A method for evaluating the damage degree of a building in a natural disaster by using a remote sensing image

Abstract

The application discloses a method for evaluating the damage degree of buildings in natural disasters by using remote sensing images. The method performs pre-alignment correction and collaborative super-resolution on multi-source remote sensing images before and after a disaster area is affected by disasters, and then uses the obtained images with the same resolution to detect the target feature change. According to the extracted target difference change feature and the target boundary, the damage degree of buildings in natural disasters and the affected area are detected, and the damage level of natural disasters and the economic loss caused by natural disasters are evaluated. By using multi-source remote sensing images, the application solves the difference information caused by different remote sensing devices, increases the detail features of remote sensing images, and enables the detection model to detect the damaged buildings from large-scale remote sensing images. By analyzing the damage quantity and the area range of buildings in the disaster area, the damage level of natural disasters and the economic loss caused by natural disasters are evaluated, and the regional distribution of the disaster area affected by natural disasters can be directly observed.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and remote sensing technology, and specifically relates to a method for assessing the degree of damage to buildings in natural disasters using remote sensing images. Background Technology

[0002] Earthquakes, mudslides, and other natural disasters occur frequently worldwide. If timely rescue efforts are not implemented, the resulting loss of life and property can be incalculable. Therefore, it is essential to quickly identify damaged buildings to pinpoint disaster areas and develop precise rescue plans.

[0003] Due to my country's vast territory, complex terrain, and mountainous and forested landscape, the task of detecting damaged buildings in disaster areas presents stringent requirements. Large-scale disaster detection based on remote sensing images is a popular research direction, and deep learning technology plays a crucial role in this field. Breakthroughs lie in small-target detection techniques at large scales, real-time modeling, and target detection techniques with weak differential features. Currently, building damage detection and assessment are mainly based on satellite or aerial remote sensing images, which have certain advantages in large-scale analysis. A. Zhang et al. (2017) labeled buildings based on satellite remote sensing images, designed separate strategies for training set labels and test set labels, and trained a weakly supervised detection model with classification labels and a semi-supervised method with a small number of pixel-level labels. M. Aamir et al. (2019) proposed a method for detecting buildings from low-contrast satellite images. They first used singular value decomposition (SVD) based on discrete wavelet transform (DWT) to optimize image contrast to represent target information, and finally detected the complete outline of the building. Gonzalez-Drigo et al. (2019) collected data on severely damaged buildings from the May 11, 2011 Lorca earthquake (Murcia, Spain) and analyzed the extent of damage using empirical and numerical methods. S. Fujita et al. (2021) developed an estimation system for damaged roofs, using deep learning techniques based on aerial photographs to obtain an overview of damage in the disaster area. To provide rapid estimation, the system utilized a pruning algorithm and automatically generated roof image data using the location information of building polygons in GIS. Q. Li et al. (2021) proposed a method for assessing building damage based on coseismic deformation, using the May 12, 2008 Wenchuan earthquake in China as an example. They collected 51 building samples and analyzed earthquake damage survey data.

[0004] Existing research has made various contributions to natural disaster detection and building identification in disaster areas. However, significant challenges remain in disaster detection in complex terrain: 1) Buildings in remote sensing images are small and have low resolution, making direct identification of building damage extremely difficult; 2) Accurately calculating the changing characteristics of damaged buildings by aligning pre-disaster remote sensing images with post-disaster images is also very difficult. Therefore, developing a technical solution that can accurately detect and assess damaged buildings in disaster areas over a large area using high-resolution remote sensing data is a problem urgently needing to be solved by those skilled in the art. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for assessing the extent of building damage during natural disasters using remote sensing images. By analyzing remote sensing image data of the affected area before and after the disaster, the method first corrects the misalignment of target features in the data, then calculates feature differences to enhance the target features of damaged buildings, and finally utilizes target boundary constraints to enable the model to more accurately detect damaged buildings.

