A method for registering remote sensing images

Abstract

The present application relates to the technical field of image annotation registration, and particularly relates to a remote sensing image registration method, and the present application proposes the following scheme: through change detection on post-disaster remote sensing images and historical annotation data, an area is divided into a first area with changed ground object morphology and a second area with stable ground object morphology; in the first area, local structure features are extracted based on ground object deformation information and adaptive anchor point matching is performed with direction and range constraints, and in the second area, global geometric features are extracted based on historical annotation data and global anchor point matching is performed; annotation accuracy registration data is generated by comprehensively matching local and global matching results, and accurate mapping of historical annotation to a new base map is realized. The present application takes into account the local authenticity of the changed area and the overall consistency of the stable area, effectively avoiding error propagation problems caused by artificial block-by-block correction and single global registration.

Description

Technical Field

[0001] This invention relates to the field of image annotation and registration technology, and in particular to a registration method for remote sensing images. Background Technology

[0002] In post-disaster emergency assessment, reconstruction planning, and spatial information updates, remote sensing imagery is frequently used as fundamental data for feature identification and spatial analysis. To improve operational efficiency, numerous feature annotations already created on historical remote sensing images are often directly transferred or reused in newly acquired post-disaster remote sensing images. However, with continuous updates to the base map, inconsistencies often arise between historical annotations and new imagery, leading to annotation offsets, morphological distortions, or local misalignments, directly impacting the reliability of subsequent analysis and decision-making results.

[0003] In existing technologies, common annotation and registration methods often rely on overall image registration or a unified geometric transformation model to correct historical annotations as a whole. When processing post-disaster scenes, these methods are prone to situations where local structures are stretched, compressed, or misaligned, causing annotations that were accurate in stable areas to lose spatial realism in changed areas. Meanwhile, some technologies attempt to filter or redraw annotations through change detection, but the results often exhibit problems such as changed areas being fragmented into scattered pieces, discontinuous region boundaries, or stable regions being misidentified as changed areas, making it difficult to produce coherent and usable registration results.

[0004] Furthermore, existing methods typically employ uniform feature types and search strategies during the feature matching stage, failing to distinguish the differences between stable and deformed regions in post-disaster images. Matching results are easily affected by local noise, stitching seams, or image radiometric differences, resulting in uneven distribution of matching points and a high rate of mismatches. This leads to spatial jumps or local distortions in the annotation and registration results. These issues make it difficult for existing technologies to simultaneously ensure overall consistency and local realism during the accurate registration of post-disaster remote sensing image annotations, limiting the stability and reliability of the annotation correction results.

[0005] To address the above issues, this application proposes a registration method for remote sensing images. Summary of the Invention

[0006] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a remote sensing image registration method. This method involves detecting changes in post-disaster remote sensing images and historical annotation data, dividing the region into a first region where landform morphology has changed and a second region where landform morphology remains stable. Within the first region, local structural features are extracted based on landform deformation information, and adaptive anchor point matching is performed with directional and range constraints. Within the second region, global geometric features are extracted based on historical annotation data, and global anchor point matching is performed. The combined results of local and global matching generate accurate registration data, achieving precise mapping of historical annotations to the new base map. This application balances the local realism of changed regions with the overall consistency of stable regions, effectively avoiding the error propagation problems caused by manual block-by-block correction and single global registration.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for registering remote sensing images, the method comprising: Acquire post-disaster remote sensing images and historical labeled data; Change detection is performed based on the post-disaster remote sensing images and historical annotation data to identify a first region and a second region, wherein the first region represents the region where the land cover morphology has changed, and the second region represents the region where the land cover morphology is stable. In the first region, local features are extracted and adaptive feature matching is performed based on ground deformation information. In the second region, global geometric features are extracted and matched based on historical annotation data. The ground deformation information is obtained by analyzing the difference information between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image corresponding to the historical annotation data. The difference information includes at least structural difference information. Based on the matching results of the first and second regions, generate registration data for annotation accuracy and output the registered annotation results.

[0008] The post-disaster remote sensing images are acquired through a remote sensing imaging platform, which includes at least one of a satellite remote sensing platform, an airborne remote sensing platform, and an unmanned aerial vehicle (UAV) remote sensing platform. The historical annotation data is used to characterize the annotation information of ground features on the base map of the previous version of the post-disaster remote sensing image, wherein the annotation information of ground features includes at least one of vector features, and the vector features include point, line and polygon features.

[0009] Change detection is performed based on the post-disaster remote sensing images and historical labeled data, including: Spatial alignment processing is performed between the previous version of the post-disaster remote sensing image corresponding to the historical annotation data and the post-disaster remote sensing image. On the base map of the previous version of the post-disaster remote sensing image corresponding to the spatially aligned historical annotation data, multiple detection units are determined based on the annotation information of ground features. Each detection unit is a local area that covers the annotation information of ground features, and the coverage area of ​​each detection unit is the same. The coverage area is set by the resolution of the remote sensing imaging platform. Select a detection unit as the current detection unit, calculate the first similarity between the current detection unit corresponding to the post-disaster remote sensing image and the current detection unit corresponding to the previous version of the post-disaster remote sensing image. If the first similarity is greater than or equal to a preset first threshold, mark the current detection unit as a change seed unit, and perform an iterative operation on the change seed unit until there are no candidate expansion units that meet the expansion conditions. The set of candidate expansion units obtained by the change seed unit through the iterative operation is determined as the first region, wherein the candidate expansion unit represents the neighborhood unit of the change seed unit. If the first similarity is less than the first threshold, the current detection unit is determined as the second region, wherein the first threshold is used to characterize the similarity determination threshold for determining the current detection unit as a changed seed unit; Once all detection units have finished processing, the first and second regions are output.

[0010] The determination of multiple detection units based on feature annotation information includes: Based on the resolution of the remote sensing imaging platform, the previous version of the post-disaster remote sensing image corresponding to the spatially aligned historical annotation data is divided into grids to obtain multiple candidate units, where each candidate unit has the same coverage area. Based on the aforementioned feature annotation information, candidate units that have a spatial overlap relationship with at least one feature annotation are identified; Candidate units that have spatial overlap with the labeled ground features are identified as the detection units.

[0011] The calculation method for the first similarity includes: For the current detection unit, extract the corresponding image blocks from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image; A similarity metric for characterizing the degree of image variation is calculated based on the image patch. The similarity metric includes differences in brightness distribution, texture features, structural features, and frequency domain features. The similarity metric is normalized, and a first similarity is generated using a weighted algorithm.

[0012] Performing an iterative operation on the changed seed unit includes: The change trend is calculated based on the change seed unit, wherein the change trend is obtained by performing directional analysis on the difference information between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image within the corresponding image block of the change seed unit; Based on the change trend, candidate expansion units are selected from the neighboring units of the change seed unit; Calculate the second similarity between the candidate expansion unit corresponding to the post-disaster remote sensing image and the candidate expansion unit corresponding to the previous version of the post-disaster remote sensing image, and compare the second similarity with a preset second threshold; When the second similarity is greater than or equal to the second threshold, the candidate expansion unit is included in the first region and the candidate expansion unit is marked as a change seed unit.

[0013] The detection unit further includes an access status, which is used to characterize whether the corresponding detection unit has participated in similarity calculation, wherein: During the iterative operation on the changed seed unit, the detection unit for which the second similarity has been calculated is set to the visited state; When a detection unit is in an visited state, the first similarity calculation is not repeated for the corresponding detection unit.

[0014] The step of extracting local features and performing adaptive feature matching includes: For each first region, at least one local feature extraction window is determined within the first region, and local structural features are extracted within the local feature extraction window. The local structural features include at least one of corner features, edge features, and line segment features. Based on the deformation information of the ground features, the matching search direction constraint and matching search range constraint of the local structural features are determined; Based on the matching search direction constraint and the matching search range constraint, the local structural features extracted from the post-disaster remote sensing image are matched with the local structural features of the corresponding region in the previous version of the post-disaster remote sensing image to obtain the local anchor point matching result of the first region.

[0015] The extraction and matching of global geometric features includes: For each second region, at least one global feature extraction range is determined within the second region based on the historical annotation data; Within the scope of global feature extraction, global geometric features are extracted from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image, respectively. The global geometric features include at least anchor point features. Global geometric feature matching is performed between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image to obtain the global anchor point matching result for the second region.

