Community infrastructure toughness measurement method based on remote sensing image segmentation

Abstract

The invention relates to a community infrastructure toughness measurement method based on remote sensing image segmentation, and the method comprises the steps: carrying out the data preprocessing of an obtained community aerial image, segmenting the image into a plurality of small blocks, carrying out the recognition of the segmented image data through employing an MPViT-DeepLab network, obtaining a building segmentation image, and carrying out the recognition of the building segmentation image, meanwhile, an MPViT-DeepLab network model subjected to three-stage transfer learning is adopted to obtain a segmented image of a community green infrastructure CSGI in the image; and the obtained building and CSGI segmented images are subjected to denoising and combination processing to be restored into a complete community segmented image, and the toughness index of the city is calculated. Compared with the prior art, the method can quickly adapt to new environmental conditions, can maintain high segmentation accuracy in different communities or different remote sensing image conditions, and provides powerful technical support for community planning, disaster management and maintenance and upgrade of infrastructures.

Description

Technical Field

[0001] The present invention relates to the technical field of geographic information data processing and evaluation based on remote sensing images, and in particular to a community infrastructure resilience measurement method based on remote sensing image segmentation. Background Art

[0002] Remote sensing image acquisition technology is the basis for measuring the resilience of community infrastructure. Modern remote sensing technology uses drones, satellites or aerial photography platforms to provide high-resolution images that capture the detailed geographic and architectural features of the community. With the development of remote sensing technology, the resolution and quality of images continue to improve, making the information extracted from the images more accurate, providing the possibility for detailed analysis of community infrastructure.

[0003] Image segmentation technology is the process of decomposing remote sensing images into multiple regions or objects. Modern image segmentation technology uses machine learning and deep learning algorithms, such as convolutional neural networks (CNN), to automatically identify and segment different components in images, such as buildings, roads, and vegetation. The application of these technologies has greatly improved the accuracy and automation of segmentation.

[0004] Geographic Information System (GIS) is a versatile geographic data management platform that can store, query, analyze and visualize geospatial data. The GIS platform integrates remote sensing image segmentation results and provides a powerful tool to analyze the spatial structure and infrastructure distribution of communities. Although it is not difficult to obtain aerial remote sensing images of communities, it is not easy to quickly obtain community components from remote sensing images and understand the complex relationships between community components. Existing large-scale GIS platforms, such as ArcGIS, can provide community building outline information, but community green infrastructure (Community-Scale Green Infrastructure CSGI) changes frequently, and it is difficult for GIS platforms to update in a timely manner and obtain specific information on community green infrastructure (CSGI) related goals. If researchers want to examine the resilience level of a community in the current time period, they need to visit records or manually mark them from aerial remote sensing images. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a community infrastructure resilience measurement method based on remote sensing image segmentation, which realizes a comprehensive and in-depth analysis of community infrastructure and provides strong technical support for urban planning, disaster management and infrastructure maintenance and upgrading.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A community infrastructure resilience measurement method based on remote sensing image segmentation, comprising the following steps:

[0008] Perform data preprocessing on the acquired community aerial images and divide the images into multiple small blocks;

[0009] For the segmented image data, the MPViT-DeepLab network recognition is used to obtain the building segmentation image, and the MPViT-DeepLab network model after three-stage transfer learning is used to obtain the segmentation image of the community green infrastructure CSGI in the image;

[0010] The acquired building and CSGI segmentation images are denoised and combined to restore them to complete community segmentation images;

[0011] Based on the restored complete community segmentation image, the city’s resilience index is calculated.

[0012] As a preferred technical solution, the specific process of the three-stage transfer learning is as follows:

[0013] In the pre-training stage, MPViT is used as the backbone network for feature extraction, and the dataset is used to train the classification feature extraction capability of the MPViT network;

[0014] In the first transfer learning phase, all parameters of the pre-trained MPViT network model are used as the initialization parameters of the feature extraction network in MPViT-DeepLab, and the community aerial images are used as the input of the MPViT-DeepLab model to obtain the segmentation results of the community buildings;

[0015] In the second transfer learning stage, the model weight parameters of the first transfer learning are used as initialization parameters and passed into MPViT-DeepLab. The batch normalization layer parameters are locked and the community aerial images with CSGI annotations are used as the input of the MPViT-DeepLab model to obtain the CSGI segmented images.

