Large model-based high-resolution remote sensing image building vector drawing method and system

Abstract

The invention discloses a high-resolution remote sensing image building vector drawing method and system based on a large model, belongs to the field of high-resolution satellite remote sensing image building vector drawing, and solves the problems that in the prior art, a building vector cannot be directly obtained, so that a large gap exists between a converted building vector and a manual drawing result, and the drawing efficiency is poor. And generalization is relatively low. Collecting a self-supervised data set for constructing a basic large model; building a building mask data set; constructing a building mask extraction model, and performing mask extraction on the to-be-extracted remote sensing image by using the building mask extraction model to obtain a building mask; constructing a building corner extraction model, and performing feature extraction on the building mask by using the building corner extraction model to obtain a Gaussian probability graph of corners; and obtaining a building vector based on the Gaussian probability graph of the building mask and the angular points, and completing drawing. The method is used for realizing efficient, flexible and accurate high-resolution satellite remote sensing image building vector mapping.

Description

Technical Field

[0001] This invention relates to the field of high-resolution satellite remote sensing image building vector mapping, specifically to a method and system for high-resolution remote sensing image building vector mapping based on a large model. Background Technology

[0002] Extracting polygonal buildings from high-resolution satellite remote sensing images is a crucial task for large-scale mapping, 3D city modeling, and various geographic information system applications. Traditional building mapping methods, based on manual field measurements and hand-drawn sketches from remote sensing images, are limited in their application for large-scale building mapping due to their high cost and low efficiency. Furthermore, with the increasing high-frequency monitoring capabilities of high-resolution satellite remote sensing images, traditional methods are insufficient to meet the demands of high-frequency building mapping.

[0003] With the development of artificial intelligence, deep learning-based building extraction from satellite remote sensing imagery has made significant progress in both efficiency and effectiveness. Currently, for efficiency reasons, the most advanced method for vector mapping of buildings in high-resolution satellite remote sensing imagery involves extracting building masks using deep learning algorithms, followed by vector post-processing techniques to vectorize and orthogonalize the extracted masks. This method requires a large number of labeled building samples for training the deep learning model and still faces challenges in generalizing the extraction of buildings from satellite remote sensing imagery. Furthermore, current deep learning-based building extraction often only obtains building masks; the regularized vectors needed for building mapping require post-processing of the building mask raster. This process is cumbersome, and while vector post-processing heavily relies on the quality of the extracted building masks, the algorithm itself struggles to handle complex building structures. This results in vectors that often do not conform to the structural characteristics of buildings, generating numerous meaningless corner points, leading to unsatisfactory final vector mapping results that differ significantly from traditional manual mapping. Secondly, the black-box nature of deep learning models makes it difficult to analyze the problems theoretically, hindering targeted adjustments to the model structure.

[0004] Another approach is to directly extract key points of buildings using deep learning algorithms and connect them to form regular building outlines. The challenge lies in establishing the topological structure between these key points. Currently, there is no universally stable algorithm capable of effectively establishing the topological structure between a large number of key points, resulting in chaotic and illogical vectorized building mappings on high-resolution satellite remote sensing imagery. Therefore, for quality considerations, manual work remains the mainstream method, but its efficiency limits its large-scale implementation.

[0005] In the prior art, Chinese patent document CN112348836A discloses "a method and apparatus for automatically extracting building outlines," which constructs a convolutional neural network model for extracting building outlines; uses the model to extract the building location and reference outline from remote sensing images to obtain the building mask image corresponding to the building; processes the building mask image based on a combination of linear features and image segmentation algorithms to obtain corresponding line segment information, corner point information, and orientation information; uses the corresponding line segment information, corner point information, and orientation information to perform feature vector mapping optimization and outline correction on the building location and reference outline to obtain the revised building outline; converts the revised building outline into geographic coordinates and outputs vectors to complete map drawing. However, this technical solution can only obtain building masks. The regularized vectors required for building mapping need to be obtained by post-processing the building mask raster. This process is not only cumbersome, but the converted vectors often do not conform to the structural characteristics of buildings, generating a large number of meaningless corner points, which is significantly different from the results of traditional manual drawing.

