is efficient, robust, and scalable, and can be widely applied in fields such as urban construction monitoring, ecological environment change analysis, agricultural area monitoring, and natural disaster assessment, providing a novel and efficient technical path for remote sensing intelligent analysis based on the metaverse.
[0160] The present invention also provides a remote sensing change detection system based on metaverse space and particle-driven system, which includes a memory and a processor;
[0161] The memory is used to store computer-executable instructions;
[0162] The processor is used to execute the computer-executable instructions to implement the remote sensing change detection method based on metaverse space and particle-driven as described above.
[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A remote sensing change detection method based on metaverse space and particle-driven dynamics, characterized in that, include: Step 1: Metaverse Modeling By mapping the coordinates of the actual geographic space to the metaverse virtual space through mapping functions, the overall construction of the metaverse scene is realized. Furthermore, by interpolating the radial basis function based on the set of elevation sampling points, a continuous terrain surface is obtained, thus realizing digital twin terrain modeling. In step 1, the actual geographic coordinate system is denoted as... : in, Latitude refers to the angular coordinates of a geographical location in the north-south direction. Longitude refers to the angular coordinates of a geographical location in the east-west direction. Earth height refers to the vertical height of a feature point relative to the Earth's reference ellipsoid, and is used for three-dimensional terrain construction. The unified virtual geographic coordinate system established in the metaverse space is denoted as... : in, For virtual space Axis coordinates; For virtual space Axis coordinates; For virtual space Axis coordinates, i.e., height values; A unified mapping function is used to establish the coordinate correspondence between virtual space and actual geographic space. The specific mapping formula is as follows: ， in, They refer to virtual space. The scaling factor of the axis; The Earth's average radius; These refer to the reference points in the virtual scene; Let the mapping function be denoted as : This mapping function converts real coordinate values into virtual coordinates in the metaverse space. To accurately construct the digital twin terrain in the metaverse, the following terrain interpolation model is used, given a set of discrete elevation sampling points: After obtaining the elevation sampling points, the actual geographic coordinates are transformed into coordinates in the metaverse virtual space according to the mapping function above. Then, the continuous terrain surface is obtained through radial basis function interpolation. That is, the height value of the m-th point: in, This represents the total number of sampling points used to construct the digital twin terrain or the virtual surface of the metaverse; From 1 to , indicating a total of One real geographic sampling point was used for digital twin terrain reconstruction; The coordinates of any point on the Earth's surface in the metaverse whose height is to be calculated; The planar coordinates of known terrain sampling points used to construct digital twin terrain; Refers to the RBF kernel function; This represents the interpolation weights obtained by solving the RBF system; Step 2: Data Preprocessing Two remote sensing images of the same target area at two different times are input into the metaverse. The inverse transformation of the mapping function is used to load the remote sensing images at the two times onto the virtual surface, thereby generating the images at the two times in the metaverse. The images at the two moments in the metaverse are processed by image registration, image normalization and other processes to generate the processed images at the two moments in the metaverse. Step 3: Pellet Formation The processed images at two moments in the metaverse are divided into and stitched together to obtain the set of particles corresponding to each of the two moments. The particle division in step 3 includes: For each image in the processed images at two moments in the metaverse, perform the following operations: Using the center of the image as a reference, the image is divided equally along the horizontal and vertical directions, so that the image is divided into four equal-sized sub-regions. Calculate the purity of each sub-region image; Determine whether the purity of each sub-region image is greater than or equal to the purity threshold. If it is, the sub-region image is no longer subdivided. Otherwise, determine whether the sub-region image has been refined to the minimum limit. If so, no further subdivision is required. Otherwise, the sub-region image is divided equally along the horizontal and vertical directions to generate four new sub-region images of equal size. The purity of each new sub-region image is recalculated. The above process is repeated until the image division termination condition is met. The termination condition is that the purity of the new sub-region images is greater than the purity threshold or the number of pixels in the new sub-region images has reached the minimum limit. After the image segmentation is completed, each sub-image region corresponds to a sphere unit, thus obtaining the sphere segmentation result corresponding to the image; Step 3, the splicing of granules, includes: For each image in the processed images at two moments in the metaverse, perform the following operations: For the sphere units in the sphere segmentation result of the image, when multiple sphere units satisfy the spatial adjacency relationship and spatial scale conditions, the multiple sphere units are spliced together. During the splicing process, boundary smoothing constraints and feature similarity weighting strategies need to be combined to ensure that the fused spheres are consistent in spectral features, texture structure and spatial layout, thereby generating a sphere set corresponding to the image. Step 4: Feature Extraction The encoder in the DDPM model is used to extract features of each sphere in the two sphere sets; Step 5: Denoising and Reconstruction Using the denoising and reconstruction module in the DDPM model, the noise component in each particle is gradually removed through the reverse denoising process to obtain the particle representation corresponding to each particle. Step 6: Generation of change detection results Based on the particle representation of each particle in the particle set corresponding to the two time points, a change detection network is used to generate the change detection results of the target region in the metaverse virtual space.
2. The remote sensing change detection method based on metaverse space and particle-driven approach according to claim 1, characterized in that, Also includes: The change detection results are back-projected onto the metaverse virtual 3D surface model to achieve a visual overlay display of the change detection results in virtual space; Based on the imaging parameters of the virtual camera model, the changing state in the three-dimensional space of the metaverse is converted into a planar image according to the pixel mapping rules. The planar image is in a binarized form, where one value in the binarization represents that there is a change on the ground surface, and the other value in the binarization represents that there is no change on the ground surface.
3. The remote sensing change detection method based on metaverse space and particle-driven approach according to claim 1, characterized in that, The calculation of the purity of each sub-region image includes: Determine the difference between the pixel value of each pixel in the sub-region image and the pixel value of the center pixel in the sub-region image; Count the number of pixels in the sub-region image whose differences satisfy a preset pixel threshold; The ratio of the number of pixels that meet a preset pixel threshold to the total number of pixels in the sub-region image is determined, and the ratio is the purity of the sub-region image.
4. The remote sensing change detection method based on metaverse space and particle-driven approach according to claim 1, characterized in that, Step 4 specifically includes: The spheres in the sphere set corresponding to each of the two time points are input into the DDPM model. The DDPM model is used to perform multi-time-step noise addition processing on each sphere, and multi-scale features of multiple time steps corresponding to each sphere are extracted. For each sphere, at each spatial scale, the features from multiple time steps are stitched together along the channel dimension to obtain the stitched features at each spatial scale, which are the features of each sphere.
5. The remote sensing change detection method based on metaverse space and particle-driven approach according to claim 1, characterized in that, Step 6 specifically includes: The spheres in the sphere sets corresponding to the two time points are denoted as the biphase sphere characteristics. The feature fusion is achieved by spatial alignment, splicing or attention weighting strategies for the dual-temporal particle-sphere features, generating particle-sphere fused features; Contextual modeling is performed on the particle-sphere fusion features to obtain particle-sphere enhancement features; The enhanced features of the particles are mapped to a binary classification decision space, and the probability of the changed state of each particle is output, thus obtaining the change detection result based on the particles.
6. The remote sensing change detection method based on metaverse space and particle-driven approach according to claim 1, characterized in that, The optimization of the classifier in the change detection network uses the binary cross-entropy loss function as the optimization objective.
7. A remote sensing change detection system based on metaverse space and particle-driven systems, characterized in that, Including memory and processor; The memory is used to store computer-executable instructions; The processor is configured to execute the computer-executable instructions to implement the method of any one of claims 1 to 6.