A method for slope cross-time and space image registration and change detection based on skeleton constraint

Abstract

The present application relates to the technical field of geological disaster remote sensing monitoring, and more particularly to a method for slope cross-time-space image registration and change detection based on skeleton constraint. The method first extracts image stable area mask using a semantic segmentation network to filter vegetation and shadow interference; then constructs a hierarchical skeleton network through a multi-scale algorithm to lock a single-pixel core rigid skeleton representing the long-term stable structure of the slope; under the constraint of the skeleton buffer zone, a sub-pixel level fine registration is realized by integrating SuperPoint and LightGlue neural networks, and an optimal homography matrix is solved for high-precision time-space registration; finally, the aligned double-time images are input into the change detection network, the pixel-level changes are mapped to specific disaster semantic labels such as landslides and collapses in combination with the skeleton topological deformation field, and a risk warning report is output. The present application overcomes the problem of registration failure caused by non-steady-state texture under seasonal field changes of the slope, and significantly improves the robustness and semantic discrimination accuracy of cross-time-space monitoring.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology for geological disasters, and in particular to a method for cross-temporal and spatial image registration and change detection of slopes based on skeleton constraints. Background Technology

[0002] Landslides, spalling, and other geological disasters are sudden and destructive, posing a serious threat to transportation arteries and the safety of people's lives and property. In the field of geological disaster monitoring, traditional slope monitoring methods mainly rely on manual inspections and ground displacement monitoring equipment. However, these methods have many limitations when dealing with slopes with complex vegetation cover and significant topographic relief. Manual inspections are not only inefficient and costly, but also difficult to implement in severe weather or rugged terrain. While ground displacement monitoring equipment can provide high-precision local deformation information, its deployment and maintenance costs are extremely high, and its coverage area is limited.

[0003] With the rapid development of drone technology, its application in slope monitoring is becoming increasingly widespread. Drones can quickly acquire high-resolution images of slopes, providing a new means for dynamic slope monitoring. However, single-temporal drone image analysis cannot capture the dynamic changes of slopes over time, making it difficult to detect potential geological hazards in a timely manner. Automated anomaly detection through continuous drone imagery of the same slope at different times and comparison of images acquired at different times is expected to become the mainstream method for highway slope safety monitoring in the future. However, when the monitoring object becomes a natural slope with continuously growing, shedding, and swaying vegetation, achieving accurate dual-temporal image registration and change detection still faces the following technical challenges: 1) Registration failure caused by non-steady texture interference. Slope surfaces are typically covered with abundant vegetation and gravel, and are severely affected by lighting angles and shadow migration. Traditional feature matching algorithms (such as SIFT and ORB) extract feature points across the entire image, easily generating numerous spurious feature points in seasonally changing vegetation areas or dynamically shifting shadow edges. These "spurious change" signals result in extremely low signal-to-noise ratios for matched point pairs, making it difficult to calculate high-precision geometric transformation matrices, thus causing registration shift.

[0004] 2) Insufficient utilization of rigid geometric contour features. Highway slope areas contain numerous rigid structures such as drainage ditches, retaining walls, and bedrock fissures, which exhibit extremely high geometric and topological stability over long time spans. However, existing registration schemes largely rely on local pixel features, neglecting the global constraints of the slope skeleton contour. Without semantic guidance from rigid structures, the algorithm struggles to distinguish between normal seasonal land cover changes and actual slope structural displacements.

[0005] 3) Disconnect between change detection and semantic recognition. Traditional change detection often remains at the level of pixel-level grayscale or color difference comparison, which cannot effectively filter non-hazardous interference and is difficult to automatically map the detected pixel offset to specific hazard types (such as landslides, weathering and erosion). This results in unstructured and difficult-to-interpret monitoring results.

[0006] In summary, existing technologies lack a slope monitoring solution that can deeply integrate semantic information, rigid skeleton constraints, and advanced deep learning algorithms, resulting in insufficient detection accuracy and automation levels in complex field environments to meet the high standards required for highway maintenance. Summary of the Invention

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide a method for cross-temporal and spatial image registration and change detection of slopes based on skeleton constraints.

[0008] The objective of this invention can be achieved through the following technical solutions: A method for cross-temporal image registration and change detection of slopes based on skeleton constraints includes the following steps: S1: Dual-temporal sensing data acquisition and image consistency processing; Acquire images and geolocation information of the same slope scene at different time phases, and perform distortion correction and radiometric calibration on the images to be registered to eliminate geometric and radiometric deviations caused by non-environmental factors; S2: Seasonal dynamic texture masking and stable region extraction based on semantic segmentation; By using a semantic segmentation model to identify and remove dynamic texture feature regions in images that change with the seasons, a stable region mask that remains topologically stable across time phases can be locked from a complex background. S3: Multi-scale skeleton extraction and spatiotemporal coarse registration; Extract single-pixel-width contour skeleton lines within the stable region mask, and perform skeleton geometric topological similarity analysis in conjunction with geolocation information to obtain the initial alignment parameters of images taken at two different times. S4: Pixel-level image fine registration driven by skeleton constraints; Using the contour skeleton as a geometric control net constraint, feature mapping and residual optimization are performed in the skeleton-guided search space to achieve sub-pixel-level spatial alignment of images captured at two different times. S5: Multi-feature fusion for slope temporal change detection and risk warning; The aligned image is input into the change detection network, feature offsets are extracted, and pixel-level changes are mapped to specific slope disaster semantic labels by combining the skeleton topology. The deformation features of abnormal areas are statistically analyzed and input into the time series prediction model, outputting slope stability scores and executing graded early warning logic.

[0009] Specifically as follows: In step S1, in order to eliminate image inconsistencies caused by illumination shift, shooting angle and camera geometric distortion during the UAV's cross-temporal and spatial acquisition process, this method proposes an image consistency processing mechanism based on the collaborative processing of perceptual metadata.

