A disaster before and after weak corresponding area-oriented rejectable SAR image-optical image registration method and system

Abstract

The present application relates to the technical field of remote sensing image processing, and particularly relates to a kind of disaster before and after weak corresponding area-oriented SAR image-optical image registration method and system, including obtaining post-disaster SAR sensing image and pre-disaster optical reference image, according to geographic reference, image-level rough matching or external geometric information to generate geometric physical priori and form search window;Extract the local structure features of two images, generate disaster area change clues and SAR observability risk;According to the similarity of features in the search window, construct a matching cost body and determine its statistics;Features, clues, risks, priori, cost body and its statistics, and geometric residual are input into end-to-end rejectable cross-modal registration network, and joint output registration displacement field, registrable probability graph, reject probability graph and matching covariance graph;According to the above three graphs, filter reliable corresponding sub-region and determine reject sub-region;Only based on reliable corresponding area estimation registration transformation parameter and output result.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and system for registering rejectable SAR images with optical images in areas with weak correspondence before and after disasters. Background Technology

[0002] Registration of SAR (Synthetic Aperture Radar) images with optical remote sensing images is a crucial technical step in disaster emergency response. Currently, existing SAR-optical image registration methods mainly fall into three categories: The first category is region-based methods, which estimate transformation parameters by calculating similarity metrics between images. However, due to the significant radiometric differences between SAR and optical images, these methods struggle to achieve stable registration results. The second category is feature-based methods, which achieve registration by extracting geometric features such as corners and edges, or by combining and statistically analyzing these features, and then matching them. However, speckle noise and overlapping shadows in SAR images make the extraction of cross-modal stable features extremely difficult. The third category is end-to-end registration methods based on deep learning, which have emerged in recent years. These methods directly input SAR and optical images into a neural network, which outputs registration transformation parameters or dense displacement fields. This method has made some progress in conventional scenarios. However, all of the above existing methods implicitly assume that the two images to be registered have a reliable grayscale or geometric correspondence within most sub-regions.

[0003] However, in disaster emergency scenarios, the above assumptions are often difficult to uphold. After a disaster, flooding, building collapses, road damage, and the construction of temporary rescue facilities can lead to a lack of true geometric correspondence between pre-disaster optical reference images and post-disaster SAR sensing images in some areas of the disaster zone. If a matching relationship is forcibly established in these weakly corresponding areas, it is easy to generate incorrect displacement estimates and spurious matching points, thereby contaminating the global registration results. Existing methods generally lack the ability to accurately express whether reliable cross-modal correspondences exist in local areas. Traditional methods based on regions and features estimate the transformation through global optimization or local feature matching. Incorrect matching in weakly corresponding disaster areas directly interferes with the calculation of transformation parameters. Although end-to-end methods based on deep learning can output transformation parameters or displacement fields, they usually compress all information of the entire image indiscriminately into a global output. When the area of ​​weakly corresponding disaster areas is large or the proportion of SAR overlay shadow areas is high, the network output is easily influenced by erroneous textures and false targets, producing registration results that deviate from the true situation. Furthermore, the above methods usually only output transformation parameters, making it difficult to indicate which local areas participated in the transformation estimation and which areas should be rejected. Therefore, a SAR image registration method is needed that can adaptively determine whether there is a reliable correspondence between local regions during the registration process and reject weakly corresponding regions, so as to avoid the contamination of the registration results by weakly corresponding regions in disaster scenarios and improve the reliability and interpretability of SAR image and optical image registration in disaster scenarios. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a refusable SAR image-optical image registration method and system for weak correspondence regions before and after disasters. It aims to improve the problem that existing SAR image-optical image registration methods in disaster scenarios lack the ability to accurately express whether there are reliable cross-modal correspondences in local areas, leading to erroneous matching in weak correspondence regions of disasters that contaminates the global registration results.

[0005] In a first aspect, the present invention provides the following technical solution: a method for registering rejectable SAR images with optical images in areas with weak correspondence before and after a disaster, comprising:

[0006] S1. Acquire post-disaster SAR sensing images and pre-disaster optical reference images; generate a geometric-physical prior based on at least one of the following: embedded georeference in the image product, the result of image-level coarse matching of the two images, or externally input geometric information; and determine the coarse corresponding position and search radius of the optical reference image based on the geometric-physical prior to form a geometric-physical constraint search window.

[0007] S2. Extract local structural features from SAR images and optical images, and generate clues to changes in disaster areas and SAR observability risks;

[0008] S3. Within the geometric and physical constraint search window, construct a matching cost body based on the similarity between the local structural features of the SAR image and the local structural features of the optical image, and determine the matching cost body statistics accordingly.

[0009] S4. Input the local structural features of the SAR image, the local structural features of the optical image, the clues to changes in the disaster area, the SAR observability risk, the geometrical prior, the matching cost volume, the statistics of the matching cost volume, and the geometric residual determined by the difference between the control point predicted by the network and the corresponding position in the optical reference map into an end-to-end rejectable cross-modal registration network. The network jointly outputs the registration displacement field, the registration probability map, the rejection probability map, and the matching covariance map.

[0010] S5. Based on the registration probability map, the rejection probability map, and the matching covariance map, select reliable corresponding regions from the registration displacement field output by the network, and determine the rejection regions.

[0011] S6. Estimate the registration transformation parameters based only on the corresponding control points within the reliable corresponding area, and output the registration result.

[0012] Secondly, the present invention provides the following technical solution: a rejectable SAR image-optical image registration system for weakly correlated areas before and after a disaster, the system comprising:

[0013] The image acquisition and geometric-physical constraint construction module is used to acquire post-disaster SAR sensing images and pre-disaster optical reference images; based on at least one of the following: embedded georeference in the image product, the result of image-level coarse matching of two images, or externally input geometric information, a geometric-physical prior is generated, and the coarse corresponding position and search radius of the optical reference image are determined based on the geometric-physical prior to form a geometric-physical constraint search window;

[0014] The feature extraction and risk generation module is used to extract local structural features of SAR images and optical images, and generate clues to changes in disaster areas and SAR observability risks.

[0015] The matching cost body construction and statistics determination module is used to construct a matching cost body based on the similarity between the local structural features of the SAR image and the local structural features of the optical image within the geometric and physical constraint search window, and determine the matching cost body statistics accordingly.

[0016] A rejection-based cross-modal registration network module is used to input the local structural features of the SAR image, the local structural features of the optical image, the cues of changes in the disaster area, the SAR observability risk, the geometric and physical priors, the matching cost volume, the statistics of the matching cost volume, and the geometric residual determined by the difference between the network-predicted corresponding control point and the coarse corresponding position of the optical reference image into an end-to-end rejection-based cross-modal registration network. The network jointly outputs a registration displacement field, a registration probability map, a rejection probability map, and a matching covariance map.

[0017] The reliable corresponding sub-region filtering and rejection sub-region determination module is used to filter reliable corresponding regions from the registration displacement field output by the network based on the registration probability map, the rejection probability map and the matching covariance map, and determine the rejection regions.