[0006] In summary, existing technologies have the technical problem of being unable to directly obtain building vectors, resulting in a significant difference between the converted building vectors and manually drawn results, and also having low generalization ability. Summary of the Invention

[0007] This invention solves the technical problem that existing technologies cannot directly obtain building vectors, resulting in a large discrepancy between the converted building vectors and manually drawn results, and also have low generalization ability.

[0008] The present invention provides a method for vector mapping of buildings based on large-scale models using high-resolution remote sensing images, specifically as follows: Step 1: Collect data from a supervised dataset; Step 2: Build a basic large model based on the self-supervised dataset; Step 3: Construct a building mask dataset, which includes a building vector annotation dataset; Step 4: Based on the basic large model and the building mask dataset, construct a building mask extraction model, and use the building mask extraction model to extract the mask from the remote sensing image to be extracted, so as to obtain the building mask; Step 5: Based on the building mask extraction model and the building vector annotation dataset, construct the building corner extraction model, and use the building corner extraction model to extract features from the building mask to obtain the Gaussian probability map of the corner. Step 6: Based on the Gaussian probability map of the building mask and corner points, obtain the building vector and complete the mapping.

[0009] Furthermore, in this embodiment of the invention, the building mask dataset in step 3 also includes an open-source dataset.

[0010] Furthermore, in this embodiment of the invention, the construction of the building mask extraction model in step 4 specifically involves: A pre-defined building mask extraction model is used, with the basic large model as the encoder and the UpperNet model as the decoder. The pre-defined building mask extraction model is then fine-tuned based on the building mask dataset to obtain the final building mask extraction model.

[0011] Furthermore, in this embodiment of the invention, the fine-tuning of the preset building mask extraction model based on the building mask dataset specifically involves: After data augmentation of the building mask dataset, the data is input into the encoder for encoding. The encoded features output from the n / 4, n / 2, n*3 / 4 and nth layers of the basic large model are used as the input to the decoder, where n is the encoder layer number. The decoder decodes the encoded features to obtain the mask probability map. Based on the probability map, the preset building mask extraction model is fine-tuned to obtain the building mask extraction model.

[0012] Furthermore, in this embodiment of the invention, the construction of the building corner extraction model in step 5 specifically involves: A pre-defined building corner extraction model is used. The weights of the pre-defined building corner extraction model are initialized using a building mask extraction model. The initialized pre-defined building corner extraction model is then fine-tuned based on a building vector annotation dataset to obtain the final building corner extraction model.

[0013] Furthermore, in this embodiment of the invention, the fine-tuning of the preset building corner extraction model based on the building vector annotation dataset specifically involves: Obtain vector labels from the building vector annotation dataset, obtain the encoding features of the horizontal rectangle enclosing the vector labels, extract the encoding features output by the building mask extraction model, and obtain the Gaussian heatmap of the corner points based on the encoding features of the horizontal rectangle enclosing the vector labels. Fine-tune the preset building corner extraction model based on the Gaussian heatmap of the corner points to obtain the building corner extraction model.

[0014] Furthermore, in this embodiment of the invention, the step of obtaining the vector labels of the building vector annotation dataset and obtaining the encoding features of the horizontal rectangle enclosing the vector labels specifically involves: The bounding horizontal rectangle of the vector label is normalized, and the top-left and bottom-right points of the normalized bounding horizontal rectangle are encoded to obtain the encoding features of the bounding horizontal rectangle of the vector label.