[0010] Step S11: Acquisition of UAV imagery and geolocation information; First, it is necessary to obtain the original two-phase images of the slope to be monitored and its geographic location information (including longitude, latitude, and absolute altitude), mainly including: (1) Visual perception information recording: Recording the raw tensor data and meta-attributes of the image, including: dual-temporal three-channel graphic tensor (Where H and W represent the image pixel height and pixel width respectively, and rgb is the pixel value of each pixel in the red, yellow and blue channels), bit depth and exposure parameters during acquisition; (2) Spatial georeferenced information recording: synchronously extract geolocation information at the time of image acquisition, including: longitude, latitude and absolute flight altitude. ; used to narrow down the initial search domain for skeleton matching using GPS clustering in step S3; Step S12: Unification processing of UAV images; Then, it is necessary to uniformize the differences in image scale and brightness caused by the two cross-temporal and spatial acquisitions by the drone. This includes the following steps: (1) Scale alignment: The ground resolution difference between the two temporal images is calculated based on the flight altitude. The second image is resampled using the first image as a reference. The two temporal images are resampled to the same spatial resolution through bilinear interpolation to ensure that the two temporal images are consistent on the physical scale of a unit pixel. (2) Illumination compensation: To address the illumination differences at different times, a histogram specification algorithm is used to perform color and brightness mapping on dual-time phase images to eliminate radiation intensity deviations caused by seasonal solar altitude angles and cloud cover. (3) Image plane geometric flattening correction: Using the pre-calibrated camera distortion coefficients, radial and tangential distortion elimination is performed on the image, and a bilinear interpolation algorithm is used for grayscale reconstruction, outputting a standard image to be registered with geometric and color distribution alignment where the image plane geometric residual is controlled at the sub-pixel level.

[0011] Specifically, this step aims to eliminate pixel deviations caused by non-disaster factors and transform the raw data under heterogeneous acquisition conditions into standardized image pairs with geometric and color distribution alignment, serving as a unified input benchmark for subsequent semantic perception and skeleton extraction.

[0012] In step S2, due to the severe spurious noise caused by seasonal vegetation growth and instantaneous shadow migration in the field slope environment, traditional registration methods are prone to matching redundancy or deviation in dynamic texture areas. In order to enable the monitoring system to shield the interference of seasonal vegetation growth and dynamic shadows on change detection, this method proposes a dynamic texture masking strategy based on the confidence of stable regions.

[0013] Step S21: Construct the semantic segmentation encoding structure; The dual-temporal normalized image output by S1 Input a pre-trained DeepLabv3+ semantic segmentation network, utilize its dilated spatial convolutional pooling (ASPP) module to capture multi-scale contextual features, and output each pixel in the image. Class probability vector There are 10 rigid stable region categories, and each dimension represents the confidence level of a pixel belonging to the corresponding category; Step S22: Calculate the regional stability score; Then calculate each pixel. Confidence score of stable region It is defined as the sum of the probabilities of rigidly stable classes: By setting a confidence threshold Filtering satisfies The pixels are used to construct an initial stable region distribution map; Step S23: Generate a stable region mask; Morphological closing operations are performed on the initial stable regions. First, a 3×3 rectangular structuring element is used to perform a dilation-erosion closing operation on the initial mask. The 'bridging' effect of the dilation process fills texture gaps, and the erosion process smooths edge burrs. Then, the connected area of ​​all stable regions in the image is calculated, and small connected regions that do not have global constraint significance are filtered out by area thresholding. Finally, a rigid stable region mask spanning multiple time phases is locked. , which serves as the computational domain for skeleton extraction in step S3.

[0014] Specifically, this mask not only provides semantic boundaries for the skeleton refinement of S3, but also provides a prior probability distribution for the subsequent elimination of "spurious change" signals.

[0015] Specifically, the attribute classification of regional stability in step S2 is mainly based on the geometric deformation characteristics of various semantic features across time and space dimensions, and the semantic categories are defined by attribute: (1) Rigid stable category area: Bedrock, highway slope, drainage ditch and artificial protective structure are defined as rigid stable category. These features have visual edge consistency across time phases and are the core carriers of skeleton constraints. The 10 categories used to calculate confidence are as follows: {Outcropped bedrock, concrete facing, masonry, drainage ditch, intercepting ditch, prestressed anchor frame beam, curbstone, crash barrier, retaining wall, maintenance walkway}; The edge consistency refers to the fact that the connectivity and geometric shape of the edges of ground features maintain a mapping relationship in the binary space under different time phases. (2) Unstable interference area: Vegetation, sky and dynamic shadows are defined as unstable categories. These features exhibit drastic texture fluctuations with the season and lighting conditions and do not participate in the subsequent skeleton constraint construction. The unstable interference area also includes occlusion edges that change with the shooting angle.

[0016] The weight space is defined as a probability distribution matrix aligned with the pixel scale of the input image. For the 10 classes of pixels identified as rigid stable regions, their corresponding weight space values ​​are set to... For pixels identified as unstable interference regions, their corresponding weight space values ​​are set to 0.

[0017] The final weight space is used as the "attention map" and is multiplied pixel by pixel with the original image. The resulting weighted gradient field is used as the direct input for skeleton extraction in the subsequent S3 step, ensuring that the extracted visual edges have physical stability.

[0018] This operation aims to guide the algorithm to focus computational resources on the edges of rigid structures with highly consistent geometry, thereby improving the robustness of skeleton constraints in cross-temporal scenarios.

[0019] In step S3, in order to establish robust initial topological associations in field environments lacking ground targets, this method proposes a rigid skeleton extraction and coarse registration mechanism based on multi-scale pyramids.

[0020] Step S31: Multi-scale skeleton extraction; The stable region mask output from step S2 is obtained using the pyramid algorithm. Perform multi-scale decomposition to construct a sequence of image pyramids from coarse to fine. ; Specifically, These represent skeleton lines at different scales extracted using the method described in S2 when the image size is scaled to different magnifications. The scaling factors used in this method are as follows: {0.1, 0.2, 0.25, 0.4, 0.5}.