[0016] Based on the matching results of the first and second regions, annotation accuracy registration data is generated, including: Based on the local anchor point matching results of the first region, a local registration relationship is determined to characterize the historical annotation data in the first region to the post-disaster remote sensing image. Based on the global anchor point matching results of the second region, a global registration relationship is determined to characterize the historical annotation data in the second region to the post-disaster remote sensing image; Based on the local registration relationship and the global registration relationship, annotation accuracy registration data is generated.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention performs regional differentiation processing on post-disaster remote sensing images and historical annotation data. In areas where landforms have changed, it introduces local adaptive anchor point matching based on deformation information, while in areas with stable landforms, it employs global geometric anchor point matching to achieve regional collaborative annotation registration. This method effectively constrains the propagation of local deformation in changed areas while maintaining overall consistency in stable regions, avoiding error diffusion and annotation breakage. It improves the accuracy and continuity of mapping historical annotations to the new base map, reduces manual correction costs, and enhances the reliability and usability of post-disaster remote sensing annotation data in practical applications. Attached Figure Description

[0018] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 An exemplary application scenario diagram provided for an embodiment of this application; Figure 2 A schematic flowchart illustrating a remote sensing image registration method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the detection unit provided in the embodiments of this application; Figure 4 This is a schematic diagram illustrating the selection of candidate expansion units provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] The technical practice addressed in this application stems from a typical yet long-standing problem in post-disaster emergency mapping and reconstruction assessment that has been masked by engineering compromises: Remote sensing base maps of disaster-stricken areas are often rapidly acquired and updated in multiple batches, across multiple platforms, and at multiple time phases. Inevitably, during the processes of orthorectification, mosaic reconstruction, coordinate reference transformation, tile slicing, and rendering, complex geometric errors such as overall translation, local distortion, and seam misalignment occur. For command and dispatch, damage assessment, road access determination, temporary construction site selection, and post-disaster reconstruction planning, updating the base map is not the end goal. More crucially, it is whether the spatial element annotations on the base map can be continuously reused and maintain geometric consistency. In the context of this application, annotations primarily refer to spatial semantic carriers used to express disaster-related features, such as building outlines, road centerlines and boundaries, water bodies and inundation areas, landslide accumulation boundaries, bridges and key nodes, temporary resettlement sites, and work areas—vector elements (points / lines / areas) that can also be extended to raster masks or detection frames. These types of annotations typically have strong business attributes: they may come from existing geographic information data before the disaster or from rapid interpretation results on the previous version of the post-disaster base map; their value lies in their ability to be directly used for statistics, analysis, navigation and engineering verification, without having to be recreated from scratch after each base map update.

[0022] Those skilled in the art will understand that a common dilemma arises in rapid post-disaster map updates: on the one hand, annotations must be updated in tandem with the base map; otherwise, misalignment of the same element on the base map will directly lead to misjudgments, such as roads being judged as fractures, buildings being judged as collapsed with expanded areas, and flooding line shifts causing errors in the affected area. On the other hand, the frequency and coverage of post-disaster base map updates often far exceed the capacity for manual editing. Traditional manual correction, element-by-element and map-by-map sheet, is not only costly and time-consuming but also introduces new inconsistencies due to differences in the subjective interpretation of local textures and shadows by annotators. To address this, common compromises in existing projects include: performing a global registration on the entire map, for example, estimating the overall translation / rotation based on a small number of control points, or using a unified transformation model to migrate the previous version of annotations across the entire image. However, in the post-disaster scenario where change and stability coexist, the above approach often presents two unavoidable problems: First, the actual changes in the morphology of features in the post-disaster area (collapse, erosion, landslides, accumulation, etc.) make the selection of control points for global registration unreliable, and the registration model is easily dragged by the changed areas, causing the labels of originally stable areas to be pulled off. Second, the errors in the base map are not consistent everywhere. The splicing seams, orthorectification residuals from different flight sorties or different satellite strips will create non-uniform misalignments in local areas, which a single global model cannot take into account. This results in phenomena where the overall map matches but some parts do not, or where some parts match but the seams on the other side collapse.

[0023] The core logic of this application is naturally derived from the aforementioned engineering dilemmas: registration of annotation accuracy is viewed as an alignment process using historical annotations as semantic anchors and the new base map as a geometric carrier. Its basic principle is not simply to pursue overall similarity between images, but rather to first spatially identify stable regions that can serve as geometric constraints and changed regions that require local deformation interpretation, and then establish matching anchor point mechanisms for each, thereby generating a registration relationship that can be directly applied to historical annotations. Specifically, after a post-disaster base map update, the misalignment of historical annotations is essentially due to the drift in the coordinate-pixel mapping relationship of the annotations as the base map product version evolves; the purpose of registration is to restore this mapping relationship, ensuring that historical annotations maintain a consistent spatial orientation at the same geographical location on the new base map. To achieve this objective, this application introduces difference information as a source of ground feature deformation information, with particular emphasis on the stability of structural difference information in post-disaster scenarios: compared to easily fluctuating factors such as brightness changes, illumination and shadows, and wet reflections, structural elements (edges, corners, line intersections, road frameworks, building corners, etc.) can still maintain alignable geometric observability in many post-disaster stable areas, while in areas where actual morphological changes have occurred, they will exhibit significant structural damage or reorganization. Based on this observation, this application does not rely on exhaustive pixel-by-pixel search of the entire map in the change detection stage. Instead, it combines the spatial distribution of historical annotations, discretizes the image into candidate units at a uniform resolution scale, and then selects detection units that spatially overlap with the ground feature annotations. This allows change determination to focus on the areas that the annotations are truly concerned with, reducing invalid calculations at the source.

[0024] Building upon this, this application further considers that post-disaster changes often exhibit a spatial pattern of linear diffusion, spreading along road / river / slope boundaries, or directional extension around the center of damage. If the approach of indiscriminately expanding each seed unit to its entire neighborhood is still adopted, a large number of candidate branches will be generated in complex scenarios, leading to redundant calculations, thus slowing down the registration process and affecting emergency response timeliness. Therefore, this application introduces a directional constraint on the direction of change during the iterative expansion of change units: the direction of change is obtained by directional analysis of the differences between the post-disaster image and the previous version image within corresponding image blocks, thereby limiting the selection direction and range of candidate expansion units, making the expansion more aligned with the geometric trend of change propagation, and avoiding unnecessary omnidirectional traversal. Simultaneously, to further suppress duplicate access, this application sets an access status for detection units, preventing units already involved in similarity calculations from being triggered repeatedly, thus significantly reducing the overhead of redundant calculations during the iteration process while ensuring the integrity of the first region extraction.

[0025] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0026] like Figure 1As shown, the scene to be updated is the target area of ​​interest for post-disaster emergency response or reconstruction assessment, and its base map is updated on a rolling basis with each batch of data collection. A remote sensing imaging platform is used to remotely acquire data from the scene to be updated, generating post-disaster remote sensing images; the remote sensing imaging platform can be any one or a combination of satellite remote sensing platforms, airborne remote sensing platforms, or UAV remote sensing platforms. A processor connected to the remote sensing imaging platform is used to preprocess, spatially align, and detect changes in the post-disaster remote sensing images, thereby identifying a first region where land cover morphology has changed and a second region where land cover morphology remains stable; a processor connected to the historical annotation data storage side is used to retrieve the previous version of the post-disaster remote sensing image base map and its land cover feature annotation information, and performs adaptive anchor point matching of local features and global geometric feature matching based on the first and second regions respectively, thereby generating annotation accuracy registration data and outputting the registered annotation results. It should be noted that... Figure 1 The processor shown can be a logical processing module on the same computing node, or it can be multiple computing nodes deployed on the edge side and the center side; this application is not limited to the specific deployment form of the processor.

[0027] Next, with reference to the accompanying drawings, the registration method for remote sensing images provided in the embodiments of this application will be further elaborated. Figure 2 The methods shown include: S1: Acquire post-disaster remote sensing images and historical annotation data; In this embodiment, the post-disaster remote sensing image is used to characterize the spatial state of the scene to be updated at the current moment. It can be acquired by the remote sensing imaging platform during post-disaster emergency response, reconstruction assessment, or rolling inspections. Historical annotation data is used to characterize the annotation results of the ground feature elements of the scene to be updated on the previous version of the post-disaster remote sensing image base map. The annotation results of the ground feature elements can be in the form of points, lines, or polygons, and are used to express the spatial semantic information of disaster-related ground features or infrastructure. In this embodiment, the historical annotation data and the post-disaster remote sensing image are not acquired at the same time, and they may have overall or partial misalignment in space due to the update of the base map product. By acquiring the post-disaster remote sensing image and the historical annotation data simultaneously, subsequent processing can register the geometric consistency between the annotation and the base map while maintaining the semantic continuity of the annotation, thereby avoiding the need for repeated production or large-scale manual revision of the annotation data after the base map is updated.

[0028] S2: Based on the post-disaster remote sensing images and historical annotation data, change detection is performed to identify the first region and the second region; The first region represents areas where landforms change, and the second region represents areas where landforms remain stable. In this embodiment, change detection is used to distinguish between areas where the morphology of ground features has undergone real changes and areas that remain stable in a post-disaster scene. Specifically, based on the difference information between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image corresponding to historical annotation data, the spatial range related to the annotation of ground feature elements is analyzed to identify a first region and a second region. The first region is used to characterize areas where the morphology of ground features has changed due to disaster effects, secondary impacts, or construction disturbances, while the second region is used to characterize areas where the ground feature structure remains stable and can serve as a geometric constraint reference. By introducing a region differentiation mechanism in the change detection stage, the subsequent annotation accuracy registration no longer relies on a single global geometric assumption, thereby avoiding interference from changed regions with the overall matching results.