[0016] As a preferred technical solution, the MPViT-DeepLab network includes the feature extraction backbone network MPViT and the DeepLabV3+ architecture;

[0017] The MPViT network performs multiple iterations of decreasing size and increasing sequence number on the input remote sensing image;

[0018] The original coarse feature data obtained after the first iteration of the MPViT network is input into the decoding layer of the MPViT-DeepLab network architecture;

[0019] The fine feature data obtained after the last iteration of the MPViT network is input into the encoding layer of the MPViT-DeepLab network architecture and subjected to dilated convolution processing with different dilation rates;

[0020] The features processed by the dilated convolution of the coding layer are fused with the original coarse features and upsampled to the resolution of the original image;

[0021] The upsampled features are decoded to generate building segmentation results.

[0022] As a preferred technical solution, the MPViT network includes multiple pairs of MS-PatchEmbed modules and MP-Transformer modules; in each iterative cycle, an MS-PatchEmbed module is used to segment the input image, and then an MP-Transformer module is used to process image features using a self-attention mechanism.

[0023] As a preferred technical solution, the MS-PatchEmbed module uses overlapping patch convolution to design a 2D convolution operation F k×k (·), including multiple convolution kernels of 3×3, 5×5 or 7×7 sizes; the MS-PatchEmbed module processes the convolution operations of features of different scales into visual tags of the same sequence length and then passes them to the MS-PatchEmbed module.

[0024] As a preferred technical solution, after receiving the multi-scale image sequence features from the MS-PatchEmbed module, the MP-Transformer module inputs them into the Transformer encoder and uses the efficient decomposition self-attention head for further processing; the MP-Transformer module also introduces a deep residual neck, and places 1×1 point-by-point convolutions for compression and expansion before and after the deep separable convolution to process the relationship between channels;

[0025] The MP-Transformer module aggregates the multi-scale features processed by the Transformer encoder, reduces the dimension through 1×1 convolution and outputs them as input data for the next iterative module.

[0026] As a preferred technical solution, the decoder outputs the feature R decoder for:

[0027] R decoder =F 1×1 (X coarse )

[0028] In the formula, F 1×1 (·) represents a 2D convolution operation with a kernel size of 1×1; Xcoarse It is the original rough feature data obtained after the first iteration of the MPViT network;

[0029] The encoder uses dilated convolutions with different dilation rates to process fine features, as follows:

[0030] R encoder =B(F 1×1 (Concat([F 1×1 (X fine ),Y1,Y2,Y3,P]))

[0031] Where B(·) represents the upsampling function; F 1×1 (·) represents a 2D convolution operation with a convolution kernel size of 1×1; Concat(·) represents a connection function that directly connects features through matrix operations; X fine is the fine feature data obtained after the last iteration of the MPViT network; Y1, Y2, Y3 are feature X fiie The result after dilated convolution with different dilation rates; P represents feature X fine The pooling result of

[0032] The features processed by the dilated convolution of the coding layer are fused with the original coarse features and upsampled as follows:

[0033] R=B(Concat([R decoder ,R encoder ])

[0034] Where B(·) represents the upsampling function; R decoder , R encoder are the output features of the decoder and encoder respectively.

[0035] As a preferred technical solution, the upsampling function adopts a value bilinear upsampling function, and uses multiple closest pixel points X(i,j) to perform weighted averaging on the output pixel B(x,y).

[0036] As a preferred technical solution, the method only performs denoising on the building segmentation result, including using an opening operation to first corrode and then dilate to remove small noise in the building segmentation object, while smoothing the boundary of the main segmentation target.