[0021] Edge detection operators are executed at each resolution layer, and a hierarchical skeleton network with spatial continuity is generated by cross-scale feature fusion to capture the rigid topology of the slope in different dimensions. Step S32: Hilditch skeleton refinement; In the hierarchical skeleton network, the Hilditch thinning algorithm is used to iteratively peel off the skeleton lines; for pixels within the mask... By determining the connectivity of its 8-neighborhood and the number of 0-1 transformations, it is thinned into a fine skeleton with a single pixel width while maintaining the original topological connectivity. Step S33: Skeleton filtering based on path length function; Assuming there are a total of n skeletons, { }, Then define each skeleton. Path length function : in skeleton The number of pixels.

[0022] Set the length threshold to The endpoints and bifurcation points in the fine skeleton are traversed by length: If the path length of a certain skeleton branch ,but The stable rigid framework was determined to be retained. If the path length of a certain skeleton branch ,but False edges caused by seasonal texture remnants are identified and recursive pruning operators are executed. For isolated and short branches in the single-pixel skeleton network, repeated pruning is performed by iteratively scanning the endpoints until the length of all branches in the network is not less than [a certain value]. ; In one embodiment of the present invention, a length threshold is set as follows: ; Finally, all eligible skeletons were integrated to identify the core rigid skeleton representing the long-term stable structure of the slope.

[0023] Step S34: Combine image geolocation information with search space constraints; After extracting the skeleton Subsequently, coarse registration is performed based on the geometric topology information and GPS information of the skeleton, specifically including the following steps: Using the latitude, longitude, and altitude geolocation information of the images recorded in S1, the spatial Euclidean distance between the center points of the images acquired at two different times is calculated. By using spatial clustering analysis to identify overlapping regions, the initial search domain for skeleton matching is limited to the neighborhood determined by location information, thus avoiding false matches caused by global search. Step S35: Coarse registration based on Fourier shape descriptors; The contour skeleton The spatial coordinate sequence is transformed into a complex field representation. First, the skeleton is extracted. The coordinates of a series of consecutive points are represented as ( , ), ( , ), …., ( , ).

[0024] Then combine the two-dimensional coordinates of each point into a complex number. , where i is the imaginary unit. This simplifies the two-dimensional planar motion coordinates into a one-dimensional complex sequence.

[0025] To ensure comparability between different images, skeletons of different lengths are usually resampled to a fixed number of points of 128 through interpolation.

[0026] Performing a Fourier transform (DFT) on a complex sequence s(k) transforms it from the spatial-temporal domain to the frequency domain, decomposing the spatial signal into shape components of different frequencies, and obtaining Fourier coefficients at different frequency exponents. : Low-frequency coefficients (n<5): represent the overall shape of the object; High-frequency coefficients (n≥5): represent the detailed texture of the object; Discard DC component (It only represents the center position of the contour), and the final graphic feature vector F is defined as: Using the method described above, calculate the values ​​of the two images respectively. Corresponding graphical feature vector .

[0027] Then, the cosine similarity between the skeleton feature vectors of the two images is calculated: The calculated cosine similarity value S ranges from [0, 1]. Set a preset threshold T = 0.85, when At that time, it was initially determined that the two had a shape correlation.

[0028] The "Nearest Neighbor Distance Ratio (NNDR)" principle is adopted, which means that among all images that may point to the same slope, if the ratio of the maximum similarity to the second largest similarity exceeds the preset safety margin of 0.8, then the physical features represented by the feature vector pair are finally locked as the same slope structure, thus completing the initial closure of the cross-temporal and spatial images.

[0029] This step aims to establish a "natural control network" between spatiotemporal images, achieve the initial closure of the global coordinate system, complete the spatiotemporal association of UAV dual-time images, provide a structured reference for the fine registration in step S4, and thus perform subpixel-level alignment within a smaller search range.

[0030] In step S4, to achieve the pixel-level alignment accuracy required for slope disease monitoring, this method proposes a skeleton-constrained feature matching and image registration method. This mechanism uses the skeleton from step S3 as a geometric control network to guide a deep neural network for high-precision feature alignment.

[0031] Step S41: Construct the skeleton adaptive constraint band; First, the core rigid frame is locked by S3. Based on this, a width of [missing information] is constructed using the dilation operator. Pixel-based strip search constraint band ,in Adaptive adjustments are made based on the image resolution and the initial alignment residual. Step S42: Skeleton feature extraction based on Superpoint; Then, a pre-trained SuperPoint convolutional neural network is used to perform a strip search on the constraint band. Feature perception is performed on the intersection region with the stable region described in step S2, and feature point pixel coordinates with sub-pixel precision and corresponding local high-dimensional descriptors are extracted simultaneously. ; Because the SuperPoint network is robust to light and shadow across scales, it ensures the consistency of feature detection of slopes under different seasons and lighting conditions. Step S43: LightGlue attention-driven feature matching; Feature point pairs from two temporal images are input into the LightGlue deep matching network. Its multi-headed attention mechanism is used to calculate the correlation strength between feature points in the two temporal images. Combined with an early-exit strategy, low-confidence matches are dynamically eliminated to improve matching efficiency while ensuring accuracy. Finally, the matching point pair index and matching confidence score are output based on the matching results. That is, the correspondence between specific pixels in the two images and the reliability score of each matching pair. When the matching score is >0.5, the matching pair is considered to be a valid matching point pair.