[0029] S3: In the first region, local features are extracted and adaptive feature matching is performed based on ground deformation information; in the second region, global geometric features are extracted and matched based on historical annotation data. The deformation information of the land features is obtained by analyzing the difference information between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image corresponding to the historical annotation data. The difference information includes at least structural difference information. In this embodiment, different feature matching strategies are employed for the first and second regions to adapt to the objective characteristics of coexistence of change and stability in post-disaster scenarios. For the first region, due to changes in the terrain structure, directly applying global geometric constraints can easily introduce incorrect matching. Therefore, based on the terrain deformation information obtained from difference information analysis, structural features are extracted within a local range, and adaptive feature matching is performed by setting matching search direction and range constraints to obtain anchor point matching results that reflect local geometric relationships. For the second region, since the terrain morphology remains stable, it is suitable to extract global geometric features based on historical annotation data and perform matching to obtain anchor point matching results that characterize the overall geometric drift relationship. By binding the feature matching strategy with regional attributes, this embodiment can avoid incorrect alignment in changing regions and fully utilize existing structural information in stable regions, improving the reliability and overall consistency of annotation registration.

[0030] S4: Based on the matching results of the first and second regions, generate registration data for annotation accuracy and output the registered annotation results; In this embodiment, based on the local anchor point matching results of the first region and the global anchor point matching results of the second region, corresponding local and global registration relationships are constructed respectively, and these registration relationships are used to generate annotation accuracy registration data. The annotation accuracy registration data describes the position correction method of historical annotation data in post-disaster remote sensing images, ensuring that historical annotations remain consistent with the actual geographical locations on the new base map. The registered annotation results maintain the semantic continuity of historical annotations while avoiding geometric misalignment problems caused by base map updates, making it suitable for scenarios such as post-disaster emergency analysis, reconstruction assessment, and continuous monitoring. Those skilled in the art will understand that the output format of the registration results can be selected as updated vector annotations or other spatial data formats according to business needs, as long as it meets the basic spatial accuracy requirements of subsequent applications; this embodiment does not further limit this.

[0031] Before detailing the specific technical aspects of the steps, this application's embodiments need to reiterate: In the context of post-disaster base map updates, the evolution of image products does not necessarily equate to actual changes in ground features. There are two types of phenomena in the image that appear to have changed in appearance: one type comes from changes in the actual morphology of ground features, such as building collapses, road washouts, and the expansion of debris, which presents continuous destruction or reconstruction at the spatial structure level; the other type comes from geometric and radial differences in the image imaging link and productization process, such as local drift at stitching seams, local stretching caused by orthorectification residuals, and edge illusions under different lighting / shadow conditions, which also appear as changes in texture or boundaries, but do not mean that the annotation semantics should change accordingly. A common challenge for those skilled in the art in practical application is that if change detection is used as the sole entry point, regardless of whether differential, correlation, or structural similarity metrics are employed, both types of phenomena may be mixed into the same set of changes, leading to a dilemma in subsequent registration: if too conservative, real change regions will be treated as stable regions and participate in global constraints, resulting in overall registration shift; if too aggressive, a large number of non-real changes will be included in the change regions, triggering unnecessary local processing and amplifying computational overhead, and even introducing instability in local matching.

[0032] Based on the aforementioned engineering facts, this application embodiment does not interpret the first region as a region with low similarity or large appearance differences. Instead, it defines the first region as: a set of regions within the labeled relevant spatial range where structural difference information exhibits an interpretable deformation trend, and this trend possesses spatial directionality. In other words, the identification of the first region depends not only on the existence of differences but also on how the differences are organized spatially. This point constitutes the core processing logic source of this application embodiment: by introducing directional analysis of difference information, the change region is transformed from a static binary classification problem into a region generation problem with directional constraints, enabling the formation process of the change region to gradually expand along the dominant trend of structural disturbance, rather than indiscriminately enumerating the entire neighborhood. Those skilled in the art will understand that structural difference information can be obtained from changes in edges, line segments, corners, and their spatial topological relationships, as long as it can at least reflect the misalignment or damage characteristics of the post-disaster image and the previous version image at the structural level. This application embodiment does not limit this.

[0033] Furthermore, the directional expansion of the first region is not aimed at increasing the size of the changed area, but rather at quickly identifying the spatial bands that truly require local adaptive matching within limited computational resources. Post-disaster changes often exhibit spatial patterns of propagation or extension: for example, expansion along road corridors, spread along river boundaries, spillover in a strip around the damage center, or the formation of long, narrow disturbances along the slope slip direction. If the conventional omnidirectional expansion mechanism of region growth is used, candidate units will generate numerous branches at complex boundaries, causing similarity calculations to be repeatedly triggered and rapidly accumulated, resulting in a sharp increase in computational load with each expansion round. Conversely, by using the direction of change as the expansion guide and only verifying neighboring units consistent with that direction, the candidate set of expansion can be converged into a controlled set of directions. This makes the iterative process closer to tracing along the disturbance band rather than generalizing the search in all directions. This significantly reduces the proportion of redundant calculations while maintaining the connectivity of the changed region, preventing the changed region extraction stage from becoming the performance bottleneck of the entire process.

[0034] Furthermore, after the first region is formed with directional constraints, the boundary of the local adaptive matching becomes clearer: local matching is not blindly initiated at all signs of change, but only carried out in areas where structural differences have been confirmed as deformation-related, using ground deformation information as the constraint basis for the search space, thus ensuring interpretable geometric consistency in the matching process of local anchor points. Those skilled in the art will understand that the so-called matching search direction constraints and matching search range constraints are not required to be fixed in a specific form; they can be expressed as directional window restrictions on candidate matching positions, adaptive shrinkage of the search radius, or offset constraints on candidate alignment paths, as long as the deformation trend is incorporated into the search space limitation of local matching. This application embodiment does not impose further limitations. Through the above mechanism, the matching behavior in the first region and the global geometric matching in the second region are logically complementary: the former emphasizes local consistency along the deformation trend, while the latter emphasizes overall geometric consistency within stable regions, thereby enabling the annotation accuracy registration to possess both robustness and engineering feasibility under complex post-disaster conditions.

[0035] Next, we will further elaborate on the technical aspects of the regional division method in this application.

[0036] In one example, post-disaster remote sensing images can be acquired by a remote sensing imaging platform, which includes at least one of a satellite remote sensing platform, an airborne remote sensing platform, and a UAV remote sensing platform. Historical annotation data is used to characterize the annotation information of ground features on the base map of the previous version of the post-disaster remote sensing image. The annotation information of ground features includes at least vector features in the form of points, lines, or polygons. Based on this type of input condition, the regional division is not based on uniform scanning of the entire image area, but rather prioritizes establishing detection objects around the spatial coverage of the ground feature annotations, so that the change determination falls within the "area of ​​actual interest in the annotation," thereby reducing the repeated calculation of background areas unrelated to the annotations. Those skilled in the art will understand that the types of features involved in the ground feature annotation can be configured according to the business task, such as roads, buildings, water body boundaries, landslide ranges, etc., as long as they can provide the minimum spatial constraints of the annotation's area of ​​interest. This application does not impose further limitations.

[0037] It should be noted that the first region in this application does not refer to areas with significant image differences in general, but rather to a set of regions within the labeled spatial range where the structural morphology of ground features exhibits an interpretable trend of change between the post-disaster image and the previous version of the image, and this trend has spatial directionality. Correspondingly, the second region refers to a set of regions within the labeled spatial range where the structural topological relationships of ground features remain stable and can provide reliable constraints for geometric alignment. Structural morphology focuses more on the stability of edges, line segments, corners, and their spatial topological relationships, rather than the brightness texture appearance that is significantly affected by differences in illumination, shadows, or radiation. Those skilled in the art will understand that structural difference information can be characterized by the distribution consistency, boundary continuity, or linear skeleton changes of structural elements in image blocks, as long as it can reflect the stability and destructive characteristics of ground feature structures in engineering; this application does not limit the specific implementation method.

[0038] In another example, change detection based on the post-disaster remote sensing images and historical annotation data includes: S2.1: Spatial alignment processing is performed between the previous version of the post-disaster remote sensing image corresponding to the historical annotation data and the post-disaster remote sensing image; Specifically, the productization process of post-disaster remote sensing images often includes orthorectification, mosaic reconstruction, projection conversion, and tile publishing. Even if the previous version of the base map and the current base map cover the same geographical area, there may be inconsistencies in coordinate references, deviations in geographic reference metadata, or pixel origin drift caused by tile conversion. If the differences or similarities are directly calculated between two temporal image blocks, the amount of difference will be masked by alignment errors, causing change detection to misclassify a large number of stable features as changed features.