[0037] As a preferred technical solution, the city's resilience indicators include:

[0038] Building density:

[0039]

[0040] In the formula, ρ buildingRepresents the actual building area, which is calculated by the ratio of segmented building pixels to the total image pixels in the community remote sensing image; A building A is the floor area of ​​the building. total__b Indicates the area of ​​the total land;

[0041] Greening rate

[0042]

[0043] In the formula, ρ greenery Indicates greening rate; A greenery is the green area of ​​the site, represented by the number of CSGI pixels in the community segmentation image; A total_g is the total area of ​​the site, expressed using the total pixels of the community segmentation image;

[0044] Heat Island Effect:

[0045] I=b G ·ρ greenery +b D ·ρ building

[0046] Where, I represents the community heat island effect index; b G is the regression coefficient of green space ratio; b D is the regression coefficient of building density; ρ greenery and ρ building They represent greening rate and building density respectively.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1) This paper proposes the MPViT-DeepLab network, which can efficiently identify and segment community aerial images, significantly improving the recognition accuracy of key infrastructure such as buildings. Compared with traditional segmentation methods, MPViT-DeepLab can better handle complex structures and details in images, thereby obtaining more accurate and clear segmentation results.

[0049] 2) The present invention uses a three-stage transfer learning model, through which the system can quickly adapt to new data sets and environmental conditions, improving the generalization ability and adaptability of the model. This means that even in different communities or different remote sensing image conditions, the system can maintain high segmentation accuracy.

[0050] 3) The present invention introduces morphological operations, especially opening operations, to remove small noise in the segmentation results and smooth the boundaries. This method not only improves the quality of the segmented image, but also enhances the reliability of the subsequent modeling steps. Through this morphological processing, the segmentation results are more consistent with the actual community structure, providing more accurate data for the modeling and analysis of community infrastructure. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of resilience infrastructure measurement based on remote sensing image segmentation of the present invention;

[0052] Figure 2 It is the community infrastructure resilience measurement modeling process based on remote sensing image segmentation of the present invention;

[0053] Figure 3 It is a schematic diagram of the MPViT-DeepLab network framework of the present invention;

[0054] Figure 4 It is a specific flow chart of the three-stage transfer learning scheme of the present invention;

[0055] Figure 5 This is a diagram of the first transfer learning process of the present invention;

[0056] Figure 6 This is a diagram of the second transfer learning process of the present invention. DETAILED DESCRIPTION

[0057] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0058] Example 1

[0059] The present invention proposes a community infrastructure resilience measurement method based on remote sensing image segmentation, which can quickly obtain segmentation maps of community buildings and CSGI, and use the segmentation maps to directly evaluate some community infrastructure resilience indicators.

[0060] First, the community aerial images need to be preprocessed and cut into small pieces of uniform size to lay the foundation for subsequent automated processing and analysis. This step ensures the consistency and processability of the image data, while also improving processing efficiency.

[0061] Next, the advanced MPViT-DeepLab network is used to identify the cut image blocks and accurately distinguish key infrastructure such as buildings, roads, and green spaces in the community. Using the powerful recognition ability of the deep learning network, it is able to extract useful information from complex images and provide accurate data for subsequent analysis. The three-stage transfer learning is used to further enhance the adaptability and accuracy of the system. With this technology, the model can quickly adapt to new data sets and environmental conditions, and maintain high recognition accuracy regardless of whether it is in different communities or different remote sensing image conditions. This step can significantly improve the generalization ability of the model, enabling it to be widely used in a variety of different scenarios.

[0062] It is also necessary to use morphological methods to denoise only the building segmentation results, mainly using the opening operation, which effectively removes small noise in the image by first corroding and then dilating, and smoothes the boundaries of the main segmentation targets. This step significantly improves the quality of the segmentation results, making the segmented image clearer and more accurate. The restoration and optimization of the segmentation results ensure the integrity of all segmented objects.

[0063] Finally, it is necessary to judge the resilience of community infrastructure. Based on the two-dimensional community model, the resilience of infrastructure is measured, and some urban resilience indicators such as building density, greening rate, and urban heat island effect index are calculated to evaluate its adaptability and resilience in the face of pressure. This step provides a scientific basis for the sustainable development and resilience building of the community, and is of great significance for improving the community's disaster resistance and resilience.