[0006] To achieve the above objectives, the present invention provides a method for assessing the degree of damage to buildings during natural disasters using remote sensing images, comprising the following steps:

[0007] Step 1: Obtain high-resolution multi-source remote sensing images of the disaster area before and after the disaster.

[0008] Step 2: Perform pre-alignment correction on the acquired multi-source remote sensing images to reduce spatial position differences and distortions in images acquired by different remote sensing devices;

[0009] Step 3: Perform collaborative super-resolution on the multi-source remote sensing images of different resolutions after pre-alignment and correction in Step 2, so that the images obtained by different remote sensing devices can achieve the same level of resolution.

[0010] Step 4: Use the same resolution image obtained in Step 3 to detect changes in target features and extract the differences in subtle features of the same target at different times in a large-scale scene.

[0011] Step 5: Based on the target difference change characteristics extracted in Step 4, use target boundary constraints to detect the degree of damage to buildings and the affected area in natural disasters;

[0012] Step 6: Using the information on damaged buildings detected in Step 5, assess the severity of the natural disaster and the resulting economic losses.

[0013] Furthermore, the high-resolution multi-source remote sensing images obtained in step 1 include satellite remote sensing images, spaceborne remote sensing images, and UAV remote sensing images in different bands. The spatial resolution of the images must be less than n1 m / pix, the image resolution must be between n2pix×n2pix and n3pix×n3pix, and the noise content of the images must not exceed n4%, where n1-n4 are set thresholds.

[0014] Furthermore, in step 2, the key feature points in the entire image are first calculated using a feature point extraction operator. Then, feature matching is performed on the common areas of the pre-disaster and post-disaster images. After a successful match, the geometric feature difference of the common areas cannot exceed n5%, i.e., MR≤n5%, where n5 is a set threshold. The matching index MR for aligning remote sensing images before and after the disaster is calculated as follows:

[0015]

[0016] In the formula, MR represents the matching index for image pre-alignment before and after the disaster, f(·) represents the non-disaster differences caused by background filtering and changes in lighting and seasons, and I pre I post This represents pre-disaster and post-disaster remote sensing images that have not undergone pre-alignment correction, I′ pre 、I′ post This represents pre-alignment and post-disaster remote sensing images.

[0017] Furthermore, in step 3, collaborative super-resolution is used to comprehensively analyze the difference information of multi-source images and predict the texture information in high-resolution cases, so that the resolution of all multi-source images reaches the highest resolution level, that is:

[0018]

[0019] c = Downsmaple(I′) i ) / I i -1,c→0 (3)

[0020] In the formula, I′ i The image represents the super-resolution image, S represents the super-resolution operation, and I represents the super-resolution image. i U represents a remote sensing image without over-resolution. i I represents the information set of multiple source images. i This represents the information difference between the super-resolution remote sensing image and other source images. The source images are obtained by downsampling the high-resolution images. N indicates that there are N source images. Downsmaple(·) represents the downsampling process. c represents the constraint of image super-resolution, that is, it means that the information of the original image cannot be destroyed. When c approaches 0, it means that the super-resolution image retains more information of the original image.

[0021] Furthermore, in step 4, the pre-aligned and coordinated super-resolution pre-disaster image is represented as X, and the post-disaster image is represented as Y. A transformation matrix T controlled by parameter W is used for image correction to align them. A normal distribution normalization function of the transformation matrix T is introduced to correct minor differences in the images, that is:

[0022] X′=1 / (1-exp(-Cov 3×3 ((TXT -1 -μ) / b))) (4)

[0023] Y′=1 / (1-exp(-Cov 3×3 ((TYT -1 -μ) / b))) (5)

[0024] T=f(W) (6)

[0025] W=D((B(X)-B(Y)),(B(X down )+B(Y down )) >1 (7)