[0039] In this embodiment, spatial alignment processing includes two stages: coordinate reference unification and pixel-level alignment. Coordinate reference unification can be completed based on the georeferenced information (projection, reference ellipsoid, resolution, affine transformation parameters) of the two images. If metadata is missing or unreliable, vector elements in historical annotation data can be used as auxiliary constraints. For example, several anchor points are selected at structurally stable locations such as road intersections, bridge endpoints, and building corners, and their positions in the previous and current images are established to correspond to each other. Based on this, the translation and rotation amounts for coarse alignment are estimated, so that the two images fall into the same pixel coordinate frame macroscopically. Pixel-level alignment is used to eliminate remaining sub-pixel deviations. Multiple local windows can be selected in the overlapping area of ​​the previous and current images, and local alignment correction is performed using the edge or line segment features corresponding to the structural difference information, so that the structural contours within the local windows overlap as much as possible at the pixel level.

[0040] S2.2: On the base map of the previous version of the post-disaster remote sensing image corresponding to the spatially aligned historical annotation data, multiple detection units are determined based on the ground feature annotation information. The detection unit is a local area covering the ground feature annotation information, wherein the coverage area of ​​each detection unit is the same, and the coverage area is set by the resolution of the remote sensing imaging platform. Specifically, if change detection is performed pixel-by-pixel or uniformly across the entire map, it will include a large amount of background areas unrelated to the annotations in the calculation, resulting in unnecessary overhead. Furthermore, it is easily interfered with by non-target factors such as river surface reflection, shadow movement, and seasonal vegetation. The focus of post-disaster annotation accuracy registration is whether the land features covered by historical annotations have undergone geometric drift or local deformation on the new base map. Therefore, the construction of detection units needs to revolve around the annotation spatial range while maintaining a uniform discrete scale for subsequent iterative expansion and access status control.

[0041] In this embodiment, a unified grid scale is first set based on the resolution of the remote sensing imaging platform. The spatially aligned post-disaster remote sensing image of the previous version is then divided into grids at the current processing scale, resulting in multiple candidate units with consistent coverage areas. The side length of the candidate unit can be set according to the resolution and the size of the target element. For example, when the resolution is 0.5 meters, the side length of the candidate unit can be set to 16 meters corresponding to 32 pixels to cover a road node or a partial structure of a building; when the resolution is 1 meter, the side length can be set to 32 meters corresponding to 32 pixels to ensure that the unit contains sufficient structural information for stable similarity calculation. After the unified division is completed, detection units are screened using historical annotation data: each vector element (point / line / area) is projected onto the grid, candidate units that spatially overlap with the vector elements are identified, and these candidate units are included in the detection unit set. For linear features such as road centerlines, a buffer zone can be applied to cover the road width and edge structure. For point features such as bridge nodes and key facility points, a fixed window centered on the point can be used to cover its neighborhood structure. For area features such as building outlines and water body boundaries, candidate units with an overlap ratio exceeding a preset ratio with area features can be selected.

[0042] In one example, the determination of multiple detection units based on feature annotation information includes: Based on the resolution of the remote sensing imaging platform, the previous version of the post-disaster remote sensing image corresponding to the spatially aligned historical annotation data is divided into grids to obtain multiple candidate units, where each candidate unit has the same coverage area. Based on the aforementioned feature annotation information, candidate units that have a spatial overlap relationship with at least one feature annotation are identified; Candidate units that have spatial overlap with the labeled ground features are identified as the detection units.

[0043] refer to Figure 3 , Figure 3 This is a schematic diagram of the detection unit provided in an embodiment of this application.

[0044] like Figure 3 As shown, the overall rectangular area represents the effective coverage of the previous version of the post-disaster remote sensing image after spatial alignment at the current processing scale. This area is first uniformly gridded according to the spatial resolution of the remote sensing imaging platform, forming multiple candidate units with consistent coverage areas, represented in the figure as regular squares. Each square corresponds to a candidate unit, and its geometric dimensions remain consistent throughout the entire image, thus providing a unified spatial discretization basis for subsequent similarity calculation, neighborhood relationship determination, and iterative expansion. Figure 3 The light gray squares indicate detection units that spatially overlap with feature labels. Spatial overlap does not require the detection unit to be completely covered by the feature; it can be selected as long as its spatial extent overlaps with at least one feature label (point, line, or polygon) in a planar position. For example, when the feature is a line feature, candidate units that intersect with the line feature or fall within its buffer zone can be considered to have spatial overlap; when it is a polygon feature, candidate units that intersect with the boundary or interior region of the polygon feature satisfy the condition.

[0045] S2.3: Select a detection unit as the current detection unit, calculate the first similarity between the current detection unit corresponding to the post-disaster remote sensing image and the current detection unit corresponding to the previous version of the post-disaster remote sensing image. If the first similarity is greater than or equal to a preset first threshold, mark the current detection unit as a change seed unit, and perform an iterative operation on the change seed unit until there are no candidate expansion units that meet the expansion conditions. Determine the set of candidate expansion units obtained by the change seed unit through the iterative operation as the first region, wherein the candidate expansion units represent the neighborhood units of the change seed unit. Specifically, if the change determination at the detection unit level is directly classified into all units at once using a global threshold, two types of results are likely to occur: First, the change intensity within the actual change region is uneven, and the edges may be fragmented due to being below the threshold; second, misalignment of seams or local alignment residuals will create high differences in a small number of units, and if these are directly diffused into change regions, non-real changes will be introduced into the local matching stage. To balance the integrity of the change region and the control of false detections, this embodiment adopts a seed expansion method: first, change seeds are obtained through a more rigorous determination, and then iterative expansion using directional constraints is used to complete the connected parts of the change region, making the first region more consistent with the spatial organization of post-disaster changes.

[0046] In this embodiment, for the selected current detection unit, corresponding image blocks are cropped from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image, and a first similarity is calculated. The first similarity can be obtained by normalizing and weighting multiple similarity measures. The measure terms include at least the structural consistency evaluation corresponding to the structural difference information, and can also combine brightness distribution differences, texture differences, and frequency domain differences to improve adaptability to changes in imaging conditions. To enable the first similarity to be compared across regions, normalization can be performed based on intra-unit statistics. For example, each measure term can be mapped to a unified 0 to 1 interval, where 0 indicates that the two image blocks are completely consistent in this dimension, and 1 indicates that they are completely inconsistent. The weighting coefficient can be set according to the stability requirements of the post-disaster scene, so that the structural difference dimension has a higher proportion. The first threshold is used to determine the change seed unit. Its determination method can adopt the stable region self-registration method: randomly select a number of candidate units that are considered to be structurally stable within the historical annotation coverage area, such as units close to the road skeleton but far from the disaster center, calculate the first similarity distribution of these units, and use the higher quantile as the reference value of the first threshold. For example, on a similarity scale of 0 to 1, if the first similarity of most stable units falls between 0.05 and 0.15, then 0.20 can be taken as the first threshold.

[0047] Furthermore, the iterative expansion does not attempt to expand every single neighboring unit around the seed unit. Instead, it uses a candidate selection process based on the direction of change to control computational load and reduce erroneous expansion. The direction of change is obtained through directional analysis of the difference information within the seed unit image block. This direction can be determined by the edge displacement trend corresponding to structural differences, the direction of line segment misalignment, or the principal axis direction of the difference energy distribution, and is used as a priority guide for expansion. During expansion, neighboring units located in the forward direction of the change direction are preferentially selected from the neighboring units of the seed unit as candidate expansion units. A second similarity is calculated for the candidate expansion units and compared with a second threshold. The second threshold differs from the first threshold and is used for expansion determination. It can be set in a boundary gradient manner: for example, a lower proportion of the first threshold is taken as the expansion threshold, so that units at the edge of the change region still have a chance to be included, thereby avoiding the break in the first region. In the above example, if the first threshold is 0.20, the second threshold can be 0.15 or 0.12, so that the expansion can cover the transition region of the change zone, but still excludes normal fluctuations in the stable region. When a candidate expansion unit meets the second threshold condition, it is incorporated into the first region and updated as a new change seed unit so that it can continue to expand along the change direction; when the condition is not met, it stops expanding in that direction so that the growth of the change region is constrained by the difference trend rather than unbounded diffusion.

[0048] In one example, the first similarity is calculated as follows: For the current detection unit, extract the corresponding image blocks from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image; A similarity metric for characterizing the degree of image variation is calculated based on the image patch. The similarity metric includes differences in brightness distribution, texture features, structural features, and frequency domain features. The similarity metric is normalized, and a first similarity is generated using a weighted algorithm.

[0049] In yet another example, performing an iterative operation on the changed seed unit includes: S2.3.1: Calculate the change trend based on the change seed unit, wherein the change trend is obtained by performing directional analysis on the difference information between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image within the corresponding image block of the change seed unit; Specifically, while the changing seed unit can characterize local structural differences, structural differences in post-disaster scenarios are not always uniformly diffused in all directions: road washouts often extend along road corridors, landslide deposits often form strips along the slope direction, and splicing joint residuals are distributed in a strip-like pattern along the joint line. If equivalent attempts are made on all neighboring units of the seed unit during the iterative expansion stage, not only will a large number of invalid candidates be formed at the boundary, but also occasional differences without spatial organization rules will be included in the expansion link, causing the boundary of the first region to appear jagged and abrupt at the grid scale.