[0064] like Figure 1 The specific implementation of each step is as follows:

[0065] Step S1: Data preprocessing is performed on the community aerial image, and the image is divided into multiple small blocks. In this embodiment, the cutting size is 300×300 pixels. This step divides the image into easy-to-process units, laying the foundation for subsequent feature extraction and analysis.

[0066] Step S2: Figure 2 As shown in the figure, after cutting, the building segmentation image is obtained through MPViT-DeepLab network recognition, and the specific distribution of CSGI in the image is obtained through the three-stage transfer learning model processing.

[0067] Figure 3This is a schematic diagram of the MPViT-DeepLab network framework, which mainly consists of the following parts: MPViT, a network that can process image data of different sizes and resolutions, including the MS-PatchEmbed module and the MP-Transformer module; DeepLabV3+ architecture. In order to better adapt to the remote sensing image building segmentation task and obtain better segmentation results, the DeepLabV3+ encoder-decoder structure is used to further process the complex relationship between the image sequence features and the sequences, and output a segmentation map with clearer edges and more accurate. The core idea of ​​MPViT is to use two structures, the multi-scale patch embedding module and the multi-path transformer module, to build a parallel multi-path training framework, and capture both fine and coarse features for dense prediction tasks. The MPViT network is designed with four size transformation iteration cycles. In each cycle, the width and height of the image sequence are reduced, the number of sequences increases, and the relationship between the sequences becomes more complex. The image sequence vectors processed by the MS-PatchEmbed module and the MP-Transformer module for the first time have large width and height dimensions, small number, and relatively simple relationship between sequences. The MPViT-DeepLab network architecture uses a decoding layer to receive these rough feature data. The image sequences processed by the last MS-PatchEmbed module and the MP-Transformer module are small in width and height, large in number, and have complex relationships between sequences. These fine feature data are received by the encoding layer of the MPViT-DeepLab network architecture for further processing by atrous convolution with different expansion rates. The features refined by atrous convolution are fused with the original coarse features to form a richer feature representation; the feature map obtained by multi-scale feature fusion is upsampled to the resolution of the original image through the Upsample operation. The upsampled features are decoded using the DeepLabV3+ network to generate segmentation results.

[0068] like Figure 4 The figure shows a specific flow chart of the three-stage transfer learning scheme, which mainly consists of the following parts: In the pre-training stage, MPViT is used as the backbone network for feature extraction, and its powerful representation ability is used to process image data. In this embodiment, the ISVR2012 data set is used to train the classification feature extraction ability of the backbone network for MPViT feature extraction;

[0069] like Figure 5 As shown, in the first transfer learning stage, all parameters of the pre-trained MPViT network model are used as the initialization parameters of the feature extraction network in MPViT-DeepLab, and the community aerial images are used as the input of the MPViT-DeepLab model to obtain the segmentation results of the community buildings;

[0070] like Figure 6As shown in the figure, the second transfer learning stage uses the model weight parameters of the first transfer learning as initialization parameters and passes them into MPViT-DeepLab, locks the batch normalization layer parameters, and uses community aerial images with CSGI annotations as the input of the MPViT-DeepLab model to obtain CSGI segmented images.

[0071] In the pre-training stage, the MPViT backbone feature extraction network is used, including the MS-PatchEmbed module and the MP-Transformer module. The core of the MS-PatchEmbed module is a function F designed using overlapping patch convolutions. k×k (·).F k×k (·) represents a 2D convolution operation with a kernel size of k×k, a stride of s, and padding of p. Given a set of features As input, F k×k (·) The size of the convolution result of the function Among them, H i With W i The calculation formula is shown in (1)(2).

