Sequential InSAR coherence graph building damage detection method based on generative model

Abstract

The invention relates to the technical field of remote sensing image anomaly detection, in particular to a sequential InSAR coherence graph building damage detection method based on a generative model, which comprises the following steps: generating a building area InSAR coherence image time sequence by using a building area SAR image time sequence; training a WGAN generation model by using the InSAR coherent images of the building area from the second moment to the (T-1) th moment to obtain a trained WGAN generation model; subtracting the InSAR coherent image, generated by the trained WGAN generation model, of the building area at the (T + 1) th moment when the disaster does not occur from the InSAR coherent image, generated by the trained WGAN generation model, of the building area at the (T + 1) th moment after the disaster occurs, so as to obtain an abnormal image; and acquiring abnormal information based on the abnormal image, and comparing and screening the building damage area by using a set threshold value. According to the invention, a more accurate detection method is provided for the post-disaster building damage condition, and the building damage detection performance is improved.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image anomaly detection, and in particular to a sequential InSAR coherence image building damage detection method based on a generation model. Background Art

[0002] In recent years, with the rapid development of satellite remote sensing technology, the high-resolution and wide-width remote sensing images obtained can provide detailed and extensive observations of the land surface, which is an important means of assessing the damage to large-scale buildings after natural disasters. By comparing the changes in satellite images before and after the disaster, the extent of damage to buildings can be accurately assessed, the scope and severity of the disaster can be quickly understood, and effective rescue and reconstruction plans can be formulated. Since synthetic aperture radar (SAR) can penetrate clouds and fog, it can obtain clear images under severe weather conditions, providing an important basis for assessing the damage to buildings caused by disasters such as earthquakes, floods and landslides; with its all-weather and all-day imaging capabilities, the application value of remote sensing technology in disaster emergency response is further enhanced.

[0003] Traditional building damage detection methods based on SAR images usually use the intensity value of SAR images, that is, by comparing the changes in the scattering intensity of the building area in the pre-disaster and post-disaster images, the corresponding damaged area is detected; however, this backscattering intensity information reflects the electromagnetic characteristics and roughness of the surface objects, and it is not sensitive to subtle changes such as building collapse and surface cracks, resulting in low accuracy in reflecting the disturbance of the land surface, and thus causing the problem of low accuracy in detecting the damaged area of ​​the building; therefore, in order to solve this problem, the prior art uses interference processing on two SAR images before and after the disaster, and the resulting coherent image can reflect the phase consistency of the land surface before and after the disaster, and it is very sensitive to slight changes in the land surface, and can reflect slight deformation and structural changes of the land surface. At the same time, compared with the backscattering intensity, the coherence change can more accurately reflect the disturbance of the land surface, especially for subtle changes such as building collapse and surface cracks, and has higher detection capabilities. In addition, sequential InSAR coherent images can also further improve the accuracy of building damage detection.

[0004] Deep learning has become the mainstream technology for time series data analysis. Among them, the use of recurrent neural networks (RNNs) as probabilistic anomaly detectors has been applied to the field of abnormal change detection in sequential interferometric synthetic aperture radar (InSAR) coherent images. However, in order to simplify the mathematical model, the existing methods assume that the coherent data obeys a Gaussian distribution, while the actual coherent data cannot meet the above probability density assumption, which in turn affects the final building damage detection effect. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the problem that the prior art is not based on the actual distribution of coherent data, and the processing of sequential InSAR coherent images results in the final abnormal image being of low accuracy and having no reference value, thus affecting the results of building damage detection.

[0006] In order to solve the above technical problems, the present invention provides a method for building damage detection based on sequential InSAR coherence images of a generative model, comprising:

[0007] Acquire the SAR images of the building area from time 1 to time T+1 to form a SAR image time series of the building area; wherein the SAR images of the building area from time 1 to time T are all SAR images of the building area before the disaster occurs; the SAR image of the building area at time T+1 is the SAR image of the building area after the disaster occurs;

[0008] Perform registration operation and coherence calculation on the SAR image time series of the building area to obtain the InSAR coherent images of the building area from the 2nd to the T+1st time, forming the InSAR coherent image time series of the building area;

