A multi-source data fusion collaborative emergency intelligent sensing and early warning method

Abstract

The application discloses a multi-source data fusion cooperative emergency intelligent sensing and early warning method, and relates to the technical field of computer processing. The method comprises the following steps: S01, acquiring land point cloud and images collected by a UAV, underwater topographic point cloud collected by an unmanned ship, and deformation monitoring data collected by a side slope radar of a target area, and arranging multi-source sensing data based on a time stamp; S02, respectively pre-processing the multi-source sensing data to obtain pre-processed land point cloud and pre-processed deformation data of pre-processed underwater point cloud in a unified coordinate system; S03, performing water-land data fusion based on the pre-processed land point cloud and the pre-processed underwater point cloud to generate a water-land integrated digital elevation model. The application realizes global high-precision seamless fusion of water-land-air multi-source heterogeneous data, automatic quantitative identification of disaster characteristics and integrated grading closed-loop decision of emergency response.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to a collaborative emergency intelligent sensing and early warning method that integrates multi-source data. Background Technology

[0002] Currently, emergency mapping and monitoring early warning methods for sudden geological disasters such as reservoir flood control and landslides mainly rely on single technologies and equipment such as UAV aerial surveying, ground slope radar, and underwater measurement. However, the existing technology system has significant shortcomings:

[0003] First, the data perception dimensions are fragmented. High-precision land topography, underwater topography and slope deformation data are usually collected and processed independently, lacking a unified spatiotemporal benchmark, which makes it impossible to form a three-dimensional monitoring view covering the entire space of "land-water-slope".

[0004] Second, the data fusion capability is weak. Point cloud and image data from different sources have large errors in coordinate system matching, feature registration, and seamless stitching. In particular, data discontinuity or model "jaggedness" is easily generated at the boundary between land and water, making it difficult to build a high-precision, continuous and consistent digital twin model, which directly affects the reliability of accurate reservoir capacity calculation and landslide boundary delineation.

[0005] Third, the level of intelligent early warning is insufficient. Key risk features such as dam cracks and minor slope deformations still rely heavily on manual identification and experience-based judgment, resulting in problems such as low efficiency, high missed detection rate, and delayed response. Furthermore, there is a lack of an integrated closed-loop technical process from multi-source data fusion to intelligent risk assessment and emergency auxiliary decision-making.

[0006] In summary, overcoming the bottlenecks in collaborative perception, accurate fusion, and intelligent early warning of multi-source heterogeneous data in emergency scenarios is a key technical issue for improving disaster emergency response capabilities. Summary of the Invention

[0007] To address the aforementioned technical problems, the present invention employs a collaborative emergency intelligent sensing and early warning method based on multi-source data fusion, the method comprising the following steps:

[0008] S01. Acquire land point clouds and images collected by UAVs, underwater terrain point clouds collected by UAVs, and deformation monitoring data collected by slope radar in the target area, and process them based on timestamps to obtain multi-source sensing data.

[0009] S02. The multi-source sensing data are preprocessed to obtain preprocessed land point cloud and preprocessed underwater point cloud deformation data in a unified coordinate system.

[0010] S03. Based on the preprocessed land point cloud and the preprocessed underwater point cloud, perform land-water data fusion. The land-water data fusion includes: coarse registration using the shoreline determined by dynamic water level as a feature, fine registration using coherent point drift algorithm, and distance-weighted smooth fusion in the land-water boundary buffer zone to generate a land-water integrated digital elevation model.

[0011] S04. The preprocessed deformation data is mapped to the integrated land and water digital elevation model through a self-supervised deep learning network to generate a three-dimensional dynamic twin scene that integrates slope deformation information.

[0012] S05. Based on the three-dimensional dynamic twin scene, perform intelligent disaster identification and risk quantification to obtain results. The results include at least identifying dam cracks and calculating their geometric parameters based on the YOLOv11 framework, calculating the current reservoir capacity and capacity loss rate based on the integrated land and water digital elevation model, and calculating the slope deformation displacement and displacement change rate based on the three-dimensional dynamic twin scene.

[0013] S06. Based on the results, the geometric parameters, storage capacity loss rate, and displacement change rate are compared with the corresponding preset early warning thresholds in real time to generate and output hierarchical collaborative emergency early warning information and auxiliary decision-making suggestions.

[0014] Preferably, the preprocessing in step S02 includes the following steps:

[0015] S21. A local fitting-based denoising method is adopted. For each target point in the point cloud, K neighboring points in its neighborhood are selected. The local plane or quadratic surface is fitted by the least squares method, and the vertical distance from each point in the neighborhood to the fitted surface is calculated.

[0016] S22. Using at least twice the average spacing of the point cloud as a distance threshold, if the vertical distance is greater than the distance threshold, the point is identified as a noise point and removed.

[0017] S23 refers to using at least three joint control points to calculate a seven-parameter transformation model and unify all point cloud data to the CGCS2000 coordinate system.

[0018] Preferably, the coarse registration step S03, which uses the shoreline determined by the dynamic water level as a feature, includes:

[0019] S31. Obtain the real-time monitoring water level elevation and elevation tolerance threshold, and extract the set of points whose elevation values ​​fall within the range of water level elevation ± tolerance threshold on the land digital elevation model, as the feature point set of the waterfront boundary line.

[0020] S32. Simultaneously, extract the set of contour points corresponding to the water level elevation from the underwater topographic data as auxiliary features;

[0021] S33. An improved iterative nearest point algorithm is adopted to assign a weight coefficient higher than that of ordinary points to the feature point set of the waterfront boundary line. The initial rotation matrix and initial translation vector are calculated using the weighted feature point set to eliminate the scale offset and rotation difference between the land point cloud and the underwater point cloud, so that the coarse registration accuracy is controlled within half a meter.

[0022] Preferably, the fine registration using the coherence point drift algorithm in step S03 includes:

[0023] S34. Treat the coarsely registered land point cloud as the centroid set of the Gaussian mixture model, and the underwater point cloud as the set of observation data points sampled from the Gaussian mixture model, and construct the probability density function.

[0024] S35. Iteratively estimate the standard deviation and rigid transformation parameters of the Gaussian mixture model using the expectation-maximization algorithm until the probability density function converges. The rigid transformation parameters include the final rotation matrix and the final translation vector.

[0025] The coherent point drift algorithm uses uniformly distributed components to model noise points and outliers.

[0026] Preferably, the step S04 of generating a three-dimensional dynamic twin scene that integrates slope deformation information includes:

[0027] S41. Extract terrain curvature feature points from the two-dimensional cloud map of the slope radar using a self-supervised deep learning network. The self-supervised deep learning network is trained in a contrastive learning manner and does not require manual annotation.