[0072]

[0073] By adjusting the size of k, s, and p parameters, F k×k (·) function can be used to generate features of the same scale or reduce the feature size. When s=1 and When F k×k The (·) function can process patches of different sizes into features of the same size. For a piece of image features, regardless of the convolution operation of 3×3, 5×5 or 7×7 convolution kernel, the output size is the same as the original size. This way, the MS-PatchEmbed module can process the convolution operations of features of different scales into visual markers of the same sequence length, and then pass them to the Transformer encoder. If you need to reduce the feature size, you only need to increase the stride. For example, when the stride s is set to 2, the width and height of the output feature will be reduced to half of the input feature. Before the image is input into the MS-PatchEmbed module for the first time, two convolutions with s=2 are performed to reduce the original feature width and height size H×W to In each MS-PatchEmbed module, a convolution of s=2 is performed. So given the input feature After being processed by the MS-PatchEmbed module, the output feature X i The size is The width and height dimensions of the feature are reduced to half of their original size.

[0074] The convolution blocks in the MS-PatchEmbed module all use batch normalized depthwise separable convolutions. Each convolution block includes a 3×3 depthwise convolution (DWConv), a 1×1 pointwise convolution (PWConv), followed by a BN layer, and finally the Hardswish activation function.

[0075] The purpose of the MP-Transformer module is to capture features with long-range dependencies while focusing on local feature relationships. After receiving the multi-scale image sequence features from the MS-PatchEmbed module, the MP-Transformer module inputs them into the Transformer encoder for further processing.

[0076] In the Transformer encoder, MPViT uses an efficient decomposition self-attention head to replace the multi-head self-attention in ViT to reduce the computational burden.

[0077]

[0078] In formula (3), Represents the query, key, and value in the linear projection, and N and C represent the number of image sequences and channels. In addition to the Transformer encoder, the MP-Transformer module also introduces a deep residual neck, placing 1×1 point-by-point convolutions for compression and expansion before and after the depthwise separable convolution to handle the relationship between channels.

[0079] The MP-Transformer module aggregates the multi-scale features processed by the Transformer encoder, reduces the dimension through 1×1 convolution and outputs them as input data for the next iterative module.

[0080] The MPViT network proposed in the present invention sets four size transformation iteration cycles and designs a multi-path visual converter network: a multi-path visual converter is developed that can process image blocks of different sizes and convert them into sequence blocks of the same size. The network can process these sequence blocks through a multi-head self-attention mechanism, while capturing fine and coarse features for intensive prediction tasks and extracting rich contextual information.

[0081] In the two transfer learning stages, the MPViT-DeepLab network was used, namely Figure 3The architecture shown, the yellow arrow represents the reduction of feature size, and the gray arrow represents data transfer. DeepLabV3+ is an encoder-decoder architecture as a whole. In DeepLabV3+, the low-level features extracted by the feature extractor are input into the decoder, and the high-level features (output features) are input into the encoder. The encoder uses dilated convolutions with different expansion rates to calculate the output features, and the calculation results are aggregated with the low-level features, so that the features finally learned by the DeepLabV3+ network retain the contextual information of the image.

[0082] For a set of input features After the dilated convolution, the output features The dimension is H′×W′×C′, where H′, W′ and the hole convolution output feature Y i The calculation formula is shown in (4)(5)(6).

[0083]

[0084]

[0085] In formulas (4), (5), and (6), k represents the size of the convolution kernel, r is the dilation rate, and W k is the corresponding weight.

[0086] The image sequence features extracted by the first MS-PatchEmbed module and the MP-Transformer module are denoted as X coarse The image sequence features extracted by the last MS-PatchEmbed module and the MP-Transformer module are denoted as X fine , the contact function F k×k (·) and formula (6), then the decoder output feature result R in MPViT-DeepLab decoder , the encoder outputs the feature result R encoder The calculation formula of the final output feature result R can be expressed as

[0087] R decoder =F 1×1 (X coarse ) (7)

[0088] R encoder =B(F 1×1 (Concat([F 1×1 (X fine ),Y1,Y2,Y3,P])) (8)

[0089] R=B(Concat([R decoder ,R encoder ]) (9)

[0090] In formulas (7), (8), and (9), Concat(·) represents the connection function, which directly connects the features through matrix operations. Y1, Y2, and Y3 are feature X fine The result after the dilation rate [6, 12, 18] of the dilated convolution, P represents the feature X fine The pooling result of , B(·) is a bilinear upsampling function with a value of 4.

