A water-blocking obstacle dynamic monitoring system based on remote sensing stereoscopic imaging

Abstract

The application discloses a kind of water-blocking obstacles dynamic monitoring system based on remote sensing stereoscopic imaging, specifically relates to computer vision and remote sensing monitoring technical field, including multi-source remote sensing data acquisition module, remote sensing data adaptive preprocessing module, remote sensing stereoscopic imaging modeling module, water-blocking obstacles intelligent identification module, water-blocking obstacles dynamic monitoring module and dynamic monitoring risk assessment and feedback module;The multi-source remote sensing data acquisition module is obtained by remote sensing data acquisition module, and the multi-source stereoscopic remote sensing data of monitoring water area is obtained according to preset monitoring period, and the first multi-source remote sensing data set is obtained;The application is changed by water-blocking obstacles dynamic monitoring module, and through space-time registration, multidimensional quantification and hierarchical early warning link, realizes automatic determination of change type, improves displacement monitoring precision;While hierarchical early warning mechanism improves early warning pertinence, provides direct basis for disposal decision, is helpful to receive early warning and make corresponding disposal measures.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and remote sensing monitoring technology, specifically to a dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging. Background Technology

[0002] Water-blocking obstacles (such as construction waste, illegal embankments, etc.) are significant hidden dangers that reduce the flood discharge capacity of water bodies and cause floods. Accurate and dynamic monitoring of these obstacles is a core requirement for the safety management of water conservancy projects. Remote sensing stereo imaging, through multi-view data acquisition and three-dimensional reconstruction technology, has achieved high-precision positioning, three-dimensional morphological analysis, dynamic change tracking, and efficient data collection in the dynamic monitoring of water-blocking obstacles, providing key technical support for river management, flood control and disaster reduction, and engineering planning.

[0003] Existing methods for dynamic monitoring of water-blocking obstacles include: first, traditional manual inspection, where staff directly observe the shape, location, and impact on water flow to make judgments and records; second, monitoring methods based on single remote sensing images, which identify obstacles through orthophotos; and third, existing stereoscopic mapping monitoring technologies, which can acquire three-dimensional data of water-blocking obstacles, quantify their displacement, deformation, or erosion, and achieve dynamic monitoring.

[0004] Current water-blocking obstacle monitoring technologies still have some limitations. For example, manual inspections are greatly restricted by terrain, have blind spots, and are inefficient and costly. Two-dimensional images cannot accurately obtain the three-dimensional dimensions of obstacles, resulting in large deviations in water-blocking risk assessment, and have low accuracy in identifying obstacles due to blurred boundaries. Existing stereoscopic mapping monitoring technologies have problems such as poor multi-source data fusion and lack of dynamic change analysis capabilities, making it difficult to achieve real-time dynamic tracking and monitoring of obstacles. Therefore, existing technologies generally suffer from technical defects such as limited monitoring range, insufficient accuracy in obtaining three-dimensional information, lack of dynamic monitoring capabilities, and poor adaptability to complex environments. There is an urgent need for a water-blocking obstacle monitoring solution that can achieve large-scale, high-precision, and dynamic monitoring. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging, comprising: Multi-source remote sensing data acquisition module: Through the remote sensing data acquisition module, multi-source stereo remote sensing data of the monitored water area are acquired according to the preset monitoring cycle to obtain the first multi-source remote sensing dataset; Remote sensing data adaptive preprocessing module: Through data preprocessing technology, the first multi-source remote sensing dataset is preprocessed to obtain the second multi-source remote sensing dataset after the data quality verification is passed; Remote sensing stereo imaging modeling module: Based on the second multi-source remote sensing dataset, the pixel displacement of corresponding points is obtained through an improved matching and fitting function, and the accuracy requirement is judged. Based on the judgment, the initial elevation is obtained through the result. Based on the generated orthophoto and the initial elevation, a digital surface model is generated, and finally, the global orthophoto and the global digital surface model are obtained. The intelligent water-blocking obstacle recognition module includes a water datum surface generation unit, an obstacle candidate area detection unit, and a three-dimensional feature fusion and extraction unit. Based on orthophotos and digital surface models, it identifies water-blocking obstacles through a deep learning segmentation model and outputs the intelligent recognition results of water-blocking obstacles. Dynamic monitoring module for water-blocking obstacles: Based on the intelligent identification results of water-blocking obstacles, it performs three-dimensional entity time-series data alignment, dynamically monitors and quantifies the change type, and issues early warning level information to the management terminal for human-computer interaction based on the quantified abnormal results; Dynamic monitoring risk assessment and feedback module: Combines obstacle extraction parameters and water datum data to calculate water resistance coefficient and classify risk levels, and feeds back abnormal risk warning information to the management terminal for human-computer interaction.