[0026] In the formula, X′ and Y′ represent the pre-disaster image and post-disaster image after deviation correction, respectively; exp(·) represents an exponential function with the natural constant e as the base; cov 3×3 (·) indicates a convolution operation using a 3×3 kernel; T -1 This represents the inverse of T; μ and b are adaptive parameters for offset and scaling, respectively; T is a transformation matrix controlled by the image unalignment factor W, trained using a loss metric constructed through a structural similarity function, and its size is consistent with the image feature size; f(·) is a 5×5 convolutional kernel filter used to filter out outliers in W; B(·) represents feature binarization; X down Y down B(X) represents the downsampling features of X and Y, and downsampling can reduce the misalignment differences in images; B(X)-B(Y) represents the misalignment features of X and Y; (B(X)-B(Y ... down )+B(Y down )) >1 The sum of the downsampled features of X and Y is greater than 1; D(A,B) represents the Euclidean distance between a feature point in A and the nearest feature point in B.

[0027] First, X′ and Y′ are obtained using the transformation matrix T. Then, the structural similarity function is used to evaluate the similarity between X′ and Y′. If the similarity is less than n6%, it is considered misaligned. Formulas (4)-(7) are executed again until the alignment error is less than n7%, achieving accurate alignment. n6 and n7 are the set thresholds. After accurately aligning the images before and after the disaster, the feature differences between the images before and after the disaster are calculated to obtain the image change information. The feature differences include the differences between the features of the images before and after the disaster, the differences between the fusion features of the images before and after the disaster, and the differences between the fusion features of the images after the disaster and the images before and after the disaster. These three differences converge into the overall feature differences of the images before and after the disaster.

[0028] The specific process for calculating feature differences is as follows: ① Calculate the fusion features of pre-disaster and post-disaster images: R0 = Cat(X,Y), where Cat(X,Y) represents concatenating X and Y together; ② Perform channel dimensionality reduction processing on the fusion features of pre-disaster and post-disaster images: R1 = Cov 1×1 (R0), where Cov 1×1 (·) indicates a convolution operation using a 1×1 kernel; ③ Calculate the difference features between pre-disaster and post-disaster images: D1=1 / (1+e X-Y ); ④ Calculate the difference between R1 and the post-disaster image: ⑤ Calculate the difference between R1 and the pre-disaster image: ⑥ Fuse the post-disaster image with the obtained differential features: M′=Cat(D1,D2,D3,Y), where Cat(D1,D2,D3,Y) represents concatenating D1, D2, D3, and Y together; ⑦ Perform channel dimensionality reduction on the fused differential features: M=Cov 1×1 (M′), where Cov 1×1 (·) indicates a convolution operation using a 1×1 convolution kernel.

[0029] Furthermore, step 5 introduces shallow boundary features into deep features, providing boundary constraints for location and category prediction. The location box detection method based on target boundary constraints first predicts the pixel-level boundary region of the target, and then predicts the building target location and damage type based on the boundary constraints. Its mathematical form is as follows:

[0030] E′=Cov 3×3 (R0) (8)

[0031] E = Cat(Cov) 3×3 (M×w >a (E′)),Cov 3×3 (M×(1-w <a (E′)))) (9)

[0032] In the formula, E′ represents the boundary feature; cov 3×3(·) represents a convolution operation using a 3×3 kernel; R0 represents the fusion feature of pre-disaster and post-disaster images; Cat(X,Y) represents concatenating X and Y together; M represents the feature obtained by fusing the difference information; w(·) is a threshold function that represents extracting boundary features greater than or less than the threshold a from E′; E represents the feature of the target fusion boundary information.

[0033] Deep learning models are used to predict the types of building damage, which are divided into four levels: no damage, minor damage, moderate damage, and severe damage. The area affected by natural disasters is estimated based on the predicted number and area of ​​damaged buildings.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] 1) The terrain in the disaster area is complex and varied. By utilizing multi-source remote sensing images, the differences in information brought by different remote sensing equipment can be resolved, and the detailed features of the remote sensing images can be increased, so that the detection model can detect damaged buildings from large-scale remote sensing images.