[0018] The registration transformation parameter estimation and result output module is used to estimate the registration transformation parameters based only on the corresponding control points within the reliable corresponding area, and output the registration results.

[0019] The present invention has the following beneficial effects:

[0020] 1. In this invention, a registration probability map is jointly output through an end-to-end network, enabling the system to adaptively determine whether there is a reliable cross-modal correspondence in each local region while outputting the registration displacement field. The system rejects real weak correspondence regions caused by disasters, thus avoiding the contamination of the registration results by erroneous textures and false targets in weak correspondence regions.

[0021] 2. In this invention, by simultaneously outputting the registration displacement field, the registration probability map, the rejection probability map, and the matching covariance map through a cross-modal registration network that can refuse, cross-modal image registration is not only achieved, but also provides multi-dimensional reliability basis for subsequent disaster mapping, change interpretation, and manual verification. Among them, the registration probability map indicates which regions participated in the transformation estimation, the rejection probability map indicates which regions were excluded due to significant changes, and the matching covariance map indicates the uncertainty of the displacement estimation.

[0022] 3. In this invention, the covariance head outputs a matching covariance map and calculates the uncertainty accordingly. During training, the negative log-likelihood loss of the covariance is used as a constraint, so that the predicted covariance can effectively reflect the distribution scale of the actual displacement error, preventing the network from making overconfident predictions of incorrect matches, and providing a reliable confidence basis for reliable region screening and global registration rejection judgment. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall process of a rejectable SAR image-optical image registration method for weakly correlated areas before and after disasters, as proposed in this invention.

[0024] Figure 2 This is a schematic diagram of the architecture of a rejectable SAR image-optical image registration system for weakly correlated areas before and after a disaster, as proposed in this invention. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] It should be noted that, in this invention, "rejection" refers to excluding local areas with low registration probability, strong post-disaster changes, or high matching uncertainty from the estimation of registration transformation parameters or reducing their matching weight in the transformation estimation during cross-modal registration of post-disaster SAR sensing images and pre-disaster optical reference images. "Rejection" in this invention does not refer to environments where Global Navigation Satellite System signals are restricted or unavailable.

[0027] Example 1

[0028] In a first embodiment of the present invention, the present invention provides a method for rejectable SAR image-optical image registration for weakly correlated areas before and after a disaster, such as... Figure 1 As shown, it includes the following steps:

[0029] S1. Acquire post-disaster SAR sensing images and pre-disaster optical reference images; generate a geometric-physical prior based on at least one of the following: embedded georeference in the image product, the result of image-level coarse matching of the two images, or externally input geometric information; and determine the coarse corresponding position and search radius of the optical reference image based on the geometric-physical prior to form a geometric-physical constraint search window.

[0030] Specifically, post-disaster SAR sensing images and pre-disaster optical reference images are acquired. If the acquired images are in multi-channel format, they are converted to single-channel grayscale format to reduce computational complexity. The post-disaster SAR sensing images and pre-disaster optical reference images are resampled to a uniform spatial resolution, and a uniform matching grid is established. Geometric physical priors are generated based on at least one of the following: embedded georeferenced information in the image product, the result of image-level coarse matching between the two images, or externally input geometric information. The geometric physical priors include the coarse corresponding position of the optical reference image, the search radius, and the geometric offset term.

[0031] The system determines whether the product georeferences of the post-disaster SAR sensing image and the pre-disaster optical reference image can be read. The product georeference includes the map projection parameters, geographic affine transformation parameters, pixel resolution, and image coverage embedded in the image file. When the product georeference can be read, the pixel coordinates of the post-disaster SAR sensing image are mapped to the coordinates of the pre-disaster optical reference image based on the product georeference, obtaining the roughly corresponding position on the optical reference image. Specifically, the pixel coordinate x in the post-disaster SAR sensing image is subjected to its geographic affine transformation. Mapped to geographic coordinates, and then the geographic coordinates are subjected to the inverse geographic affine transformation of the pre-disaster optical reference image. Mapping to pre-disaster optical reference image coordinates, the calculation formula is as follows:

[0032] ;

[0033] Where x represents the pixel coordinates in the post-disaster SAR sensing image. This represents the geographic affine transformation of post-disaster SAR sensing images. Represents the geographic affine transformation of the pre-disaster optical reference image. This indicates the coarse corresponding position on the optical reference map. The search radius is determined based on the geocoding error and pixel resolution of the product's georeferenced data, calculated using the following formula:

[0034] ;

[0035] in, This indicates the geocoding error of post-disaster SAR sensing images. This indicates the geocoding error of the pre-disaster optical reference image. This indicates the pixel resolution of the post-disaster SAR sensing image. This indicates the pixel resolution of the pre-disaster optical reference image. This is a preset proportionality coefficient, with a value ranging from 1 to 3. In a preferred embodiment, The value is 2.

[0036] When product georeferenced images cannot be read, image-level coarse matching is performed on post-disaster SAR sensing images and pre-disaster optical reference images to obtain a coarse transform. Image-level coarse matching is performed on a low-resolution pyramid, specifically searching for the transform that maximizes the structural feature similarity between the two images from a set of candidate transforms (translation, similarity, affine, or homography). The number of layers in the low-resolution pyramid is determined based on the image size; for a 1024×1024 pixel image, three layers can be used, with each layer's size halved. The formula for calculating the coarse transform is:

[0037] ;

[0038] in, Represents the set of candidate transformations for translation, similarity, affine, or homography. and These represent low-resolution structural feature extraction operators for post-disaster SAR sensing images and pre-disaster optical reference images, respectively. Represents the normalized similarity function. This represents the image domain. The pixel coordinates of the post-disaster SAR sensing image are mapped to the coordinates of the pre-disaster optical reference image using a coarse transformation. The calculation formula is as follows:

[0039] ;

[0040] in, This is the coarse transform obtained from image-level coarse matching. The search radius is determined based on the matching confidence of the image-level coarse matching. The matching confidence is characterized by the difference between the maximum and second-largest peak values ​​of the coarse matching cost. The formula for calculating the search radius is:

[0041] ;

[0042] in, The preset wide search radius benchmark value, This is the difference between the maximum and second-largest peak values ​​of the coarse matching cost. In a preferred embodiment, The value is 10% of the width of the post-disaster SAR sensing image.

[0043] For airborne or UAV-borne SAR platforms, the externally input geometric information includes Position and Attitude Measurement System (POS) parameters, SAR range parameters, azimuth parameters, side-looking direction, and a digital elevation model (DEM / DSM). In this case, the coarse corresponding position is given by the SAR geometric projection operator and the optical projection operator, calculated using the following formula:

[0044] ;

[0045] in, This represents an operator that performs geometric projection based on SAR pixel coordinates, platform position and attitude parameters, SAR imaging geometric parameters, and elevation information. This represents a projection operator that maps geographic coordinates to coordinates in a pre-disaster optical reference image. Indicates the platform's position and attitude parameters. Represents the geometric parameters of SAR imaging. This indicates the elevation value of a digital elevation model or digital surface model.