[0028] S42. Map the coordinates of the two-dimensional radar cloud image to the three-dimensional space of the integrated land and water digital elevation model through affine transformation to generate three-dimensional deformed heat map coordinates.

[0029] S43. The cumulative deformation and instantaneous deformation rate of each terrain grid point are obtained by interpolation using the transformed three-dimensional coordinates within the coordinates of the three-dimensional deformation heat map;

[0030] S44. Dynamically update the deformation color spectrum layer in the 3D dynamic twin scene.

[0031] Preferably, step S05, which involves identifying dam cracks and calculating their geometric parameters based on the YOLOv11 framework, includes:

[0032] S51. The YOLOv11 model was trained using the publicly available dam crack dataset Crack200 and self-built crack samples. The confidence threshold was set to ≥0.85 and the non-maximum suppression cross-union ratio threshold was set to 0.5.

[0033] S52. Input the high-definition images of the dam body obtained by the UAV during close-range flight into the trained YOLOv11 model, and output the bounding box and class probability of the crack target.

[0034] S53. The Zhang-Suen thinning algorithm is used to iteratively peel off edge pixels from the bounding box of the detected crack target and extract the crack center skeleton line with a single pixel width.

[0035] S54. Accumulate the length pixel by pixel along the skeleton line at the center of the crack to obtain the crack pixel length; calculate the crack width pixel number along the normal direction of the skeleton line, and take the maximum value as the crack pixel width.

[0036] S55. Combining the physical size and pixel size ratio of the checkerboard calibration patch preset on the dam surface, the crack pixel length and crack pixel width are converted into actual length and actual width, and the maximum crack length and maximum width values ​​are output.

[0037] Preferably, step S05, which involves calculating the current reservoir capacity and capacity loss rate based on the integrated land-water digital elevation model, includes:

[0038] Based on real-time monitoring of water level elevation, extract all grid cells whose elevation values ​​are lower than the water level elevation in the integrated land and water digital elevation model;

[0039] The current effective reservoir capacity is calculated by multiplying the area of ​​all grid cells by the water depth at the grid center point and summing the results.

[0040] Read the original design capacity curve of the reservoir and obtain the design capacity corresponding to the design water level through linear interpolation;

[0041] Calculate storage capacity loss = Designed storage capacity - Current effective storage capacity; Storage capacity loss rate = Storage capacity loss / Designed storage capacity × 100%.

[0042] When the reservoir capacity loss rate exceeds the preset siltation warning threshold, a siltation warning signal is generated.

[0043] Preferably, step S05, which involves calculating the slope deformation displacement and displacement rate based on a three-dimensional dynamic twin scene, includes:

[0044] In a three-dimensional dynamic twin scene, regular deformation monitoring grids are divided, and each grid point is associated with the deformation time series corresponding to the slope radar.

[0045] Calculate the displacement difference between the current moment and the previous moment to obtain the instantaneous displacement.

[0046] Calculate the cumulative displacement over the past N consecutive time intervals to obtain the cumulative displacement.

[0047] Calculate the rate of change of displacement per unit time, where the rate of change of displacement = (current displacement - previous displacement) / time interval;

[0048] When the cumulative displacement exceeds the deformation accumulation threshold or the displacement change rate exceeds the rate threshold, a landslide classification early warning is triggered.

[0049] Preferably, step S06, which involves generating and outputting tiered collaborative emergency early warning information and auxiliary decision-making suggestions, includes:

[0050] Level 1 warning: When the maximum width of the crack exceeds the first crack threshold or the displacement change rate exceeds the first rate threshold, a yellow warning is issued, and it is recommended to strengthen inspections;

[0051] Level 2 warning: When the maximum width of the crack exceeds the second crack threshold or the cumulative displacement exceeds the cumulative threshold, an orange warning is issued, and it is recommended to organize the evacuation of downstream personnel.

[0052] Level 3 warning: When the reservoir capacity loss rate exceeds the siltation threshold and is superimposed with an accelerating displacement trend, a red warning will be issued, and it is recommended to immediately initiate flood discharge and emergency rescue.

[0053] Preferably, between steps S02 and S03, the calculation of the preprocessed land point cloud density and underwater point cloud density is also included. If the point cloud density of a certain local area is lower than the preset coverage density threshold, a supplementary measurement command is triggered to control the UAV or unmanned vessel to conduct supplementary data collection in that area.

[0054] Following step S06, the generated early warning information and auxiliary decision-making suggestions are pushed to the emergency command terminal, and on-site handling feedback is received. The early warning threshold and monitoring frequency are dynamically adjusted based on the feedback information.

[0055] The present invention has at least the following beneficial effects:

[0056] 1. In the coarse registration stage, assigning high weights to shoreline feature points effectively suppresses interference from water ripples and vegetation noise. In the fine registration stage, the coherent point drift algorithm is used to transform point cloud registration into a probability estimation problem of Gaussian mixture model, which is naturally robust to noise and outliers. In the fusion stage, cosine smoothing and Laplacian smoothing are introduced to completely solve the data discontinuity and model "jagged" problems in the water-land interface area.

[0057] 2. This invention integrates multiple intelligent analysis functions within a unified 3D dynamic twin scenario, including automatic detection and quantification of dam cracks, slope deformation tracking, and dynamic reservoir capacity calculation. Specifically, the YOLOv11 crack identification model, combined with the Zhang-Suen refinement algorithm and calibration patch conversion, achieves fully automated, high-precision measurement of crack length and width. Meanwhile, slope radar data is upscaled to a 3D heatmap through self-supervised deep learning, enabling real-time calculation and visualization of deformation displacement without requiring a large number of labeled samples. Compared to traditional assessment methods that rely on manual experience and are fragmented and independent, this invention achieves an automated, integrated processing flow from multi-source data input to risk feature quantification output. It solves the problems of low efficiency, high false negative rate, and insufficient decision support associated with traditional methods, providing objective, accurate, and visualized scientific evidence for emergency command. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the seamless integration of land and water processing provided in Embodiment 1 of the present invention;

[0060] Figure 2 This is a schematic diagram of slope radar monitoring data provided in Embodiment 1 of the present invention;

[0061] Figure 3 This is a schematic diagram of slope radar data fusion provided in Embodiment 1 of the present invention;

[0062] Figure 4 This is a schematic diagram of the monitoring capacity calculation provided in Embodiment 1 of the present invention;

[0063] Figure 5 This is a schematic diagram of the results of AI dam crack identification provided in Embodiment 1 of the present invention;

[0064] Figure 6 This is a flowchart provided for Embodiment 1 of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0067] Example 1

[0068] This embodiment provides a collaborative emergency intelligent sensing and early warning method based on multi-source data fusion. The method includes the following steps: Figure 6 As shown:

[0069] S01. Acquire land point clouds and images collected by UAVs, underwater terrain point clouds collected by UAVs, and deformation monitoring data collected by slope radar in the target area, and process them based on timestamps to obtain multi-source sensing data.