[0036] 2) By mining the characteristic change information of damaged buildings affected by different natural disasters from remote sensing images of disaster areas, the location of damaged building targets and the extent of damage can be quickly located and assessed.

[0037] 3) By analyzing the number and extent of damage to buildings in the disaster area, assessing the level of natural disaster hazard and the resulting economic losses, we can get a comprehensive and intuitive view of the regional distribution of the disaster area affected by natural disasters. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0039] This invention provides a method for assessing the extent of building damage during natural disasters using remote sensing images. The technical solution of this invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] like Figure 1 As shown, the present invention provides a method for assessing the degree of damage to buildings in natural disasters using remote sensing images, comprising the following steps:

[0041] Step 1: Obtain high-resolution multi-source remote sensing images of the disaster area before and after the disaster.

[0042] The acquired high-resolution multi-source remote sensing images include satellite remote sensing images, spaceborne remote sensing images, and UAV remote sensing images in different bands. The spatial resolution of the images must be less than 10 m / pixel, the image resolution must be between 1000pix×1000pix and 2000pix×2000pix, and the noise content of the images must not exceed 10%.

[0043] Step 2: Perform pre-alignment correction on the acquired multi-source remote sensing images to reduce spatial position differences and distortions in images acquired by different remote sensing devices.

[0044] Because buildings in images before and after the disaster exhibit different characteristics, such as collapsed or covered structures, there are significant feature differences in the images to be matched, which may affect the extraction of key feature points during image matching. Therefore, pre-alignment correction is performed before image change detection. This involves using a feature point extraction operator to calculate the key feature points in the entire image, and then performing feature matching on the common areas of the pre- and post-disaster images. After successful matching, the geometric feature difference in the common areas cannot exceed 30% (i.e., MR ≤ 30%). The matching index MR for aligning remote sensing images before and after the disaster is calculated as follows:

[0045]

[0046] In the formula, MR represents the matching index for image pre-alignment before and after the disaster, f(·) represents the non-disaster differences caused by background filtering and changes in lighting and seasons, and I pre I post This represents pre-disaster and post-disaster remote sensing images that have not undergone pre-alignment correction, I′ pre 、I′ post This represents pre-alignment and post-disaster remote sensing images.

[0047] Step 3: Perform collaborative super-resolution on the multi-source remote sensing images of different resolutions after pre-alignment and correction in Step 2, so that the images obtained by different remote sensing devices can achieve the same level of resolution.

[0048] Because multi-source images exhibit significant resolution differences, and the image information captured by different devices varies, especially in terms of texture and other detailed information, collaborative super-resolution utilizes the differential information from multiple source images for comprehensive analysis and predicts texture information at high resolutions, ensuring that all multi-source images reach the highest possible resolution.

[0049]

[0050] c = Downsmaple(I′) i ) / I i -1,c→0 (3)

[0051] In the formula, I′ i The image represents the super-resolution image, S represents the super-resolution operation, and I represents the super-resolution image. i U represents a remote sensing image without over-resolution. i This represents a set of information from multiple image sources. This represents the information difference between the super-resolution remote sensing image and other source images. The source images are obtained by downsampling the high-resolution images. N indicates that there are N source images. Downsmaple(·) represents the downsampling process. c represents the constraint of image super-resolution, that is, it means that the information of the original image cannot be destroyed. When c approaches 0, it means that the super-resolution image retains more information of the original image.

[0052] Collaborative super-resolution does not destroy the structural information of low-resolution images, nor does it change the information content of the original images; that is, it does not add or reduce the target or target attributes of the image.

[0053] Step 4: Use the same resolution image obtained in Step 3 to detect changes in target features and extract the differences in subtle features of the same target at different times in a large-scale scene.

[0054] Since there are still slight deviations between the pre-disaster images and the post-disaster images after pre-alignment, such as image target distortion caused by different camera parameters, resulting in misalignment of target size and shape, the pre-aligned images of the same resolution obtained in step 3 are used to fine-tune the alignment using a transformation matrix to achieve accurate alignment between the pre-disaster images and the post-disaster images.

