Satellite image building intelligent identification method and system based on deep learning and GIS

Abstract

The invention relates to the technical field of building identification, and particularly discloses a satellite image building intelligent identification method and system based on deep learning and GIS, and the method comprises the steps: obtaining high-resolution satellite image data through a remote sensing satellite, capturing geographic space data corresponding to a satellite image from the Internet through employing a crawler technology, and carrying out the recognition of the high-resolution satellite image data; the collected data are input into a GIS database; the method comprises the following steps: performing coordinate mapping and region division on a satellite image by using a GIS system, and generating an image segmentation region with a geographic position identifier in combination with existing urban planning and infrastructure data; according to the invention, the improved Mask R-CNN network and the depth point convolutional network are introduced, so that the positioning, segmentation and classification of the building can be carried out more accurately. The spatial analysis function of the GIS system is combined with deep learning, so that the generation of the building distribution diagram is accurate, and the urban expansion and building growth trend can be effectively predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of building recognition, and particularly relates to an intelligent building recognition method and system for satellite images based on deep learning and GIS. Background Art

[0002] With the continuous development of remote sensing technology and deep learning, using satellite image data for building recognition and analysis has become an important application in the fields of urban planning, environmental monitoring, disaster management, etc. Satellite images have become an important data source for urban building recognition due to their wide coverage and convenient data acquisition. Traditional building recognition methods mostly rely on manual annotation and image processing techniques, which have problems such as high annotation costs, low efficiency, and insufficient accuracy.

[0003] With the rise of deep learning technology, especially the application of convolutional neural networks and region-based convolutional neural networks (such as Mask R-CNN), the accuracy and automation level of building recognition have been significantly improved. However, existing building recognition methods are usually limited to simple image segmentation or object detection, and it is difficult to make full use of the complex geographical information and spatio-temporal data in satellite images for efficient building recognition, classification, and distribution analysis. In addition, existing technical methods generally lack the ability of effective multi-source data fusion and deep feature extraction, which leads to low recognition accuracy of buildings in complex urban environments and cannot meet the requirements of generating accurate building distribution maps and intelligent urban planning. And the building analysis method based on GIS (Geographic Information System) usually lacks the combination with deep learning models, resulting in limited prediction accuracy of building distribution and urban expansion. The key technical problems to be solved in building recognition based on satellite images include the processing of image quality, the diversity and complex shapes of buildings, the effective integration of spatio-temporal data, etc.

[0004] Therefore, it is necessary to propose an intelligent building recognition method and system for satellite images based on deep learning and GIS to solve the problem of low accuracy in building recognition in complex urban environments existing in the prior art.

[0005] The above information disclosed in this background art is only used to increase the understanding of the background art of the present invention. Therefore, it may include prior art that is not known to those of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent building recognition method and system for satellite images based on deep learning and GIS to solve the problems raised in the above background art.

[0007] To achieve the above purpose, the present invention provides the following technical solutions:

[0008] A satellite image building intelligent recognition method based on deep learning and GIS includes the following steps:

[0009] Obtain high-resolution satellite image data through a remote sensing satellite, and use web crawler technology to capture geospatial data corresponding to the satellite image from the Internet;

[0010] Based on the geospatial data, use the GIS system to perform coordinate mapping and regional division on the satellite image, and combine the existing urban planning and infrastructure data to generate an image segmentation area with geographical location identification;

[0011] Use the improved Mask R-CNN network to locate and segment buildings in the image segmentation area, and generate a building area segmentation result;

[0012] Based on the building area segmentation result, use the deep point convolution network to extract the deep features of the building shape, identify and classify the buildings, and generate corresponding identification and classification results;

[0013] Use the spatial analysis function of the GIS system to perform refined analysis on the identification and classification results, and generate a building distribution map of the current image segmentation area;

[0014] Use the GIS system to perform spatial coordination and synthesis processing on the building distribution maps of all generated image segmentation areas to obtain a complete building distribution map.