[0007] Preferably, the initial elevation includes: B1: inputting the epipolar image I_epi and grayscale correction coefficient γ(x,y) from the second multi-source remote sensing dataset into the grayscale correction type covariance matching function, where (p,q) are the horizontal and vertical displacement parameters of the image, to obtain the matching value R(p,q) corresponding to the image displacement; B2: Within the preset search interval, select discrete combinations of (p, q) one by one according to the preset pixel step size to obtain corresponding R(p, q). If the largest R(p0, q0) is less than the set threshold, the region is determined to be a match failure, and the elevation mean interpolation of adjacent windows is used to fill the gap. At the same time, if the pixel step size is greater than or equal to the required accuracy, the combination (p0, q0) with the largest R(p0, q0) and greater than or equal to the corresponding threshold is the pixel displacement of the corresponding point. If the pixel step size is less than the required accuracy, sub-pixel interpolation is performed. Taking the horizontal disparity p as the center, one point is taken on both the left and right sides according to the pixel step size, resulting in three sets of points p and corresponding R(p). Then, based on these three sets of data, a quadratic interpolation function R(p) = ap is constructed. 2 +bp+c, and solve for the coefficients a, b, and c. Differentiate R(p) with respect to 0 to obtain the extreme point of R(p), which is the optimal disparity p0. Then, convert the horizontal disparity p0 of the corresponding points into the initial elevation Z. ini ; I_epi imagery of satellite and UAV epipolar lines respectively sat and I_epi uavPerform the B1-B2 matching operation to obtain the initial elevation Z of the satellite and the UAV. ini,sat and Z ini,uav .

[0008] Preferably, the orthophoto and digital surface model: B3: Convert the optical image pixel coordinates (x,y) into geodetic coordinates (X,Y,Z) including elevation using an RPC function; then convert the geodetic coordinates (X,Y,Z) into the orthophoto pixel coordinates (x1,y1) using the projection function proj(); finally, iterate through each pixel coordinate using the resampling function Resample() to obtain the initial satellite and UAV orthophoto I_DOM respectively. sat,in and I_DOM uav,in ; B4: Based on the planar pixel coordinates (x, y) of the initial orthophoto, according to the pre-divided windows, the overlap rate of the windows must meet the corresponding threshold. For the position coordinates (x, y) of all pixels within each window, and combined with radar data I_rad qw The corresponding elevation value Z_rad qw Initial elevation value Z ini The elevation values ​​Z of the DSM (Digital Surface Model) are generated using a radar-constrained moving surface fitting function, along with a terrain slope correction factor θ. Accuracy verification is then performed based on the requirement that the absolute difference ΔZ ≤ the corresponding threshold for all DSM elevation values ​​corresponding to ground control points. The elevation values ​​Z obtained from all windows are then stitched together to obtain a DSM elevation grid covering the entire area, thus outputting a satellite and UAV digital surface model (DSM) that meets the required accuracy. sat and DSM uav By combining the optical image pixel coordinates (x, y) and the DSM elevation Z... DSM A second orthorectification is performed to obtain the optimized orthorectified images I_DOM of the satellite and UAV. sat,opt and I_DOM uav,opt .