[0015] Furthermore, in this embodiment of the invention, the process of obtaining the building vector based on the Gaussian probability map of the building mask and corner points in step 6 is specifically as follows: The Gaussian probability map of the corner points is filtered, and the pixel positions with the same pixel value before and after filtering are taken as corner points. The nearest pixel on the boundary of the building mask to each corner point is obtained, and the pixel points are connected to obtain the building vector.

[0016] Furthermore, in this embodiment of the invention, the building vector annotation dataset specifically comprises: High-resolution satellite remote sensing image samples were manually delineated, and the vector labels of the delineated high-resolution satellite remote sensing image samples were converted into raster labels. The high-resolution satellite remote sensing image samples were then cropped to obtain a building vector annotation dataset.

[0017] The present invention discloses a high-resolution remote sensing image building vector mapping system based on a large model. The system is constructed according to the above method and includes the following modules: The basic data acquisition module collects data from supervised datasets; The basic model building module constructs a basic large model based on a self-supervised dataset. The mask data acquisition module constructs a building mask dataset, which includes a building vector annotation dataset. The mask model construction module, based on the basic large model and the building mask dataset, constructs a building mask extraction model. The building mask extraction model is then used to extract the mask from the remote sensing image to be extracted, thus obtaining the building mask. The corner model construction module constructs a building corner extraction model based on the building mask extraction model and the building vector annotation dataset. It then uses the building corner extraction model to extract features from the building mask and obtains the Gaussian probability map of the corner. The vector mapping module uses Gaussian probability maps of building masks and corner points to obtain building vectors and complete the mapping.

[0018] This invention solves the technical problem in existing technologies where building vectors cannot be directly obtained, leading to significant discrepancies between the converted building vectors and manually drawn results, and also exhibiting low generalization ability. Specific beneficial effects include: This invention proposes a high-resolution remote sensing imagery building vector mapping method based on a large model. First, a basic large model of Jilin-1 high-resolution satellite remote sensing imagery is pre-trained using a massive amount of Jilin-1 high-resolution satellite remote sensing imagery combined with a self-supervised learning algorithm. Then, based on this basic large model, a building mask extraction model is fine-tuned using 100,000 building mask samples and supervised learning. Finally, a building corner extraction model is trained using 100,000 manually labeled building instances as samples and building mask cues. This not only improves the generalization ability of the building extraction model but also avoids the creation of complex corner topologies by combining individual building masks with deep learning corner learning. The extracted results can be directly used for building vector mapping, achieving efficient, flexible, and accurate high-resolution satellite remote sensing imagery building vector mapping. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a high-resolution remote sensing image building vector mapping method based on a large model, as described in Implementation Method 1. Figure 2 It is the 0.5-meter resolution remote sensing image of Jilin-1 satellite described in Implementation Method Six; Figure 3 It is the result of manually drawing vectors as described in Implementation Method Six; Figure 4 It is the binarized label described in Implementation Method Six; Figure 5 This is a diagram of the building corner extraction model architecture described in Implementation Method 4; Figure 6 This is a vector drawing rendering of buildings in Jilin City as described in Implementation Method Seven; Figure 7 This is a vector drawing rendering of buildings in Dalian City as described in Implementation Method Seven; Figure 8 This is a vector drawing rendering of buildings in Changchun City as described in Implementation Method Seven; Figure 9 This is a vector drawing rendering of buildings in Hangzhou City as described in Implementation Method Seven; Figure 10 This is a vector drawing rendering of buildings in Yizhou City as described in Implementation Method Seven. Detailed Implementation

[0020] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0021] Implementation Method 1. A method for vector mapping of buildings based on large-scale model-based high-resolution remote sensing images, specifically as follows: Step 1: Collect data from a supervised dataset; Step 2: Build a basic large model based on the self-supervised dataset; Step 3: Construct a building mask dataset, which includes a building vector annotation dataset; Step 4: Based on the basic large model and the building mask dataset, construct a building mask extraction model, and use the building mask extraction model to extract the mask from the remote sensing image to be extracted, so as to obtain the building mask; Step 5: Based on the building mask extraction model and the building vector annotation dataset, construct the building corner extraction model, and use the building corner extraction model to extract features from the building mask to obtain the Gaussian probability map of the corner. Step 6: Based on the Gaussian probability map of the building mask and corner points, obtain the building vector and complete the mapping.