[0055] The pre-disaster image, after pre-alignment and coordinated super-resolution, is represented as X, and the post-disaster image, after pre-alignment and coordinated super-resolution, is represented as Y. A transformation matrix T, controlled by parameter W, is used for image correction to align them. A normal distribution normalization function of the transformation matrix T is introduced to correct minor differences in the images, i.e.:

[0056] X′=1 / (1-exp(-Cov 3×3 ((TXT -1 -μ) / b))) (4)

[0057] Y′=1 / (1-exp(-Cov 3×3 ((TYT -1 -μ) / b))) (5)

[0058] T=f(W) (6)

[0059] W=D((B(X)-B(Y)),(B(X down )+B(Y down )) >1 (7)

[0060] In the formula, X′ and Y′ represent the pre-disaster image and post-disaster image after deviation correction, respectively; exp(·) represents an exponential function with the natural constant e as the base; cov 3×3 (·) indicates a convolution operation using a 3×3 kernel; T -1This represents the inverse of T; μ and b are adaptive parameters for offset and scaling, respectively; T is a transformation matrix controlled by the image unalignment factor W, trained using a loss metric constructed through a structural similarity function, and its size is consistent with the image feature size; f(·) is a 5×5 convolutional kernel filter used to filter out outliers in W; B(·) represents feature binarization; X down Y down B(X) represents the downsampling features of X and Y, and downsampling can reduce the misalignment differences in images; B(X)-B(Y) represents the misalignment features of X and Y; (B(X)-B(Y ... down )+B(Y down )) >1 The sum of the downsampled features of X and Y is greater than 1; D(A,B) represents the Euclidean distance between a feature point in A and the nearest feature point in B.

[0061] First, X′ and Y′ are obtained using the transformation matrix T. Then, the structural similarity function is used to evaluate the similarity between X′ and Y′. If the similarity is less than 90%, it is considered unaligned. Formulas (4)-(7) are executed again until the alignment error is less than 10%, thus achieving accurate alignment. After accurately aligning the images before and after the disaster, the feature differences between the images before and after the disaster are calculated to obtain the image change information. The feature differences include the differences between the features of the images before and after the disaster, the differences between the fusion features of the images before and after the disaster and the images before and after the disaster, and the differences between the fusion features of the images after the disaster and the images before and after the disaster. The three differences converge into the overall feature differences of the images before and after the disaster. The specific process of feature difference calculation is as follows: ① Calculate the fusion features of the images before and after the disaster: R0 = Cat(X,Y), where Cat(X,Y) means stitching X and Y together; ② Perform channel dimensionality reduction processing on the fusion features of the images before and after the disaster: R1 = Cov 1×1 (R0), where Cov 1×1 (·) indicates a convolution operation using a 1×1 kernel; ③ Calculate the difference features between pre-disaster and post-disaster images: D1=1 / (1+e X-Y ); ④ Calculate the difference between R1 and the post-disaster image: ⑤ Calculate the difference between R1 and the pre-disaster image: ⑥ Fuse the post-disaster image with the obtained differential features: M′=Cat(D1,D2,D3,Y), where Cat(D1,D2,D3,Y) represents concatenating D1, D2, D3, and Y together; ⑦ Perform channel dimensionality reduction on the fused differential features: M=Cov 1×1 (M′), where Cov 1×1 (·) indicates a convolution operation using a 1×1 convolution kernel.

[0062] Step 5: Based on the target difference change features extracted in Step 4, use target boundary constraints to detect the degree of damage to buildings and the affected area in natural disasters.