[0009] Input the InSAR coherent images of the building area from the 2nd to T-1st time into the WGAN generation model to generate the InSAR coherent images of the building area from the 3rd to T time; wherein the WGAN generation model includes a generator and a discriminator; the InSAR coherent images of the building area from the 3rd to T time generated by the WGAN generation model are used as generated image samples, and the InSAR coherent images of the building area from the 3rd to T time are used as real image samples; with the goal of minimizing the Wasserstein distance between the real image sample and the generated image sample and minimizing the absolute error loss, the loss function of the generator is constructed; with the goal of minimizing the Wasserstein distance between the generated image sample and the real image sample, the loss function of the discriminator is constructed to train the WGAN generation model;

[0010] The InSAR coherent images of the building area from the 2nd to Tth moment are input into the trained WGRN model to generate the InSAR coherent images of the building area from the 3rd to T+1th moment when no disaster occurs; the generated InSAR coherent image of the building area at the T+1th moment when no disaster occurs is subtracted from the InSAR coherent image of the building area at the T+1th moment after the disaster occurs to obtain the abnormal image; the building area map is converted into a binary mask image of the building area and then multiplied with the abnormal image to extract multiple abnormal information values ​​of the building area; each abnormal information value of the building area is compared with the set threshold. If the current abnormal information is greater than the threshold, the area corresponding to the current abnormal information is determined to be a building damaged area.

[0011] Preferably, performing a registration operation on the SAR image time series of the building area to obtain the registered SAR image time series of the building area includes:

[0012] The SAR image at the first moment in the SAR image time series of the building area is used as the main image, and the remaining images in the SAR image time series of the building area are used as the SAR image time series to be registered;

[0013] The registration operation includes: a coarse registration stage and a fine registration stage;

[0014] The coarse registration stage includes:

[0015] Taking the center point of the main image as the initial control point, through the iterative algorithm, find the same-name point corresponding to the initial control point in the SAR image to be registered at the current moment; based on the coordinates of the initial control point and the coordinates of the same-name point corresponding to the initial control point in the SAR image to be registered at the current moment, calculate the coordinate offset between the SAR image to be registered at the current moment and the main image;

[0016] The fine registration stage includes:

[0017] Select multiple control points in the main image, and determine the same-name points corresponding to the SAR image to be registered at the current moment and each control point based on the coordinate offset between the SAR image to be registered at the current moment and the main image obtained in the coarse registration stage;

[0018] The coordinates of the points with the same names corresponding to the SAR image to be registered at the current moment and each control point are screened, and the optimal coordinate transformation model corresponding to the SAR image to be registered at the current moment is fitted; based on the optimal coordinate transformation model corresponding to the SAR image to be registered at the current moment, the coordinate transformation and resampling of the SAR image to be registered at the current moment are performed to obtain the registered SAR image at the current moment;

[0019] The registration operation is performed on the SAR images to be registered at each moment in turn to obtain the registered SAR images at each moment, and a time series of SAR images of the registered building area is formed.

[0020] Preferably, according to the coherence coefficient calculation formula, the coherence calculation is performed on the time series of the SAR images of the building area after registration to obtain the time series of the InSAR coherent images of the building area, which includes:

[0021] The coherence coefficient calculation formula is expressed as: ; in, , They are the first Time, Coordinates in the SAR image at time The complex value at ; Represents the first SAR image at the moment and The pixels of the SAR image at the time are conjugate multiplied to obtain the Interference images of moments; , Respectively represent the number of rows and columns of the data block during coherence calculation; ; The InSAR coherent images of the building area from time 2 to time T+1 are obtained in sequence to form a time series of InSAR coherent images of the building area.

[0022] Preferably, the loss function of the construction generator is expressed as: ; in, represents the loss of the generator; Indicates The generated image samples at time instant; Represents the sample distribution generated by the generator; represents the mean; represents the discriminator network; Denotes the discriminator for The score of the generated image sample at the moment; The weight hyperparameter representing the loss; Indicates The absolute error loss between the real image sample and the generated image sample at each moment; represents a real image sample.