[0032] Step S44: Projection transformation and image registration; For the initial set of matching pairs, the RANSAC algorithm is used to perform a consistency check; Calculate the reprojection error of the matching point pairs under the fundamental matrix, and remove outliers that do not conform to the epipolar geometry constraints, i.e., the positional relationship between the two points does not conform to the geometric logic of camera imaging. Set a successful matching criterion: the number of valid matching point pairs (i.e., the set of corresponding points with high confidence that pass feature matching and geometric consistency tests) If the average confidence level is higher than the preset threshold, the current region registration and matching is confirmed to be successful; otherwise, return to step S3 to readjust the skeleton similarity search parameters. Based on the set of n high-quality feature matching points after successful confirmation ,in As a reference image point, For the image points to be registered, a homogeneous linear equation system is established using the direct linear transform (DLT). The homography matrix is ​​calculated as follows: The least squares criterion is used to minimize the reprojection cost function. Iterative optimization of matrix coefficients is performed to eliminate the interference of outliers on alignment accuracy: Using the solved optimal homography matrix The coordinates of each pixel in the image to be registered Perform a perspective transformation mapping to obtain its corresponding position in the reference image coordinate system. : That is, the pixel mapping relationship satisfies: Step S45: Registration accuracy evaluation and quality control; Bilinear interpolation is used to resample the transformed non-integer pixel coordinates to obtain the pixel grayscale values ​​of the aligned image; the transformed skeleton is then calculated. With reference skeleton The spatial residual between them ensures that the edge alignment accuracy reaches the sub-pixel level; After meeting the accuracy requirements, output the registered spatiotemporally aligned image. This serves as the benchmark for detecting abnormal changes in step S5.

[0033] In step S5, to transform abstract image differences into quantifiable disaster prevention decision-making criteria, this method proposes a mechanism that integrates skeleton deformation field change detection and semantic discrimination. This mechanism takes the aforementioned aligned image pairs as input and generates semantic early warning reports through temporal feature fusion.

[0034] The registered dual-temporal images output in step S4 are paired Inputting the pre-trained ChangeFormer Change Detection Network, the system utilizes a two-stream (Siamese) architecture to extract multi-scale spatial feature differences, generating a normalized change confidence heatmap. ; By utilizing the Transformer structure in the ChangeFormer network, global long-range dependencies between two temporal images are captured, improving the completeness of identifying anomalous change regions. Using the unstable interference region mask generated in step S2, the heatmap is... Perform pixel-by-pixel zeroing to eliminate false change signals caused by vegetation decay and light and shadow migration, ensuring that the detection target is focused on the bedrock, retaining wall and main road area; The core rigid skeleton extracted in step S3 Using geometric reference, calculate and Coordinate offset vector of skeleton pixels under time phase ; By constructing a skeleton topological deformation field, the overall displacement trend of key slope structures is quantified to determine whether structural sliding exists. Combined with pixel-level change intensity The skeleton is a vector A multi-criteria decision-making mechanism is used to map the detection results into specific semantic tags for slope hazards: (1) Landslide / collapse: manifested as a sudden change in confidence of large-area connectivity within a stable region, accompanied by severe fracture or displacement of the skeletal structure; (2) Weathering and erosion: manifested as small, scattered clusters of changes within the bedrock area, with the main skeleton maintaining topological consistency; (3) Damage to ancillary structures: manifested as local morphological distortion of specific skeleton lines such as drainage ditches and crash barriers; Finally, based on the equivalent area of ​​the changed region Combined with the deformation rate factor, the disaster risk index is calculated, and a slope safety monitoring report is generated, which is marked with the disaster type, location and degree of damage.

[0035] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly improves the robustness of registration in complex slope environments. By locking the stable region mask through the semantic segmentation module in step S2, this invention effectively eliminates non-steady texture interference caused by vegetation decay, seasonal color changes, and light and shadow migration in slope scenes. Combined with the depth feature extraction of SuperPoint and LightGlue in S4, it solves the pain point that traditional SIFT / ORB algorithms are prone to matching failure in low-texture or high-repetition texture areas in the field.

[0036] (2) Achieved precise alignment based on physical rigid constraints. This invention innovatively introduces a multi-scale feature pyramid and skeleton refinement algorithm to accurately extract core rigid skeletons with long-term topological stability, such as drainage ditches and retaining walls, from massive terrain features. By using the skeleton as a geometric reference system and setting a strip-shaped search buffer zone, the search space for pixel-level fine registration is limited to the skeleton neighborhood, which greatly reduces the mismatch rate and ensures sub-pixel level alignment accuracy of images at key engineering structures.

[0037] (3) An automated mapping mechanism from pixel offset to semantic disasters was established. This invention is no longer limited to simple pixel difference statistics, but constructs a skeleton topological deformation field and combines it with skeleton displacement vectors. The change confidence heatmap generated by ChangeFormer can automatically identify specific disaster types such as landslides, collapses, weathering and erosion, and structural damage. This gives the monitoring results clear physical meaning and semantic labels, greatly improving the intuitiveness of disaster early warning and the efficiency of decision-making.

[0038] (4) It possesses high engineering adaptability and closed-loop quality control. In step S4, this invention designs a successful matching confirmation mechanism and a backtracking strategy. If the current feature cannot meet the accuracy requirements of homography matrix solving, the system will automatically return to S3 to adjust the skeleton search parameters. This adaptive closed-loop design enhances the system's adaptability to images of different resolutions and different slope types, ensuring the stability of long-term monitoring data.

[0039] (5) High-efficiency task architecture under edge-cloud collaboration. This invention adopts an edge-cloud collaborative architecture, where the UAV is only responsible for collecting images and transmitting data back, while all complex computational tasks are completed on the cloud server. This architecture design makes full use of the UAV's maneuverability and the cloud server's powerful computing capabilities, ensuring both high efficiency in data acquisition and accuracy and speed in data processing. At the same time, the system is compatible with various types of sensors (such as visible light cameras, multispectral cameras, lidar, etc.), supports multi-source data fusion, further enhancing the comprehensiveness and adaptability of slope monitoring, and can be widely applied in the field of slope geological disaster monitoring under different geographical environments and climatic conditions. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a structural diagram of the slope anomaly area identification system of the present invention (① is the UAV image acquisition module, ② is the cloud server); Figure 3 This is a schematic diagram of the multi-scale skeleton pyramid proposed in this invention; Figure 4 This is a flowchart of the rigid skeleton extraction and spatiotemporal coarse registration method proposed in this invention; Figure 5 This is a flowchart of the skeleton-constrained image registration process proposed in this invention; Figure 6 This is a schematic diagram of skeleton-constrained image registration proposed in this invention; Figure 7 This is a schematic diagram of the dual-time slope image change detection and abnormal area identification method proposed in this invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0042] This embodiment provides a method for cross-temporal image registration and change detection of slopes based on skeleton constraints, aiming to solve the problems of low image registration accuracy and difficulty in capturing early disaster signals in the existing technology under the scenario of seasonal dynamic changes in slopes.