[0009] Preferably, the global orthophoto image and the global digital surface model: B5: First, the projection transformation function is called to convert the original coordinate system src_crs of the satellite and UAV imagery into the target coordinate system tgt_crs; then, the SIFT algorithm is used to extract the imagery from I_DOM. sat,opt and I_DOM uav,opt Feature points for registration are extracted, and their coordinates and descriptors are output. Descriptors are matched using Euclidean distance, and mismatched pairs are eliminated using the RANSAC algorithm to obtain valid matching pairs (Match_pairs). The affine transformation matrix M is then calculated based on Match_pairs, and I_DOM is... uav,opt DSM uav The aligned I_DOM is obtained through the M-transformation.uav,al DSM uav,al For non-overlapping spatial regions: directly stitch together satellite high-precision product sets and UAV high-precision product sets to retain single-source data; for spatially overlapping regions, use satellite and UAV DSM elevation accuracy ΔZ. sat and ΔZ uav Calculate the fusion weights w for satellites and drones sat and w uav The DOM and DSM fusion functions are used respectively. fu and DSM fu This yields the global orthophoto image I_DOM and the global digital surface model DSM.

[0010] Preferably, the water datum generation unit: collects the average value μ_Z and the standard deviation σ_Z of historical water level data from multiple time periods, and combines the confidence coefficient α and the seasonal correction function f_sea(t), where t is the monitoring time, to generate a water datum model Z_ba(x,y), which represents the digital water level datum elevation of the target water area (x,y).

[0011] Preferably, the obstacle candidate region detection unit: performs pixel grayscale value standardization processing on the input orthophoto I_DOM, inputs the standardized full I_DOM image into the pre-trained deep learning segmentation model, outputs a pixel-level prediction map, sets a confidence threshold, marks the pixel regions with prediction probabilities ≥ the threshold as suspected water-blocking obstacle candidate regions, generates a raster format candidate region map, and determines the position coordinates and confidence of suspected water-blocking obstacles on the DOM.

[0012] Preferably, the three-dimensional feature fusion extraction unit first combines the location coordinates (x, y) of the suspected water-blocking obstacle with the elevation Z at the (x, y) location in the DSM. DSM Align (x,y) with the water level datum elevation Z_ba(x,y); then set the elevation threshold Z. th , will Z DSM The difference between (x,y) and Z_ba(x,y) is ≥ Z th The target is marked as a water-blocking obstacle; then, a morphological rule constraint function SS is constructed using the planar area S_obj of the candidate water-blocking obstacle region, the minimum area threshold S_min, and the aspect ratio AR of the suspected region, to filter regions whose shapes conform to the fixed obstacle characteristics; then, the difference between the DSM elevation and the reference elevation at each (x,y) position in the suspected region is accumulated, and then multiplied by the ground area pixel_size corresponding to the DOM pixel. 2 The actual volume V_ini is obtained; finally, the intelligent identification results of the water-blocking obstacle are obtained, including the location, boundary, elevation and volume of the water-blocking obstacle.

[0013] Preferably, the dynamic monitoring module for water-blocking obstacles includes: C1: collecting the point cloud coordinates (x, y, z), corresponding DSM, and monitoring area coordinates of the three-dimensional entities of obstacles at different time phases t1, t2, ..., tn; based on the point cloud registration method ICP, integrating the Euclidean distance attenuation weight and time interval weight w_t of the matching point pairs of N point clouds into the error function EE; then, when EE is less than a set threshold, obtaining the optimal rotation matrix R and translation vector T, and applying it to the point cloud P_i of the target period to obtain the aligned target period obstacle point cloud PQ_i; C2: Compare point cloud PQ_i with reference period point cloud Q_i, and calculate volume change ΔV, area change ΔS, and spatial displacement change ΔL respectively; when S_in(S_t2,S_t1) / S_t2 < area first threshold S_th1 and volume V_t2 at period t2 ≥ minimum effective volume threshold V_min, the obstacle change type is new; when S_in(S_t2,S_t1) / S_t1 < S_th1 and volume V_t2 at period t2 < V_min, the obstacle change type is demolition; when ΔV > 0 and S_in(S_t2,S_t1) / S_t1 ≥ area second threshold S_th2, the obstacle change type is expansion; when ΔV < 0 and S_in(S_t2,S_t1) / S_t2 ≥ S_th2, the obstacle change type is contraction; S_in(S_t2,S_t1) is the intersection of the planar areas of the two time phase obstacles; C3: When the absolute value of the volume change |ΔV| ≥ the first volume warning threshold ΔV_th1 or the absolute value of the spatial displacement change |ΔL| ≥ the first area warning threshold ΔL_th1, a red level one warning message is issued. When the second volume warning threshold ΔV_th2 ≤ |ΔV| < ΔV_th1 or the second area warning threshold ΔL_th2 ≤ |ΔL| < ΔL_th1, an orange level two warning message is issued. When |ΔV| < V_th2 or |ΔL| < ΔL_th2, no warning is issued, ΔV_th1 > ΔV_th2, ΔL_th1 > ΔL_th2. The warning information includes the warning level, trigger threshold, change location, change type, quantitative indicators, three-dimensional change comparison chart, and handling measures for different warning levels.