[0022] Existing technologies have the technical problem of being unable to directly obtain building vectors, resulting in a significant difference between the converted building vectors and manually drawn results, and low generalization ability.

[0023] To address the aforementioned technical problems, this embodiment provides a method for vector mapping of buildings based on large-scale remote sensing images, specifically including the following steps: Step 1: Collect data from a supervised dataset; First, based on the 0.75 resolution global imagery of Jilin-1 satellite in 2024, a self-supervised dataset of Jilin-1 was collected over the global land area using a 518-pixel window, resulting in a total of 10 million samples.

[0024] Step 2: Build a basic large model based on the self-supervised dataset; Using the Jilin-1 self-supervised dataset and a contrastive learning algorithm, a basic large-scale model of Jilin-1 with a Vision Transformer backbone was constructed, specifically as follows: The images from the Jilin-1 self-supervised dataset were transformed by random scaling, random cropping, random rotation, and random color enhancement before being input into a Siamese network with Vision Transformer as its backbone. The image features after different data enhancements were used to optimize the model using contrastive learning loss. The entire optimization process iterated for 300 rounds.

[0025] Step 3: Construct a building mask dataset, which includes a building vector annotation dataset; The building mask dataset contains 100,000 samples, of which 10,000 samples contain 100,000 manually annotated building vector instances by experts.

[0026] Step 4: Based on the basic large model and the building mask dataset, construct a building mask extraction model, and use the building mask extraction model to extract the mask from the remote sensing image to be extracted, so as to obtain the building mask; Step 5: Based on the building mask extraction model and the building vector annotation dataset, construct the building corner extraction model, and use the building corner extraction model to extract features from the building mask to obtain the Gaussian probability map of the corner. Step 6: Based on the Gaussian probability map of the building mask and corner points, obtain the building vector and complete the mapping.

[0027] Therefore, this implementation utilizes large-scale Jilin-1 satellite remote sensing imagery to construct a foundational large model, enhancing the stability and generalization ability of the model backbone. It also collects and processes building mask datasets and building vector annotation datasets, training building mask extraction and building corner extraction models respectively. After multiple rounds of optimization, an end-to-end building vector mapping framework based on this large model is obtained, as follows: Figure 1 As shown, it has high practical value and solves the technical problem that existing technologies cannot directly obtain building vectors, resulting in a large gap between the converted building vectors and the manually drawn results, and low generalization.

[0028] Implementation Method 2. This implementation method further defines the high-resolution remote sensing image building vector mapping method based on a large model described in Implementation Method 1. The building mask dataset in step 3 also includes open-source datasets.

[0029] First, a high-resolution satellite remote sensing imagery building mask dataset was collected, sourced from WHUBuilding Dataset, Aerial Imagery Dataset, Massachusetts Duilding Dataset, Vaihingen, Potsdam, and LoveDA. For original samples larger than 800x800 pixels, a 518-pixel sliding window was used to re-crop the samples. For multi-class sample labels, non-building labels were remapped to 0, and building labels were remapped to 1. Finally, an open-source dataset containing 90,000 open-source samples was constructed. Implementation Method 3. This implementation method further defines the high-resolution remote sensing image building vector mapping method based on a large model described in Implementation Method 1. Specifically, the construction of the building mask extraction model in step 4 is as follows: A pre-defined building mask extraction model is used, with a basic large model serving as the encoder and an UpperNet model as the decoder. After data augmentation of the building mask dataset, the data is input into the encoder for encoding. The encoded features output from layers n / 4, n / 2, n*3 / 4, and n of the basic large model are used as the input to the decoder, where n is the encoder layer number. The decoder decodes the encoded features to obtain a mask probability map. Based on the probability map, the pre-defined building mask extraction model is fine-tuned to obtain the final building mask extraction model.