[0046] The geometrical physical constraint search window is determined based on the coarse corresponding position of the optical reference image and the search radius. The geometrical physical constraint search window is defined as a circular region centered on the coarse corresponding position of the optical reference image and with a radius equal to the search radius. The calculation formula is as follows:

[0047] ;

[0048] in, The x-axis represents the geometric and physical constraint search window at position x, and the y-axis represents the candidate pixel coordinates in the pre-disaster optical reference image. This represents the Euclidean norm. The geometric bias term is used for the modulation matching cost, and its calculation formula is:

[0049] ;

[0050] in, To prevent division by zero errors, a preset small constant is used. In a preferred embodiment, Values The geometric bias term is 0 at the center of the search window and gradually decreases as the candidate position deviates from the coarse corresponding position. It is used to constrain the matching search range and suppress false matches that are far from the coarse corresponding position.

[0051] S2. Extract local structural features from SAR images and optical images, and generate clues to changes in disaster areas and SAR observability risks.

[0052] Specifically, the intensity map of the post-disaster SAR sensing image is obtained as the basic input to the SAR image. Logarithmic magnitude map, local mean map, local variance map, range coordinate map, and azimuth coordinate map are generated from the post-disaster SAR sensing image as derived inputs to the SAR image. The logarithmic magnitude map is obtained by taking the logarithm of the intensity map and is used to compress the dynamic range of speckle noise in the SAR image. The local mean map and local variance map are calculated by sliding a fixed-size window on the image; the window size can be set to 7×7 pixels, and are used to describe the local texture statistics of the SAR image. The range coordinate map and azimuth coordinate map are directly generated from the image grid coordinates, representing the spatial location information of each pixel in the range and azimuth directions, respectively, and are used to provide spatial location priors for the network. The basic SAR image input and the derived SAR image input are jointly input into the SAR image feature coding branch.

[0053] The SAR image feature coding branch employs a convolutional neural network structure, consisting of four sequentially connected convolutional residual blocks. The first convolutional residual block has a 3×3 kernel, 32 output channels, and a stride of 2; the second has a 3×3 kernel, 64 output channels, and a stride of 2; the third has a 3×3 kernel, 96 output channels, and a stride of 1; and the fourth has a 3×3 kernel, 128 output channels, and a stride of 1. Each convolutional residual block includes a convolutional layer, a batch normalization layer, and a Gaussian error linear unit activation function. The SAR image feature coding branch downsamples the input and extracts features, ultimately outputting local structural features of the SAR image with a spatial resolution of 1 / 4 of the input image and 128 channels.

[0054] The pre-disaster optical reference image is converted into a grayscale image. The gradient magnitude map, edge response map, and corner line structure response map of the grayscale image are calculated as the optical basis input. The gradient magnitude map is obtained by calculating the gradient of the image in the horizontal and vertical directions using the Sobel operator and then taking the modulus, used to describe the intensity of grayscale changes in the image. The edge response map is calculated using traditional image edge detection algorithms (such as Canny edge detection) or deep learning-based convolutional edge detection operators, used to extract edge structure information in the image. The corner line structure response map is calculated using corner detection operators or line structure extraction operators, used to extract corner and line structure information in the image. The optical basis input is then fed into the optical image feature coding branch. The optical image feature coding branch adopts a convolutional neural network structure with the same scale as the SAR image feature coding branch, but the network parameters are not shared. The optical image feature coding branch consists of four convolutional residual blocks connected sequentially. The kernel size, number of output channels, stride, and activation function settings of each convolutional residual block are consistent with those of the SAR image feature coding branch. The optical image feature coding branch downsamples and extracts features from the input, and finally outputs the local structural features of the optical image with a spatial resolution of 1 / 4 of the input image and 128 channels.

[0055] Obtain the coarse corresponding position of the optical reference image in the geometrical prior. Perform discrete wavelet transform on the post-disaster SAR sensing image and the pre-disaster optical reference image respectively to obtain their respective multi-scale wavelet coefficients. The discrete wavelet transform uses the Daubechies wavelet basis, with a decomposition level of 3. Each level decomposition generates three directional sub-bands: a horizontal sub-band, a vertical sub-band, and a diagonal sub-band. Based on the coarse corresponding position of the optical reference image, compare the wavelet coefficients of the post-disaster SAR sensing image with the wavelet coefficients of the pre-disaster optical reference image at the corresponding position. For each scale level and each directional sub-band, calculate the normalized absolute difference of the wavelet coefficients of the two images, and then sum the differences of each scale level and each directional sub-band with weights to obtain the wavelet multi-scale structural difference response. The formula for calculating the wavelet multi-scale structural difference response is:

[0056] ;

[0057] Where L represents the number of wavelet decomposition layers, with a value of 3; H, V, and D represent the horizontal sub-band, vertical sub-band, and diagonal sub-band, respectively. This represents the preset weighting coefficient of the b-th directional sub-band in the l-th layer. The weighting coefficient of each directional sub-band in each layer can be taken from 0.5 to 2.0. In a preferred embodiment, each weighting coefficient is taken as 1.0. This represents a local normalization function used to adjust wavelet coefficients to the same dimension. This represents the wavelet coefficient at position x in the b-th direction sub-band of the l-th layer of the post-disaster SAR sensing image; This indicates the coarse corresponding position of the optical reference image in the b-th direction sub-band of the l-th layer of the pre-disaster optical reference image. Wavelet coefficients at; This means truncating the calculation result to the range of 0 to 1.

[0058] Obtain disaster type codes, prior probability of disaster area change, and stable structure probabilities. Disaster type codes are vectors or scalars obtained by numerically encoding disaster categories; for example, flood disaster is coded as 1, earthquake disaster as 2, and fire disaster as 3. Prior probability of disaster area change can be provided by externally input disaster extent products or by satellite-borne SAR pre- and post-disaster change detection results. Stable structure probability represents the probability that a location belongs to a stable structure (such as roads, bridges, building outlines, etc.), which can be provided by stable structure prior layers in historical multi-period remote sensing images or geographic information data. In the absence of external stable structure priors, stable structure probability can be estimated online from matching evidence of the current image pair or set to a neutral default value of 0.5. The wavelet multi-scale structural difference response, disaster type codes, prior probability of disaster area change, and stable structure probabilities are fused to generate clues about disaster area changes. The fusion can use a weighted summation method; the calculation formula for disaster area change clues is:

[0059] ;

[0060] in, This represents the Sigmoid function. ; This is the bias term, and its value is 0. , , , The weighting coefficients are preset and can range from 0.1 to 2.0. In a preferred embodiment, The value is 1.0. The value is 0.8. The value is 1.2. The value is 0.5; This represents the prior probability of a change in the spaceborne SAR at location x; This represents the probability of a stable structure at position x; This indicates the disaster type code.