[0023] Preferably, the loss function of constructing the discriminator is expressed as: ; in, represents the loss of the discriminator; represents a real image sample; Represents the real image sample distribution; Indicates The generated image samples at time instant; Represents the sample distribution generated by the generator; represents the mean; represents the discriminator network; represents the score of the discriminator for the real image sample; Denotes the discriminator for The score of the generated image sample at time t.

[0024] Preferably, the generator is composed of a GRU model with three hidden layers of size 128 and a fully connected layer.

[0025] Preferably, the update expression of the GRU model is: ; ; ; ; in, represents the update gate; Represents the Sigmoid activation function, which maps the model input to the interval [0,1]; Indicates The input sample corresponding to the moment, that is, A true image of the moment; , Respectively represent the weight matrices of the update gate input state and hidden state; represents the update gate bias term; Reset gate. , Respectively represent the weight matrices of the reset gate input state and hidden state; Represents the reset gate bias term; Indicates Candidate hidden states at time instant; Indicates The hidden state of the moment; Indicates The hidden state corresponding to the moment; , Represent the input state of the candidate hidden state and the weight matrix of the hidden state respectively; A bias term representing the candidate hidden state; Represents element-wise multiplication;

[0026] No. The update expression of the hidden state at the moment is: ;in, Represents the GRU model.

[0027] Preferably, the expression of the fully connected layer is: ;

[0028] in, Indicates The output sample at time is the output of the fully connected layer, that is, Generate images at the moment; represents a fully connected layer; Indicates The hidden state at the moment is the input of the fully connected layer.

[0029] Preferably, the discriminator is composed of a one-dimensional CNN model with a convolution kernel of 1×1 and 64 output channels and a fully connected layer.

[0030] Preferably, during the training process of the WGAN generative model, in each training cycle, the generator is updated once after the discriminator is cyclically updated five times.

[0031] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0032] The invention discloses a method for detecting building damage using sequential InSAR coherence images based on a generative model. The method performs registration, interference operation, and coherence calculation on sequential SAR images of the land surface of a building area within a certain period of time, so as to obtain sequential InSAR coherence images of the land surface within a certain period of time, namely, a time series of InSAR coherence images. In the model training process, in order to enable the generator to better learn the time changes of the InSAR coherence images of the building area before the disaster occurs and to ensure that the images generated by the generator have a higher similarity with the real images, a loss function of the generator is constructed with the goal of minimizing the Wasserstein distance between the real image samples and the generated image samples and minimizing the absolute error loss. In order to allow the discriminator to focus only on the difference between the generated samples and the real samples during the training process, a loss function of the discriminator is constructed with the goal of minimizing the Wasserstein distance between the generated image samples and the real image samples. At the same time, since the time before the disaster occurs The sequence of InSAR coherent images are all free of anomalies. Therefore, if they are used as training data, the predicted coherent images generated by the generator in the trained WGAN generation model will not be affected by anomalies. In addition, the Wasserstein distance is considered in the loss of the WGAB generation model, so that the WGAB generation model can better capture the complex image distribution, thereby generating more realistic and higher-resolution coherent images. Therefore, the coherent images predicted and generated by the WGAB generation model before the disaster are highly accurate, providing more accurate data support for the subsequent assessment of the damage to buildings after the disaster. The first predicted coherent image after the disaster is subtracted from the first real coherent image to obtain an abnormal image, and the post-disaster building damage area is screened out using the set threshold, which can more accurately distinguish between changes caused by disasters and changes caused by natural development, further improving the detection accuracy of abnormal coherence changes and the accuracy of the detection results of building damage areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0034] Figure 1 It is a flow chart of a method for detecting building damage using sequential InSAR coherence images based on a generative model provided by the present invention;

[0035] Figure 2 This is an abnormal image of the Italian town of Norcia after the disaster, obtained by using the sequential InSAR coherence image building damage detection method based on the generative model;

[0036] Figure 3It is a schematic diagram of the statistical distribution of true coherence values ​​and the statistical distribution of coherence values ​​generated by WGAN after the disaster in the small town of Norcia, Italy;

[0037] Figure 4 This is a comparison chart of the precision-recall curves of the damage detection effect after the disaster in the small town of Norcia, Italy, obtained based on different methods. DETAILED DESCRIPTION