[0050] In this embodiment, for the image block corresponding to the change seed unit, difference information is first constructed, which includes at least structural difference information. Structural difference information can be obtained using edge / segment consistency analysis: for example, extracting edge maps or line segment sets from the post-disaster image block and the previous version image block respectively, and overlaying them in spatially aligned coordinates to obtain edge displacement concentration areas and edge missing areas; alternatively, corner distribution consistency analysis can be used: extracting corner sets from the two image blocks and statistically analyzing their spatial density changes to obtain directional bands with more concentrated structural reconstruction. After obtaining the difference information, directional analysis is performed to obtain the change direction: specifically, several directional sectors can be established within the image block, for example, dividing directional intervals with a step size of 15° or 22.5°, statistically analyzing the energy concentration or spatial extension length of the difference information in each directional sector, and selecting the direction with the highest concentration or longest extension length as the change direction; if two opposite directions both show high concentration, they are recorded in a bidirectional direction form, facilitating subsequent expansion verification along both the forward and reverse directions. To reduce the impact of local noise, directional analysis can prioritize statistical analysis of continuous linear segments or connected difference clusters corresponding to structural difference information, rather than statistical analysis of discrete pixels one by one, so that the direction of change is more in line with the spatial organization of structural changes.

[0051] S2.3.2: Based on the change trend, candidate expansion units are selected from the neighboring units of the change seed unit; Specifically, the neighboring units of the changing seed unit typically include four- or eight-neighboring units. If the second similarity is calculated for each of the neighboring units, the computational cost of each expansion increases proportionally to the boundary length. When the changing region extends in a strip shape, the boundary length increases continuously as the expansion progresses, causing a large number of lateral candidates unrelated to the orientation to be repeatedly evaluated, resulting in a lengthened iteration chain. To make the expansion process more in line with the spatial organization of differences and to avoid invalid calculations for neighbors in directions unrelated to the orientation, a directional window mechanism is used to select candidate expansion units: with the changing orientation as the center, only neighboring units falling within that directional window are selected to enter the candidate set.

[0052] In this embodiment, after determining the set of neighboring units of the changing seed unit, a direction window is constructed based on the direction of change, and the neighboring units are filtered by direction. The direction window can be defined as an angular range centered on the direction of change, for example, extending 45° to the left and right of the direction of change to form a window; in the grid neighborhood, the orientation of each neighboring unit relative to the seed unit can be mapped to a discrete direction, such as up, down, left, right, upper left, upper right, lower left, and lower right, and it is determined whether it falls into the direction window.

[0053] It should be noted that, since the directional window in this embodiment is not limited to a single discrete direction, but covers a certain angular range centered on the direction of change, such as a 90-degree range, the number of candidate expansion units falling into the directional window can be one or more under different change patterns and grid scales. This number is not a fixed value, but is adaptively determined according to the stability of the direction of change and the spatial discretization method.

[0054] In this embodiment, when the direction of change is highly consistent with the grid direction, for example, when the direction of change is close to the horizontal or vertical direction, the direction window typically contains only one forward neighboring cell. In this case, the expansion process manifests as a gradual advancement along a single main direction. When the direction of change is diagonal or lies between multiple discrete directions, the direction window can simultaneously contain two or three neighboring cells, such as a forward cell and its adjacent deflection cell, thereby allowing the expansion path to extend in a band-like or polygonal shape at the grid scale. In this way, the expansion process can maintain directional constraints while being compatible with the directional quantization errors caused by grid discretization, avoiding the breakage of the change region at the grid level due to overly rigid directional mapping.

[0055] refer to Figure 4 , Figure 4 This is a schematic diagram illustrating the selection of candidate expansion units provided in an embodiment of this application.

[0056] like Figure 4As shown, the gray square represents the current seed unit of change, and the eight squares surrounding it constitute the set of neighboring units of that seed unit, located in the top, bottom, left, right, and four diagonal directions, specifically including neighboring units one through eight. Without introducing constraints on the direction of change, these eight neighboring units could theoretically all be included in the extended verification range, requiring the calculation of the second similarity and threshold determination for each. This introduces a large amount of lateral calculation unrelated to the main trend of change when the change region extends in a strip or diagonally.

[0057] like Figure 4 As indicated by the middle arrow, the direction of change is obtained from the directional analysis of the difference information within the image block corresponding to the change seed unit. Its direction is not necessarily completely consistent with the orthogonal direction of the grid, but may fall between adjacent discrete directions. Based on this, a direction window is constructed centered on the direction of change, and the spatial orientation of neighboring units relative to the change seed unit is mapped to discrete directions for filtering. Only neighboring units falling within the range of this direction window are retained as candidate expansion units. In the example condition where the direction window covers an area of ​​approximately 90 degrees, neighboring unit 1 and neighboring unit 8, located in the forward direction of the change direction and its adjacent deflection direction, are filtered as candidate expansion unit 1 and candidate expansion unit 2, respectively. The remaining neighboring units located laterally or in the opposite direction are not included in the candidate set.

[0058] In one optional implementation, when there are multiple candidate expansion units, they can be merged into a single candidate expansion unit using an image stitching algorithm.

[0059] S2.3.3: Calculate the second similarity between the candidate expansion unit corresponding to the post-disaster remote sensing image and the candidate expansion unit corresponding to the previous version of the post-disaster remote sensing image, and compare the second similarity with a preset second threshold; Specifically, the first similarity used in the seed determination stage is typically more focused on strictly distinguishing between stable fluctuations and anomalous changes to ensure the reliability of seed units. The expansion determination stage, however, deals with the edges or transition regions of the change zone, where the difference intensity may be lower than the first threshold but still belong to the connected parts of the same change region. Directly reusing the first similarity and first threshold would fragment the change region into multiple discrete segments, forcing subsequent local adaptive matching windows to repeatedly activate on multiple fragments, resulting in a lengthy processing chain. Therefore, when the second similarity and second threshold are used for expansion determination, they need to be differentiated from those used in seed determination in terms of calculation content and threshold caliber: the second similarity emphasizes the judgment of the continuity of structural difference trends, while the second threshold is used to determine whether candidate units can serve as boundary extension units of the change region.

[0060] In this embodiment, for each candidate expansion unit, corresponding image blocks are cropped from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image, and a second similarity is calculated. The second similarity can use the multi-dimensional framework of the first similarity, but the weight configuration and metric selection are more inclined towards structural difference information to ensure that the expansion proceeds along the structural change zone: for example, in the second similarity, the weight of structural feature differences is increased and the weight of brightness differences is decreased to reduce the impact of shadow changes or radiation differences on the expansion decision; at the same time, a difference continuity metric can be introduced, that is, comparing the degree of consistency between the difference information of the candidate unit and the difference information of its parent seed unit in the direction of movement, as a supplementary factor to the second similarity, so that the expansion is more inclined to continue along the same structural perturbation direction.

[0061] S2.3.4: When the second similarity is greater than or equal to the second threshold, the candidate expansion unit is included in the first region and the candidate expansion unit is marked as a change seed unit; Specifically, at the grid scale, the changed region is often composed of multiple adjacent detection units. The expansion process is essentially to gradually push the confirmed changed connected pieces outward until the boundary converges. If a candidate expansion unit, after meeting the expansion judgment condition, is simply marked as the first region without becoming a new expansion starting point, the expansion will stop near the initial seed and cannot cross the long strip region along the change direction, causing the first region to break and requiring local adaptive matching to be repeated in multiple fragmented regions. Updating the candidate expansion units that meet the conditions as new change seed units allows the expansion link to continue advancing along the change direction until the difference continuity terminates, thereby forming a coherent changed region boundary at the grid discrete scale.

[0062] In this embodiment, when the second similarity is greater than or equal to the second threshold, the corresponding candidate expansion unit is included in the first region, and the candidate expansion unit is marked as a change seed unit to enter the next iteration. To ensure the controllability of the expansion, the boundary set can be updated while including the first region: the newly included seed unit is added to the boundary queue, and the units in its neighborhood that have not yet been processed and are not in the access state are used as subsequent possible candidates; at the same time, the units that have completed the expansion verification and do not meet the second threshold are recorded as path blocking units. If the same path pointing to the blocking unit is encountered again in subsequent iterations, the second similarity calculation will not be triggered again to reduce duplicate verification.

[0063] In some optional implementations, after a candidate expansion unit is updated to a change seed unit, the change direction of the parent seed unit can be inherited as an initial value, and the change direction can be recalculated or corrected based on the difference information of the new seed unit, so that the expansion direction can be adaptively adjusted when the change zone bends or forks: for example, when the road direction turns, the main axis direction of the difference information will also change accordingly, so that the expansion can continue to advance along the road turn without breaking.

[0064] Furthermore, if the second similarity is less than the second threshold, it indicates that although the corresponding candidate expansion unit has certain differences between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image, these differences fail to maintain a continuous or consistent trend with the confirmed change seed units at the structural level. They are more likely to originate from local imaging noise, residual registration errors, or non-disaster-related apparent changes. In this case, excluding the candidate expansion unit from the first region can prevent the introduction of units without real deformation correlation into the change region, thereby preventing the change region from being widened or shifted in direction by irrelevant disturbances during the expansion process. Simultaneously, this unit is recorded as a verified unit that failed the expansion judgment. This record is used to constrain repeated verification behavior from the same change trend or adjacent change seed units in subsequent iterations, ensuring that the candidate expansion unit is no longer repeatedly triggered for second similarity calculation under the same or similar expansion conditions.