[0030] The building mask extraction model follows an encoder-decoder structure, where the encoder originates from the basic large model backbone in step 2, and the decoder uses the UpperNet model. The training data comes from the 100,000 building mask dataset in step 3. Before being input into the building mask extraction model, the building mask dataset undergoes random combinations of data augmentation, including horizontal flipping, vertical flipping, random rotation, and affine transformation. After encoding by the encoder, the features of the building mask dataset are obtained. The output features from layers n / 4, n / 2, n*3 / 4, and n of the basic large model backbone are used as the input to the decoder, where n is the encoder layer number. Specifically, n=768 for Vision Transformer Base, n=24 for Vision Transformer Large, and n=32 for Vision Transformer Huge. After decoding by the decoder, a probability map of the sample size is obtained. A cross-entropy loss is constructed using the building mask labels and the probability map, and the model weights are optimized along the negative gradient direction using a gradient differentiation algorithm. The building mask model underwent 100 rounds of training and optimization on 100,000 samples.

[0031] Implementation Method 4. This implementation method further defines the high-resolution remote sensing image building vector mapping method based on a large model described in Implementation Method 1. Specifically, the construction of the building corner extraction model in step 5 is as follows: A pre-defined building corner extraction model is used. The pre-defined building corner extraction model is then weighted and initialized using a building mask extraction model. Vector labels from the building vector annotation dataset are obtained. The circumscribed horizontal rectangle of the vector labels is normalized. The top-left and bottom-right points of the normalized circumscribed horizontal rectangle are encoded to obtain the encoded features of the circumscribed horizontal rectangle of the vector labels. Based on the encoded features output by the building mask extraction model and the encoded features of the circumscribed horizontal rectangle of the vector labels, a Gaussian heatmap of the corners is obtained. The pre-defined building corner extraction model is then fine-tuned based on the Gaussian heatmap of the corners to obtain the final building corner extraction model.

[0032] Building upon step 4, a building corner extraction model based on building mask cues was further constructed. The architecture of the building corner extraction model is as follows: Figure 5 As shown. The building corner extraction model includes a VisionTransformer feature extraction backbone network for extracting sample image features; a cue box encoding structure for encoding the outer horizontal bounding box of the building; a cross-attention fusion structure for fusing multi-level sample image features and cue features; and a decoding structure for decoding the fused multi-level features. Taking a 518×518×3 (i.e., image length × image width × number of bands) sample as an example, the calculation process of the building corner extraction model is as follows: Using a building mask, the image range of a single building is cropped out. During the cropping process, the range of a single building is randomly expanded, and the cropped sample image is sampled to 196×196 using bilinear interpolation. The sample labels are processed in the same way.

[0033] Using the building mask model weights obtained in step 4, the VisionTransformer in the building corner point extraction model is used. Figure 5 The layers (Encoder Layer n / 4, Encoder Layer n / 2, Encoder Layer n*3 / 4, and Encoder Layer n) in the UpperNet module are initialized with weights. Layers with mismatched weight shapes are initialized using bilinear interpolation resampling.

[0034] The Vision Tranformer backbone performs feature encoding on the sample images, obtaining feature codes from four layers: Encoder Layer n / 4, Encoder Layer n / 2, Encoder Layer n*3 / 4, and Encoder Layer n. The encoding shape is (embedding_dim, 14, 14), where embedding_dim represents the encoding dimension. The embedding_dim of Vision Tranformer Base is 768, the embedding_dim of Vision Tranformer Large is 1024, and the embedding_dim of Vision Tranformer Huge is 1280.