[0063] Deep learning models predict the location of building targets based on the predicted offset between the building target and the anchor frame. In the extracted image features, boundary features are represented at a shallow level, while deeper boundary features are replaced by abstract features. Because target features are small in large-scale images, adjacent target features can interfere with each other after feature extraction by a deep network, thus affecting the detection of building target locations. This invention introduces shallow boundary features into deeper features, providing boundary constraints for location and category prediction. The location box detection method based on target boundary constraints first predicts the pixel-level boundary region of the target, and then predicts the building target location and damage type based on the boundary constraints. Its mathematical form is as follows:

[0064] E′=Cov 3×3 (R0) (8)

[0065] E = Cat(Cov) 3×3 (M×w >a (E′)),Cov 3×3 (M×(1-w <a (E′)))) (9)

[0066] In the formula, E′ represents the boundary feature; cov 3×3 (·) represents a convolution operation using a 3×3 kernel; R0 represents the fusion feature of pre-disaster and post-disaster images; Cat(X,Y) represents concatenating X and Y together; M represents the feature obtained by fusing the difference information; w(·) is a threshold function that represents extracting boundary features greater than or less than the threshold a from E′; E represents the feature of the target fusion boundary information.

[0067] Deep learning models are used to predict the types of building damage, which are divided into four levels: no damage, minor damage, moderate damage, and severe damage. The area affected by natural disasters is estimated based on the predicted number and area of ​​damaged buildings.

[0068] Step 6: Using the information on damaged buildings detected in Step 5, assess the severity of the natural disaster and the resulting economic losses.

[0069] In natural disasters, damage to buildings can indirectly indicate the harm caused by the disaster. The severity of the natural disaster and the economic loss can be assessed based on the type of building damage and the area affected.

[0070] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for assessing the extent of building damage during natural disasters using remote sensing images, characterized in that, Includes the following steps: Step 1: Obtain high-resolution multi-source remote sensing images of the disaster area before and after the disaster. Step 2: Perform pre-alignment correction on the acquired multi-source remote sensing images to reduce spatial position differences and distortions in images acquired by different remote sensing devices; First, a feature point extraction operator is used to calculate the key feature points in the entire image. Then, feature matching is performed on the common regions of the pre-disaster and post-disaster images. After a successful match, the geometric feature difference of the common regions cannot exceed a certain threshold. ,Right now , The matching index for aligning remote sensing images of the disaster area before and after the disaster, based on a set threshold. The calculation method is as follows: （1） In the formula, MR This represents the matching index for image pre-alignment before and after the disaster in the affected area. This indicates differences in image filtering background, as well as non-hazardous differences caused by lighting and seasonal variations. , This refers to pre-disaster and post-disaster remote sensing images that have not undergone pre-alignment correction. , This represents pre- and post-disaster remote sensing images after pre-alignment and correction. Step 3: Perform collaborative super-resolution on the multi-source remote sensing images of different resolutions after pre-alignment and correction in Step 2, so that the images obtained by different remote sensing devices can achieve the same level of resolution. Step 4: Use the same resolution image obtained in Step 3 to detect changes in target features and extract the differences in subtle features of the same target at different times in a large-scale scene. The specific process for calculating feature differences is as follows: ① Calculate the fusion features of pre-disaster and post-disaster images: ,in Indicates will X and Y ① The images are stitched together; ② Channel dimensionality reduction is performed on the fusion features of pre-disaster and post-disaster images: ,in Indicates use Convolution operation of convolution kernel; ③ Calculate the difference features between pre-disaster and post-disaster images: ④ Calculation Differences from post-disaster images: ⑤ Calculation Differences from pre-disaster images: ; ⑥ Fuse the post-disaster images with the obtained differential features: ,in Indicates will , , and Y ⑦ Segment the data together; ⑦ Perform channel dimensionality reduction on the fused differential features: ,in Indicates use Convolution operation of convolution kernel; Step 5: Based on the target difference change characteristics extracted in Step 4, use target boundary constraints to detect the degree of damage to buildings and the affected area in natural disasters; Step 6: Using the information on damaged buildings detected in Step 5, assess the severity of the natural disaster and the resulting economic losses.