[0065] S2.4: If the first similarity is less than the first threshold, the current detection unit is determined as the second region, wherein the first threshold is used to characterize the similarity determination threshold for determining the current detection unit as a changed seed unit; In this embodiment, when the first similarity of the current detection unit is less than the first threshold, the detection unit is identified as a second region unit and its spatial index is recorded. It should be noted that since the second region is in a relatively stable state by default, no further expansion is necessary.

[0066] S2.5: After all detection units have finished processing, output the first region and the second region.

[0067] For example, the following is a numerical example to illustrate how to convert the calculation process to obtain the specific area division. This example is only used to explain the relationship between the calculation link and the dimensions. The selected parameters and values ​​are illustrative and do not represent the actual calibration results or engineering recommended values.

[0068] In this example, we assume the remote sensing imaging platform has a resolution of 0.5 m / pixel, and the candidate unit after gridding has a side length of 32 pixels, corresponding to a ground coverage of approximately 16 m × 16 m. One candidate unit that spatially overlaps with historical annotation line features (e.g., road centerline buffer zone) is selected as the current detection unit, and image patches corresponding to this unit are cropped from both the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image. To generate the first similarity score, four similarity measures are calculated for this image patch, and each measure is mapped to a uniform scale from 0 to 1, where 0 indicates high consistency between the two time phases in that dimension, and 1 indicates significant difference in that dimension; the difference in brightness distribution is set to 0.12, the difference in texture features to 0.18, the difference in structural features to 0.34, and the difference in frequency domain features to 0.15. Considering that structural differences are more indicative of real-world morphological changes in post-disaster scenarios, the weights are set as follows: brightness 0.15, texture 0.20, structure 0.45, and frequency domain 0.20. Based on this, the first similarity is obtained by weighted summation of the above four normalized metrics, resulting in a conversion of 0.23. This means that structural differences contribute more to the summation value, making the overall similarity more sensitive to changes at the structural level. This conversion refers only to the weighted synthesis of normalized multi-dimensional quantities, which can be implemented by those skilled in the art.

[0069] For example, 20 candidate units considered structurally stable are selected within the historical annotation coverage area, such as areas far from the disaster center and with continuous road skeletons. Their first similarity distribution is calculated, and it is found that the first similarity of these stable units mostly falls between 0.06 and 0.16, with the higher quantile (e.g., the 90th percentile) being 0.17. To make the determination of change seeds more stringent, a margin can be added to this quantile value to cover a small amount of alignment residuals and illumination differences. For example, a margin of 0.03 is taken, and the first threshold is 0.20. Since the first similarity of the current detected unit is 0.23, which is greater than 0.20, the current detected unit is marked as a change seed unit and enters the iterative expansion link; if the first similarity is less than 0.20, the current detected unit is directly assigned to the second region and no longer triggers subsequent expansion determination.

[0070] For example, difference information is constructed for the image block corresponding to the seed unit of change, and directional analysis is performed to obtain the change direction pointing to the "upper right" direction, which can be understood as the change zone extending to the upper right along the road corridor. A directional window is constructed with this direction as the center, taking a range of 45° to the left and right, i.e., a directional window of approximately 90°; in the eight neighborhoods, there may be more than one neighboring unit falling into this directional window. Suppose that two candidate expansion units are obtained at this time: candidate expansion unit A (neighborhood in the right direction) and candidate expansion unit B (neighborhood in the upper right direction). Image blocks corresponding to A and B in the post-disaster image and the previous version image are cropped respectively, and the second similarity is calculated. The second similarity still uses multi-dimensional normalization and weighted synthesis, but in the expansion judgment, more emphasis is placed on the continuity of the structural difference trend. Therefore, for example, the weights are adjusted to brightness 0.10, texture 0.15, structure 0.55, and frequency domain 0.20, and the difference continuity is allowed to be introduced as a supplementary representation of the structural difference. For example, the consistency of the extension of the structural difference in the direction is included in the statistics of the structural difference. In this example, the normalized metrics for candidate expansion unit A are 0.10 for brightness, 0.16 for texture, 0.24 for structure, and 0.12 for frequency domain, resulting in a second similarity of 0.19. Similarly, the normalized metrics for candidate expansion unit B are 0.08 for brightness, 0.14 for texture, 0.11 for structure, and 0.10 for frequency domain, also resulting in a second similarity of 0.12. The second threshold is used for expansion determination, and its value is generally lower than the first threshold to allow transition units at the edges of changing regions to be included. For example, a lower value of the first threshold of 0.20 can be used as the expansion threshold, such as 0.15. Alternatively, it can be fine-tuned by combining the upper limit of local stability fluctuations, but in this example, a second threshold of 0.15 is used for ease of explanation. Since the second similarity of candidate expansion unit A is 0.19, which is greater than 0.15, candidate expansion unit A is included in the first region and marked as a new seed unit for variation, so that its neighboring units can continue to be screened along the same direction. Since the second similarity of candidate expansion unit B is 0.12, which is less than 0.15, candidate expansion unit B is not included in the first region, and is recorded as a unit that has been verified but failed the expansion decision. This is used to constrain subsequent repeated verification from the same direction, thereby avoiding repeated calculation of the second similarity for this unit during the iteration process.

[0071] It should be noted that after the above expansion determination is completed, if none of the changed seed units (including the newly added seed unit A) have candidate expansion units that satisfy the second threshold within their directional window, the expansion in that direction naturally terminates, and the boundary of the changed region converges in that direction. When all candidate expansion units of all seed units can no longer pass the second threshold determination, the iteration operation ends, and the first region is formed by the initial changed seed units and the set of units included through expansion. In contrast, detection units that do not meet the first threshold and are not included in the first region through expansion can be identified as the second region. The second region is a stable region that can be used as a global geometric feature matching constraint by default, and no further expansion processing is performed. This allows the first and second regions to form a clear division of labor at the grid scale and provides a direct boundary for subsequent local adaptive matching and global geometric matching.

[0072] In yet another example, the detection unit further includes an access state, which characterizes whether the corresponding detection unit has participated in similarity calculation, wherein: During the iterative operation on the changed seed unit, the detection unit for which the second similarity has been calculated is set to the visited state; When a detection unit is in an visited state, the first similarity calculation is not repeated for the corresponding detection unit.

[0073] It should be emphasized that the second region of this application consists entirely of the corresponding detection units, that is, it does not include candidate units that have not been included in the feature annotation information. However, due to changes in the terrain, there may be some candidate units in the first region that did not originally include feature annotation information.

[0074] Next, we will further elaborate on the technical aspects of local and global matching in the method of this application.

[0075] In one example, the extraction of local features and adaptive feature matching includes: S3.1: For each first region, at least one local feature extraction window is determined within the first region, and local structural features are extracted within the local feature extraction window. The local structural features include at least one of corner features, edge features, and line segment features. Specifically, the spatial units corresponding to the first region have been expanded through variation seeds and directional constraints. Within them, there are often both locally stable structures that can be aligned and structural fragments that have been deformed or destroyed. If features are extracted directly across the entire first region and global matching is attempted, a large number of features will come from newly formed structures after deformation or structures that have disappeared, leading to inconsistencies in matching pairs. This increases the probability of mismatches and causes jumps in local registration relationships.

[0076] In this embodiment, for each first region, the set of grid cells in that region is first clustered to obtain one or more connected sub-regions. Local feature extraction windows are then deployed using these connected sub-regions as the basic objects. The local feature extraction windows can employ either a sliding window or an adaptive window approach: in the sliding window approach, windows are deployed within the bounding rectangle of the connected sub-region at preset step sizes; in the adaptive window approach, the center of the candidate window is chosen based on the location where the difference gradient in the structural difference information changes more gently, allowing the window to preferentially cover alignable structures rather than completely destroyed areas. The window size is set in relation to the image resolution. For example, at a resolution of 0.5m / pixel, the window can be 64×64 pixels, covering approximately 32m×32m; at a resolution of 1m / pixel, the window can be 64×64 pixels, covering approximately 64m×64m, ensuring that the window contains structural context such as road bends, building corners, and river boundaries that can be used for matching. Overlap between windows is allowed, with an overlap ratio of 25% or 50%, to avoid features being truncated due to local structures being located precisely at the window boundaries. In this embodiment, local structural features are extracted within each window. These local structural features include at least one of the following: corner features, edge features, and line segment features. Corner features are obtained by performing corner response calculations on local image blocks and selecting key points with high response values ​​and sparse spatial distribution. Edge features are obtained by edge detection to acquire a set of edge pixels and extract their main direction and connected contours. Line segment features are obtained by line segment detection to acquire a set of line segments and record their endpoints, lengths, directions, and intersecting relationships. To improve feature stability, intensity normalization and denoising processing can be performed on the image blocks before extraction to avoid deviations in structural feature extraction caused by local brightness differences.