[0043] This identification method is integrated into a slope anomaly area identification system, which includes a UAV image acquisition module, a semantic segmentation module, a skeleton contour extraction module, a feature matching and registration module, a change detection module, and a visualization module. The UAV is only responsible for acquiring images and transmitting data; the subsequent data processing module is embedded in the server.

[0044] The slope anomaly area identification method of this invention is mainly applied to geological hazard monitoring of highway slopes. Taking a steep slope next to a highway as an example, the slope is about 30 meters high, with a slope of about 60 degrees and a surface vegetation coverage of about 70%, mainly consisting of herbaceous plants and a small number of shrubs, and there are potential geological hazards such as landslides and spalling. The specific implementation steps are as follows: Step S1: Dual-time image acquisition and preprocessing by UAV; A drone (DJI Mavic 4 Pro) flew at a speed of 12 meters per second at an altitude of 120 meters above the slope along a preset flight path. Simultaneously, a high-resolution camera (Hasselblad H6D-100c, 100-megapixel resolution) was activated to capture images of the slope's surface and record spatiotemporal synchronization tags such as GPS location, flight altitude, and camera attitude angle. The data acquisition method is as follows: Figure 2 As shown, the acquired images are first preprocessed on the drone, including using a Gaussian filter to remove noise and normalizing the image pixel values ​​to the 0-1 range to improve image quality and the accuracy of subsequent processing. The preprocessed images are then transmitted to the backend server via a wireless communication module.

[0045] Step S2: Semantic segmentation and stable region extraction; On the backend server, the DeepLabV3+ semantic segmentation model was first used to analyze the acquired images, identifying different regions such as vegetation, shadows, earthwork grids, drainage ditches, anchor bolts, and bedrock. This model was trained on a manually constructed slope dataset containing 2000 labeled images, covering categories such as vegetation, earthwork grids, drainage ditches, anchor bolts, and bedrock. By using a reverse approach, regions that change over time, such as vegetation and shadows, were excluded, while stable regions that remain unchanged across time, such as earthwork grids, drainage ditches, anchor bolts, and exposed bedrock, were identified as the benchmark for subsequent skeleton extraction and image matching. For example, in the acquired images, vegetation areas accounted for 68% of the images, while stable areas accounted for 32%.

[0046] Step S3: Multi-scale skeleton contour extraction and coarse registration; For the locked stable regions, a multi-scale morphological skeletonization algorithm is performed to extract the main topological skeleton. Specifically, a multi-scale image pyramid is first constructed using the Laplacian pyramid algorithm, such as... Figure 3As shown, the image size is scaled to different ratios ({0.1, 0.2, 0.25, 0.4, 0.5}), and skeleton lines at different scales are extracted using the method described in S2. These skeleton lines are then fused to generate a hierarchical skeleton network. An improved Hilditch thinning algorithm is then used to extract the skeleton from stable regions, with an iteration limit of 500. Short branches less than 30 pixels in length are removed, and the main skeleton is retained. Finally, a continuous and smooth path is fitted using master curve analysis to eliminate jagged noise and generate a refined skeleton network. For example, in the stable region of a slope, a skeleton network with a total length of approximately 1200 pixels is extracted. The rigid skeleton extraction and spatiotemporal coarse registration method flow is as follows: Figure 4 .

[0047] Step S4: Feature matching and pixel-level image registration with skeleton constraints; After extracting the feature vectors of the contours, the similarity transformation parameters of the image pairs are calculated using an image matching algorithm. A "shape-space" joint similarity framework is adopted. First, the geometric similarity of the skeleton itself is measured using shape descriptors (including feature parameters such as the total length of the skeleton, the number of branches, and the node connectivity). Then, the overall consistency between the two temporal images is measured using spatial transformation parameters. Principal component analysis (PCA) is used to align the contour skeletons, eliminating spatial deviations caused by differences in shooting perspective and target object rotation, ensuring the accuracy of similarity calculation. In this implementation case, the shape similarity of the two-temporal slope image pairs is calculated to be 0.85, and the spatial similarity is 0.92, indicating basic overlap. The flowchart of skeleton-constrained image registration is shown below. Figure 5 The visualization results of image matching are as follows: Figure 6 As shown, feature extraction and matching are performed based on the extracted slope contour skeleton; the white line represents the slope contour skeleton, the red dots represent feature points extracted within the skeleton constraint space, and the green line represents successfully paired feature point pairs. Based on the similarity evaluation results, geometric projection transformation parameters, including the rotation matrix R, translation vector t, and scale factor s, are calculated for successfully matched image pairs.

[0048] Based on the image-space transformation parameters output in step S4, homography projection transformation is performed on the temporal image I2 of phase T2, and it is resampled to the coordinate system of phase T1 to generate the registered image I2'. The overlapping areas of the slopes in the dual-phase images are rasterized to align the images with a unified grid structure, improving registration accuracy and change detection efficiency. Registration accuracy is detected through bidirectional residual error; if the feature point alignment error exceeds a threshold, local affine optimization adjustment is initiated. For example, when a large local error is detected, the skeleton nodes are used as control points to adaptively refine the grid in areas of severe deformation. The size of the refined grid is dynamically adjusted according to the deformation gradient; a grid spacing of 0.1 meters is used for areas with a deformation gradient greater than 0.5, and a grid spacing of 0.5 meters is used for areas less than or equal to 0.5. Sparse sampling is performed on flat areas, with the sampling interval expanded to 1 meter to balance registration accuracy and computational efficiency.