[0015] Preferably, the step of obtaining high-resolution satellite image data through a remote sensing satellite and using web crawler technology to capture geospatial data corresponding to the satellite image from the Internet includes:

[0016] Introduce an ORB feature point-based algorithm to achieve precise alignment between different satellite images and correct the geometric distortion of the satellite image;

[0017] Use a denoising encoder model based on a convolutional neural network to remove noise in the satellite image while retaining the structural details of the building area;

[0018] Use a generative adversarial network to synthesize diverse satellite images for data augmentation, and combine self-adaptive histogram equalization to enhance the contrast of the satellite image.

[0019] Preferably, the step of using the GIS system to perform coordinate mapping and regional division on the satellite image includes:

[0020] Combine the region proposal network of Mask R-CNN to generate candidate regions for image segmentation, and optimize the accuracy of the candidate region proposals for image segmentation through online learning;

[0021] Combined with the Feature Pyramid Network, fuse satellite image features at different scales to accurately detect candidate regions for image segmentation from small to large;

[0022] Utilize the Convolutional Long Short-Term Memory Network to capture the spatio-temporal patterns of changes in candidate regions for image segmentation through spatio-temporal analysis of time series in space.

[0023] Preferably, use the improved Mask R-CNN network to locate and segment buildings in the image segmentation region, generating the building region segmentation result, including:

[0024] Input the image segmentation region into the pre-trained EfficientNet backbone network to extract regional image features, and generate corresponding feature maps through the FPN network;

[0025] Use the Region Proposal Network to perform sliding window scanning on the feature maps to find regions where buildings may exist and generate candidate regions for buildings;

[0026] The loss function formula of the Region Proposal Network:

[0027]

[0028] In the formula, i is the index of the anchor points in the training batch, is the loss weight, p i is the class prediction value corresponding to each anchor point, p i ′ is the class label, t i is the vector of the coordinates of the 4 corner points of the predicted bounding box, t i ′ is the vector of the corner point coordinates corresponding to the labeled bounding box, L f is the classification loss, using cross-entropy loss, N f is the number of anchor points for the classification loss, L r is the regression loss, using SmoothL1 loss, N r is the number of anchor points for the regression loss;

[0029] Use the Softmax classifier combined with non-maximum suppression to perform binary classification and screening on the candidate regions of buildings to obtain the regions of interest;

[0030] Input the feature maps and the regions of interest into ROIAlign to generate aligned feature maps of a fixed size for each region of interest;

[0031] Input the aligned feature maps into the localization branch, and perform localization regression of the bounding box for the building through the fully connected layer to obtain the building position;

[0032] Input the aligned feature map and building location into the segmentation branch, and perform pixel-level segmentation on each building through per-pixel classification to generate a binary mask of the building;

[0033] The loss function of the localization branch and the segmentation branch is composed of the formula:

[0034] L loss = L f + L r + L mask

[0035] In the formula, L f is the classification loss, using cross-entropy loss, L r is the regression loss, using Smooth L1 loss, L mask is the segmentation loss, using binary cross-entropy loss;

[0036] Process the binary mask of the building using bilinear interpolation to restore it to the same dimension as the image segmentation region;

[0037] Extract the building area from the image segmentation region using the restored binary mask of the building to generate the building area segmentation result.

[0038] Preferably, based on the building area segmentation result, use a depth point convolutional network to extract the depth features of the building shape, identify and classify the building, and generate the corresponding identification and classification results, including:

[0039] Combine the depth residual network of ResNet-101 or DenseNet to extract the depth features of the building shape from the building area segmentation result;

[0040] Among them, use the convolutional network formula to classify and associate analyze the building, and optimize the effect of the depth point convolutional network in processing the adjacent relationship of the building;

[0041] H (l+1) = σ(AH (l) W (l) )

[0042] In the formula, H( l ) is the node feature matrix of the l-th layer, A is the normalized adjacency matrix of the graph, W( l ) is the learning weight matrix of each layer, and σ is the activation function;

[0043] Combine self-supervised learning to perform unsupervised feature learning on the building image, and optimize the recognition effect of the graph convolutional network on buildings with complex shapes;

[0044] Based on the depth features of the extracted building shapes, a support vector machine is used to identify and classify the buildings, generating corresponding identification and classification results.