[0077] S3.2: Based on the deformation information of the ground features, determine the matching search direction constraint and matching search range constraint of the local structural features; Specifically, the challenges of local matching in the first region lie in two aspects: firstly, pose deviations caused by overall translation and seam misalignment; and secondly, non-rigid changes caused by local distortion or damage. If the search range is set too large, the matching will be interfered with by a large number of irrelevant structures; if the search range is set too small, it may not be able to cover the true offset, leading to matching failure. Similarly, without directional constraints, matching candidates will spread in all directions, especially in near-linear structures such as road edges and riverbanks, where erroneous correspondences along the structural direction are prone to occur. Ground feature deformation information comes from the directional and amplitude characteristics of the differences between post-disaster images and previous versions of images. It not only represents the location of the change but also carries the direction of change propagation and the trend of local misalignment. Therefore, it can be used to converge the matching search space from a two-dimensional plane to a restricted region consistent with the deformation trend, so that local matching can still maintain a controllable candidate scale in complex change areas.

[0078] In this embodiment, the deformation information of ground features includes at least the directional statistical results of structural difference information, and may further include local displacement trend estimation and deformation intensity classification information. The matching search direction constraint can be determined by the deformation direction: for example, when the deformation direction is consistent with or approximately consistent with the aforementioned change trend, the matching search direction constraint is set as a fan-shaped directional window centered on the deformation direction, restricting candidate matching points to be mainly distributed within this fan; when the deformation extends in both directions or has a bifurcation trend, it is allowed to set two directional windows and verify them separately. The matching search range constraint is used to limit the search radius or search window size, which can be determined by deformation intensity estimation. Deformation intensity can be obtained by the concentration of local structural differences or the statistics of local alignment residuals. For example, the structural difference intensity is divided into three levels: low, medium, and high, and different search radii are set for each level. For example, at a resolution of 0.5 m / pixel, the search radius for low-level deformation can be set to 6 pixels (approximately 3 m), for medium-level deformation to 12 pixels (approximately 6 m), and for high-level deformation to 20 pixels (approximately 10 m), so that the search range can cover local misalignments while avoiding unbounded expansion. If there is an obtained global geometric drift in the first region (e.g., a global anchor point matching result from the second region), the search range can be superimposed with a local deformation margin on the drift. For example, if the global drift is estimated to be 2 m east and 1 m north, the local search center can first be shifted according to the drift, and then a local search can be performed around it according to the above radius, thereby reducing invalid searches.

[0079] S3.3: Based on the matching search direction constraint and the matching search range constraint, the local structural features extracted from the post-disaster remote sensing image are matched with the local structural features of the corresponding area in the previous version of the post-disaster remote sensing image to obtain the local anchor point matching result of the first area. Specifically, the local anchor point matching results will be directly used to construct the local registration relationship of the first region. Therefore, the matching process not only needs to find feature pairs, but also needs to perform consistency verification on the matching pairs to avoid non-physical jumps in the registration relationship caused by a small number of local mismatches. In the first region, some structures may actually disappear or be newly formed, resulting in some features not corresponding between the two phases; if the matching still forcibly outputs a correspondence, the local registration relationship will be pulled by erroneous feature pairs.

[0080] In this embodiment, within the local feature extraction window of the post-disaster remote sensing image, descriptive information is constructed for each local structural feature and retrieved within the corresponding region of the previous version image. The determination of the corresponding region follows the directional and range constraints of S3.2: first, the search center is determined based on the global drift or deformation direction, then candidate regions are cropped based on the search radius, and within the candidate regions, only candidate positions falling into the directional window are matched. The matching calculation can adopt the nearest neighbor matching method based on descriptor similarity: for example, the local neighborhood gradient distribution description is calculated for corner features, the line segment direction, length, and endpoint neighborhood structure description is calculated for line segment features, and the direction histogram and curvature change description of edge segments are calculated for edge features; several candidates with the closest descriptors within the candidate range are selected as initial matching pairs. To reduce ambiguity, a bidirectional consistency check can be performed on the initial matching pairs: that is, while the post-disaster image feature matches the previous version image feature, the reverse match should also point back to the same feature, and matching pairs that fail the consistency check are eliminated. Subsequently, a local geometric consistency check is performed on the matching pairs that pass the consistency check: for example, selecting several matching pairs within the same window and checking whether their relative displacements are concentrated within a certain finite range, or checking whether the angular relationship and relative distance relationship between line segment matching pairs are consistent; matching pairs that do not conform to local geometric consistency are marked as outliers and removed. The set of remaining matching pairs constitutes the local anchor point matching result of the first region, which can be output in the form of anchor point coordinate pairs for use in subsequent generation of local registration relationships.

[0081] In yet another example, the extraction and matching of global geometric features includes: S3.4: For each second region, determine at least one global feature extraction range within the second region based on the historical annotation data; Specifically, the second region is identified as a set of areas with stable landforms, exhibiting higher structural consistency between post-disaster and previous image versions, and can serve as a reliable reference for inferring overall geometric drift relationships. However, the second region is typically not a continuous block, but rather composed of multiple dispersed stable units, with variations in structural information density and repeatability among different units. If all pixels of the second region are indiscriminately included in global feature extraction, it will introduce a large number of low-information areas, such as large areas of uniform ground, spectrally singular open spaces, and water surfaces, reducing the distinguishability of anchor point features. Furthermore, it will include transitional units near the boundaries of the second region in the calculation, where deformation effects or alignment residuals may still remain, thus causing the global matching results to be influenced by local anomalies.

[0082] In this embodiment, the method for determining the global feature extraction range based on historical annotation data includes: within the second region, using the annotation information of land feature elements as clues, constructing several global feature extraction ranges, so that each global feature extraction range covers at least one type of structurally stable and observable annotation object in both temporal phases. For example, when the historical annotations include line features (road centerline, embankment line), a buffer zone can be constructed on the spatial range of the line feature and intersected with the second region unit set to obtain a strip-shaped extraction range distributed along the line; when the historical annotations include point features (bridge nodes, road intersections, key facility points), a square or circular neighborhood range of fixed scale can be constructed with the point feature as the center, and intersected with the second region to obtain the extraction range; when the historical annotations include surface features (building outlines, stable water body boundaries), the extraction range can be constructed on the boundary zone of the surface feature, making the range more concentrated on the structural edge rather than the low-texture area inside the surface. To ensure spatial coverage of global constraints, the number of extraction ranges can be set according to the labeled coverage area and image scale: for example, at least 5 global feature extraction ranges should be selected within a map sheet and spatially distributed across different quadrants of the map sheet; when the labeled objects are dense, the number of extraction ranges can be increased to 10-20 to improve matching redundancy. The scale of each extraction range can be set according to the resolution. For example, when the resolution is 0.5m / pixel, a single extraction range can be 128×128 pixels, approximately 64m×64m; when the resolution is 1m / pixel, it can be 128×128 pixels, approximately 128m×128m, to ensure that the range contains multiple types of structural elements for matching.

[0083] S3.5: Within the scope of the global feature extraction, extract global geometric features from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image, respectively. The global geometric features include at least anchor point features. Specifically, global geometric features are used to describe geometric entities or relationships in the second region that can stably correspond between two temporal images. They need to possess both repeatable detection and distinguishable matching attributes. Post-disaster images may differ from previous versions in terms of radiometric conditions, such as variations in illumination angle, shadow projection, and changes in reflection due to surface moisture. If global matching relies solely on brightness or texture features, mismatches of homogeneous texture areas or insufficient feature point repetition are likely to occur. In contrast, geometric structures (road intersections, bridge endpoints, building corners, shoreline inflections, etc.) have greater cross-temporal repeatability in stable regions. Therefore, global geometric features should at least include anchor point features to ensure that global matching is based on structurally stable geometric references.

[0084] In this embodiment, within each global feature extraction range, anchor point features are extracted from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image, respectively. Anchor point features can be one or more combinations of keypoint anchors, line intersection anchors, or boundary fold anchors: keypoint anchors can be obtained by corner detection, and descriptive information with rotation and scale adaptation capabilities can be calculated for their surrounding local neighborhood; line intersection anchors can first detect line segment sets, then identify line segment intersections or nearest-neighbor convergence points as anchors, and record geometric descriptions such as the angle combination at the intersection and the direction sequence of adjacent line segments; boundary fold anchors can extract turning points with significant curvature changes on boundaries such as road boundaries, riverbanks, and building outlines, and record the direction of their two sides and the local boundary shape encoding. To ensure spatial uniformity of anchor points within the range, suppression strategies can be implemented in the anchor point candidate set. For example, a minimum spacing constraint can be used to ensure that the pixel distance between any two anchor points is not less than a preset distance. For instance, the minimum spacing can be 20 pixels (approximately 10m) at a resolution of 0.5m / pixel and 20 pixels (approximately 20m) at a resolution of 1m / pixel, to avoid excessive concentration of anchor points at a local edge, which could lead to a global estimation bias. In addition to anchor point features, this embodiment also allows the simultaneous extraction of global geometric relationship features, such as the relative distance ranking between anchor points, the shape category of triangles formed by three points, and the topological relationship summary of line segment networks, to provide stronger geometric consistency constraints during subsequent matching.