[0014] The technical effects and advantages of this invention are as follows: 1. This invention improves the accuracy of matching corresponding points in environments with water surface reflection and fog by using an improved covariance function, filtering, and environmental correction coefficients; at the same time, the radar-constrained fitting function combined with the reference surface correction improves the accuracy of DSM, and the water blocking coefficient is introduced into the relative height of the reference surface to reduce risk assessment error, thus solving the defect of not being able to quantify water blocking risk. 2. The intelligent water-blocking obstacle recognition module of this invention, through the generation of water datum surface, candidate area detection and three-dimensional fusion extraction link, and through adaptive preprocessing and improved deep learning segmentation model, effectively copes with complex environments such as water surface reflection, fog and vegetation obstruction; at the same time, combined with height threshold and morphological rules, it filters out moving targets such as ships and floating objects, reduces the false detection rate and improves the recognition accuracy of water-blocking obstacles. 3. This invention utilizes a dynamic monitoring module for water-blocking obstacles, employing spatiotemporal registration, multi-dimensional quantification, and a hierarchical early warning system to automatically determine the type of change, reduce volume quantification errors, and improve displacement monitoring accuracy. Simultaneously, the hierarchical early warning mechanism enhances the targeting of early warnings, providing direct evidence for decision-making in conjunction with early warning information, thus facilitating the early receipt of warnings and the implementation of corresponding measures. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0016] Figure 2 This is a flowchart illustrating the remote sensing stereo imaging modeling module of the present invention.

[0017] Figure 3 This is a flowchart illustrating the intelligent water-blocking obstacle recognition module of the present invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1 As shown, the present invention provides a dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging, including a multi-source remote sensing data acquisition module, a remote sensing data adaptive preprocessing module, a remote sensing stereo imaging modeling module, a water-blocking obstacle intelligent identification module, a water-blocking obstacle dynamic monitoring module, and a dynamic monitoring risk assessment and feedback module. The multi-source remote sensing data acquisition module is connected to the remote sensing data adaptive preprocessing module, the remote sensing stereo imaging modeling module is connected to the remote sensing data adaptive preprocessing module and the water-blocking obstacle intelligent identification module, and the water-blocking obstacle dynamic monitoring module is connected to the water-blocking obstacle intelligent identification module and the dynamic monitoring risk assessment and feedback module.

[0020] Multi-source remote sensing data acquisition module: Through the remote sensing data acquisition module, multi-source stereo remote sensing data of the monitored water area are acquired according to the preset monitoring cycle to obtain the first multi-source remote sensing dataset; This embodiment requires specific explanation regarding multi-source stereo remote sensing data, including but not limited to image data from high-resolution optical stereo mapping satellites (such as WorldView and Gaofen-7), UAV oblique photogrammetry systems, and radar data (using Interferometric Synthetic Aperture Radar (InSAR) data, acquired using radar satellites such as Sentinel-1 and RADARSAT-2); the optical satellite stereo images employ forward-looking, backward-looking, and front-looking three-line array stereo image pairs; the ground resolution (GSD) of the UAV oblique photogrammetry images is no greater than the corresponding threshold (e.g., 5 cm / pixel); and the forward overlap is not... Less than 80%, with a lateral overlap of not less than 75%; synchronously record image location information (GPS / IMU data), with the flight altitude H calculated inversely using the formula GSD=(f×H) / S×k, where f is the focal length of the UAV camera (an inherent parameter of the camera), S is the single pixel size of the camera sensor, and k is the environmental correction coefficient (generally taken as 0.9 to 1.1); simultaneously import the boundary vector of the monitored water area, GPU memory capacity, and environmental parameters (fog, rainfall); and use the GPU memory (such as the video memory of NVIDIA Tesla series GPUs) of the hardware (industrial-grade server or high-performance computer) used for data processing in the system backend.