[0045] Preferably, using the spatial analysis function of the GIS system, a refined analysis is performed on the identification and classification results to generate a building distribution map of the current image segmentation area, including:

[0046] Based on the identification and classification results, a clustering analysis of the spatial distribution of buildings is performed using the DBSCAN or K-means clustering algorithm to identify the building density and distribution characteristics of different regions;

[0047] Combined with a temporal convolutional neural network for predicting the distribution of regional buildings, by analyzing the temporal data of urban expansion, predicting the building growth pattern and trend, and generating a prediction result;

[0048] Temporal convolutional neural network prediction formula:

[0049]

[0050] In the formula, y t is the predicted building growth pattern, x t-j is the input feature at time t-j, ω j is the convolutional kernel, and f is the activation function;

[0051] According to the prediction result and combined with the building density and distribution characteristics of different regions, a building distribution map of the current image segmentation area is generated through the GIS system.

[0052] The satellite image building intelligent recognition system based on deep learning and GIS includes:

[0053] A data management module for obtaining high-resolution satellite image data through a remote sensing satellite and scraping geospatial data corresponding to the satellite image from the Internet using web crawler technology;

[0054] An image segmentation module for performing coordinate mapping and regional division on the satellite image using the GIS system based on the geospatial data, and generating an image segmentation area with geographical location identifiers in combination with existing urban planning and infrastructure data;

[0055] A building detection module for locating and segmenting the buildings in the image segmentation area using an improved MaskR-CNN network to generate a building area segmentation result;

[0056] A building classification module for extracting the depth features of the building shapes based on the building area segmentation result using a depth point convolutional network to identify and classify the buildings, generating corresponding identification and classification results;

[0057] A building space analysis module, which is used to utilize the spatial analysis function of the GIS system to conduct refined analysis on the recognition and classification results, generate a building distribution map of the current image segmentation area, and predict the building growth trend;

[0058] A building distribution synthesis module, which is used to utilize the GIS system to perform spatial coordination and synthesis processing on the building distribution maps of all generated image segmentation areas to obtain a complete building distribution map.

[0059] Preferably, the building detection module further includes:

[0060] A generative adversarial network, which is used to synthesize diverse satellite images for data augmentation and enhance the image contrast by combining self-adaptive histogram equalization;

[0061] Combine the region proposal network of Mask R-CNN to generate candidate regions for image segmentation, and optimize the accuracy of candidate regions through online learning;

[0062] Combine the feature pyramid network and the convolutional long short-term memory network to perform fusion and spatio-temporal pattern analysis on satellite image features at different scales, and accurately detect buildings in the image segmentation area.

[0063] Preferably, the building classification module further includes:

[0064] Combine the deep residual network of ResNet-101 or DenseNet to extract the deep features of the building shape from the building area segmentation results;

[0065] Among them, use a convolutional network to conduct classification and correlation analysis on buildings, and optimize the effect of the deep point convolutional network in processing the adjacent relationship of buildings;

[0066] Combine self-supervised learning to perform unsupervised feature learning on building images, and optimize the recognition effect of the graph convolutional network on buildings with complex shapes.

[0067] Preferably, the building space analysis module further includes:

[0068] Conduct clustering analysis of building distributions based on the DBSCAN or K-means clustering algorithm to identify the building density and distribution characteristics in different regions;

[0069] Predict the building growth pattern and trend through a temporal convolutional neural network for the temporal data of urban expansion, and generate a building distribution map.

[0070] Compared with the prior art, the beneficial effects of the present invention are:

[0071] The present invention introduces an improved Mask R-CNN network and a depth point convolution network to more accurately locate, segment, and classify buildings. Through multi-source data fusion and deep feature extraction, the accuracy of building recognition is improved. By combining the spatial analysis function of the GIS system with deep learning, the generation of building distribution maps is not only accurate but also can effectively predict urban expansion and building growth trends, providing data support for urban planning and management. In addition, through the application of advanced algorithms such as temporal convolutional neural networks and graph convolutional networks, the ability to handle complex building shapes and adjacent relationships is enhanced, achieving the efficiency and accuracy of building recognition and distribution analysis. Brief Description of the Drawings

[0072] Figure 1 It is a flowchart of the intelligent building recognition method for satellite images based on deep learning and GIS of the present invention;