[0038] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0039] Reference Figure 1 As shown, Figure 1 The present invention provides a flow chart of a sequential InSAR coherence image building damage detection method based on a generative model; specifically comprising:

[0040] S1: Acquire SAR images of the building area from time 1 to time T+1 to form a SAR image time series of the building area; wherein the SAR images of the building area from time 1 to time T are all SAR images of the building area before the disaster occurs; the SAR image of the building area at time T+1 is the SAR image of the building area after the disaster occurs;

[0041] S2: Perform registration operation and coherence calculation on the SAR image time series of the building area to obtain the InSAR coherent images of the building area from the 2nd to the T+1st time, forming the InSAR coherent image time series of the building area; wherein, the InSAR coherent images of the building area from the 2nd to the Tth time are all the InSAR coherent images of the building area before the disaster occurs; the InSAR coherent image of the building area at the T+1th time is the InSAR coherent image of the building area at the T+1th time after the disaster occurs, that is, the first real coherent image of the building area after the disaster occurs, including:

[0042] S21: performing a registration operation on the SAR image time series of the building area to obtain a registered SAR image time series of the building area, including:

[0043] The SAR image at the first moment in the SAR image time series of the building area is used as the main image, and the remaining images in the SAR image time series of the building area are used as the SAR image time series to be registered;

[0044] The registration operation includes: a coarse registration stage and a fine registration stage;

[0045] The coarse registration stage includes:

[0046] Taking the center point of the main image as the initial control point, through the iterative algorithm, find the same-name point corresponding to the initial control point in the SAR image to be registered at the current moment; based on the coordinates of the initial control point and the coordinates of the same-name point corresponding to the initial control point in the SAR image to be registered at the current moment, calculate the coordinate offset between the SAR image to be registered at the current moment and the main image;

[0047] The fine registration stage includes:

[0048] Select multiple control points in the main image, and determine the same-name points corresponding to the SAR image to be registered at the current moment and each control point based on the coordinate offset between the SAR image to be registered at the current moment and the main image obtained in the coarse registration stage;

[0049] The coordinates of the points with the same names corresponding to the SAR image to be registered at the current moment and each control point are screened, and the optimal coordinate transformation model corresponding to the SAR image to be registered at the current moment is fitted; based on the optimal coordinate transformation model corresponding to the SAR image to be registered at the current moment, the coordinate transformation and resampling of the SAR image to be registered at the current moment are performed to obtain the registered SAR image at the current moment;

[0050] The registration operation is performed on the SAR images to be registered at each moment in turn to obtain the registered SAR images at each moment, and a time series of SAR images of the registered building area is formed;

[0051] S22: According to the coherence coefficient calculation formula, the coherence calculation is performed on the time series of the SAR images of the building area after registration to obtain the time series of the InSAR coherent images of the building area, including:

[0052] The coherence coefficient calculation formula is expressed as:

[0053] ;

[0054] in, , They are the first Time, Coordinates in the SAR image at time The complex value at ; Represents the first SAR image at the moment and The pixels of the SAR image at the time are conjugate multiplied to obtain the Interference images of moments; , Respectively represent the number of rows and columns of the data block during coherence calculation; ;

[0055] The InSAR coherent images of the building area from the 2nd to the T+1st time are obtained in sequence to form a time series of InSAR coherent images of the building area;

[0056] In summary, assuming that there are T+1 SAR images in total, T InSAR coherent images will be obtained; interferometric coherence is used to describe the similarity between the regions with the same name in the two images in the interferometric pair, and is distributed in the interval [0,1]; a coherence of 0 means that the two images are completely irrelevant, and a coherence of 1 means that the two images are completely consistent;

[0057] S3: Input the InSAR coherent images of the building area from the 2nd to T-1st time into the WGAN generation model to generate the InSAR coherent images of the building area from the 3rd to T time; wherein the WGAN generation model includes a generator and a discriminator; the InSAR coherent images of the building area from the 3rd to T time generated by the WGAN generation model are used as generated image samples, and the InSAR coherent images of the building area from the 3rd to T time are used as real image samples; with the goal of minimizing the Wasserstein distance between the real image sample and the generated image sample and minimizing the absolute error loss, the loss function of the generator is constructed; with the goal of minimizing the Wasserstein distance between the generated image sample and the real image sample, the loss function of the discriminator is constructed to train the WGAN generation model;