[0061] Based on the intensity gradient, local variance, and texture statistics of post-disaster SAR sensing images, weak texture risk and repetitive texture risk are calculated. The intensity gradient is calculated using the Sobel operator, the local variance is calculated within a 7×7 pixel sliding window, and the texture statistics include statistics such as gray-level co-occurrence matrix contrast, energy, and entropy within the local window. Weak texture risk characterizes the degree to which the texture information in the region is scarce and stable matching features are difficult to extract. Regions with intensity gradients less than a preset gradient threshold or local variance less than a preset variance threshold have a higher weak texture risk value. Repetitive texture risk characterizes the degree to which the region has periodic repeating structures and is prone to ambiguous matching. Regions with periodic changes in local variance or multi-peaked distributions in texture statistics have a higher repetitive texture risk value.

[0062] Based on the intensity gradient, local variance, and texture statistics of post-disaster SAR sensing images, and combined with at least one of the following: local incident angle, slope of digital elevation model or digital surface model, side-looking direction, and overlay / shadow geometry, the risks of overlay, shadowing, low coherence, and strong specular reflection saturation are calculated. The local incident angle is calculated from SAR imaging geometry parameters and the digital elevation model / digital surface model; the slope of the digital elevation model or digital surface model is calculated from external DEM / DSM data; the side-looking direction is provided by SAR imaging metadata; and the overlay / shadow geometry is calculated jointly by SAR imaging geometry and DEM / DSM. Overlay risk characterizes the degree of overlap of echo signals from different elevation features in the SAR image due to terrain undulation, with higher values ​​on slopes with greater slope and facing the radar beam. Shadowing risk characterizes the degree to which terrain occlusion causes some areas in the SAR image to lack echo signals, with higher values ​​on slopes with greater slope and facing away from the radar beam. Low coherence risk characterizes the degree of low coherence and poor interferometric phase quality of SAR signals, with higher values ​​in low coherence areas such as water bodies and vegetation. Strong specular reflection saturation risk characterizes the degree to which specular reflection from flat terrain surfaces (such as water surfaces and smooth roads) causes echo signal saturation. The unified calculation method for each risk component in SAR observability risk is as follows:

[0063] ;

[0064] Where j represents the risk type index, corresponding to overlay risk, shadow risk, low coherence risk, strong specular reflection saturation risk, weak texture risk, and repeated texture risk, respectively; to The preset weighting coefficients and bias terms are for the j-th type of risk. Each coefficient can take values ​​in the range of -2.0 to 2.0. In a preferred embodiment, the weighting coefficients and bias terms for each type of risk are obtained by calibration of the validation set. This represents the local angle of incidence at position x; This represents the slope of the digital elevation model or digital surface model at location x. This represents the magnitude of the intensity gradient in the post-disaster SAR sensing image at location x. This represents the local window variance of the post-disaster SAR sensing image at location x, and the window size can be set to 7×7 pixels. This represents the initial geometric risk value generated by the digital elevation model or digital surface model and the SAR side-looking direction.

[0065] Six risk categories—overlap risk, shadow risk, low coherence risk, strong specular reflection saturation risk, weak texture risk, and repetitive texture risk—are combined to generate SAR observability risk. The combination method can be either a weighted sum of the six risk components mapped to the 0-1 interval using a Sigmoid function, or a method that takes the maximum value. In a preferred embodiment, a weighted sum of the six risk components is used, with each risk component having a weight coefficient of 1 / 6. The comprehensive calculation formula for SAR observability risk is as follows:

[0066] ;

[0067] in, The preset weighting coefficients represent the risk type j, and each weighting coefficient can take values ​​from 0.1 to 0.5. In a preferred embodiment, each... The value is 1 / 6; This means truncating the summation result to the range of 0 to 1.

[0068] S3. Within the geometric and physical constraint search window, construct a matching cost volume based on the similarity between the local structural features of the SAR image and the local structural features of the optical image, and determine the matching cost volume statistics accordingly.

[0069] Specifically, local structural features of the SAR image are input into the local feature projection layer, and after 1×1 convolution and L2 norm normalization, the SAR image projection features projected onto the matching feature space are obtained. Similarly, local structural features of the optical image are input into the local feature projection layer, and after 1×1 convolution and L2 norm normalization, the optical image projection features projected onto the matching feature space are obtained. The number of 1×1 convolution output channels in the local feature projection layer is set to 128.

[0070] Within the geometrically constrained search window, the similarity between the SAR image projection features at each SAR image grid location and the corresponding optical image projection features at each optical candidate location is calculated to obtain the basic matching cost. The similarity is measured using cosine similarity, and the calculation formula is as follows:

[0071] ;

[0072] Where x represents the SAR image grid position, and y represents the position in the optical reference image within the geometrically constrained search window. Candidate optical positions within, This represents the SAR image projection features at location x. This represents the optical image projection feature at position y. and These represent optional linear transformations of SAR image projection features and optical image projection features, respectively. In a preferred embodiment... and All are identity mappings, where T denotes vector transpose. Describing the L2 norm, In a preferred embodiment, a preset small constant is used to prevent division by zero errors. Values Basic matching cost The value ranges from -1 to 1. The larger the value, the more similar the local structure at grid position x in the SAR image is to the local structure at optical candidate position y.

[0073] A geometric bias term is generated based on geometrical priors. This term constrains the matching search range, taking a value of 0 at the center of the search window and gradually decreasing as the candidate position deviates from its corresponding position. The formula for calculating the geometric bias term is:

[0074] ;

[0075] in, Indicates the corresponding position in the optical reference diagram. Indicates the search radius. In a preferred embodiment, a preset small constant is used to prevent division by zero errors. Values .

[0076] A disaster change bias term is generated based on disaster area change clues. This term is used to reduce the matching cost in areas with a high probability of disaster change and increase the matching cost in areas with a high probability of stable structures. The formula for calculating the disaster change bias term is:

[0077] ;

[0078] in, This represents the probability of post-disaster change at location x. This represents the wavelet multi-scale structural difference response at location x. This represents the prior probability of spaceborne SAR change at location x. This represents the probability of a stable structure at position x. , , , The weighting coefficients are preset and can range from 0.1 to 2.0. In a preferred embodiment... The value is 0.5. The value is 0.3. The value is 0.2. The value is 0.8.

[0079] An observation bias term is generated based on SAR observability risk. This term is used to reduce matching costs in areas with high SAR observability risk. The formula for calculating the observation bias term is:

[0080] ;

[0081] in, This indicates the risk of overlay at position x. This indicates the risk of shadowing at position x. This indicates a low coherence risk at position x. This indicates the risk of strong specular reflection saturation at location x. This indicates the risk of weak textures at position x. This indicates the risk of duplicate textures at position x; , , , , , The weighting coefficients are preset and can range from 0.1 to 2.0. In a preferred embodiment... , , , , , All values ​​are 0.5.

[0082] The basic matching cost is superimposed with the geometric bias, hazard change bias, and observation bias to obtain the matching cost volume. The formula for calculating the matching cost volume is:

[0083] ;

[0084] in, Matching cost body The higher the value of , the higher the reliability of the match between the SAR image grid position x and the optical candidate position y.