[0070] Specifically, before data acquisition, the three main pieces of equipment—UAVs, unmanned surface vessels (USVs), and slope radars—are synchronized with high precision. This can be achieved using the timing module of the BeiDou Navigation Satellite System or a Network Time Protocol (NTP) server, ensuring that the time deviation of each system is controlled within milliseconds. During data acquisition, the UAVs perform aerial surveying tasks along a pre-set parallel strip route. Their onboard lidar and high-resolution cameras continuously collect land area point clouds and images, with each frame recorded with a precise GPS timestamp. The USVs conduct multi-beam underwater topographic scanning along a grid-like survey line network, with each depth sounding point also associated with a real-time time tag. The slope radar performs azimuth scanning of the landslide body at a fixed frequency (10 minutes / time for normal operations, 5 minutes / time for emergency situations), generating a two-dimensional radar cloud image for each scan and marking the start and end timestamps. All equipment transmits the timestamped raw data back to the data server at the ground command center in real time via 4G / 5G or microwave communication links. During data processing, the system uses a unified time base (such as whole seconds or whole minutes) as a reference window to extract data packets collected by each device within the same time window, and aligns and packages them according to time order.

[0071] S02. Preprocess the multi-source sensing data separately to obtain preprocessed land point cloud and preprocessed underwater point cloud preprocessed deformation data under a unified coordinate system.

[0072] Furthermore, the preprocessing in the above embodiments includes the following steps:

[0073] S21. A local fitting-based denoising method is adopted. For each target point in the point cloud, K neighboring points in its neighborhood are selected. The local plane or quadratic surface is fitted by the least squares method, and the vertical distance from each point in the neighborhood to the fitted surface is calculated.

[0074] S22. Use at least twice the average spacing of the point cloud as the distance threshold. If the vertical distance is greater than the distance threshold, the point is identified as a noise point and removed.

[0075] S23 refers to using at least three joint control points to calculate a seven-parameter transformation model and unify all point cloud data to the CGCS2000 coordinate system.

[0076] Specifically, for each target point in the point cloud, a KD-tree or octree spatial index is first used to quickly search for its K nearest neighbors (K is typically 20-50, depending on the point cloud density). Then, the least squares method is used to fit these neighbors to a local plane or quadratic surface. Taking plane fitting as an example, the fitting model can be expressed as:

[0077] (1);

[0078] in, , , , , These are the fitting parameters. For each point in the point cloud... The quadratic surface is fitted using the least squares method, with the goal of minimizing the error function.

[0079] (2);

[0080] in, This represents the number of points within the neighborhood. The perpendicular distance from each point within the neighborhood to the fitted surface is calculated. As shown in formula (7), this time we use twice the average spacing of the point cloud as the threshold. .like If the point is identified as a noise point, it will be removed.

[0081] (3);

[0082] It should be noted that at least three joint control points need to be set up between the land and water data to serve as coordinate transformation benchmarks. The transformation parameters are calculated using formula (4) to achieve accurate registration of the point cloud:

[0083] (4);

[0084] In the formula: These are the transformed coordinates; Coordinates before transformation; Scale factor; They are respectively Rotation angles along the three coordinate axes, where Around Matrix of axis rotation Around Matrix of axis rotation Around The matrices for axis rotation are shown in equations (5), (6), and (7); There are 3 translation parameters.

[0085] (5);

[0086] (6); (7);

[0087] To solve for these seven parameters, at least three non-collinear high-precision control points need to be established in both the target and source coordinate systems, and their coordinates in both systems need to be measured. An error equation is established using the least squares method, and the optimal seven parameters are calculated. Then, all point cloud coordinates in the source coordinate system are substituted into the above formula, and their coordinates in the CGCS2000 coordinate system are calculated point by point to achieve unification.

[0088] S03. Based on the preprocessed land point cloud and the preprocessed underwater point cloud, the land and water data fusion is performed. The land and water data fusion includes: coarse registration using the shoreline determined by the dynamic water level as a feature, fine registration using the coherent point drift algorithm, and distance-weighted smooth fusion in the land-water boundary buffer zone to generate a land-water integrated digital elevation model.

[0089] Furthermore, the coarse registration using the shoreline determined by dynamic water level as a feature in the above embodiments includes:

[0090] S31. Obtain the real-time monitoring water level elevation and elevation tolerance threshold, and extract the set of points whose elevation values ​​fall within the range of water level elevation ± tolerance threshold on the land digital elevation model, as the feature point set of the waterfront boundary line.

[0091] S32. Simultaneously, extract the set of contour points corresponding to the water level elevation from the underwater topographic data as auxiliary features;

[0092] S33. An improved iterative nearest point algorithm is adopted to assign a weight coefficient higher than that of ordinary points to the feature point set of the waterfront boundary line, and the initial rotation matrix and initial translation vector are calculated using the weighted feature point set.

[0093] Furthermore, the above embodiments perform fine registration using a coherence point drift algorithm, including:

[0094] S34. Treat the coarsely registered land point cloud as the centroid set of the Gaussian mixture model, and the underwater point cloud as the set of observation data points sampled from the Gaussian mixture model, and construct the probability density function.

[0095] S35. Iteratively estimate the standard deviation and rigid transformation parameters of the Gaussian mixture model using the expectation-maximization algorithm until the probability density function converges. The rigid transformation parameters include the final rotation matrix and the final translation vector.

[0096] Among them, the coherent point drift algorithm models noise points and outliers using uniformly distributed components.

[0097] Specifically, the shoreline is a crucial transitional zone connecting land and water, characterized by stable geometry and distinct spectral features, making it the preferred feature for land-water data registration. This scheme employs an automatic shoreline extraction method based on elevation thresholds. Using synchronously acquired real-time water level data, it accurately delineates the shoreline boundary on the land digital elevation model. Simultaneously, it extracts isobaths from underwater topographic data as auxiliary features to construct the correspondence between land and water features.

[0098] Formula for determining shoreline: (8);

[0099] In the formula: The set of shoreline features, i.e., the set of planar coordinates of the boundary between land and water. Let these be the planar coordinates of the ground points. The elevation value at this point is obtained by interpolation from the land area DEM. In order to monitor water level elevation in real time, The elevation tolerance threshold, taking into account water level fluctuations and DEM errors, is generally taken as 0.1-0.3 meters.