[0058] Among them, the loss function of the generator is expressed as:

[0059] ;

[0060] in, represents the loss of the generator; Indicates The generated image samples at the moment; Represents the sample distribution generated by the generator; represents the mean; represents the discriminator network; Denotes the discriminator for The score of the generated image sample at the moment; The weight hyperparameter representing the loss; Represents the absolute error loss between the real image sample and the generated image sample; represents a real image sample; in a specific embodiment of the present invention, ;

[0061] Among them, the expression of the loss function of the discriminator is:

[0062] ;

[0063] in, represents the loss of the discriminator; represents a real image sample; Represents the real image sample distribution; Indicates The generated image samples at time instant; Represents the sample distribution generated by the generator; represents the mean; represents the discriminator network; represents the score of the discriminator for the real image sample; Denotes the discriminator for The score of the generated image sample at the moment;

[0064] Among them, the update expression of the GRU model is:

[0065] ;

[0066] ;

[0067] ;

[0068] ;

[0069] in, Represents the update gate, which is used to control how much of the state information at the previous moment should be retained in the current state, and how much new information in the current input should be added to the current state, so that the network can flexibly adjust its memory capacity to effectively capture long-term dependencies; Represents the Sigmoid activation function, which maps the model input to the interval [0,1]; Indicates The input sample corresponding to the moment, that is, A true image of the moment; , Respectively represent the weight matrices of the update gate input state and hidden state; represents the update gate bias term; Represents the reset gate, which is used to control the extent to which the hidden state of the previous moment can be ignored. If the reset gate believes that past experience is not relevant to the current task, it will make the network rely more on the current input to make decisions; , Respectively represent the weight matrices of the reset gate input state and hidden state; Represents the reset gate bias term; Indicates Candidate hidden states at time instant; Indicates The hidden state of the moment; Indicates The hidden state corresponding to the moment; , Represent the input state of the candidate hidden state and the weight matrix of the hidden state respectively; A bias term representing the candidate hidden state; Represents element-by-element multiplication; using the update gate and reset gate, GRU can effectively ignore data with large seasonal differences when predicting InSAR coherence, and focus on the sequence patterns of historically similar conditions, thereby improving the prediction accuracy;

[0070] No. The update expression of the hidden state at the moment is: ;in, Represents the GRU model;

[0071] Among them, the expression of the fully connected layer of the generator is:

[0072] ;

[0073] in, Indicates The output sample at time is the output of the fully connected layer, that is, The generated image samples at time instant; represents a fully connected layer; Indicates The hidden state at the moment is the input of the fully connected layer;

[0074] In summary, assuming that the image before the disaster is , then the training is mainly used To predict coherence time series;

[0075] The WGAN network consists of a generator and a discriminator. The generator is trained by using coherent images without abnormalities to learn the temporal changes of coherence without abnormalities. The generated coherence map is also without abnormalities. The discriminator is used to represent the Wasserstein distance. When training WGAN, the input is normalized to [-1,1] and the discriminator is used to train the generator. Indicates The InSAR coherent image at time The real image of the moment, Indicates the generated The InSAR coherent image at time The generated image at the moment; the calculation formula of Wasserstein distance is:

[0076] ;

[0077] in, Represents the true sample distribution The sample distribution generated by the generator Wasserstein distance between them; express and The set of all possible joint distributions; express and A possible joint distribution of

[0078] The advantage of Wasserstein distance is that it can reflect the distance between two distributions even if they do not overlap. The Wasserstein distance is converted into the following formula:

[0079] ;

[0080] Define the transformed loss as the generator loss; where express The Lipschitz constant does not exceed K; during the training process, all parameters of the neural network are set Not exceeding a certain range [-c,c], construct a The discriminator network ; Based on the discriminator network , in the parameters Without exceeding the range [-c,c], To maximize as much as possible, The Wasserstein distance between the approximate real image sample distribution and the generated image sample distribution;

[0081] Since the generator needs to minimize the Wasserstein distance between the real image sample distribution and the generated image sample distribution, that is, minimize , and because The first term of has nothing to do with the generator, so the generator should minimize the following formula, namely: ;

[0082] Therefore, the generator loss and discriminator loss of the WGAN model are:

[0083] ;

[0084] ; in, Represents the original generator loss of the WGAN model;

[0085] The generator and discriminator of WGAN are trained using the time series coherent images before the disaster to predict the time series coherent images. Since the coherent images before the disaster are not affected by the anomaly, the generator only generates coherent images that are not affected by the anomaly.