[0035] Using the sample vector labels, calculate the minimum and maximum x and y coordinates of the vector coordinates to obtain the circumscribed horizontal rectangle of the vector labels. ,in, The coordinates of the top right corner of the horizontal frame are the image coordinates. Defines the width and height of the horizontal frame. The bounding horizontal rectangle of the vector label. Normalization will be performed, that is... Mapped to , Mapped to ,in, Let the width and height of the sample image be given. Then, cosine position encoding is used to determine the top left point of the normalized circumscribed horizontal rectangle. and the bottom right point Encode to obtain eigenvectors.

[0036] The four-layer feature encoding of the sample image is processed by a combined module of Convolution(embedding_dim, 256, Kernel=1) + Batch Normalization + ReLU to obtain four ( The features of each dimension are then cross-attention calculated with the encoded features of the horizontal rectangle bounding the building vector, resulting in four enhanced features. Dimensional features.

[0037] Four enhanced The dimensional features are input into the UpperNet module for feature fusion and decoding, resulting in... The corner Gaussian heatmap after size decoding is then resampled using bilinear interpolation. .

[0038] The coordinates of the points constituting the building vector are converted into a Gaussian heatmap using the OpenCV library and used as labels, which are then compared with the data obtained from the previous decoding step. The Gaussian heatmap for corner points was used, the model loss was calculated using the cross-entropy loss function, and the model parameters were differentiated using the chain rule. Finally, the model weights were updated and optimized using the stochastic gradient descent algorithm. The building corner point extraction model underwent 100 rounds of training and optimization on 10,000 samples and 100,000 building instances.

[0039] Implementation Method 5. This implementation method further defines the high-resolution remote sensing image building vector mapping method based on a large model described in Implementation Method 1. Specifically, in step 6, the building vector is obtained based on the Gaussian probability map of the building mask and corner points. The Gaussian probability map of the corner points is filtered, and the pixel positions with the same pixel value before and after filtering are taken as corner points. The nearest pixel on the boundary of the building mask to each corner point is obtained, and the pixel points are connected to obtain the building vector.

[0040] First, the building mask extraction model obtained in step 4 is used for inference to obtain the building mask. Then, the building corner extraction model obtained in step 5 is used to calculate the Gaussian probability map of the corner from the building mask and the image. The Gaussian probability map is filtered using a MaxPooling layer with a window size of 5x5. The pixel positions that are equal to those before filtering are the corners. Finally, the nearest pixel of the corner to the building mask is calculated. The connection order between the nearest pixels is the corner topology. After connection, the building vector is obtained.

[0041] Implementation Method Six. This implementation method further defines the high-resolution remote sensing image building vector mapping method based on a large model described in Implementation Method One or Four. Specifically, the building vector annotation dataset is as follows: High-resolution satellite remote sensing image samples were manually delineated, and the vector labels of the delineated high-resolution satellite remote sensing image samples were converted into raster labels. The high-resolution satellite remote sensing image samples were then cropped to obtain a building vector annotation dataset.

[0042] The building vector annotation dataset was constructed from manual drawings by professionals with over 3 years of experience in the field of remote sensing. The drawing process followed these principles: ① Building roofs were the primary targets for delineation; ② Only key nodes constituting the building roof were delineated; ③ The dataset included various types of building roof forms, and roofs larger than 100 square meters were all delineated. Specifically: High-resolution remote sensing imagery from the Jilin-1 satellite with a resolution of 0.5 meters was selected as the sample, covering urban areas with different geographical distributions across the country. In terms of building morphology, the sample included various building types such as commercial buildings, residential buildings, factories, stadiums, and rural buildings. Figure 2 As shown. During the drawing process, the vector lines follow the outline of the building's roof, retaining only the key nodes that constitute the roof and avoiding meaningless redundant nodes, such as... Figure 3 As shown. After the drawing is completed, a vector-to-raster algorithm is used to convert the vector labels into raster labels, where the pixel values ​​of building labels are mapped to 1, and the background label values ​​are mapped to 0, as shown. Figure 4 As shown, transparent represents the background and red represents the buildings. The samples were re-cropped using a 518-pixel sliding window to obtain 10,000 samples and 100,000 building vector instances.