[0100] Furthermore, the coarse registration stage aims to achieve initial alignment between the land point cloud and the water point cloud, eliminating large-scale offsets and rotational differences between them. This scheme employs an improved Iterative Closest Point (ICP) algorithm, introducing a feature weighting mechanism. It prioritizes the use of reliable feature points such as shorelines to calculate transformation parameters, avoiding interference from water surface points (easily fluctuating) and vegetation points (unstable) on the registration results. This ensures that the coarse registration accuracy is controlled within half a meter, providing a good initial pose for subsequent fine registration.

[0101] (9);

[0102] The registration error function measures the overall difference between two sets of point clouds. A 3D rotation matrix, describing the pose difference between two sets of point clouds, is a 3×3 orthogonal matrix. A three-dimensional translation vector that describes the positional offset between two sets of point clouds. , The three-dimensional coordinates of the i-th point in the land point cloud. , The nearest point corresponding to pi in the water point cloud. For the nearest point index, The number of corresponding point pairs involved in the calculation.

[0103] The fine registration stage employs the Coherent Point Drift (CPD) algorithm, transforming the point cloud registration problem into a parameter estimation problem for a Gaussian mixture model. This algorithm treats land point clouds as the centroids of a Gaussian distribution and water point clouds as observation data sampled from these centroids. Through expectation-maximization (EM) iterative optimization, it solves for the optimal rigid transformation parameters (rotation and translation), achieving precise alignment between the two point sets. The CPD algorithm is robust to noise and outliers, making it suitable for handling situations where land and water point clouds have uneven density and large distribution differences.

[0104] Gaussian mixture probability model:

[0105] (10);

[0106] In the formula: Water point The probability of its occurrence, The number of centroids (i.e., Gaussian components) in land point clouds. Spatial dimension, in three-dimensional point cloud , The standard deviation of a Gaussian distribution controls the degree of "fatness" or "thinness" of the distribution. The three-dimensional coordinates of the j-th point in the water point cloud. The three-dimensional coordinates of the i-th centroid point in the land point cloud. A rigid transformation with parameter θ acts on the center of mass. The result is that θ is the transformation parameter to be determined, which includes rotation R and translation t.

[0107] rigid transformation expression:

[0108] (11);

[0109] In the formula: Rotation matrix, satisfying t is the translation vector.

[0110] S04. Map the preprocessed deformation data to the integrated land and water digital elevation model through a self-supervised deep learning network to generate a 3D dynamic twin scene that integrates slope deformation information. This includes the following steps (e.g.) Figure 1 (as shown)

[0111] S41. Topographic curvature feature points in the two-dimensional cloud map of the slope radar are extracted by a self-supervised deep learning network. The self-supervised deep learning network is trained by contrastive learning and does not require manual annotation.

[0112] S42. Map the coordinates of the two-dimensional radar cloud image to the three-dimensional space of the integrated land and water digital elevation model through affine transformation to generate three-dimensional deformed heat map coordinates.

[0113] S43. Within the coordinates of the three-dimensional deformation heatmap, the cumulative deformation and instantaneous deformation rate of each terrain grid point are obtained by interpolation using the transformed three-dimensional coordinates, and a deformation color spectrum layer is generated in the three-dimensional dynamic twin scene.

[0114] Specifically, after precise registration of land and water data, it is necessary to eliminate the "jagged" phenomenon at the boundary between the two areas and construct a visually continuous and measurement-consistent seamless digital elevation model. This employs a distance-weighted fusion strategy, dynamically adjusting the weights of land and water data within a buffer zone (typically 5-10 meters wide) based on the distance from points to the shoreline: UAV aerial survey data dominates near the land side, while unmanned surface vessel bathymetry data dominates near the water side, with a smooth transition in the intermediate area. After fusion, the Laplacian smoothing algorithm is used to eliminate local elevation abrupt changes, ensuring continuous slope and eliminating artificial traces.

[0115] Distance weighting function (cosine smoothing):

[0116] (12);

[0117] In the formula: —Weighting coefficients for land data, with values ​​ranging from 0 to 1; The vertical distance from the current point to the shoreline, in meters. ; The fusion buffer width is typically 5-10, and the weight of the water area data is... .

[0118] Elevation-weighted fusion:

[0119] (13);

[0120] In the formula: The merged elevation values Land area DEM interpolated elevation values, Elevation values ​​interpolated from underwater DEM.

[0121] S05. Based on the three-dimensional dynamic twin scene, perform intelligent disaster identification and risk quantification to obtain results. The results include at least the identification of dam cracks and calculation of their geometric parameters based on the YOLOv11 framework, the calculation of the current reservoir capacity and capacity loss rate based on the integrated land and water digital elevation model, and the calculation of slope deformation displacement and displacement change rate based on the three-dimensional dynamic twin scene.

[0122] Furthermore, in the above embodiments, the identification of dam cracks and calculation of their geometric parameters based on the YOLOv11 framework include (such as...) Figure 3 and Figure 4 ):

[0123] S51. The YOLOv11 model is trained using the publicly available dam crack dataset Crack200 and self-built crack samples to obtain a well-trained YOLOv11 model.

[0124] S52. Input the high-definition images of the dam body obtained by the UAV during close-range flight into the trained YOLOv11 model, and output the bounding box and class probability of the crack target.

[0125] S53. The Zhang-Suen thinning algorithm is used to iteratively peel off edge pixels from the bounding box of the detected crack target and extract the crack center skeleton line with a single pixel width.

[0126] S54. Accumulate the length pixel by pixel along the skeleton line at the center of the crack to obtain the crack pixel length; calculate the crack width pixel number along the normal direction of the skeleton line, and take the maximum value as the crack pixel width.

[0127] S55. Combining the physical size and pixel size ratio of the checkerboard calibration patch preset on the dam surface, the crack pixel length and crack pixel width are converted into actual length and actual width, and the maximum crack length and maximum width values ​​are output.

[0128] Specifically, firstly, during the model training phase, researchers need to construct a dataset containing thousands of labeled images. This includes both the publicly available Crack200 crack dataset (covering samples with different lighting, backgrounds, and crack morphologies) and self-built crack samples collected from the target reservoir dam (such as images of fine cracks of different widths, mesh cracks, and cracks with water stains). By performing transfer learning training on the YOLOv11 model and setting a confidence threshold (e.g., 0.85) and a non-maximum suppression cross-union ratio (CUI) threshold (e.g., 0.5), the model can accurately learn the texture, edge, and shape features of cracks, ultimately outputting a lightweight detection network that can quickly locate crack targets in actual dam images and provide class probabilities. For example, when a drone acquires a high-resolution image of the dam body with a resolution of 5472×3648 pixels by flying close to the dam surface at a distance of less than 10 meters, the trained YOLOv11 model will perform convolution and feature pyramid processing on the image in milliseconds, and output multiple bounding boxes containing crack targets and their confidence scores (such as "crack: 0.92"), thereby automatically selecting the approximate area of ​​each crack in the image.