[0073] Figure 2 It is a framework diagram of the intelligent building recognition system for satellite images based on deep learning and GIS of the present invention. Detailed Embodiments

[0074] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0075] Embodiment 1:

[0076] Please refer to Figure 1 As shown, the intelligent building recognition method for satellite images based on deep learning and GIS includes the following steps:

[0077] Obtain high-resolution satellite image data through a remote sensing satellite, use web crawler technology to scrape geospatial data corresponding to the satellite images from the Internet, and input the collected data into the GIS database;

[0078] Among them, introduce an algorithm based on ORB feature points to achieve precise alignment between different satellite images and correct the geometric distortion of the satellite images;

[0079] Use a denoising encoder model based on a convolutional neural network to remove noise in the satellite images while retaining the structural details of the building areas;

[0080] Use a generative adversarial network to synthesize diverse satellite images for data augmentation, and combine self-adaptive histogram equalization to enhance the contrast of the satellite images.

[0081] Furthermore, by integrating technologies such as remote sensing satellite imagery, geospatial data, ORB feature point algorithm, convolutional neural network denoising, and generative adversarial network data augmentation, precise alignment between different satellite images is achieved, geometric distortion is corrected, noise is removed while retaining details of the building areas, and at the same time, through data augmentation and contrast enhancement, the clarity and distinguishability of the images are improved, thus significantly enhancing the accuracy and stability of building recognition.

[0082] Based on geospatial data, the satellite imagery is subjected to coordinate mapping and regional division using a GIS system, and by combining existing urban planning and infrastructure data, image segmentation regions with geographical location identifiers are generated.

[0083] Among them, in combination with the region proposal network of Mask R-CNN, candidate regions for image segmentation are generated, and the accuracy of the candidate region proposals for image segmentation is optimized through online learning.

[0084] In combination with the feature pyramid network, fusion is performed on satellite image features at different scales to accurately detect candidate regions for image segmentation from small to large.

[0085] Using a convolutional long short-term memory network, through spatio-temporal analysis in space, the spatio-temporal patterns of changes in candidate regions for image segmentation are captured.

[0086] Furthermore, by integrating a GIS system, urban planning data, and advanced deep learning algorithms, precise image segmentation and regional recognition are achieved in high-resolution satellite imagery. By optimizing the segmentation accuracy through online learning, fusing features at different scales, and capturing spatio-temporal change patterns, the accuracy and robustness of image segmentation are significantly improved, thus more accurately identifying and annotating regions such as buildings, enhancing the application value and practicality of satellite imagery data.

[0087] An improved Mask R-CNN network is used to locate and segment buildings in the image segmentation regions, generating building region segmentation results.

[0088] The image segmentation regions are input into a pre-trained EfficientNet backbone network to extract regional image features, and corresponding feature maps are generated through an FPN network.

[0089] The region proposal network is used to perform a sliding window scan on the feature maps to find regions where buildings may exist, generating candidate regions for buildings.

[0090] A Softmax classifier combined with non-maximum suppression is used to perform binary classification and screening on the candidate regions for buildings to obtain regions of interest.

[0091] Input the feature map and the region of interest into ROIAlign to generate an aligned feature map of a fixed size for each region of interest;

[0092] Input the aligned feature map into the localization branch, and perform localization regression of the bounding box of the building through a fully connected layer to obtain the building location;

[0093] Input the aligned feature map and the building location into the segmentation branch, and perform pixel-level segmentation on each building through per-pixel classification to generate a binary mask of the building;

[0094] Process the binary mask of the building using bilinear interpolation to restore it to the same dimension as the image segmentation region;

[0095] Extract the building area from the image segmentation region using the restored binary mask of the building to generate the building area segmentation result.

[0096] Furthermore, by combining the improved Mask R-CNN network with the EfficientNet and FPN networks, high-precision object detection and pixel-level segmentation are achieved in the localization and segmentation tasks of buildings. This method has high accuracy and robustness in building extraction and segmentation, provides refined building area segmentation results, and is widely applicable to fields such as urban planning and building monitoring.