[0085] S3.6: Perform global geometric feature matching between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image to obtain the global anchor point matching result of the second region; Specifically, the global anchor point matching results in the second region are used to characterize the overall geometric drift relationship between the previous version of the image and the post-disaster image. This requires not only matching a sufficient number of anchor point pairs but also ensuring that the anchor point pairs are geometrically consistent to avoid a few incorrect matches interfering with the overall relationship estimation. In remote sensing images, even in stable regions, false matches may occur due to local texture similarities, especially in scenes with regular road grids, repetitive building layouts, or striped farmland textures. Without geometric consistency constraints, the matching results may appear locally reasonable but globally contradictory, making subsequent global registration relationships unstable for annotation transfer.

[0086] In this embodiment, global anchor point features of the post-disaster image and the previous version image are matched. The matching process may include candidate generation, mutual consistency verification, and geometric consistency screening. In the candidate generation stage, for each previous version anchor point, several candidates with the closest descriptors are retrieved from the post-disaster anchor point set, such as the top 2 or 3, forming an initial candidate pair. In the mutual consistency verification stage, it is required that after a previous version anchor point matches a post-disaster anchor point, the reverse match should also point back to the same previous version anchor point; otherwise, the candidate pair is discarded to reduce ambiguity caused by one-to-many or many-to-one relationships. In the geometric consistency screening stage, a portion of the candidate pairs that pass the mutual verification are selected as fitting samples, and their consistency is evaluated. For example, a global pose relationship can be estimated by randomly selecting several candidate pairs, such as a relationship model with translation, rotation, and scale changes as the main components, and the number of candidate pairs that can be simultaneously explained is counted. Finally, anchor point pairs consistently explained by this relationship model are retained as the inlier set. To ensure the output can be directly used for subsequent registration, this embodiment outputs the inlier set in the form of "anchor point coordinate pairs + matching confidence attributes." The confidence attributes may include descriptor similarity, cross-checking consistency markers, and geometric consistency inlier markers. If the number of matches is insufficient, range supplementation can be triggered: for example, if the number of inliers is less than a preset lower limit, the global feature extraction range is added from the second region, and extraction and matching are repeated. This lower limit can be set according to the task accuracy and model degrees of freedom; for example, at least 8 pairs of anchor point inliers can be required for stable estimation of the overall relationship.

[0087] In yet another example, based on the matching results of the first and second regions, annotation precision registration data is generated, including: Based on the local anchor point matching results of the first region, a local registration relationship is determined to characterize the historical annotation data in the first region to the post-disaster remote sensing image. Based on the global anchor point matching results of the second region, a global registration relationship is determined to characterize the historical annotation data in the second region to the post-disaster remote sensing image; Based on the local registration relationship and the global registration relationship, annotation accuracy registration data is generated.

[0088] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for registering remote sensing images, characterized in that, The method includes: Acquire post-disaster remote sensing images and historical labeled data; Change detection is performed based on the post-disaster remote sensing images and historical annotation data to identify a first region and a second region, wherein the first region represents the region where the land cover morphology has changed, and the second region represents the region where the land cover morphology is stable. In the first region, local features are extracted and adaptive feature matching is performed based on ground deformation information. In the second region, global geometric features are extracted and matched based on historical annotation data. The ground deformation information is obtained by analyzing the difference information between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image corresponding to the historical annotation data. The difference information includes at least structural difference information. Based on the matching results of the first and second regions, generate registration data for annotation accuracy and output the registered annotation results.

2. The registration method for remote sensing images according to claim 1, characterized in that, The post-disaster remote sensing images are acquired through a remote sensing imaging platform, which includes at least one of a satellite remote sensing platform, an airborne remote sensing platform, and an unmanned aerial vehicle (UAV) remote sensing platform. The historical annotation data is used to characterize the annotation information of ground features on the base map of the previous version of the post-disaster remote sensing image, wherein the annotation information of ground features includes at least one of vector features, and the vector features include point, line and polygon features.

3. The registration method for remote sensing images according to claim 1, characterized in that, Change detection is performed based on the post-disaster remote sensing images and historical labeled data, including: Spatial alignment processing is performed between the previous version of the post-disaster remote sensing image corresponding to the historical annotation data and the post-disaster remote sensing image. On the base map of the previous version of the post-disaster remote sensing image corresponding to the spatially aligned historical annotation data, multiple detection units are determined based on the annotation information of ground features. Each detection unit is a local area that covers the annotation information of ground features, and the coverage area of ​​each detection unit is the same. The coverage area is set by the resolution of the remote sensing imaging platform. Select a detection unit as the current detection unit, calculate the first similarity between the current detection unit corresponding to the post-disaster remote sensing image and the current detection unit corresponding to the previous version of the post-disaster remote sensing image. If the first similarity is greater than or equal to a preset first threshold, mark the current detection unit as a change seed unit, and perform an iterative operation on the change seed unit until there are no candidate expansion units that meet the expansion conditions. The set of candidate expansion units obtained by the change seed unit through the iterative operation is determined as the first region, wherein the candidate expansion unit represents the neighborhood unit of the change seed unit. If the first similarity is less than the first threshold, the current detection unit is determined as the second region, wherein the first threshold is used to characterize the similarity determination threshold for determining the current detection unit as a changed seed unit; Once all detection units have finished processing, the first and second regions are output.

4. The remote sensing image registration method according to claim 3, characterized in that, The determination of multiple detection units based on feature annotation information includes: Based on the resolution of the remote sensing imaging platform, the previous version of the post-disaster remote sensing image corresponding to the spatially aligned historical annotation data is divided into grids to obtain multiple candidate units, where each candidate unit has the same coverage area. Based on the aforementioned feature annotation information, candidate units that have a spatial overlap relationship with at least one feature annotation are identified; Candidate units that have spatial overlap with the labeled ground features are identified as the detection units.

5. The remote sensing image registration method according to claim 3, characterized in that, The calculation method for the first similarity includes: For the current detection unit, extract the corresponding image blocks from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image; A similarity metric for characterizing the degree of image variation is calculated based on the image patch. The similarity metric includes differences in brightness distribution, texture features, structural features, and frequency domain features. The similarity metric is normalized, and a first similarity is generated using a weighted algorithm.

6. The remote sensing image registration method according to claim 3, characterized in that, Performing an iterative operation on the changed seed unit includes: The change trend is calculated based on the change seed unit, wherein the change trend is obtained by performing directional analysis on the difference information between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image within the corresponding image block of the change seed unit; Based on the change trend, candidate expansion units are selected from the neighboring units of the change seed unit; Calculate the second similarity between the candidate expansion unit corresponding to the post-disaster remote sensing image and the candidate expansion unit corresponding to the previous version of the post-disaster remote sensing image, and compare the second similarity with a preset second threshold; When the second similarity is greater than or equal to the second threshold, the candidate expansion unit is included in the first region and the candidate expansion unit is marked as a change seed unit.

7. The remote sensing image registration method according to claim 6, characterized in that, The detection unit further includes an access status, which is used to characterize whether the corresponding detection unit has participated in similarity calculation, wherein: During the iterative operation on the changed seed unit, the detection unit for which the second similarity has been calculated is set to the visited state; When a detection unit is in an visited state, the first similarity calculation is not repeated for the corresponding detection unit.

8. The registration method for remote sensing images according to claim 1, characterized in that, The step of extracting local features and performing adaptive feature matching includes: For each first region, at least one local feature extraction window is determined within the first region, and local structural features are extracted within the local feature extraction window. The local structural features include at least one of corner features, edge features, and line segment features. Based on the deformation information of the ground features, the matching search direction constraint and matching search range constraint of the local structural features are determined; Based on the matching search direction constraint and the matching search range constraint, the local structural features extracted from the post-disaster remote sensing image are matched with the local structural features of the corresponding region in the previous version of the post-disaster remote sensing image to obtain the local anchor point matching result of the first region.

9. The method for registering remote sensing images according to claim 8, characterized in that, The extraction and matching of global geometric features includes: For each second region, at least one global feature extraction range is determined within the second region based on the historical annotation data; Within the scope of global feature extraction, global geometric features are extracted from the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image, respectively. The global geometric features include at least anchor point features. Global geometric feature matching is performed between the post-disaster remote sensing image and the previous version of the post-disaster remote sensing image to obtain the global anchor point matching result for the second region.

10. The method for registering remote sensing images according to claim 9, characterized in that, Based on the matching results of the first and second regions, annotation accuracy registration data is generated, including: Based on the local anchor point matching results of the first region, a local registration relationship is determined to characterize the historical annotation data in the first region to the post-disaster remote sensing image. Based on the global anchor point matching results of the second region, a global registration relationship is determined to characterize the historical annotation data in the second region to the post-disaster remote sensing image; Based on the local registration relationship and the global registration relationship, annotation accuracy registration data is generated.

Images