[0043] In summary, the proposed high-resolution remote sensing image building vector mapping method based on a large model, built upon a large-scale, high-resolution Jilin-1 satellite remote sensing image pre-trained basic model, collected a dataset of 90,000 open-source remote sensing image building masks, and manually created a dataset of 10,000 vector-annotated buildings based on Jilin-1 satellite remote sensing imagery for fine-tuning the building mask extraction model. Building upon the building mask extraction model, a corner learning algorithm for mask cues is constructed using the 10,000 vector-annotated building dataset containing 100,000 building instances, thereby directly obtaining structured building vectors for building vector mapping.

[0044] Implementation Method Seven. This implementation method, based on the high-resolution remote sensing image building vector mapping method described in Implementation Methods One to Six, trained for 720 hours on 10 million samples acquired from the Jilin-1 global imagery satellite at 0.75 resolution using 8 A100 graphics cards, for 48 hours on 100,000 building mask samples, and for 24 hours on 100,000 building vector instances. In terms of inference efficiency, a single RTX 3090 GPU can map 144 building instances.

[0045] like Figure 6 , Figure 7 , Figure 8 , Figure 9 and Figure 10 As shown, this implementation method was tested in multiple regions including Jilin City, Dalian City, Changchun City, Hangzhou City, and Yizhou City, using vector mapping of buildings. The images were sourced from Jilin-1 satellite remote sensing images with a resolution of 0.5 meters. The test results showed excellent visualization mapping effects, with significant effects on the fit of building roofs, the construction of key corner points, and the suppression of redundant nodes, demonstrating strong applicability.

[0046] Implementation Method 8. This implementation method describes a high-resolution remote sensing image building vector mapping system based on a large model. The system is constructed according to the method described in Implementation Method 1 and includes the following modules: The basic data acquisition module collects data from supervised datasets; The basic model building module constructs a basic large model based on a self-supervised dataset. The mask data acquisition module constructs a building mask dataset, which includes a building vector annotation dataset. The mask model construction module, based on the basic large model and the building mask dataset, constructs a building mask extraction model. The building mask extraction model is then used to extract the mask from the remote sensing image to be extracted, thus obtaining the building mask. The corner model construction module constructs a building corner extraction model based on the building mask extraction model and the building vector annotation dataset. It then uses the building corner extraction model to extract features from the building mask and obtains the Gaussian probability map of the corner. The vector mapping module uses Gaussian probability maps of building masks and corner points to obtain building vectors and complete the mapping.

[0047] The above provides a detailed description of the method and system for high-resolution remote sensing image building vector mapping based on a large model proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for vector mapping of buildings based on large-scale remote sensing images, characterized in that, Specifically: Step 1: Collect data from a supervised dataset; Step 2: Build a basic large model based on the self-supervised dataset; Step 3: Construct a building mask dataset, which includes a building vector annotation dataset; Step 4: Based on the basic large model and the building mask dataset, construct a building mask extraction model, and use the building mask extraction model to extract the mask from the remote sensing image to be extracted, so as to obtain the building mask; Step 5: Based on the building mask extraction model and the building vector annotation dataset, construct the building corner extraction model, and use the building corner extraction model to extract features from the building mask to obtain the Gaussian probability map of the corner. Step 6: Based on the Gaussian probability map of the building mask and corner points, obtain the building vector and complete the mapping.

2. The method for vector mapping of buildings based on large-scale remote sensing images according to claim 1, characterized in that, The building mask dataset in step 3 also includes open-source datasets.

3. The method for vector mapping of buildings based on large-scale remote sensing images according to claim 1, characterized in that, The construction of the building mask extraction model in step 4 is specifically as follows: A pre-defined building mask extraction model is used, with the basic large model as the encoder and the UpperNet model as the decoder. The pre-defined building mask extraction model is then fine-tuned based on the building mask dataset to obtain the final building mask extraction model.