[0021] The remote sensing data adaptive preprocessing module: This module preprocesses the first multi-source remote sensing dataset using data preprocessing techniques to obtain a second multi-source remote sensing dataset that has passed data quality verification, including: A1: The first multi-source remote sensing dataset is classified into a main dataset {I_sat,I_uav,I_rad} and an auxiliary parameter set {V_bnd,M_gpu,P_env}, which includes satellite imagery I_sat, UAV imagery I_uav, radar data I_rad, water boundary vector data V_bnd, GPU memory M_gpu, and environmental parameters P_env. In this embodiment, it should be specifically noted that the optical images include satellite images I_sat and UAV images I_uav.

[0022] A2: First, spatially match the water boundary vector data V_bnd (usually a polygon vector in shapefile format, containing the geographic coordinates of the water area and shoreline) with the currently processed tile (a rectangular area represented by geographic coordinates). If the tile and V_bnd have no intersection, determine that the water area pixel area A_wa=0 and the shoreline pixel area A_ba=0, and treat it as an invalid tile. If the tile and V_bnd have an intersection, convert the vector format V_bnd into raster data pi_co with the same pixel resolution as the current tile. Geo, Til, and GSD represent the geodetic coordinates (e.g., longitude / latitude) of each point in V_bnd, the geographic coordinates of the top-left corner of the current tile, and the ground resolution of the tile, respectively. To ensure rounding down, geographic coordinates are converted to integer pixel coordinates. Geo-Til is in length, and GSD is in length / pixel. After V_bnd rasterization, V_bnd will form pixel sets for the water area and shoreline area in the tile's pixel coordinate system. The number of pixels at the two sets is counted to obtain the water area pixel area A_wa and the shoreline pixel area A_ba. Finally, an invalid tile determination function R is constructed based on A_wa and A_ba. Invalid tiles are removed by setting R < the corresponding threshold (e.g., 0.2, which can be reduced to 0.15 in foggy / rainy weather). The tile size is then adjusted using the tile size adaptive function Size. ti The selected tiles are segmented to obtain the effective satellite and drone tiles, T_eff. sat and T_eff uav Invalid tile determination function R=[(A_wa+A_ba) / A_tot]×w env A_tot is the total area of ​​the tile in pixels, w env The environmental impact weight is determined by the environmental parameter P_env, such as 0.85 to 0.95 for foggy / rainy days and 1 for sunny days, used to strengthen the removal of invalid data under harsh environments. Size ti M_gpu is the tile side length (typically 1024-2048 pixels), M_gpu is the GPU memory capacity (unit converted to bytes), 3 is the number of channels in the RGB image, and 2 is the number of bytes per pixel. A3: First, calculate the grayscale value I_opt of the satellite and drone imagery at (x,y). sat (x,y) and I_opt uav Radiometric correction is performed on (x,y) to obtain the corrected optical image's grayscale value I_opt1(x,y) at (x,y), where I_opt1(x,y) = G × I_opt(x,y) + B, G is the gain coefficient provided by the satellite / UAV sensor calibration parameters, and B is the bias coefficient, also provided by the sensor calibration parameters; then the radiometrically corrected grayscale value I_opt1(x,y) is... cal (x,y) is atmospherically corrected using conventional atmospheric correction methods (such as the 6S atmospheric correction model) to obtain the atmospherically corrected grayscale values ​​I_opt2 for the satellite and UAV. sat (x,y) and I_opt2 uav (x,y), and finally iterate through all pixel coordinates to obtain the complete satellite and UAV image correction data I_caL sat and I_caL uavFinally, based on the RPC file of the stereo image pair, the exterior orientation elements are calculated. Based on the exterior orientation elements, the slope 'a' and intercept 'b' corresponding to each epipolar line are calculated using the photogrammetric epipolar geometry, ensuring that the epipolar lines are parallel to the x-axis (i.e., a=0, at which point the epipolar line equation simplifies to y=b, and corresponding points differ only in the x-direction). The atmospherically corrected optical stereo image pair is then resampled along the epipolar lines, and pixels located on the same epipolar line in the original image are rearranged into the same row of the epipolar line image (the y-coordinate is fixed at b), ultimately yielding the satellite and UAV epipolar line image I_epi. sat and I_epi uav ; This embodiment needs to be specifically explained because the mainstream optical sensors of satellites / UAVs have a basically linear response, which can meet the preprocessing accuracy requirements and has higher computational efficiency; RPC files refer to rational polynomial coefficient files. For optical satellite imagery, RPC files are distributed by satellite operators along with the satellite imagery and are standard supporting data for satellite imagery; for UAV oblique photography imagery, RPC files can be automatically generated from the UAV's POS data (shooting position and attitude recorded by GPS / IMU) + camera parameters (focal length, pixel size) and are the standard output of UAV image preprocessing.