[0097] Based on the building area segmentation result, use a deep point convolutional network to extract the deep features of the building shape, identify and classify the building, and generate the corresponding identification and classification results;

[0098] Combine the deep residual network of ResNet-101 or DenseNet to extract the deep features of the building shape from the building area segmentation result;

[0099] Among them, use the convolutional network formula to classify and perform correlation analysis on the building, and optimize the effect of the deep point convolutional network in processing the adjacent relationship of the building;

[0100] Combine self-supervised learning to perform unsupervised feature learning on the building image, and optimize the recognition effect of the graph convolutional network on buildings with complex shapes;

[0101] Based on the extracted deep features of the building shape, use a support vector machine to identify and classify the building, and generate the corresponding identification and classification results.

[0102] Furthermore, by combining a depth point convolutional network, a ResNet-101 or DenseNet deep residual network, and a graph convolutional network, deep features of building shapes are efficiently extracted for accurate recognition and classification. By optimizing the graph convolutional network to handle the adjacent relationships of buildings and applying self-supervised learning, the method improves the recognition ability for complex building shapes. Meanwhile, a support vector machine is used to classify buildings, further enhancing the recognition accuracy and robustness.

[0103] Using the spatial analysis function of the GIS system, the recognition and classification results are refined and analyzed to generate a building distribution map of the current image segmentation area;

[0104] Based on the recognition and classification results, the DBSCAN or K-means clustering algorithm is used to perform clustering analysis on the spatial distribution of buildings to identify the building density and distribution characteristics in different regions;

[0105] Combined with a temporal convolutional neural network for regional building distribution prediction, by analyzing the temporal data of urban expansion, the building growth pattern and trend are predicted to generate prediction results;

[0106] According to the prediction results and combined with the building density and distribution characteristics in different regions, a building distribution map of the current image segmentation area is generated through the GIS system.

[0107] Furthermore, by combining the spatial analysis function of the GIS system with the clustering algorithm, a refined analysis of the building recognition and classification results is achieved, effectively identifying the building density and distribution characteristics in different regions. Meanwhile, a temporal convolutional neural network is used for temporal prediction of urban building distribution, accurately predicting the building growth pattern and trend. Finally, a building distribution map is generated through the GIS system to help urban planning and management personnel understand the urban expansion dynamics, optimize resource allocation and planning decisions, which has important practical application value.

[0108] The GIS system is used to perform spatial coordination and synthesis processing on the building distribution maps of all generated image segmentation areas to obtain a complete building distribution map.

[0109] Example 2:

[0110] Please refer to Figure 2 As shown, the satellite image building intelligent recognition system based on deep learning and GIS includes:

[0111] A data management module, which is used to obtain high-resolution satellite image data through a remote sensing satellite, crawl geospatial data corresponding to the satellite image from the Internet using web crawler technology, and input the collected data into the GIS database;

[0112] An image segmentation module, which is used to perform coordinate mapping and regional division on satellite images using a GIS system based on geospatial data, and combine existing urban planning and infrastructure data to generate image segmentation regions with geographical location identifiers;

[0113] A building detection module, which is used to locate and segment buildings in the image segmentation region using an improved Mask R-CNN network to generate a building region segmentation result;

[0114] A generative adversarial network, which is used to synthesize diverse satellite images for data augmentation and enhance the image contrast by combining self-adaptive histogram equalization;

[0115] Combine the region proposal network of Mask R-CNN to generate candidate regions for image segmentation, and optimize the accuracy of candidate regions through online learning;

[0116] Combine a feature pyramid network and a convolutional long short-term memory network to perform fusion and spatio-temporal pattern analysis on satellite image features at different scales, and accurately detect buildings in the image segmentation region.

[0117] A building classification module, which is used to extract deep features of building shapes based on the building region segmentation result using a deep point convolutional network, identify and classify buildings, and generate corresponding identification and classification results;

[0118] Combine the deep residual network of ResNet-101 or DenseNet to extract deep features of building shapes from the building region segmentation result;

[0119] Among them, use a convolutional network to perform classification and correlation analysis on buildings, and optimize the effect of the deep point convolutional network in processing building adjacent relationships;

[0120] Combine self-supervised learning to perform unsupervised feature learning on building images, and optimize the recognition effect of the graph convolutional network on buildings with complex shapes.