[0086] The loss of the generator in the present invention is: ;

[0087] in, It is the L1 loss (absolute error loss) of the generated data and the actual data. Adding the L1 loss to the generator's loss can improve the quality and stability of the generated images, as well as increase the similarity in content between the generated images and the real images.

[0088] In a specific embodiment of the present invention, the generator is composed of a GRU model with three hidden layers of size 128 and a fully connected layer; the discriminator is composed of a one-dimensional CNN model with a convolution kernel of 1×1 and an output channel of 64 and a fully connected layer; the GRU network can learn long-term dependencies in time series, and the discriminator uses a one-dimensional CNN network with a convolution kernel of 1×1 and a fully connected layer for feature extraction and fusion; the one-dimensional CNN network can optimize input features through channel transformation without changing the time dimension of the feature map, and learn richer feature representations; a fully connected layer is used for feature fusion, and the diversified features extracted by the CNN network are integrated, and the generated data and the real data are added to the discriminator respectively to realize the representation of Wasserstein distance;

[0089] During the training process, the present invention updates the discriminator 5 times in a cycle and then updates the generator once, so that the negative discriminator loss can better represent the Wasserstein distance between the generated sequence and the actual sequence; after multiple iterations, the generator can better generate data until the loss of the discriminator is close to stability, and the training is stopped;

[0090] Since the coherent images before the disaster were normal, the trained WGAN network was used to generate the coherent images after the disaster. The predicted coherent images included the influence of temporal decorrelation but were not affected by abnormalities. Since the loss of the discriminator is the negative Wasserstein distance, its absolute value can be used to indicate the training results.

[0091] S4: Input the InSAR coherent images of the building area from the 2nd to Tth moment into the trained WGRN model to generate the InSAR coherent images of the building area from the 3rd to T+1th moment when no disaster occurs; subtract the generated InSAR coherent image of the building area at the T+1th moment when no disaster occurs from the InSAR coherent image of the building area at the T+1th moment after the disaster occurs to obtain the abnormal image; convert the building area map into a binary mask image of the building area, and then multiply it with the abnormal image to extract multiple abnormal information values ​​of the building area; compare each abnormal information value of the building area with the set threshold. If the current abnormal information is greater than the threshold, the area corresponding to the current abnormal information is determined to be the building damaged area.

[0092] To verify the effectiveness of the proposed building damage detection method, data from the central Italian earthquake on August 24, 2016 were used in this experiment. This event destroyed most of the small town of Norcia, Italy. The Copernicus Emergency Management Service compiled a damage map to assess the extent of damage to each building. This comprehensive damage assessment can quantitatively verify the effectiveness of using the WGAN method for anomaly detection based on the existing loss level. In the experiment, 35 coherent images were actually generated using Sentinel-1 VV-polarized SLC image processing, of which 34 were pre-disaster coherent images and 1 was post-disaster coherent image.

[0093] Figure 2 The results of building damage detection using the WGAN method are shown. The present invention uses real damage data from the Copernicus Management Service to select the optimal threshold for damage.

[0094] The WGAN method mainly makes predictions by learning the distribution of data, without making any assumptions about the data distribution of the image. Figure 3 As can be seen in the figure, the distribution of coherent images does not strictly follow the Gaussian distribution, and the WGAN method can learn the distribution of coherent images more accurately.