[0023] A4: The grayscale value I_rad at (x,y) of the radar data is denoised using an improved Lee filter denoising function to obtain the denoised radar data I_rad. q I_rad q =I_rad(x,y)×(1-C)+μ×C×exp(-σ 2 / σ0 2 C is the gray-level variation coefficient of the pixels within the filtering window (the ratio of the standard deviation to the mean of the pixels within the current window), μ is the average gray-level value of the pixels within the filtering window, and σ is the average gray-level value of the pixels within the filtering window. 2 and σ0 2 These represent the grayscale variance and baseline variance of the pixels, respectively. These values ​​are set according to the radar band, for example, 100 for C-band radar, to control the strong noise suppression effect of the exponential term. The algorithm iterates through all pixels in the image to generate complete denoised radar data I_rad. qw The signal-to-noise ratio (SNR) is calculated based on the average gray values ​​of the signal and noise regions in the radar data. The SNR must satisfy a threshold (e.g., 15 dB). Finally, the second multi-source remote sensing dataset, after passing data quality verification, is obtained, including effective satellite and UAV tiles T_eff, optical image correction data I_caL, epipolar image I_epi, and denoised radar data I_rad. qw ; Please see Figure 2As shown, the remote sensing stereo imaging modeling module: Based on the second multi-source remote sensing dataset, it obtains the pixel displacement of corresponding points and judges the accuracy requirements through an improved matching and fitting function. Based on the judgment, it obtains the initial elevation. Based on the generated orthophoto and the initial elevation, it generates a digital surface model. Finally, it obtains a global orthophoto and a global digital surface model, including: B1: Input the epipolar image I_epi and grayscale correction coefficient γ(x,y) from the second multi-source remote sensing dataset into the grayscale correction covariance matching function, where (p,q) are the horizontal and vertical displacement parameters of the image, and obtain the matching value R(p,q) corresponding to the image displacement. The grayscale correction coefficient γ(x,y) = 1 - λ × |g(x,y) - g avg (x,y)| / 255, γ1(x,y)=1-λ×|g1(x+p,y+q)-g1 avg (x,y)| / 255, g(x,y) and g1(x+p,y+q) are the gray values ​​at (x,y) and (x+p,y+q) in the epipolar image I_epi, respectively, λ is the correction intensity (generally taken as 0.3~0.5, the gray uniformity λ of satellite image is smaller than that of UAV image which is easily affected by reflection), g avg (x,y) and g1 avg (x, y) represent the average grayscale values ​​of the image within the adaptive matching window D (generally a core area of ​​7×7 pixels, with the edge area extended to 9×9), where (x, y) ∈ D; the grayscale correction type covariance matching function... ; B2: Within the preset search interval, according to the preset pixel step size (e.g., a step size of 1 pixel), select discrete combinations of (p, q) one by one to obtain the corresponding R(p, q). If the largest R(p0, q0) is less than the set threshold (e.g., 0.8), the matching of this region is determined to be unsuccessful, and the elevation mean interpolation of adjacent windows is used to fill the gap. At the same time, if the pixel step size is greater than or equal to the required accuracy, the combination (p0, q0) with the largest R(p0, q0) and greater than or equal to the corresponding threshold is the pixel displacement of the corresponding point. If the pixel step size is less than the required accuracy, sub-pixel interpolation is performed. Taking the horizontal disparity p as the center, one point is taken on the left and right sides according to the pixel step size, resulting in three sets of points p and corresponding R(p). Then, based on these three sets of data, a quadratic interpolation function R(p) = ap is constructed. 