[0129] After obtaining the bounding box of the crack, the system enters the crack skeleton extraction and geometric quantization stage. Since the bounding box output by YOLOv11 only provides a rough location of the crack and cannot directly calculate its length and width, the Zhang-Suen thinning algorithm is required. This algorithm is a parallel iterative edge stripping method. Its operating principle is as follows: for the binary crack image within the bounding box, the system scans pixel by pixel and judges the 8-neighborhood conditions of the current pixel. If the pixel is a crack foreground point and satisfies constraints such as "two endpoints, no interruption of connectivity, and no excessive erosion", it is marked as a deletable edge point. After multiple iterations (usually 10 to 30 times), until no pixels can be deleted, a crack center skeleton line with a width of one pixel is finally obtained. For example, for an arc-shaped crack with a width of about 5 mm and a length of about 1.2 m, the crack area in the original image may occupy hundreds of pixels. After processing by the Zhang-Suen algorithm, only the skeleton line with a width of 1 pixel along the crack direction is retained, clearly reflecting the extension trajectory of the crack.

[0130] Subsequently, the system accumulates the length pixel by pixel along the skeleton line—that is, counts the total number of pixels on the skeleton line, and then multiplies it by the ground resolution corresponding to each pixel (for example, when the drone flies at an altitude of 10 meters and the camera focal length is 35mm, the ground resolution corresponding to a single pixel is approximately 2mm) to obtain the actual length of the crack. For calculating the crack width, at each pixel on the skeleton line, the system searches outwards along its normal direction (perpendicular to the skeleton line) for crack edge pixels, records the number of foreground pixels of the crack in that direction, takes the maximum number of pixels in all normal directions as the crack pixel width, and then multiplies it by the ground resolution to convert it into the actual width.

[0131] For example, slope radar equipment can be installed in the geologically stable bedrock area on the opposite bank of the reservoir to build a non-contact, all-weather slope deformation monitoring network. The equipment adopts an azimuth scanning mode to fully cover landslide areas and high-risk slope areas. Under normal circumstances, a scan is completed every 10 minutes, and in emergency situations, the frequency can be increased to once every 5 minutes to capture minute displacements on the mountain surface in real time.

[0132] Let the two-dimensional radar cloud image be , Represents pixel coordinates Rotation transformation operator / rotation matrix corresponding to the position:

[0133] Image pixel horizontal (width) index

[0134] Image pixel vertical (height) index

[0135] Through self-supervised encoder network (in (These are network weight parameters), which are used as a whole to convert a two-dimensional radar cloud image into a three-dimensional deformable thermal field.

[0136]

[0137] in, For planar coordinates, The thermal field represents the deformation amplitude.

[0138] Self-supervised training of a network is achieved by minimizing the following loss function:

[0139]

[0140] Reconstruction loss : Ensure that information is not lost after the dimensional upgrade.

[0141]

[0142] in It is a radar cloud image reconstructed from three-dimensional coordinates through back projection. The width and height of the image are in pixels.

[0143] Smoothing Regularized Loss : Constraining the continuity of the deformation field.

[0144]

[0145] in For image gradient operators, This is the regularization weighting coefficient.

[0146] 2. Extraction of terrain curvature feature points (for coordinate registration)

[0147] For terrain surfaces Calculate its first-order partial derivatives , and second-order partial derivatives , ,

[0148] 1. Basic quantities of curved surfaces

[0149]

[0150] 2. Gaussian curvature and mean curvature

[0151]

[0152] 3. Principal curvature (extreme values ​​of curvature)

[0153]

[0154] 4. Feature point selection and setting curvature threshold Points that meet the following conditions are selected as matching feature points:

[0155] or

[0156] 3. Core Formulas for Affine Coordinate Transformation

[0157] Basic form

[0158] Heatmap coordinates output by the network Mapped to real terrain 3D coordinates

[0159]

[0160]

[0161] Real terrain 3D coordinates

[0162]

[0163] It is a 3x3 three-dimensional rotation orthogonal matrix

[0164]

[0165] Homogeneous coordinate complete matrix form

[0166]

[0167] Selection and translation matrix solution

[0168] By matching Using curvature feature points, the optimal rotation matrix is ​​solved using the least squares method. and

[0169]

[0170] 4. Formulas for displacement and rate of change of displacement

[0171] 1. Change in three-dimensional resultant displacement at a single point

[0172] Comparison with baseline time and the current moment Calculate the total displacement using the three-dimensional coordinates.

[0173]

[0174] 2. Rate of change of displacement

[0175]

[0176] in

[0177] The time difference between the two

[0178] 5. Deformation Visualization Early Warning Judgment Formula

[0179] Setting engineering safety displacement threshold and velocity-displacement threshold An early warning will be triggered when any of the following conditions are met.

[0180] Furthermore, the calculation of the current reservoir capacity and capacity loss rate based on the integrated land-water digital elevation model in the above embodiments includes (such as...) Figure 4 (as shown)

[0181] Based on real-time monitoring of water level elevation, all grid cells with elevation values ​​lower than the water level elevation are extracted from the integrated land and water digital elevation model.

[0182] The current effective reservoir capacity is calculated by multiplying the area of ​​all grid cells by the water depth at the grid center point and summing the results.

[0183] Read the original design capacity curve of the reservoir and obtain the design capacity corresponding to the design water level through linear interpolation;

[0184] Calculate storage capacity loss = Designed storage capacity - Current effective storage capacity; Storage capacity loss rate = Storage capacity loss / Designed storage capacity × 100%.

[0185] When the reservoir capacity loss rate exceeds a preset siltation warning threshold, a desilting early warning signal is generated.

[0186] Specifically, first, relying on the high-precision integrated water-land digital elevation model (DEM) generated in the previous steps, which has been unified to the CGCS2000 coordinate system and eliminates elevation abrupt changes at the water-land junction, it can truly reflect the continuous distribution of underwater terrain and land terrain in the reservoir area. In real-time operation, the system obtains the current reservoir water level elevation value Hwater through a water level gauge or remote sensing means, then performs a spatial query in the integrated water-land DEM to extract all grid cells with elevation values less than or equal to Hwater, and these cells constitute the complete reservoir area terrain below the current water surface. Since the DEM usually adopts a regular grid structure, each grid cell has a known plane area Agrid (for example, 1m×1m), and the elevation Zcell of the grid center point can be directly read from the DEM, so the current water depth hcell = Hwater - Zcell. Using the regular grid integration method, that is, accumulating the product of the water depth of all submerged grid cells and the cell area, the current effective reservoir capacity Vcurrent = Σ(Agrid × hcell) can be obtained. This principle is similar to the discrete summation of the volume of an irregular solid in calculus; the finer the grid division, the higher the calculation accuracy of the reservoir capacity.