4. The method for vector mapping of buildings based on large-scale remote sensing images according to claim 3, characterized in that, The fine-tuning of the preset building mask extraction model based on the building mask dataset specifically involves: After data augmentation of the building mask dataset, the data is input into the encoder for encoding. The encoded features output from the n / 4, n / 2, n*3 / 4 and nth layers of the basic large model are used as the input to the decoder, where n is the encoder layer number. The decoder decodes the encoded features to obtain the mask probability map. Based on the probability map, the preset building mask extraction model is fine-tuned to obtain the building mask extraction model.

5. The method for vector mapping of buildings based on large-scale remote sensing images according to claim 1, characterized in that, The construction of the building corner extraction model in step 5 is specifically as follows: A pre-defined building corner extraction model is used. The weights of the pre-defined building corner extraction model are initialized using a building mask extraction model. The initialized pre-defined building corner extraction model is then fine-tuned based on a building vector annotation dataset to obtain the final building corner extraction model.

6. The method for vector mapping of buildings based on large-scale remote sensing images according to claim 5, characterized in that, The fine-tuning of the preset building corner extraction model based on the building vector annotation dataset is as follows: Obtain vector labels from the building vector annotation dataset, obtain the encoding features of the horizontal rectangle enclosing the vector labels, extract the encoding features output by the building mask extraction model, and obtain the Gaussian heatmap of the corner points based on the encoding features of the horizontal rectangle enclosing the vector labels. Fine-tune the preset building corner extraction model based on the Gaussian heatmap of the corner points to obtain the building corner extraction model.

7. The method for vector mapping of buildings based on large-scale remote sensing images according to claim 6, characterized in that, The process of obtaining vector labels from the building vector annotation dataset and deriving the encoding features of the bounding horizontal rectangle of the vector labels specifically involves: The bounding horizontal rectangle of the vector label is normalized, and the top-left and bottom-right points of the normalized bounding horizontal rectangle are encoded to obtain the encoding features of the bounding horizontal rectangle of the vector label.

8. The method for vector mapping of buildings based on large-scale remote sensing images according to claim 1, characterized in that, In step 6, the building vector is obtained based on the Gaussian probability map of the building mask and corner points, specifically as follows: The Gaussian probability map of the corner points is filtered, and the pixel positions with the same pixel value before and after filtering are taken as corner points. The nearest pixel on the boundary of the building mask to each corner point is obtained, and the pixel points are connected to obtain the building vector.

9. A method for vector mapping of buildings based on large-scale remote sensing images according to claim 1, 5, or 6, characterized in that, The aforementioned building vector annotation dataset is specifically as follows: High-resolution satellite remote sensing image samples were manually delineated, and the vector labels of the delineated high-resolution satellite remote sensing image samples were converted into raster labels. The high-resolution satellite remote sensing image samples were then cropped to obtain a building vector annotation dataset.

10. A high-resolution remote sensing image building vector mapping system based on a large model, wherein the system is constructed according to the method described in claim 1, characterized in that, Includes the following modules: The basic data acquisition module collects data from supervised datasets; The basic model building module constructs a basic large model based on a self-supervised dataset. The mask data acquisition module constructs a building mask dataset, which includes a building vector annotation dataset. The mask model construction module, based on the basic large model and the building mask dataset, constructs a building mask extraction model. The building mask extraction model is then used to extract the mask from the remote sensing image to be extracted, thus obtaining the building mask. The corner model construction module constructs a building corner extraction model based on the building mask extraction model and the building vector annotation dataset. It then uses the building corner extraction model to extract features from the building mask and obtains the Gaussian probability map of the corner. The vector mapping module uses Gaussian probability maps of building masks and corner points to obtain building vectors and complete the mapping process.

Images