[0085] The matching cost volume is obtained, which contains the matching cost for each SAR image grid location corresponding to each optical candidate location within the geometrically constrained search window. For each SAR image grid location, the maximum value among all matching costs within the corresponding search window is determined as the maximum value of the matching cost volume. The formula for calculating the maximum value of the matching cost volume is:

[0086] ;

[0087] in, This represents the maximum value of the matching cost body at position x. The maximum value of the matching cost body reflects the confidence of the best match at that position; the larger the value, the higher the probability that a reliable correspondence exists at that position.

[0088] For each SAR image grid location, the difference between the maximum and second-largest values ​​is determined from all matching costs within the corresponding search window, and this difference is used as the maximum and second-largest value interval. The formula for calculating the maximum and second-largest value interval is:

[0089] ;

[0090] in, This represents the second largest cost among all matches within the search window at position x. The interval between the maximum and second largest values ​​reflects the sharpness of the match peak; a larger interval indicates a more precise match at that position with lower ambiguity.

[0091] For each SAR image grid location, a probability distribution is calculated based on all matching costs within the corresponding search window. The probability distribution is obtained through Softmax normalization, and the calculation formula is as follows:

[0092] ;

[0093] in, This represents the matching probability of the optical candidate position y at position x in the SAR image grid; T is a preset temperature coefficient used to control the smoothness of the probability distribution, with a value ranging from 0.1 to 1.0, and in a preferred embodiment, T is set to 0.5. The information entropy is calculated based on the probability distribution, serving as the cost volume entropy. The formula for calculating the cost volume entropy is:

[0094] ;

[0095] in, This represents the cost volume entropy at position x. In a preferred embodiment, a preset small constant is used to prevent zero values ​​from occurring in logarithmic operations. Values The cost volume entropy reflects the degree of concentration of the matching probability distribution. The smaller the entropy value, the more concentrated the matching distribution and the higher the certainty of the matching; the larger the entropy value, the more dispersed the matching distribution and the higher the ambiguity of the matching.

[0096] S4. Input the local structural features of the SAR image, the local structural features of the optical image, the clues of changes in the disaster area, the SAR observability risk, the geometric and physical priors, the matching cost volume, the statistics of the matching cost volume, and the geometric residuals determined by the difference between the corresponding control points predicted by the network and the corresponding positions of the optical reference map into the end-to-end rejectable cross-modal registration network. The network jointly outputs the registration displacement field, the registration probability map, the rejection probability map, and the matching covariance map.

[0097] Specifically, local structural features of the SAR image are input into the SAR image feature coding branch for encoding, and local structural features of the optical image are input into the optical image feature coding branch for encoding. Both the SAR image feature coding branch and the optical image feature coding branch adopt a convolutional neural network structure, each containing four convolutional residual blocks. The kernel size of each convolutional residual block is 3×3, with output channels of 32, 64, 96, and 128, and strides of 2, 2, 1, and 1, respectively. Each convolutional residual block includes a convolutional layer, a batch normalization layer, and a Gaussian error linear unit activation function. The network parameters of the SAR image feature coding branch and the optical image feature coding branch are not shared. The encoded features, along with clues to changes in disaster areas, SAR observability risks, and geometric and physical priors, are input into the local feature projection layer. The local feature projection layer projects the above multi-source features to the same matching feature space through a 1×1 convolution operation, outputting 128 channels and performing L2 norm normalization along the channel dimension. The local feature projection layer outputs SAR image projection features and optical image projection features projected to the same matching feature space.

[0098] The SAR image projection features, optical image projection features, matching cost volume, matching cost volume statistics, and geometric residuals are input into the cross-modal fusion layer. The cross-modal fusion layer employs two local window attention blocks or correlation volume aggregation blocks, with a window size of 8×8 pixels, 4 attention heads, a feedforward hidden layer dimension of 512, and a Gaussian error linear unit as the activation function. The cross-modal fusion layer performs cross-modal interaction and fusion of the SAR image projection features and optical image projection features through a self-attention mechanism or correlation volume aggregation mechanism to obtain fused features. The fused features have 256 channels and a spatial resolution of a uniform matching grid size.

[0099] The fused features are input into the displacement head. The displacement head consists of a 3×3 convolutional layer, a Gaussian error linear unit activation function, and a 1×1 convolutional layer connected sequentially. The 3×3 convolutional layer has 128 output channels, and the 1×1 convolutional layer has 2 output channels. The displacement head outputs a registration displacement field. Each grid position in the registration displacement field contains a two-dimensional offset in the horizontal and vertical directions, representing the local two-dimensional offset between the coordinates of the SAR sensed image and the corresponding coordinates of the optical reference image.

[0100] The fusion features, matching cost volume statistics, geometric residuals, hazard area change clues, and SAR observability risks are input into the registration head. The registration head consists of a multilayer perceptron or a 1×1 convolutional layer, specifically including a 1×1 convolutional layer to reduce the number of input channels from 384 to 128, a Gaussian error linear unit activation function, and another 1×1 convolutional layer to reduce the number of channels to 1, finally activating the output with a Sigmoid function. The registration head outputs a registration probability map, where each grid position takes a value between 0 and 1, representing the probability that a reliable cross-modal correspondence exists in that local area.

[0101] The fused features are input into the covariance head. The covariance head consists of a 3×3 convolutional layer, a Gaussian error linear unit activation function, and a 1×1 convolutional layer connected sequentially. The 3×3 convolutional layer has 128 output channels, and the 1×1 convolutional layer has 3 output channels. The covariance head outputs three parameters, corresponding to the horizontal standard deviation, the vertical standard deviation, and the correlation coefficient. The horizontal and vertical standard deviations are mapped to the positive interval using the Softplus function, and the correlation coefficient is mapped to the -1 to 1 interval using the Tanh function, thus forming a two-dimensional matching covariance matrix. The covariance head outputs a matching covariance map. Each grid position in the matching covariance map contains a 2×2 covariance matrix, representing the uncertainty of the registration displacement field estimation at that position.

[0102] A rejection probability map is generated based on the registration probability map, the matching covariance map, the matching cost volume statistics, and the geometric residuals. Specifically, the uncertainty scalar for each grid location is determined based on the matching covariance map; the uncertainty scalar can be the logarithm of the determinant of the covariance matrix or the trace. The complement probability, uncertainty scalar, cost volume entropy, and geometric residuals of the registration probability map are weighted and fused, and then activated by the Sigmoid function to output the rejection probability map. Each grid location in the rejection probability map takes a value between 0 and 1, representing the probability that the local region should be rejected. The rejection probability map and the registration probability map are numerically complementary or, after joint calibration, consistent.

[0103] The geometric residual is determined by the difference between the control points predicted by the network and their corresponding positions in the optical reference image. The control points predicted by the network are obtained by adding the registration displacement field to the SAR image grid positions. The formula for calculating the geometric residual is:

[0104] ;

[0105] in, Indicates the grid position of the SAR image. This indicates the control points corresponding to the network prediction. This represents the registration displacement field at position x. Indicates the corresponding position in the optical reference diagram. Denotes the Euclidean norm. A preset normalization constant is used to adjust the geometric residuals to a reasonable numerical range. In a preferred embodiment... The value is the search radius. .