[0091]

[0092] Formula (10) describes in detail the specific calculation method of the bilinear upsampling function with a value of 4. It essentially uses the four closest pixel points X(i,j) to perform a weighted average on the output pixel B(x,y), and the weight is related to the relative distance of x,y to the four points. This upsampling method can improve the image resolution and provide a smooth interpolation method, which helps to recover the features in the decoding stage.

[0093] Step S3: The acquired segmented image is subjected to denoising and combination processing, and finally restored to a complete community segmentation image. Due to the complex distribution of CSGI, there may be segmented objects with smaller targets, so the morphological method is mainly used to perform denoising only on the building segmentation results, and the opening operation is mainly used to remove small noise in the building segmentation object. The calculation method of the opening operation is erosion first and then dilation.

[0094] The purpose of the erosion operation is to reduce the size of the object. It is often used to connect adjacent elements and fill small holes. The segmentation results of remote sensing images usually have some small holes, which affect the accuracy of the segmentation results. The segmentation results can be processed through the erosion operation. The formula of the erosion operation is shown in (11).

[0095]

[0096] In formula (11), A represents the original image, B represents the structural element, (x, y) means the pixel position, and the formula indicates that the structural element B is completely contained in the foreground pixels of A.

[0097] The purpose of the dilation operation is opposite to that of the erosion operation. It is used to expand the boundaries of objects in the image. It is usually used to remove small noise points and keep the size of the main target unchanged. In the dilation operation, as long as the structural element overlaps with the foreground part, the center pixel of the original image will be activated. The formula of the dilation operation can be written as (12).

[0098]

[0099] In formula (12), A and B have the same meaning as in the corrosion operation. As long as B touches any foreground pixel in A, (x, y) will be set as the foreground.

[0100] The opening operation is to first perform corrosion and then use the dilation operation, which can help remove small noise and smooth the boundaries of the main segmentation target without changing the target area. The specific formula of the opening operation is shown in (3).

[0101]

[0102] After the opening operation, the internal noise of the segmented image will be removed, and the external boundary will become smoother and neater, further improving the accuracy of the segmentation and greatly improving the practical value of the segmentation result.

[0103] Step S4: Use the segmentation result model to calculate some urban resilience indicators, such as building density, greening rate, urban heat island effect index, etc. Community infrastructure resilience measurement involves multiple disciplines such as architecture, ecology, and sociology. It aims to use qualitative or quantitative evaluation indicators to judge the resilience of community infrastructure, so as to provide community workers, urban managers and related researchers with a certain basis for judgment.

[0104] The following will introduce some urban resilience indicators that can be calculated using the segmentation result model.

[0105] (1) Building density

[0106] Cao Yongfeng and others proposed that building density is the coverage of buildings, that is, the ratio of the sum of the areas of all buildings within a certain land use area to the total land area. It can reflect the vacant land rate and the density of buildings within the land use area to a certain extent. The calculation formula of building density is as follows:

[0107]

[0108] In formula (14), A building The meaning is the actual building area, A building Refers to the floor area of ​​the building, A total__b Refers to the total land area. ρ of the community in the remote sensing image building It can be calculated by the ratio of segmented building pixels to the total pixels of the image.

[0109] Building density plays a guiding role in improving residents' comfort, environmental protection, and urban planning. Zhou Chuanbin et al. pointed out in their research that the building density of a suitable residential community should not exceed 30%. The resolution of the database DroneDeploy Segmentation Dataset used in the present invention is 10 cm per pixel. If the number of residents in the community is investigated, the segmentation model can also be used to evaluate the per capita living area. The general recommended value is not less than 20 square meters per person.

[0110] (2) Greening rate

[0111] The greening rate is the percentage of greening in a site, that is, the ratio of the greening area to the total area of ​​a site. The greening rate can clearly reflect the greening situation of a certain area. Especially in communities, it is often used as a common indicator by planners and designers. The calculation formula of the greening rate can be expressed as (15).