[0049] Step S5: Detection and risk warning of abnormal slope changes; The registered image pair {I1, I2'} is input into the improved ChangeFormer change detection network. This network, combined with a dual-branch encoder, extracts image features before and after registration, and achieves feature interaction through a cross-attention mechanism, ultimately outputting a pixel-level change probability map. Morphological filtering (combining opening and closing operations) removes noise from the probability map and fills cracks, extracting the contours of connected change regions. Image color features are analyzed using the HSV color space difference analysis, and texture features are extracted using the Gray-Level Co-occurrence Matrix (GLCM). A multi-dimensional anomaly feature vector is constructed by combining color and texture feature changes. Based on the color, texture, spatial context, and topological relationship with the skeleton of the anomaly region, the change area is automatically labeled as "slide," "peeling," "water seepage," or "human disturbance," achieving an interpretable mapping from pixel changes to disaster semantics. Figure 7 As shown, the T1 image shows that the slope surface is evenly covered by herbaceous vegetation. In the T2 image, the vegetation disappears in some areas, exposing the soil or bedrock, which is identified as a local landslide. In the T1 image, the slope vegetation is evenly distributed, while in the T2 image, the vegetation in a specific area is abnormally tall and the texture shows a dense cluster pattern, which is identified as an excess of water caused by underground seepage points.

[0050] The system statistically analyzes the area growth rate, deformation vector direction, and spatial distribution density of abnormal areas, inputting them into a trained LSTM time-series prediction model (64 hidden layer units, 12 time steps), and outputs a slope stability score. In this specific implementation case, the calculated slope stability score is 78. Since the slope stability score is below the preset safety threshold (80 points), the system automatically generates an early warning message, indicating the slope status and providing treatment suggestions. In this embodiment, the system indicates that the slope has a landslide risk and recommends further on-site investigation and reinforcement.

[0051] A slope anomaly area identification system that implements the above method includes the following main modules: UAV image acquisition module (1): Used to acquire dual-temporal raw image pairs of the slope. It also records the GPS coordinates, flight altitude, and camera attitude angle metadata corresponding to each waypoint, providing pose constraints for subsequent spatial clustering; Semantic segmentation and stable region extraction module (2): used to perform step S2, extracting ground feature confidence probability vectors using a pre-trained DeepLabv3+ network. Based on the region stability attribute classification, output a stable region mask after removing vegetation and shadow interference. ; Multi-scale skeleton contour extraction module (3): Used to execute step S3, constructing a multi-scale feature space using the pyramid algorithm, and combining the Hilditch thinning algorithm with the path length function. Redundancy pruning is performed to output the core outline skeleton representing the rigid structure of the slope. ; Skeleton feature matching and image registration module (4): Used to perform step S4, in the skeleton buffer band Under constraints, SuperPoint and LightGlue neural networks are used to extract and match sub-pixel feature points, and the optimal homography matrix is ​​solved. To output registered images and geometric transformation parameters; Slope change detection and risk warning module (5): Used to execute step S5, inputting the aligned image into the ChangeFormer change detection network, extracting temporal difference features, and combining the skeleton topological deformation field to determine the pixel-level change intensity. Mapped to specific semantic tags for slope hazards; Multidimensional information visualization module (6): used to overlay heat map to display original image, semantic change probability map, vector outline of abnormal area and corresponding slope safety monitoring report.

[0052] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0054] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0057] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0058] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for cross-temporal image registration and change detection of slopes based on skeleton constraints, characterized in that, Includes the following steps: S1: Dual-temporal sensing data acquisition and image consistency processing; Acquire images and geolocation information of the same slope scene at different time phases, and perform distortion correction and radiometric calibration on the images to be registered to eliminate geometric and radiometric deviations caused by non-environmental factors; S2: Seasonal dynamic texture masking and stable region extraction based on semantic segmentation; By using a semantic segmentation model to identify and remove dynamic texture feature regions in images that change with the seasons, a stable region mask that remains topologically stable across time phases can be locked from a complex background. S3: Multi-scale skeleton extraction and spatiotemporal coarse registration; Extract single-pixel-width contour skeleton lines within the stable region mask, and perform skeleton geometric topological similarity analysis in conjunction with geolocation information to obtain the initial alignment parameters of images taken at two different times. S4: Pixel-level image fine registration driven by skeleton constraints; Using the contour skeleton as a geometric control net constraint, feature mapping and residual optimization are performed in the skeleton-guided search space to achieve sub-pixel-level spatial alignment of images captured at two different times. S5: Multi-feature fusion for slope temporal change detection and risk warning; The aligned image is input into the change detection network, feature offsets are extracted, and pixel-level changes are mapped to specific slope disaster semantic labels by combining the skeleton topology. The deformation features of abnormal areas are statistically analyzed and input into the time series prediction model, outputting slope stability scores and executing graded early warning logic.

2. The method for cross-temporal image registration and change detection of slopes based on skeleton constraints according to claim 1, characterized in that, Step S1 specifically involves: Step S11: Acquisition of UAV imagery and geolocation information; First, it is necessary to obtain the original two-temporal images of the slope to be monitored and its geographic location information, including: (1) Visual perception information recording: Recording the raw tensor data and meta-attributes of the image, including: dual-temporal three-channel graphic tensor The bit depth and exposure parameters during acquisition; where H and W represent the image pixel height and pixel width respectively, and rgb is the pixel value of each pixel in the red, yellow and blue channels; (2) Spatial geographic reference information recording: Synchronously extract the geographic location information at the time of image acquisition, including: longitude, latitude and absolute flight altitude. ; used to narrow down the initial search domain for skeleton matching in step S3 using GPS clustering; Step S12: Unification processing of UAV images; Then, the image scale and brightness differences caused by the two cross-temporal and spatial acquisitions by the drone are uniformized, which includes the following steps: (1) Scale alignment: The ground resolution difference between the two temporal images is calculated based on the flight altitude. The second image is resampled using the first image as a reference. The two temporal images are resampled to the same spatial resolution through bilinear interpolation to ensure that the two temporal images are consistent on the physical scale of a unit pixel. (2) Illumination compensation: To address the illumination differences at different times, a histogram specification algorithm is used to perform color and brightness mapping on dual-time phase images to eliminate radiation intensity deviations caused by seasonal solar altitude angles and cloud cover. (3) Image plane geometric flattening correction: Using the pre-calibrated camera distortion coefficients, radial and tangential distortion elimination is performed on the image, and a bilinear interpolation algorithm is used for grayscale reconstruction, outputting a standard image to be registered with geometric and color distribution alignment where the image plane geometric residual is controlled at the sub-pixel level.