[0106] The spatial resolution of the uniform matching grid is 1 / 4 of the input image size. When the input image size is 512×512 pixels, the uniform matching grid size is 128×128. The geometrically constrained search radius *r* is set to 8 grid cells, then the number of optical candidate locations K corresponding to each SAR image grid location is... That is, 289. The above network parameters are a preferred implementation. Depending on the input image resolution, sensor parameters, and computing resources, the step size, search radius, and number of channels in each layer of the uniform matching grid can be adjusted accordingly.

[0107] S5. Based on the registration probability map, rejection probability map, and matching covariance map, select reliable corresponding regions from the registration displacement field output by the network, and determine the rejection regions.

[0108] Specifically, the registerability probability map, rejection probability map, matching covariance map, and registration displacement field output by the network are obtained. These information are jointly output by the end-to-end rejectable cross-modal registration network in step S4 and serve as input data for this step. The uncertainty at each grid location is determined based on the matching covariance map. The matching covariance map records the two-dimensional covariance matrix of the estimated registration displacement field at each grid location. For each grid location, the uncertainty scalar is calculated based on its corresponding two-dimensional covariance matrix. The formula for calculating the uncertainty is:

[0109] ;

[0110] in, This represents the matching covariance matrix at position x. Represents the determinant of a matrix. The determinant value represents the natural logarithm. A larger determinant value indicates a larger area of ​​the ellipse spanned by the covariance matrix, meaning a higher uncertainty in the displacement estimate. In a preferred embodiment, the uncertainty can also be calculated using the trace of the covariance matrix, as shown in the formula: .

[0111] The grid positions where the registration probability map satisfies a first preset condition, the rejection probability map satisfies a second preset condition, and the uncertainty satisfies a third preset condition are designated as reliable corresponding regions. Specifically, the first preset condition for the registration probability map is that the registration probability is greater than a first preset threshold; the second preset condition for the rejection probability map is that the rejection probability is less than a second preset threshold; and the third preset condition for uncertainty is that the uncertainty is less than a third preset threshold. In a preferred embodiment, the first preset threshold is 0.7, the second preset threshold is 0.3, and the third preset threshold is determined through statistical analysis of the validation set. Specifically, samples with small registration errors on the validation set are selected, their uncertainty distribution is statistically analyzed, and the 80th percentile is used as the value of the third preset threshold. The mathematical expression for the reliable corresponding region is:

[0112] ;

[0113] in, Indicates a reliable corresponding region. Indicates a uniform matching grid. This represents the registration probability at position x. Let x represent the rejection probability at position x. This represents the uncertainty at position x. This represents the first preset threshold. This indicates the second preset threshold. This indicates the third preset threshold.

[0114] Grid locations that do not meet the reliable correspondence region condition are designated as rejection regions. A rejection region is the set of all grid locations in the uniformly matched grid that do not belong to a reliable correspondence region. The mathematical expression for a rejection region is:

[0115] ;

[0116] in, This indicates a rejection region. Grid locations within the rejection region are not included in subsequent registration transformation parameter estimation. For feature points within the rejection region, the system does not use them as reliable matching points for transformation estimation, thereby avoiding contamination of the registration results by areas with weak hazard correspondence, SAR overlay shadows, and low observability areas.

[0117] The values ​​of each preset threshold are determined through validation set calibration. Specifically, different threshold combinations are iterated and searched on the validation set, with the joint optimization objective of minimizing the registration error in reliable areas, the false acceptance rate in non-registerable areas, and the false rejection rate in registerable areas. The threshold combination that achieves a balance among these three factors is selected as the final calibration result. For image data with different sensor resolutions, different disaster types, and different geographical regions, each preset threshold can be adaptively adjusted within the above calibration framework.

[0118] S6. Estimate the registration transformation parameters based only on the corresponding control points within the reliable corresponding area, and output the registration results.

[0119] Specifically, the reliable corresponding region, registration displacement field, registration probability map, rejection probability map, matching covariance map, and rejection region are obtained. This information is generated and provided by steps S4 and S5, respectively. The uncertainty at each grid location is determined based on the matching covariance map. The uncertainty is calculated in the same way as in step S5, using either the logarithm of the determinant of the covariance matrix or the trace of the covariance matrix.

[0120] Based on the registration probability, uncertainty, and geometric residual of each grid location within the reliable corresponding region, a weight is determined for each grid location. The weight measures the contribution of that corresponding control point to the estimation of the registration transformation parameters. The formula for calculating the weight is:

[0121] ;

[0122] in, This represents the weight at position x. This represents the registration probability at position x. This represents the uncertainty at position x. This represents the geometric residual at position x. The preset scaling parameter is used to control the decay rate of the geometric residual on the weights. In a preferred embodiment... The value is 1.0. The higher the registration probability, the lower the uncertainty, and the smaller the geometric residual, the greater the weight of the corresponding control point.

[0123] Based on the registration displacement field and weights of each grid location within the reliable correspondence region, the registration transformation parameters are estimated. These parameters represent at least one of the following: translation, rotation, scaling, shearing, affine matrix, homography matrix, thin-plate spline control points, or local grid control parameters. The registration transformation parameters are estimated using a weighted least squares optimization method. The optimization objective is to minimize the weighted projection error of each corresponding control point within the reliable correspondence region. The formula for calculating the objective function is:

[0124] ;

[0125] in, This represents the estimated registration transformation. Indicates a reliable corresponding region. This indicates the coordinates after transformation T is applied to position x. This represents the registration displacement field at position x. This represents the Euclidean norm. The objective function is solved using a least squares solver or an iterative optimization algorithm.

[0126] The registration quality index is determined based on the proportion of reliable corresponding regions to all grid regions, the proportion of rejected regions to all grid regions, the average registration probability within reliable corresponding regions, the average uncertainty within reliable corresponding regions, and the average geometric residual within reliable corresponding regions. The registration quality index characterizes the reliability of the current registration result. The formula for calculating the registration quality index is:

[0127] ;

[0128] in, This represents the registration quality index, with a value ranging from 0 to 1; Represents the Sigmoid function; This indicates the proportion of the reliable corresponding area to the total grid area; This indicates the proportion of the rejected area to the total grid area; This represents the average registration probability within a reliable corresponding region; This represents the average uncertainty within the reliable corresponding region; This represents the average geometric residual within the reliable corresponding region; For bias terms, , , , , These are preset weighting coefficients. Each weighting coefficient can take values ​​ranging from 0.1 to 2.0. In a preferred embodiment... The value is 0. The value is 0.5. The value is 0.3. The value is 0.5. The value is 0.2. The value is 0.2.

[0129] When the registration quality index is lower than the first preset threshold, the proportion of reliable corresponding regions to all grid regions is lower than the second preset threshold, or the proportion of rejected regions to all grid regions is higher than the third preset threshold, a global registration rejection flag is output. The global registration rejection flag indicates that the current image pair is not suitable for automatic registration or requires manual review. In a preferred embodiment, the first preset threshold is 0.5, the second preset threshold is 0.1, and the third preset threshold is 0.8. The values ​​of each of the above preset thresholds are determined through a validation set calibration method.