[0187] In order to evaluate the siltation degree of the reservoir or the reservoir capacity loss situation, the system needs to compare the current effective reservoir capacity with the reservoir capacity under the original design working condition of the reservoir. The original design reservoir capacity curve of the reservoir is usually stored in the database in the form of a water level-reservoir capacity relationship table or a fitted curve. After obtaining the current water level Hwater, the system finds the design reservoir capacity Vdesign corresponding to this water level from the design reservoir capacity curve through a linear interpolation method. For example, if the known water level H1 on the design curve corresponds to the reservoir capacity V1, the water level H2 corresponds to the reservoir capacity V2, and H1 < Hwater < H2, then the design reservoir capacity Vdesign = V1 + (V2 - V1) × (Hwater - H1) / (H2 - H1). Then the reservoir capacity loss ΔV = Vdesign - Vcurrent is calculated. This difference reflects the reduction in reservoir capacity caused by sediment siltation, landslide mass intrusion or other reasons. Further calculate the reservoir capacity loss rate Rloss = (ΔV / Vdesign) × 100%. When Rloss exceeds the preset siltation warning threshold, the system automatically generates a desilting early warning signal and pushes it to the emergency management platform, prompting the management department to carry out reservoir area desilting or sediment flushing scheduling in time.

[0188] Alternatively, the reservoir capacity calculation includes the following: the effective reservoir capacity is the submerged volume between the water level surface and the terrain surface,

[0189]

[0190] where: Water level elevation (constant) :coordinate Ground elevation;

[0191] Discretization approximation divides the entire computational domain into equal-sized rectangular regular grids, each with a size of [missing information]. (Step size in the X direction)* (Step size in the Y direction), the area of ​​a single grid cell is: ;

[0192] Using the water depth at the center of the grid as an approximation of the average water depth across the entire grid, the continuous integral is transformed into a high-divergence summation (rectangular numerical integration method) to obtain an approximate reservoir capacity:

[0193] ;

[0194] in : No. The coordinates of the center point of each grid. No. Ground elevation of the center point of each grid;

[0195] when At that time, the grid is included in the effective storage capacity.

[0196] Furthermore, in the above embodiments, the calculation of slope deformation displacement and displacement rate based on a three-dimensional dynamic twin scene includes:

[0197] In a three-dimensional dynamic twin scene, regular deformation monitoring grids are divided, and each grid point is associated with the deformation time series corresponding to the slope radar.

[0198] Calculate the displacement difference between the current moment and the previous moment to obtain the instantaneous displacement.

[0199] Calculate the cumulative displacement over the past N consecutive time intervals to obtain the cumulative displacement.

[0200] Calculate the rate of change of displacement per unit time: Rate of change of displacement = (Current displacement - Displacement at the previous moment) / Time interval;

[0201] When the cumulative displacement exceeds the deformation accumulation threshold or the displacement change rate exceeds the rate threshold, a landslide classification warning is triggered.

[0202] Specifically, based on an integrated land and water digital elevation model, a regular deformation monitoring grid is divided at fixed intervals (e.g., 1m x 1m or 2m x 2m) in the target slope area (typically covering the landslide body and surrounding affected area). The three-dimensional coordinates (X, Y, Z) of each grid node are uniquely determined in the twin scene. Simultaneously, a self-supervised deep learning network is used to upscale and map the two-dimensional radar cloud images (each pixel representing radar wave reflection intensity, implicitly containing deformation phase information) obtained from slope radar scans at different times onto this three-dimensional grid. Specifically, using terrain curvature feature points as control references, an affine transformation is employed to establish the correspondence between radar image coordinates and three-dimensional spatial coordinates. Then, a deformation time series updated over time is generated for each grid node—that is, the cumulative displacement sequence of that node relative to the initial reference time at each historical monitoring moment. The sampling frequency of the sequence is synchronized with the scanning frequency of the slope radar: under normal conditions, a deformation value is recorded every 10 minutes; under emergency conditions, this can be increased to once every 5 minutes.

[0203] For any monitoring grid node, let its cumulative displacement (relative to the initial reference) at the current time t be S. t The cumulative displacement at the previous adjacent monitoring time t-Δt is S. t-Δt The difference in displacement between the current moment and the previous moment is the instantaneous displacement ΔS. t =S t -S t-Δt This instantaneous displacement reflects minute changes in the slope surface within two adjacent scanning intervals, typically reaching millimeter or even sub-millimeter levels. Simultaneously, the system maintains a cumulative displacement sequence for the past N consecutive time points (e.g., N=6, corresponding to 6 sampling points within 1 hour in standard mode), and directly reads the cumulative displacement S at the current time point. t This value represents the total deformation of the slope at that point from the start of monitoring to the current moment, and is used to determine the long-term cumulative trend.

[0204] The rate of change of displacement is defined as the instantaneous displacement divided by the corresponding time interval, i.e., v = ΔSt / Δt. Where Δt is the scanning period of the slope radar (600 seconds normally, 300 seconds in emergency situations). This rate directly characterizes the speed of slope deformation and is a key indicator for determining whether a landslide has entered the accelerated creep stage.

[0205] S06. Based on the results (e.g.) Figure 5 As shown in the figure, the geometric parameters, storage capacity loss rate, and displacement change rate are compared with the corresponding preset early warning thresholds in real time to generate and output hierarchical collaborative emergency early warning information and auxiliary decision-making suggestions.

[0206] The above embodiments generate and output hierarchical collaborative emergency early warning information and auxiliary decision-making suggestions, including:

[0207] Level 1 warning: When the maximum width of the crack exceeds the first crack threshold or the displacement change rate exceeds the first rate threshold, a yellow warning is issued, and it is recommended to strengthen inspections;

[0208] Level 2 warning: When the maximum width of the crack exceeds the second crack threshold or the cumulative displacement exceeds the cumulative threshold, an orange warning is issued, and it is recommended to organize the evacuation of downstream personnel.

[0209] Level 3 warning: When the reservoir capacity loss rate exceeds the siltation threshold and is superimposed with an accelerating displacement trend, a red warning will be issued, and it is recommended to immediately initiate flood discharge and emergency rescue.

[0210] Specifically, the system presets two core early warning parameters for each monitoring grid node (or integrates them by region): the cumulative deformation threshold T. cum (e.g., 20 mm) and rate threshold T rate (e.g., 2 mm / hour). The real-time monitoring logic is as follows:

[0211] Yellow alert (Level 1): When the rate of change of displacement v at any node > T rate But the cumulative displacement S t Not exceeding T cum When the system determines that the slope has entered an active deformation period, it issues a yellow warning and recommends increasing the frequency of monitoring and sending drones to conduct close-range verification.