[0095] To calculate the optimal loss threshold for each method, the weighted harmonic mean of precision and recall can be used , the formula is:

[0096] ;

[0097] Will change with the change of threshold, set in the experiment , thus calculating all possible thresholds of the three methods score; selected The weighted precision is twice as high as the recall because we want to direct limited emergency resources to where they are most likely to be damaged, thus supporting higher precision at the expense of lower recall; When the value is , the precision of the three methods is higher than 0.8, which means that more than 80% of the damaged points are indeed damaged, but the recall of the WGAN method is 0.6332, and the recalls of the other two methods are 0.6080 and 0.5611, which is a significant quantitative improvement; the optimal threshold of each method and the corresponding precision and recall are shown in Table 1.

[0098] Table 1 Experimental results obtained using different detection methods:

[0099] method Optimal Threshold Optimum accuracy Optimal recall Accuracy AUC CCD 0.3079 0.8505 0.5611 0.8133 RNN 2.1259 0.88018 0.6080 0.8588 WGAN 0.8473 0.8879 0.6332 0.8622

[0100] The area under the precision-recall curve can be used to quantitatively compare the effects of different detection methods; the closer the area under the precision-recall curve is to 1, the better the effect; Figure 4 It is the precision-recall curve of different methods. The area under the precision-recall curve of the WGAN method is 0.8622, while the areas under the precision-recall curves of the RNN and CDD methods are 0.8588 and 0.8133 respectively, which shows some improvement.

[0101] In summary, the sequential InSAR coherent image building damage detection system based on deep generative model proposed in the present invention can achieve better post-disaster building damage detection effect.

[0102] The present invention relates to a sequential InSAR coherent image building damage detection method based on a deep generative model, which generates a coherent image time series using a SAR image time series; calculates the loss function of the generator and discriminator of the WGAN network; trains the WGAN using the coherent image time series before the disaster occurs; uses the trained WGAN network to predict the coherent image after the disaster occurs, and compares the predicted coherent image with the actual coherent image to obtain an abnormal image. The present invention can provide a more accurate detection method for post-disaster building damage under all-day and all-climate conditions, thereby improving the performance of building damage detection.

[0103] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for building damage detection based on sequential InSAR coherence images of generative models, characterized in that: include: Acquire the SAR images of the building area from time 1 to time T+1 to form a SAR image time series of the building area; wherein the SAR images of the building area from time 1 to time T are all SAR images of the building area before the disaster occurs; the SAR image of the building area at time T+1 is the SAR image of the building area after the disaster occurs; Perform registration operation and coherence calculation on the SAR image time series of the building area to obtain the InSAR coherent images of the building area from the 2nd to the T+1st time, forming the InSAR coherent image time series of the building area; Input the InSAR coherent images of the building area from the 2nd to T-1st time into the WGAN generation model to generate the InSAR coherent images of the building area from the 3rd to T time; wherein the WGAN generation model includes a generator and a discriminator; the InSAR coherent images of the building area from the 3rd to T time generated by the WGAN generation model are used as generated image samples, and the InSAR coherent images of the building area from the 3rd to T time are used as real image samples; with the goal of minimizing the Wasserstein distance between the real image sample and the generated image sample and minimizing the absolute error loss, the loss function of the generator is constructed; with the goal of minimizing the Wasserstein distance between the generated image sample and the real image sample, the loss function of the discriminator is constructed to train the WGAN generation model; The InSAR coherent images of the building area from the 2nd to Tth moment are input into the trained WGRN model to generate the InSAR coherent images of the building area from the 3rd to T+1th moment when no disaster occurs; the generated InSAR coherent image of the building area at the T+1th moment when no disaster occurs is subtracted from the InSAR coherent image of the building area at the T+1th moment after the disaster occurs to obtain the abnormal image; the building area map is converted into a binary mask image of the building area and then multiplied with the abnormal image to extract multiple abnormal information values ​​of the building area; each abnormal information value of the building area is compared with the set threshold. If the current abnormal information is greater than the threshold, the area corresponding to the current abnormal information is determined to be a building damaged area.

2. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 1, characterized in that: The SAR image time series of the building area is registered, and the SAR image time series of the building area after registration includes: The SAR image at the first moment in the SAR image time series of the building area is used as the main image, and the remaining images in the SAR image time series of the building area are used as the SAR image time series to be registered; The registration operation includes: a coarse registration stage and a fine registration stage; The coarse registration stage includes: Taking the center point of the main image as the initial control point, through the iterative algorithm, find the same-name point corresponding to the initial control point in the SAR image to be registered at the current moment; based on the coordinates of the initial control point and the coordinates of the same-name point corresponding to the initial control point in the SAR image to be registered at the current moment, calculate the coordinate offset between the SAR image to be registered at the current moment and the main image; The fine registration stage includes: Select multiple control points in the main image, and determine the same-name points corresponding to the SAR image to be registered at the current moment and each control point based on the coordinate offset between the SAR image to be registered at the current moment and the main image obtained in the coarse registration stage; The coordinates of the points with the same names corresponding to the SAR image to be registered at the current moment and each control point are screened, and the optimal coordinate transformation model corresponding to the SAR image to be registered at the current moment is fitted; based on the optimal coordinate transformation model corresponding to the SAR image to be registered at the current moment, the coordinate transformation and resampling of the SAR image to be registered at the current moment are performed to obtain the registered SAR image at the current moment; The registration operation is performed on the SAR images to be registered at each moment in turn to obtain the registered SAR images at each moment, and a time series of SAR images of the registered building area is formed.

3. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 2, characterized in that: According to the coherence coefficient calculation formula, the coherence of the SAR image time series of the building area after registration is calculated, and the InSAR coherent image time series of the building area is obtained, including: The coherence coefficient calculation formula is expressed as: ； in, , They are the first Time, Coordinates in the SAR image at time The complex value at ; Represents the first SAR image at the moment and The pixels of the SAR image at the time are conjugate multiplied to obtain the Interference images of moments; , Respectively represent the number of rows and columns of the data block during coherence calculation; ; The InSAR coherent images of the building area from time 2 to time T+1 are obtained in sequence to form a time series of InSAR coherent images of the building area.

4. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 1, characterized in that: The loss function of the construction generator is expressed as: ； in, represents the loss of the generator; Indicates The generated image samples at time instant; Represents the sample distribution generated by the generator; represents the mean; represents the discriminator network; Denotes the discriminator for The score of the generated image sample at the moment; The weight hyperparameter representing the loss; Indicates The absolute error loss between the real image sample and the generated image sample at each moment; represents a real image sample.

5. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 1, characterized in that: The loss function of constructing the discriminator is expressed as: ； in, represents the loss of the discriminator; represents a real image sample; Represents the real image sample distribution; Indicates The generated image samples at the moment; Represents the sample distribution generated by the generator; represents the mean; represents the discriminator network; represents the score of the discriminator for the real image sample; Denotes the discriminator for The score of the generated image sample at time t.

6. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 1, characterized in that: The generator consists of a GRU model with three hidden layers of size 128 and a fully connected layer.

7. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 6, characterized in that: The update expression of the GRU model is: ； ； ； ； in, represents the update gate; Represents the Sigmoid activation function, which maps the model input to the interval [0,1]; Indicates The input sample corresponding to the moment, that is, A true image of the moment; , Respectively represent the weight matrices of the update gate input state and hidden state; represents the update gate bias term; Reset gate. , Respectively represent the weight matrices of the reset gate input state and hidden state; Represents the reset gate bias term; Indicates Candidate hidden states at time instant; Indicates The hidden state of the moment; Indicates The hidden state corresponding to the moment; , Represent the input state of the candidate hidden state and the weight matrix of the hidden state respectively; A bias term representing the candidate hidden state; Represents element-wise multiplication; No. The update expression of the hidden state at the moment is: ;in, Represents the GRU model.

8. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 6, characterized in that: The expression of the fully connected layer is: ； in, Indicates The output sample at time is the output of the fully connected layer, that is, Generate images at the moment; represents a fully connected layer; Indicates The hidden state at the moment is the input of the fully connected layer.

9. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 1, characterized in that: The discriminator consists of a one-dimensional CNN model with a convolution kernel of 1×1 and 64 output channels and a fully connected layer.

10. The method for building damage detection based on sequential InSAR coherence images of a generative model according to claim 1, characterized in that: During the training process of the WGAN generative model, in each training cycle, the discriminator is updated five times in a cycle and then the generator is updated once.

Images