[0130] The system outputs registration transformation parameters, registration displacement field, reliable correspondence regions, rejection regions, and registration quality index. Registration transformation parameters describe the global or local geometric mapping relationship between the post-disaster SAR sensed image and the pre-disaster optical reference image. The registration displacement field describes the local two-dimensional offset at each grid location. The reliable correspondence region indicates which local areas have reliable cross-modal correspondences. The rejection region indicates which local areas should be excluded from the registration transformation estimate. The registration quality index indicates the overall reliability of the current registration result. When a global registration rejection flag is output, the registration transformation parameters should not be applied to subsequent disaster mapping or multi-source data fusion applications.

[0131] Example 2

[0132] Existing SAR image registration methods, whether region-based, feature-based, or deep learning-based, all implicitly assume that the two images to be registered have a reliable grayscale or geometric correspondence in most sub-regions. However, in disaster emergency scenarios, flooding, building collapses, road washouts, and the construction of temporary rescue facilities can lead to a lack of true geometric correspondence between pre-disaster optical reference images and post-disaster SAR sensing images in some disaster areas. Existing methods generally lack the ability to accurately represent whether reliable cross-modal correspondences exist in local areas—traditional region- and feature-based methods estimate the transformation through global optimization or local feature matching. Incorrect matching in weak correspondence regions directly contaminates the estimation results of the global transformation. Although end-to-end methods based on deep learning can output transformation parameters or displacement fields, they typically compress all information from the entire image indiscriminately into a global output. When the area of ​​weak correspondence regions in disaster scenarios is large or the proportion of SAR overlay shadow regions is high, the network output is easily influenced by erroneous textures and false targets, resulting in registration results that deviate from the true situation. Furthermore, the aforementioned methods usually only output transformation parameters and cannot explain which local regions participated in the transformation estimation, which regions should be rejected, or the uncertainty of the registration results. This severely restricts the reliability and interpretability of SAR image registration in disaster scenarios. To solve the above problems, this invention provides a rejectable SAR image-optical image registration system for weak correspondence regions before and after disasters, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this system is as follows:

[0133] The image acquisition and geometrical constraint construction module acquires post-disaster SAR sensing images and pre-disaster optical reference images. It generates a geometrical prior based on at least one of the following: embedded georeferenced information within the image product, the result of image-level coarse matching between the two images, or externally input geometric information. Based on this prior, it determines the coarse corresponding position and search radius of the optical reference image to form a geometrical constraint search window. The feature extraction and risk generation module extracts local structural features from both the SAR and optical images and generates clues to changes in the disaster area and SAR observability risks. The matching cost body construction and statistic determination module constructs a matching cost body within the geometrical constraint search window based on the similarity between the local structural features of the SAR and optical images, and determines the matching cost body statistics accordingly. Cross-modal rejection is possible. The modal registration network module inputs the local structural features of the SAR image, the local structural features of the optical image, clues to changes in disaster areas, SAR observability risks, geometric and physical priors, matching cost volume, matching cost volume statistics, and geometric residuals determined by the difference between the network-predicted corresponding control points and the coarse corresponding positions of the optical reference map into an end-to-end rejectable cross-modal registration network. This network jointly outputs a registration displacement field, a registration probability map, a rejection probability map, and a matching covariance map. The reliable corresponding sub-region screening and rejection sub-region determination module screens reliable corresponding sub-regions from the registration displacement field output by the network based on the registration probability map, rejection probability map, and matching covariance map, and determines the rejection sub-regions. The registration transformation parameter estimation and result output module estimates the registration transformation parameters based only on the corresponding control points within the reliable corresponding regions and outputs the registration results.

[0134] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for registering rejectable SAR images with optical images in areas with weak correspondence before and after a disaster, characterized in that, include: S1. Acquire post-disaster SAR sensing images and pre-disaster optical reference images; Based on at least one of the following: embedded georeferences in the image product, the results obtained from image-level coarse matching of two images, or externally input geometric information, a geometric-physical prior is generated, and the coarse corresponding position and search radius of the optical reference map are determined based on the geometric-physical prior to form a geometric-physical constraint search window. S2. Extract local structural features from SAR images and optical images, and generate clues to changes in disaster areas and SAR observability risks; S3. Within the geometric and physical constraint search window, construct a matching cost body based on the similarity between the local structural features of the SAR image and the local structural features of the optical image, and determine the matching cost body statistics accordingly. S4. Input the local structural features of the SAR image, the local structural features of the optical image, the clues to changes in the disaster area, the SAR observability risk, the geometrical prior, the matching cost volume, the statistics of the matching cost volume, and the geometric residual determined by the difference between the control point predicted by the network and the corresponding position in the optical reference map into an end-to-end rejectable cross-modal registration network. The network jointly outputs the registration displacement field, the registration probability map, the rejection probability map, and the matching covariance map. S5. Based on the registration probability map, the rejection probability map, and the matching covariance map, select reliable corresponding regions from the registration displacement field output by the network, and determine the rejection regions. S6. Estimate the registration transformation parameters based only on the corresponding control points within the reliable corresponding area, and output the registration result.

2. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In S1, generating geometric-physical priors includes: Determine whether the product geographic reference of each of the post-disaster SAR sensing image and the pre-disaster optical reference image can be read; When the product georeference can be read, the pixel coordinates of the post-disaster SAR sensing image are mapped to the coordinates of the pre-disaster optical reference image based on the product georeference to obtain the corresponding position of the optical reference image, and the search radius is determined based on the geocoding error of the product georeference. When the product georeference cannot be read, image-level coarse matching is performed on the post-disaster SAR sensing image and the pre-disaster optical reference image to obtain a coarse transformation. The pixel coordinates of the post-disaster SAR sensing image are mapped to the coordinates of the pre-disaster optical reference image according to the coarse transformation to obtain the coarse corresponding position of the optical reference image. The search radius is determined according to the matching confidence of the image-level coarse matching. The geometric physical constraint search window is determined based on the corresponding position of the optical reference image and the search radius.

3. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In step S2, local structural features of the SAR image and local structural features of the optical image are extracted, including: The intensity map of the post-disaster SAR sensing image is obtained as the basic input of the SAR image, and the logarithmic amplitude map, local mean map, local variance map, range coordinate map and azimuth coordinate map are generated from the post-disaster SAR sensing image as the SAR image derived input. The SAR image base input and the SAR image derived input are input into the SAR image feature encoding branch to obtain the local structural features of the SAR image; The pre-disaster optical reference image is converted into a grayscale image, and the gradient magnitude map, edge response map, and corner line structure response map of the grayscale image are calculated as optical basis inputs. The optical fundamentals are input into the optical image feature encoding branch to obtain the local structural features of the optical image.

4. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In S2, clues to changes in disaster areas and SAR observability risks are generated, including: Obtain the coarse corresponding position of the optical reference map in the geometrical-physical prior; The multi-scale wavelet coefficients of the post-disaster SAR sensing image and the pre-disaster optical reference image are obtained respectively. Based on the corresponding position of the optical reference image, the difference between the wavelet coefficients of the post-disaster SAR sensing image and the pre-disaster optical reference image in the same direction at the same scale is calculated to obtain the wavelet multi-scale structural difference response. Based on the wavelet multi-scale structural difference response, disaster type coding, disaster area change probability prior and stable structure probability, the disaster area change clues are fused to generate the disaster area change clues; Based on the intensity gradient, local variance, and texture statistics of the post-disaster SAR sensing image, and combined with at least one of the following: local incident angle, slope of digital elevation model or digital surface model, side view direction, and overlapping shadow geometry, the overlapping risk, shadow risk, low coherence risk, strong specular reflection saturation risk, weak texture risk, and repeated texture risk are calculated, and the six types of risks are combined to generate the SAR observability risk.

5. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In step S3, a matching cost body is constructed based on the similarity between the local structural features of the SAR image and the local structural features of the optical image, including: The local structural features of the SAR image are projected onto the matching feature space to obtain the SAR image projection features; the local structural features of the optical image are projected onto the matching feature space to obtain the optical image projection features. Within the geometric and physical constraint search window, the similarity between the SAR image projection feature at each SAR image grid location and the optical image projection feature at each corresponding optical candidate location is calculated to obtain the basic matching cost. A geometric bias term is generated based on the aforementioned geometric and physical priors; a disaster change bias term is generated based on the aforementioned disaster area change clues; and an observation bias term is generated based on the aforementioned SAR observability risk. The basic matching cost is superimposed with the geometric bias, the disaster change bias, and the observation bias to obtain the matching cost body.

6. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In S3, the matching cost body statistics are determined, including: The matching cost body is obtained, which contains the matching cost of each SAR image grid position corresponding to each optical candidate position within the geometric and physical constraint search window; For each SAR image grid location, the maximum value is determined from all matching costs within the corresponding search window, and this value is taken as the maximum value of the matching cost volume. For each SAR image grid location, the difference between the maximum and the second largest value is determined from all matching costs within the corresponding search window, and is used as the interval between the maximum and the second largest value; For each SAR image grid location, a probability distribution is calculated based on all matching costs within the corresponding search window, and the information entropy is calculated based on the probability distribution as the cost volume entropy.

7. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In S4, the registration displacement field, the registration probability map, the rejection probability map, and the matching covariance map are output, including: The local structural features of the SAR image and the local structural features of the optical image are respectively input into the SAR image feature coding branch and the optical image feature coding branch for encoding. The encoded features, together with the disaster area change clues, the SAR observability risk and the geometric and physical priors, are input into the local feature projection layer to obtain the SAR image projection features and the optical image projection features projected onto the same matching feature space. The SAR image projection features, the optical image projection features, the matching cost volume, the matching cost volume statistics, and the geometric residual are input into the cross-modal fusion layer to obtain the fused features; The fusion features are input into the displacement head to obtain the registration displacement field; The fusion features, the matching cost volume statistics, the geometric residuals, the disaster area change clues, and the SAR observability risk are input into the registration head to obtain the registration probability map; The fusion features are input into the covariance head to obtain the matching covariance map; The rejection probability map is generated based on the registration probability map, the matching covariance map, the matching cost volume statistics, and the geometric residuals.

8. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In step S5, reliable corresponding regions are selected from the registration displacement field output by the network, and rejection regions are determined, including: Obtain the registration probability map, the rejection probability map, the matching covariance map, and the registration displacement field output by the network; The uncertainty of each grid location is determined based on the matching covariance plot; The grid positions where the registration probability map satisfies the first preset condition, the rejection probability map satisfies the second preset condition, and the uncertainty satisfies the third preset condition are taken as the reliable corresponding regions. Grid locations that do not meet the conditions for the reliable corresponding region are designated as the rejection regions.

9. The method for registering rejectable SAR images with optical images for weakly correlated areas before and after a disaster, as described in claim 1, is characterized in that... In step S6, the registration transformation parameters are estimated based solely on the corresponding control points within the reliable corresponding region, and the registration result is output, including: Obtain the reliable corresponding region, the registration displacement field, the registration probability map, the rejection probability map, the matching covariance map, and the rejection region; The uncertainty of each grid location is determined based on the matching covariance plot; The weight of each grid location is determined based on the registration probability, uncertainty, and geometric residual of each grid location within the reliable corresponding region. Based on the registration displacement field and the weights of each grid position within the reliable corresponding region, the registration transformation parameters are estimated; The registration quality index is determined based on the proportion of the reliable corresponding region to the total grid region, the proportion of the rejected region to the total grid region, the average registration probability within the reliable corresponding region, the average uncertainty within the reliable corresponding region, and the average geometric residual within the reliable corresponding region. When the registration quality index is lower than a preset threshold, the proportion of the reliable corresponding region to the total grid region is lower than a preset threshold, or the proportion of the rejected region to the total grid region is higher than a preset threshold, a global registration rejection flag is output. Output the registration transformation parameters, the registration displacement field, the reliable correspondence region, the rejection region, and the registration quality index.

10. A rejectable SAR image-optical image registration system for areas with weak correspondence before and after a disaster, characterized in that, A rejectable SAR-optical image registration method for weakly correlated areas before and after disasters, as described in any one of claims 1-9, the system comprising: The image acquisition and geometric-physical constraint construction module is used to acquire post-disaster SAR sensing images and pre-disaster optical reference images; based on at least one of the following: embedded georeference in the image product, the result of image-level coarse matching of two images, or externally input geometric information, a geometric-physical prior is generated, and the coarse corresponding position and search radius of the optical reference image are determined based on the geometric-physical prior to form a geometric-physical constraint search window; The feature extraction and risk generation module is used to extract local structural features of SAR images and optical images, and generate clues to changes in disaster areas and SAR observability risks. The matching cost body construction and statistics determination module is used to construct a matching cost body based on the similarity between the local structural features of the SAR image and the local structural features of the optical image within the geometric and physical constraint search window, and determine the matching cost body statistics accordingly. A rejection-based cross-modal registration network module is used to input the local structural features of the SAR image, the local structural features of the optical image, the cues of changes in the disaster area, the SAR observability risk, the geometric and physical priors, the matching cost volume, the statistics of the matching cost volume, and the geometric residual determined by the difference between the network-predicted corresponding control point and the coarse corresponding position of the optical reference image into an end-to-end rejection-based cross-modal registration network. The network jointly outputs a registration displacement field, a registration probability map, a rejection probability map, and a matching covariance map. The reliable corresponding sub-region filtering and rejection sub-region determination module is used to filter reliable corresponding regions from the registration displacement field output by the network based on the registration probability map, the rejection probability map and the matching covariance map, and determine the rejection regions. The registration transformation parameter estimation and result output module is used to estimate the registration transformation parameters based only on the corresponding control points within the reliable corresponding area, and output the registration results.

Images