2. The method for assessing the degree of building damage in natural disasters using remote sensing images as described in claim 1, characterized in that: The high-resolution multi-source remote sensing images acquired in step 1 include satellite remote sensing images, spaceborne remote sensing images, and UAV remote sensing images in different bands. The spatial resolution of the images must be less than [missing information]. m / pix, image resolution at The noise content of the image is no greater than , The threshold value is set.

3. The method for assessing the degree of building damage in natural disasters using remote sensing images as described in claim 1, characterized in that: In step 3, collaborative super-resolution is used to comprehensively analyze the difference information of multi-source images and predict the texture information in high-resolution cases, so that the resolution of all multi-source images reaches the highest resolution level, i.e.: （2） （3） In the formula, This represents the remote sensing image after super-resolution. S This indicates an over-resolution operation. This indicates a remote sensing image that has not undergone over-resolution. This represents a set of information from multiple image sources. This indicates the information difference between the super-resolution remote sensing image and other source images, which are obtained by downsampling high-resolution images. N Indicates that there is N One source image, This indicates downsampling processing. c The constraint representing image super-resolution is that the information of the original image cannot be destroyed. c tend The more information the super-resolution image retains from the original image, the better.

4. The method for assessing the degree of building damage in natural disasters using remote sensing images as described in claim 1, characterized in that: In step 4, the pre-aligned and coordinated super-resolution pre-disaster image is represented as follows: X The post-disaster image, after pre-alignment and coordinated super-resolution, is represented as follows: Y Using a parameter W Control transformation matrix T Perform graphic correction to align it; introduce a transformation matrix. T The normal distribution normalization function corrects small differences in the image, that is: （4） （5） （6） （7） In the formula, These represent the pre-disaster and post-disaster images after deviation correction, respectively. This represents an exponential function with the natural constant e as its base. Indicates use Convolution operation of convolution kernel; Indicates to Find the inverse; These are the adaptive parameters for offset and scaling, respectively; T It is affected by the image non-alignment factor The control transformation matrix is ​​trained using a loss metric constructed from structural similarity functions, and its size is consistent with the size of the image features. convolution kernel Filters used for filtering Outliers in; Binarization of features; express X , Y The downsampling feature can reduce the misalignment differences in images; express X and Y The non-aligned features; express X , Y The sum of the downsampled features is greater than 1; express A Middle feature points and B The Euclidean distance to the nearest feature point in the middle; First use the transformation matrix T get Then use structural similarity functions to evaluate and The similarity, if the similarity is less than If the alignment error is less than 0.5%, it is considered misaligned, and formulas (4)-(7) are executed again until the alignment error is less than 0.5%. %, to achieve precise alignment. and The threshold value is set.

5. The method for assessing the degree of building damage in natural disasters using remote sensing images as described in claim 4, characterized in that: In step 4, after accurately aligning the images before and after the disaster, the feature differences between the images before and after the disaster are calculated to obtain the image change information. The feature differences include three parts: the difference between the features of the images before and after the disaster, the difference between the fusion features of the images before and after the disaster, and the difference between the fusion features of the images after the disaster and the images before and after the disaster. These three differences converge into the overall feature differences of the images before and after the disaster.

6. The method for assessing the degree of building damage in natural disasters using remote sensing images as described in claim 1, characterized in that: Step 5 introduces shallow boundary features into deep features, providing boundary constraints for location and category prediction. The location box detection method based on target boundary constraints first predicts the pixel-level boundary region of the target, and then predicts the location and damage type of the building target based on the boundary constraints. Its mathematical form is as follows: （8） （9） In the formula, Indicate boundary features; Indicates use Convolution operation of convolution kernel; This indicates the fusion features of pre-disaster and post-disaster images; Indicates will X and Y spliced ​​together; This represents the feature obtained by fusing the differential information; It is a threshold function, representing from Extracting values ​​greater than or less than a threshold a Boundary features; Features representing the fusion of target boundary information; Deep learning models are used to predict the types of building damage, which are divided into four levels: no damage, minor damage, moderate damage, and severe damage. The area affected by natural disasters is estimated based on the predicted number and area of ​​damaged buildings.

Images