[0121] A building spatial analysis module, which is used to perform refined analysis on the recognition and classification results using the spatial analysis function of the GIS system, generate a building distribution map of the current image segmentation region, and predict the building growth trend;

[0122] Perform clustering analysis of building distribution based on the DBSCAN or K-means clustering algorithm to identify the building density and distribution characteristics in different regions;

[0123] Predict the building growth pattern and trend of the time series data of urban expansion through a temporal convolutional neural network to generate a building distribution map;

[0124] A building distribution synthesis module is used to perform spatial coordination and synthesis processing on the building distribution maps of all generated image segmentation regions using a GIS system to obtain a complete building distribution map.

[0125] Application example: Employees of a geography-related enterprise collect geographical image data and perform building recognition

[0126] I. Employee roles and tasks

[0127] As an employee of a geography-related enterprise, Xiao Li is assigned the task of collecting geographical image data and performing building recognition and classification. He needs to use the "Intelligent Satellite Image Building Recognition System Based on Deep Learning and GIS" provided by the company to complete this work.

[0128] II. Workflow

[0129] Xiao Li first obtains high-resolution satellite image data using a remote sensing satellite. To ensure the comprehensiveness and accuracy of the data, he also uses web crawler technology to scrape geographical spatial data corresponding to the satellite images from the Internet, such as topographic maps, urban planning maps, etc. These data will be input into the GIS database to provide a basis for subsequent processing and analysis.

[0130] With the support of the GIS system, Xiao Li preprocesses the collected satellite image data. He uses the coordinate mapping and region division functions of the GIS system, combined with existing urban planning and infrastructure data, to generate image segmentation regions with geographical location identifiers. This helps to divide the complex satellite image data into smaller regions that are easier to process and analyze.

[0131] Next, Xiao Li uses the improved Mask R-CNN network in the building detection module of the system to locate and segment the buildings in the image segmentation regions. This can accurately identify the building regions in the image and generate building region segmentation results.

[0132] Based on the building region segmentation results, Xiao Li uses the deep point convolutional network in the building classification module to extract the deep features of the building shapes and perform building recognition and classification. The system can classify the buildings into different types according to their features such as shape, size, material, etc., such as residential buildings, commercial buildings, public facilities, etc.

[0133] After the building recognition and classification are completed, Xiao Li uses the spatial analysis function of the GIS system to perform refined analysis on the recognition and classification results. He uses the DBSCAN or K-means clustering algorithm to perform clustering analysis on the spatial distribution of the buildings to identify the building density and distribution characteristics in different regions. At the same time, he also combines a temporal convolutional neural network to predict the building growth patterns and trends of the temporal data of urban expansion and generate a building distribution map.

[0134] Finally, Xiao Li used the building distribution synthesis module in the system to perform spatial coordination and synthesis processing on the building distribution maps of all the generated image segmentation regions, obtaining a complete building distribution map. He presented this map to the company's decision-making level, providing a scientific basis and data support for urban planning, environmental monitoring, etc.

[0135] III. Work Results and Impacts

[0136] By applying the satellite image building intelligent recognition method and system based on deep learning and GIS, Xiao Li not only efficiently completed the tasks of collecting geographic image data and performing building recognition and classification, but also provided the company with accurate and reliable building distribution maps and growth trend prediction results. These results are of great significance for the company's business development and decision-making, and help improve the company's competitiveness in the fields related to geography.

[0137] Example 3:

[0138] The embodiment of the present invention also provides a computer-readable storage medium, on which a program of the satellite image building intelligent recognition system based on deep learning and GIS as described in any one of the above is stored. When the program is executed by a processor, it realizes each process of the above intelligent recognition system embodiment and can achieve the same technical effects. To avoid repetition, it will not be elaborated here. Among them, the computer-readable storage medium, such as a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc, etc.

[0139] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without conflict, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0140] In the attached drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments of the present invention are involved, and other structures can refer to the general design. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other.