[0112]

[0113] In formula (15), ρ greenery is the greening rate, A greenery Indicates the green area of ​​the site, A total Indicates the total area of ​​the site. In the present invention, A greenery It can be replaced by the number of CSGI pixels in the segmented image, A total_g The total pixel representation of the segmented image can be used, so the greening rate calculation problem is transformed into a pixel ratio problem. In general, the greening rate of a resilient community is not less than 30%, and the public green area is not less than 15%. In the present invention, the public green area includes trees, grass, lakes, and low shrubs, but does not include rooftop green plants. If the total number of people in the community is counted, the per capita public green area can also be calculated. It is generally recommended to be no less than 2.5 square meters per person.

[0114] (3) Heat Island Effect

[0115] The heat island effect refers to the phenomenon that when a city develops to a certain scale, due to various factors, the temperature in the city is significantly higher than that in the suburbs, forming a high-temperature island. The present invention mainly focuses on the heat island effect at the community level. The transpiration of community green plants and the evaporation of water bodies can effectively reduce the ambient temperature. Studies have shown that when the vegetation coverage rate in a certain area reaches more than 30%, the inhibitory effect of these green spaces on the heat island effect can be maximized. In areas with high building density, roof greening has become one of the most effective ways to increase the total green area because buildings absorb a lot of heat. The average daily heat island effect of a resilient community does not exceed 1.5°C. The heat island effect is usually obtained through temperature sensors and is related to factors such as weather and season. Remote sensing image segmentation can provide qualitative assistance for the heat island effect by calculating the greening rate, etc.

[0116] Ng et al. used Hong Kong, China as a research object and compared the ambient air temperature at the pedestrian level under different greening strategies. They pointed out that the number of trees needed to reduce the pedestrian temperature by 1°C accounts for about 33% of the total urban area. The present invention makes a preliminary calculation of the community heat island effect based on the relationship between building density and greening rate. The specific calculation formula is shown in (16).

[0117] I=b G ·ρ greenery +b D ·ρ building (16)

[0118] In formula (16), I represents the community heat island effect index, b G is the regression coefficient of green space ratio, b D is the regression coefficient of building density, ρ greenery and ρ building Respectively represent the greening rate and building density. In this embodiment, b G Set to -2.25, b D Set to 1.67, the I value represents the difference between the ambient temperature and the air temperature, and a positive value means that the ambient temperature is higher than the air temperature.

[0119] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A community infrastructure resilience measurement method based on remote sensing image segmentation, characterized in that the steps include: Perform data preprocessing on the acquired community aerial images and divide the images into multiple small blocks; For the segmented image data, the MPViT-DeepLab network recognition is used to obtain the building segmentation image, and the MPViT-DeepLab network model after three-stage transfer learning is used to obtain the segmentation image of the community green infrastructure CSGI in the image; The acquired building and CSGI segmentation images are denoised and combined to restore them to complete community segmentation images; Based on the restored complete community segmentation image, the city’s resilience index is calculated.

2. According to claim 1, a community infrastructure resilience measurement method based on remote sensing image segmentation is characterized in that: The specific process of the three-stage transfer learning is as follows: In the pre-training stage, MPViT is used as the backbone network for feature extraction, and the dataset is used to train the classification feature extraction capability of the MPViT network; In the first transfer learning phase, all parameters of the pre-trained MPViT network model are used as the initialization parameters of the feature extraction network in MPViT-DeepLab, and the community aerial images are used as the input of the MPViT-DeepLab model to obtain the segmentation results of the community buildings; In the second transfer learning stage, the model weight parameters of the first transfer learning are used as initialization parameters and passed into MPViT-DeepLab. The batch normalization layer parameters are locked and the community aerial images with CSGI annotations are used as the input of the MPViT-DeepLab model to obtain the CSGI segmented images.