3. The method for cross-temporal image registration and change detection of slopes based on skeleton constraints according to claim 1, characterized in that, Step S2 specifically involves: Step S21: Construct the semantic segmentation encoding structure; The dual-temporal normalized image output by S1 Input a pre-trained DeepLabv3+ semantic segmentation network, utilize its dilated spatial convolutional pooling (ASPP) module to capture multi-scale contextual features, and output each pixel of the image. Class probability vector There are 10 rigid stable region categories, and each dimension represents the confidence level of a pixel belonging to the corresponding category; Step S22: Calculate the regional stability score; Then calculate each pixel. Confidence score of stable region It is defined as the sum of the probabilities of rigidly stable classes: By setting a confidence threshold Filtering satisfies The pixels are used to construct an initial stable region distribution map; Step S23: Generate a stable region mask Morphological closing operations are performed on the initial stable regions. First, a 3×3 rectangular structuring element is used to perform a dilation-erosion closing operation on the initial mask. The dilation process fills texture gaps, and the erosion process smooths edge burrs. Then, the connected area of ​​all stable regions in the image is calculated, and small connected regions that do not have global constraint significance are filtered out by area thresholding. Finally, a rigid stable region mask spanning the time phase is locked. , which serves as the computational domain for skeleton extraction in step S3.

4. The method for cross-temporal image registration and change detection of slopes based on skeleton constraints according to claim 3, characterized in that, In step S2, the attribute classification of regional stability is based on the geometric deformation characteristics of various semantic features across time and space dimensions, and the semantic categories are defined by attribute classification: (1) Rigid stable category area: Bedrock, highway slope, drainage ditch and artificial protective structure are defined as rigid stable category. These features have visual edge consistency across time phases and are the core carriers of skeleton constraints. The 10 categories used to calculate confidence are as follows: {Outcropped bedrock, concrete facing, masonry, drainage ditch, intercepting ditch, prestressed anchor frame beam, curbstone, crash barrier, retaining wall, maintenance walkway}; The edge consistency refers to the fact that the connectivity and geometric shape of the edges of ground features maintain a mapping relationship in the binary space under different time phases. (2) Unstable interference area: Vegetation, sky and dynamic shadows are defined as unstable categories. These features exhibit drastic texture fluctuations with the season and lighting conditions and do not participate in the subsequent skeleton constraint construction. The unstable interference area also includes occlusion edges that change with the shooting angle.

5. The method for cross-temporal image registration and change detection of slopes based on skeleton constraints according to claim 3, characterized in that, In step S2, the weight space is defined as a probability distribution matrix aligned with the pixel scale of the input image; for the 10 classes of pixels identified as rigid stable regions, their corresponding weight space values ​​are set to... For pixels identified as unstable interference regions, their corresponding weight space values ​​are set to 0. The final weight space is used as an "attention map" and is multiplied pixel-by-pixel with the original image. The resulting weighted gradient field is used as the direct input for skeleton extraction in the subsequent S3 step.

6. The method for cross-temporal image registration and change detection of slopes based on skeleton constraints according to claim 1, characterized in that, Step S3 specifically involves: Step S31: Multi-scale skeleton extraction; The stable region mask output from step S2 is obtained using the pyramid algorithm. Perform multi-scale decomposition to construct a sequence of image pyramids from coarse to fine. ; These represent skeleton lines at different scales extracted using the method described in S2 when the image size is scaled to different magnifications. Edge detection operators are executed at each resolution layer, and a hierarchical skeleton network with spatial continuity is generated by cross-scale feature fusion to capture the rigid topology of the slope in different dimensions. Step S32: Hilditch skeleton refinement; In the hierarchical skeleton network, the Hilditch thinning algorithm is used to iteratively peel off the skeleton lines; for pixels within the mask... By determining the connectivity of its 8-neighborhood and the number of 0-1 transformations, it is thinned into a fine skeleton with a single pixel width while maintaining the original topological connectivity. Step S33: Skeleton filtering based on path length function; Assuming there are a total of n skeletons, { }, Then define each skeleton. Path length function : in skeleton The number of pixels; Set the length threshold to The endpoints and bifurcation points in the fine skeleton are traversed by length: If the path length of a certain skeleton branch ,but The stable rigid framework was determined to be retained; If the path length of a certain skeleton branch ,but False edges caused by seasonal texture remnants are identified and recursive pruning operators are executed. For isolated and short branches in the single-pixel skeleton network, repeated pruning is performed by iteratively scanning the endpoints until the length of all branches in the network is not less than [a certain value]. ; Finally, all eligible skeletons were integrated to identify the core rigid skeleton representing the long-term stability structure of the slope. Step S34: Combine image geolocation information with search space constraints; After extracting the skeleton Subsequently, coarse registration is performed based on the geometric topology information and GPS information of the skeleton, specifically including the following steps: Using the latitude, longitude, and altitude geolocation information of the images recorded in S1, the spatial Euclidean distance between the center points of the images acquired at two different times is calculated. By using spatial clustering analysis to identify overlapping regions, the initial search domain for skeleton matching is limited to the neighborhood determined by the location information. Step S35: Coarse registration based on Fourier shape descriptors; The contour skeleton The spatial coordinate sequence is transformed into a complex field representation. First, the skeleton is extracted. The coordinates of a series of consecutive points are represented as ( , ), ( , ), …., ( , ); Then combine the two-dimensional coordinates of each point into a complex number. , where i is the imaginary unit, which simplifies the two-dimensional planar motion coordinates into a one-dimensional complex sequence; Performing a Fourier transform on the complex sequence s(k) converts it from the space-time domain to the frequency domain, decomposing the spatial signal into shape components of different frequencies, and obtaining the Fourier coefficients at different frequency exponents. : Low-frequency coefficients (n<5): represent the overall shape of the object; High-frequency coefficients (n≥5): represent the detailed texture of the object; Discard DC component The final graphical feature vector F is defined as: Using the method described above, calculate the values ​​of the two images respectively. Corresponding graphical feature vector ; Then, the cosine similarity between the skeleton feature vectors of the two images is calculated: The calculated cosine similarity value S ranges from [0, 1]. Set a preset threshold T = 0.85, when At that time, it was initially determined that the two had a shape correlation; The nearest neighbor distance ratio (NNDR) principle is adopted, which means that if the ratio of the maximum similarity to the second maximum similarity exceeds a preset safety margin of 0.8 among all images that may point to the same slope, the physical features represented by the feature vector pair are finally locked as the same slope structure, thus completing the initial closure of the cross-temporal and spatial images.