[0141] The flowcharts shown in the drawings are merely illustrative examples and do not necessarily include all contents and operations / steps, nor are they necessarily executed in the order described. For example, some operations / steps can be decomposed, combined or partially merged, so the actual execution order may be changed according to the actual situation.

[0142] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent recognition method for satellite image buildings based on deep learning and GIS, characterized in that, Including the following steps: Obtain high-resolution satellite image data through remote sensing satellites, and use web crawler technology to scrape geospatial data corresponding to the satellite images from the Internet; Based on the geospatial data, use the GIS system to perform coordinate mapping and regional division on the satellite images, and combine the existing urban planning and infrastructure data to generate image segmentation regions with geographical location identifiers; Use the improved Mask R-CNN network to locate and segment the buildings in the image segmentation regions, and generate the building region segmentation results; Based on the building region segmentation results, use the deep point convolutional network to extract the depth features of the building shapes, identify and classify the buildings, and generate the corresponding identification and classification results; Use the spatial analysis function of the GIS system to perform refined analysis on the identification and classification results, and generate the building distribution map of the current image segmentation region; Use the GIS system to perform spatial coordination and synthesis processing on the building distribution maps of all generated image segmentation regions to obtain a complete building distribution map.

2. The satellite image building intelligent recognition method based on deep learning and GIS according to claim 1, characterized in that The step of obtaining high-resolution satellite image data through remote sensing satellites and scraping geospatial data corresponding to the satellite images from the Internet includes: Introduce the ORB feature point algorithm to achieve precise alignment between different satellite images and correct the geometric distortion of the satellite images; Use the denoising encoder model based on the convolutional neural network to remove the noise in the satellite images while retaining the structural details of the building regions; Use the generative adversarial network to synthesize diverse satellite images for data augmentation, and combine the self-adaptive histogram equalization to enhance the contrast of the satellite images.

3. The satellite image building intelligent recognition method based on deep learning and GIS according to claim 2, characterized in that, The step of using the GIS system to perform coordinate mapping and regional division on the satellite images includes: Combine the region proposal network of Mask R-CNN to generate candidate regions for image segmentation, and optimize the accuracy of the candidate region proposals for image segmentation through online learning; Combine the feature pyramid network to fuse the satellite image features at different scales, and accurately detect the candidate regions for image segmentation from small to large; Use the convolutional long short-term memory network to capture the spatio-temporal patterns of the changes in the candidate regions for image segmentation through spatio-temporal analysis in space.

4. The method for intelligent identification of satellite image buildings based on deep learning and GIS according to claim 3, wherein: The step of using the improved Mask R-CNN network to locate and segment the buildings in the image segmentation regions and generate the building region segmentation results includes: Input the image segmentation region into the pre-trained EfficientNet backbone network to extract the regional image features, and generate the corresponding feature maps through the FPN network; Use the region proposal network to perform sliding window scanning on the feature maps to find the regions where buildings may exist and generate candidate regions for buildings; Use the Softmax classifier combined with non-maximum suppression to perform binary classification and screening on the candidate regions for buildings to obtain the regions of interest; Input the feature maps and the regions of interest into ROIAlign to generate aligned feature maps of a fixed size for each region of interest; Input the aligned feature maps into the localization branch, and perform localization regression of the bounding boxes of the buildings through the fully connected layer to obtain the building positions; Input the aligned feature map and building locations into the segmentation branch, and perform pixel-level segmentation on each building through per-pixel classification to generate a binary mask of the building; The loss function of the localization branch and the segmentation branch consists of the formula: L loss = L f + L r + L mask where, L f is the classification loss, using cross-entropy loss, L r is the regression loss, using Smooth L1 loss, L mask is the segmentation loss, using binary cross-entropy loss; Process the binary mask of the building using bilinear interpolation to restore it to the same dimension as the image segmentation region; Extract the building area from the image segmentation region using the restored binary mask of the building to generate the building area segmentation result.