3. The community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 2 is characterized in that: The MPViT-DeepLab network includes the feature extraction backbone network MPViT and the DeepLabV3+ architecture; The MPViT network performs multiple iterations of decreasing size and increasing sequence number on the input remote sensing image; The original coarse feature data obtained after the first iteration of the MPViT network is input into the decoding layer of the MPViT-DeepLab network architecture; The fine feature data obtained after the last iteration of the MPViT network is input into the encoding layer of the MPViT-DeepLab network architecture and subjected to dilated convolution processing with different dilation rates; The features processed by the dilated convolution of the coding layer are fused with the original coarse features and upsampled to the resolution of the original image; The upsampled features are decoded to generate building segmentation results.

4. A community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 2 or 3, characterized in that: The MPViT network includes multiple pairs of MS-PatchEmbed modules and MP-Transformer modules; In each iteration cycle, the input image is segmented through an MS-PatchEmbed module, and then the image features are processed using the self-attention mechanism through an MP-Transformer module.

5. The community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 4 is characterized in that: The MS-PatchEmbed module uses overlapping patch convolution to design a 2D convolution operation F k×k (·), including multiple convolution kernels of 3×3, 5×5 or 7×7 sizes; the MS-PatchEmbed module processes the convolution operations of features of different scales into visual tags of the same sequence length and then passes them to the MS-PatchEmbed module.

6. The community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 4 is characterized in that: After receiving the multi-scale image sequence features from the MS-PatchEmbed module, the MP-Transformer module inputs them into the Transformer encoder and uses the efficient decomposition self-attention head for further processing; the MP-Transformer module also introduces a deep residual neck, placing 1×1 point-by-point convolutions for compression and expansion before and after the depthwise separable convolution to process the relationship between channels; The MP-Transformer module aggregates the multi-scale features processed by the Transformer encoder, reduces the dimension through 1×1 convolution and outputs them as input data for the next iterative module.

7. The community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 3 is characterized in that: The decoder outputs the feature R decoder for: R decoder =F 1×1 (X coarse ) In the formula, F 1×1 (·) represents a 2D convolution operation with a kernel size of 1×1; X coarse It is the original rough feature data obtained after the first iteration of the MPViT network; The encoder uses dilated convolutions with different dilation rates to process fine features, as follows: R encoder =B(F 1×1 (Concat([F 1×1 (X fine ),Y1,Y2,Y3,P])) Where B(·) represents the upsampling function; F 1×1 (·) represents a 2D convolution operation with a convolution kernel size of 1×1; Concat(·) represents a connection function that directly connects features through matrix operations; X fine is the fine feature data obtained after the last iteration of the MPViT network; Y1, Y2, Y3 are feature X fine The result after dilated convolution with different dilation rates; P represents feature X fine The pooling result of The features processed by the dilated convolution of the coding layer are fused with the original coarse features and upsampled as follows: R=B(Concat([R decoder ,R encoder ]) Where B(·) represents the upsampling function; R decoder , R encoder are the output features of the decoder and encoder respectively.

8. The community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 7 is characterized in that: The upsampling function adopts a value bilinear upsampling function, and uses a plurality of closest pixel points X(i,j) to perform weighted averaging on the output pixel B(x,y).

9. The community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 1 is characterized in that: The method only performs denoising on the building segmentation result, including using an opening operation to first corrode and then dilate, thereby removing small noise in the building segmentation object and smoothing the boundary of the main segmentation target.

10. The community infrastructure resilience measurement method based on remote sensing image segmentation according to claim 1, characterized in that: The city's resilience indicators include: Building density: In the formula, ρ building Represents the actual building area, which is calculated by the ratio of segmented building pixels to the total image pixels in the community remote sensing image; A building A is the floor area of ​​the building. total__b Indicates the area of ​​the total land; Greening rate In the formula, ρ greenery Indicates greening rate; A greenery is the green area of ​​the site, represented by the number of CSGI pixels in the community segmentation image; A total_g is the total area of ​​the site, expressed using the total pixels of the community segmentation image; Heat Island Effect: I=b G ·r greenery +b D ·r building Where, ρ represents the community heat island effect index; b G is the regression coefficient of green space ratio; b D is the regression coefficient of building density; ρ greenery and ρ building They represent greening rate and building density respectively.

Images