7. The method for cross-temporal image registration and change detection of slopes based on skeleton constraints according to claim 1, characterized in that, Step S4 specifically involves, Step S41: Construct the skeleton adaptive constraint band; First, the core rigid frame is locked by S3. Based on this, a width of [missing information] is constructed using the dilation operator. Pixel-based strip search constraint band ,in Adaptive adjustments are made based on the image resolution and the initial alignment residual. Step S42: Skeleton feature extraction based on Superpoint; Then, a pre-trained SuperPoint convolutional neural network is used to perform a strip search on the constraint band. Feature perception is performed on the intersection region with the stable region described in step S2, and feature point pixel coordinates with sub-pixel precision and corresponding local high-dimensional descriptors are extracted simultaneously. ; Step S43: LightGlue attention-driven feature matching; The feature point pairs of the two-temporal images are input into the LightGlue deep matching network; The algorithm utilizes its multi-layer head attention mechanism to calculate the correlation strength between feature points in two time phases, and combines it with an inference adaptive stopping mechanism to dynamically eliminate low-confidence matches in order to improve matching efficiency while ensuring accuracy. Finally, it outputs the matching point pair index and matching confidence score based on the matching results, which is the correspondence between specific pixels in the two images and the reliability score of each matching pair. When the matching score is >0.5, the matching pair is determined to be a valid matching point pair. Step S44: Projection transformation and image registration; For the initial set of matching pairs, the RANSAC algorithm is used to perform a consistency check; Calculate the reprojection error of the matching point pairs under the fundamental matrix, and remove outliers that do not conform to the epipolar geometry constraints, i.e., the positional relationship between the two points does not conform to the geometric logic of camera imaging. Set matching success criteria: when the number of valid matching pairs... If the average confidence level is higher than the preset threshold, the current region registration and matching is confirmed to be successful; otherwise, return to step S3 to readjust the skeleton similarity search parameters. Based on the set of n high-quality feature matching points after successful confirmation ,in As a reference image point, Image points to be registered; Establish a homogeneous linear equation system using direct linear transformation Calculate the homography matrix : The least squares criterion is used to minimize the reprojection cost function. Iterative optimization of matrix coefficients is performed to eliminate the interference of outliers on alignment accuracy: Using the solved optimal homography matrix The coordinates of each pixel in the image to be registered Perform a perspective transformation mapping to obtain its corresponding position in the reference image coordinate system. : That is, the pixel mapping relationship satisfies: Step S45: Registration accuracy evaluation and quality control; Bilinear interpolation is used to resample the transformed non-integer pixel coordinates to obtain the pixel grayscale values ​​of the aligned image; the transformed skeleton is then calculated. With reference skeleton The spatial residual between them ensures that the edge alignment accuracy reaches the sub-pixel level; After meeting the accuracy requirements, output the registered spatiotemporally aligned image. This serves as the benchmark for detecting abnormal changes in step S5.

8. The method for cross-temporal image registration and change detection of slopes based on skeleton constraints according to claim 1, characterized in that, Step S5 specifically involves: The registered dual-temporal images output by S4 are paired Inputting the pre-trained ChangeFormer change detection network, the system utilizes a two-stream structure to extract multi-scale spatial feature differences and generate a normalized change confidence heatmap. ; By utilizing the Transformer structure in the ChangeFormer network, global long-range dependencies between two temporal images are captured, improving the completeness of identifying anomalous change regions. Using the unstable interference region mask generated in step S2, the heatmap is... Perform pixel-by-pixel zeroing to eliminate false change signals caused by vegetation decay and light and shadow migration, ensuring that the detection target is focused on the bedrock, retaining wall and main road area; The core rigid skeleton extracted in step S3 Using geometric reference, calculate and Coordinate offset vector of skeleton pixels under time phase ; By constructing a skeleton topological deformation field, the overall displacement trend of key slope structures is quantified to determine whether structural sliding exists. Combined with pixel-level change intensity The skeleton is a vector A multi-criteria decision-making mechanism is used to map the detection results into specific semantic tags for slope hazards: (1) Landslide / collapse: manifested as a sudden change in confidence of large-area connectivity within a stable region, accompanied by severe fracture or displacement of the skeletal structure; (2) Weathering and erosion: manifested as small, scattered clusters of changes within the bedrock area, with the main skeleton maintaining topological consistency; (3) Damage to ancillary structures: manifested as local morphological distortion of specific skeleton lines of drainage ditches and crash barriers; Finally, based on the equivalent area of ​​the changed region Combined with the deformation rate factor, the disaster risk index is calculated, and a slope safety monitoring report is generated, which is marked with the disaster type, location and degree of damage.

Images