5. The satellite image building intelligent recognition method based on deep learning and GIS according to claim 4, characterized in that: Based on the building area segmentation result, use a depth point convolutional network to extract the depth features of the building shape, identify and classify the building, and generate the corresponding identification and classification results, including: Combine the depth residual network of ResNet-101 or DenseNet to extract the depth features of the building shape from the building area segmentation result; Among them, use the convolutional network formula to classify and perform correlation analysis on the building to optimize the effect of the depth point convolutional network in processing the adjacent relationship of the building; H (l+1) = σ(AH (l) W (l) ) where \(H(\) l ) is the node feature matrix of the \(l\)-th layer, \(A\) is the normalized adjacency matrix of the graph, \(W(\) l ) is the learning weight matrix of each layer, and \(\sigma\) is the activation function; Combine self-supervised learning to perform unsupervised feature learning on the building image to optimize the recognition effect of the graph convolutional network on buildings with complex shapes; Based on the extracted depth features of the building shape, use a support vector machine to identify and classify the building, and generate the corresponding identification and classification results.

6. The satellite image building intelligent recognition method based on deep learning and GIS according to claim 5, characterized in that: Use the spatial analysis function of the GIS system to perform refined analysis on the recognition and classification results to generate a building distribution map of the current image segmentation region, including: Based on the recognition and classification results, use the DBSCAN or K-means clustering algorithm to perform clustering analysis on the spatial distribution of buildings to identify the building density and distribution characteristics in different regions; Combine a temporal convolutional neural network to predict the regional building distribution, analyze the temporal data of urban expansion, predict the building growth pattern and trend, and generate a prediction result; The prediction formula of the temporal convolutional neural network: where y t is the predicted building growth pattern, x t-j is the input feature at time t-j, ω j is the convolutional kernel, and f is the activation function; According to the prediction result and the building density and distribution characteristics in different regions, generate a building distribution map of the current image segmentation region through the GIS system.

7. The intelligent building recognition system for satellite images based on deep learning and GIS is characterized in that Including: A data management module for obtaining high-resolution satellite image data through a remote sensing satellite and scraping geospatial data corresponding to the satellite image from the Internet using web crawler technology; An image segmentation module for performing coordinate mapping and regional division on the satellite image using the GIS system based on the geospatial data, and generating an image segmentation region with geographical location identification in combination with existing urban planning and infrastructure data; A building detection module for using an improved Mask R-CNN network to locate and segment buildings in the image segmentation region to generate a building area segmentation result; A building classification module for extracting the depth features of the building shape using a depth point convolutional network based on the building area segmentation result, identifying and classifying the building, and generating the corresponding identification and classification results; A building spatial analysis module for using the spatial analysis function of the GIS system to perform refined analysis on the recognition and classification results, generate a building distribution map of the current image segmentation region, and predict the building growth trend; The building distribution synthesis module is used to perform spatial coordination and synthesis processing on the building distribution maps of all generated image segmentation regions using the GIS system to obtain a complete building distribution map.

8. The satellite image building intelligent recognition system based on deep learning and GIS according to claim 7, characterized in that, The building detection module further includes: A generative adversarial network for synthesizing diverse satellite images for data augmentation and enhancing the image contrast by combining with self-adaptive histogram equalization; Combining the region proposal network of Mask R-CNN to generate candidate regions for image segmentation and optimizing the accuracy of candidate regions through online learning; Combining the feature pyramid network and the convolutional long short-term memory network to perform fusion and spatio-temporal pattern analysis on satellite image features at different scales to accurately detect buildings in the image segmentation region.

9. The satellite image building intelligent recognition system based on deep learning and GIS according to claim 8, wherein The building classification module further includes: Combining a deep residual network such as ResNet-101 or DenseNet to extract deep features of the building shape from the building area segmentation results; Among them, a convolutional network is used to classify and perform correlation analysis on buildings to optimize the effect of the depth point convolutional network in processing the adjacent relationship of buildings; Combining self-supervised learning to perform unsupervised feature learning on building images to optimize the recognition effect of the graph convolutional network on buildings with complex shapes.

10. The satellite image building intelligent recognition system based on deep learning and GIS according to claim 9, characterized in that, The building spatial analysis module further includes: Performing clustering analysis of building distribution based on the DBSCAN or K-means clustering algorithm to identify the building density and distribution characteristics of different regions; Predicting the building growth pattern and trend through the temporal convolutional neural network for the temporal data of urban expansion to generate a building distribution map.

Images