[0212] Orange alert (Level 2): ​​When the cumulative displacement S t >T cum When the displacement change rate is still positive (continuous deformation), the system issues an orange warning, indicating an increased potential landslide risk, and recommends that people in downstream and danger zones prepare for evacuation.

[0213] Red Alert (Level 3): When the cumulative displacement, S... t >T cum If the displacement change rate v > 2 × Trate (i.e., the rate exceeds twice the threshold), or if the displacement change rate shows a monotonically increasing trend for three consecutive monitoring cycles (accelerated creep characteristics), the system will immediately issue a red alert, requiring the initiation of emergency rescue, forced evacuation, and closure of traffic in the reservoir area.

[0214] Example 2

[0215] Based on the above embodiment 1, this example further includes calculating the preprocessed land point cloud density and underwater point cloud density between steps S02 and S03. If the point cloud density of a certain local area is lower than the preset coverage density threshold, a supplementary measurement command is triggered to control the UAV or unmanned vessel to conduct supplementary data collection in that area.

[0216] Specifically, firstly, the preprocessed land and underwater point clouds are projected onto a unified regular grid. The grid size can be dynamically set according to the terrain complexity of the target area (e.g., 2m x 2m). For each grid cell, the number of point clouds falling within that cell is counted and divided by the grid area to obtain the actual point cloud density for that cell. The preset coverage density threshold is determined based on the accuracy requirements of the subsequent digital elevation model (DEM) construction. For example, when a DEM grid resolution better than 0.5 meters is required, the point cloud density threshold can be set to no less than 4 points / square meter. Next, all grid cells are traversed, and the actual density is compared with the threshold. If the density of a cell is lower than the threshold, the cell is marked as an "undersampled area." The system further analyzes the distribution characteristics of the undersampled areas—if the undersampled cells are concentrated in a continuous block (such as the shaded area of ​​a steep slope or the underwater vegetation obstruction area), a supplementary measurement task instruction is automatically generated. For undersampled land areas, the supplementary measurement command controls the UAV to follow the original planned route but reduce its flight altitude (e.g., from 100 meters to 50 meters) or increase the route overlap (heading overlap increased to 90%) for secondary data collection. For undersampled water areas, the supplementary measurement command controls the UAV to reduce the spacing between the main survey lines (e.g., from 50 meters to 25 meters) or use single-beam densification of survey lines for supplementary scanning. In terms of operating principle, this method is based on the data-driven closed-loop quality control concept: by calculating the point cloud density distribution map in real time, the actual collection effect is compared with the theoretical requirements, and the difference is used to drive the reoperation of the collection equipment, thereby avoiding the transmission of local data voids caused by factors such as terrain undulation, vegetation obstruction, and water turbidity to the subsequent fusion modeling stage.

[0217] After step S06, the generated early warning information and auxiliary decision-making suggestions are pushed to the emergency command terminal, and on-site handling feedback is received. The early warning threshold and monitoring frequency are dynamically adjusted based on the feedback information.

[0218] Specifically, firstly, the system compares the quantitative risk indicators obtained in step S05, such as crack geometric parameters, reservoir capacity loss rate, and slope displacement change rate, with preset multi-level early warning thresholds (e.g., yellow warning: crack width ≥ 2mm or displacement rate ≥ 0.5mm / h; orange warning: crack width ≥ 5mm or cumulative displacement ≥ 20mm; red warning: crack width ≥ 10mm and displacement acceleration trend lasting for more than 3 hours). It then automatically generates a structured early warning message containing the risk location, risk level, handling suggestions, and a timestamp. This message is pushed to the bound emergency command terminal (such as the command center screen, on-site mobile terminal, and handheld device for on-duty personnel) via 5G or BeiDou short message communication link. The push strategy adopts a "tiered broadcast" mode: Level 1 warnings are only pushed to on-site inspectors and technical personnel; Level 2 warnings are simultaneously pushed to the reservoir management unit and township flood control offices; and Level 3 warnings are directly reported to the county-level emergency command center and trigger an automatic voice call. Secondly, the system receives on-site handling feedback from terminals. This feedback can be submitted via a standardized form, including: whether the hazard has been confirmed, the actual handling measures taken (e.g., "downstream residents have been evacuated," "the floodgates have been opened"), the actual width of the cracks verified on-site (if the manual verification value deviates from the AI ​​recognition value by more than 30%, it is marked "recognition deviation"), and whether the warning sensitivity needs adjustment. The system stores the feedback information in a historical case database and triggers a dynamic parameter tuning module. This module uses a sliding window learning strategy to statistically analyze the "false alarm" and "missed alarm" events in the most recent N feedback reports. If the false alarm rate exceeds 5% or the missed alarm rate exceeds 1%, the system automatically adjusts the offset of the corresponding warning threshold. The adjustment rules are: if the false alarm rate is too high, the warning threshold is appropriately raised (e.g., the crack width warning value is increased from 2mm to 2.5mm); if the missed alarm rate is too high, the threshold is appropriately lowered. Meanwhile, for scenarios under emergency conditions (with a Level 3 warning triggered), the system automatically increases the frequency of slope radar monitoring from the usual 10 minutes / time to 5 minutes / time, and increases the frequency of drone patrols from once a day to once an hour, until on-site feedback confirms that the danger has been eliminated.

[0219] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A collaborative emergency intelligent sensing and early warning method based on multi-source data fusion, characterized in that, The method includes the following steps: S01. Acquire land point clouds and images collected by UAVs, underwater terrain point clouds collected by UAVs, and deformation monitoring data collected by slope radar in the target area, and process them based on timestamps to obtain multi-source sensing data. S02. The multi-source sensing data are preprocessed to obtain preprocessed land point cloud and preprocessed underwater point cloud deformation data in a unified coordinate system. S03. Based on the preprocessed land point cloud and the preprocessed underwater point cloud, perform land-water data fusion. The land-water data fusion includes: coarse registration using the shoreline determined by dynamic water level as a feature, fine registration using coherent point drift algorithm, and distance-weighted smooth fusion in the land-water boundary buffer zone to generate a land-water integrated digital elevation model. S04. The preprocessed deformation data is mapped to the integrated land and water digital elevation model through a self-supervised deep learning network to generate a three-dimensional dynamic twin scene that integrates slope deformation information. S05. Based on the three-dimensional dynamic twin scene, perform intelligent disaster identification and risk quantification to obtain results. The results include at least identifying dam cracks and calculating their geometric parameters based on the YOLOv11 framework, calculating the current reservoir capacity and capacity loss rate based on the integrated land and water digital elevation model, and calculating the slope deformation displacement and displacement change rate based on the three-dimensional dynamic twin scene. S06. Based on the results, the geometric parameters, storage capacity loss rate, and displacement change rate are compared with the corresponding preset early warning thresholds in real time to generate and output hierarchical collaborative emergency early warning information and auxiliary decision-making suggestions.

2. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 1, characterized in that, The preprocessing in step S02 includes the following steps: S21. A local fitting-based denoising method is adopted. For each target point in the point cloud, K neighboring points in its neighborhood are selected. The local plane or quadratic surface is fitted by the least squares method, and the vertical distance from each point in the neighborhood to the fitted surface is calculated. S22. Using at least twice the average spacing of the point cloud as a distance threshold, if the vertical distance is greater than the distance threshold, the point is identified as a noise point and removed. S23 refers to using at least three joint control points to calculate a seven-parameter transformation model and unify all point cloud data to the CGCS2000 coordinate system.

3. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 1, characterized in that, The coarse registration step S03, which uses the shoreline determined by the dynamic water level as a feature, includes: S31. Obtain the real-time monitoring water level elevation and elevation tolerance threshold, and extract the set of points whose elevation values ​​fall within the range of water level elevation ± tolerance threshold on the land digital elevation model, as the feature point set of the waterfront boundary line. S32. Simultaneously, extract the set of contour points corresponding to the water level elevation from the underwater topographic data as auxiliary features; S33. An improved iterative nearest point algorithm is used to assign a weight coefficient higher than that of ordinary points to the feature point set of the waterfront boundary line, and the initial rotation matrix and initial translation vector are calculated using the weighted feature point set.

4. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 3, characterized in that, The fine registration using the coherence point drift algorithm in step S03 includes: S34. Treat the coarsely registered land point cloud as the centroid set of the Gaussian mixture model, and the underwater point cloud as the set of observation data points sampled from the Gaussian mixture model, and construct the probability density function. S35. Iteratively estimate the standard deviation and rigid transformation parameters of the Gaussian mixture model using the expectation-maximization algorithm until the probability density function converges. The rigid transformation parameters include the final rotation matrix and the final translation vector. The coherent point drift algorithm uses uniformly distributed components to model noise points and outliers.

5. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 1, characterized in that, The step S04 of generating a three-dimensional dynamic twin scene that integrates slope deformation information includes: S41. Extract terrain curvature feature points from the two-dimensional cloud map of the slope radar using a self-supervised deep learning network. The self-supervised deep learning network is trained in a contrastive learning manner and does not require manual annotation. S42. Map the coordinates of the two-dimensional radar cloud image to the three-dimensional space of the integrated land and water digital elevation model through affine transformation to generate three-dimensional deformed heat map coordinates. S43. Within the coordinates of the three-dimensional deformation heatmap, the cumulative deformation and instantaneous deformation rate of each terrain grid point are obtained by interpolation using the transformed three-dimensional coordinates, and a deformation color spectrum layer in the three-dimensional dynamic twin scene is generated.

6. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 1, characterized in that, The step S05, which involves identifying dam cracks and calculating their geometric parameters based on the YOLOv11 framework, includes: S51. The YOLOv11 model is trained using the publicly available dam crack dataset Crack200 and self-built crack samples to obtain a well-trained YOLOv11 model. S52. Input the high-definition images of the dam body obtained by the UAV during close-range flight into the trained YOLOv11 model, and output the bounding box and class probability of the crack target. S53. The Zhang-Suen thinning algorithm is used to iteratively peel off edge pixels from the bounding box of the detected crack target and extract the crack center skeleton line with a single pixel width. S54. Accumulate the length pixel by pixel along the skeleton line at the center of the crack to obtain the crack pixel length; calculate the crack width pixel number along the normal direction of the skeleton line, and take the maximum value as the crack pixel width. S55. Combining the physical size and pixel size ratio of the checkerboard calibration patch preset on the dam surface, the crack pixel length and crack pixel width are converted into actual length and actual width, and the maximum crack length and maximum width values ​​are output.

7. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 6, characterized in that, The step S05, which involves calculating the current reservoir capacity and capacity loss rate based on the integrated land-water digital elevation model, includes: Based on real-time monitoring of water level elevation, extract all grid cells whose elevation values ​​are lower than the water level elevation in the integrated land and water digital elevation model; The current effective reservoir capacity is calculated by multiplying the area of ​​all grid cells by the water depth at the grid center point and summing the results. Read the original design capacity curve of the reservoir and obtain the design capacity corresponding to the design water level through linear interpolation; Calculate storage capacity loss = Designed storage capacity - Current effective storage capacity; Storage capacity loss rate = Storage capacity loss / Designed storage capacity × 100%. When the reservoir capacity loss rate exceeds the preset siltation warning threshold, a siltation warning signal is generated.

8. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 7, characterized in that, The step S05, which involves calculating the slope deformation displacement and displacement rate based on a three-dimensional dynamic twin scene, includes: In a three-dimensional dynamic twin scene, regular deformation monitoring grids are divided, and each grid point is associated with the deformation time series corresponding to the slope radar. Calculate the displacement difference between the current moment and the previous moment to obtain the instantaneous displacement. Calculate the cumulative displacement over the past N consecutive time intervals to obtain the cumulative displacement. Calculate the rate of change of displacement per unit time, where the rate of change of displacement = (current displacement - previous displacement) / time interval; When the cumulative displacement exceeds the deformation accumulation threshold or the displacement change rate exceeds the rate threshold, a landslide classification early warning is triggered.

9. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 1, characterized in that, The step S06 of generating and outputting hierarchical collaborative emergency early warning information and auxiliary decision-making suggestions includes: Level 1 warning: When the maximum width of the crack exceeds the first crack threshold or the displacement change rate exceeds the first rate threshold, a yellow warning is issued, and it is recommended to strengthen inspections; Level 2 warning: When the maximum width of the crack exceeds the second crack threshold or the cumulative displacement exceeds the cumulative threshold, an orange warning is issued, and it is recommended to organize the evacuation of downstream personnel. Level 3 warning: When the reservoir capacity loss rate exceeds the siltation threshold and is superimposed with an accelerating displacement trend, a red warning will be issued, and it is recommended to immediately initiate flood discharge and emergency rescue.

10. The collaborative emergency intelligent sensing and early warning method based on multi-source data fusion according to claim 1, characterized in that, Between steps S02 and S03, the calculation of the preprocessed land point cloud density and underwater point cloud density is also included. If the point cloud density of a certain local area is lower than the preset coverage density threshold, a supplementary measurement command is triggered to control the drone or unmanned vessel to conduct supplementary data collection in that area. After step S06, the generated early warning information and auxiliary decision-making suggestions are pushed to the emergency command terminal, and on-site handling feedback is received. The early warning threshold and monitoring frequency are dynamically adjusted based on the feedback information.

Images