2 +bp+c, and solve for the coefficients a, b, and c. Differentiate R(p) with respect to 0 to obtain the extreme point of R(p), which is the optimal disparity p0. Then, convert the horizontal disparity p0 of the corresponding points into the initial elevation Z. ini Z ini=H-(B×f / (p0×GSD)), where H is the satellite orbital altitude or UAV flight altitude, GSD is the resolution of the epipolar image, B is the baseline length of the epipolar image pair (calculated from the exterior orientation element), and f is the focal length of the optical sensor (provided by the sensor parameters); I_epi is the epipolar image of both satellite and UAV. sat and I_epi uav Perform the B1-B2 matching operation to obtain the initial elevation Z of the satellite and the UAV. ini,sat and Z ini,uav ; In this embodiment, it is necessary to specifically explain that, based on the resolution and imaging conditions of the epipolar image, a search interval [p] for the preset horizontal parallax p is defined. min ,p max ], typically a search range of ±20 to 50 pixels and vertical parallax q is taken (vertical parallax has been eliminated in epipolar imagery, so q is usually set directly to 0); for example, UAV epipolar imagery (resolution 0.1m / pixel, flight altitude 100m), with gentle terrain: set p∈[-20,20]; satellite epipolar imagery (resolution 2m / pixel, orbital altitude 500km), with complex terrain: set p∈[-5,5]; a feasible scheme [p min ,p max The theoretical parallax p of the water surface is estimated using the following model. wa =(B×f) / (H wa ×GSD), H wa p is the distance from the sensor to the water surface. min =p wa -Δp, Δp=(B×f×H) max ) / (H wa 2 ×GSD), H max p is the estimated maximum height of the water-blocking obstacle. max =p wa +δ, where δ is a small positive value used to accommodate waves, errors, etc.; the core object of the second interpolation is horizontal parallax. Since the epipolar image has eliminated vertical parallax, q=0 does not require interpolation.

[0024] B3: The optical image pixel coordinates (x, y) are converted to geodetic coordinates (X, Y, Z) including elevation using an RPC function; then, the geodetic coordinates (X, Y, Z) are converted to orthorectified image pixel coordinates (x1, y1) using a projection function proj() (such as Gauss-Kruger projection, UTM projection, or other planar projection functions), eliminating image distortion caused by terrain undulations; finally, the resampling function Resample() iterates through each pixel coordinate to obtain the initial satellite and UAV orthorectified images I_DOM respectively. sat,in and I_DOM uav,in, I_dom=Resample(I_caL(x,y),(x1,y1)), where I_caL(x,y) is the pixel grayscale value of the optical image corresponding to (x1,y1); It should be specifically noted in this embodiment that the resampling function for satellite imagery typically uses bilinear interpolation, while that for UAVs typically uses bicubic interpolation.

[0025] B4: Based on the planar pixel coordinates (x, y) of the initial orthophoto, according to a pre-divided window (e.g., 31×31 pixels), the overlap rate of the window must meet the corresponding threshold (e.g., 50%), the position coordinates (x, y) of all pixels within each window, combined with radar data I_rad. qw The corresponding elevation value Z_rad qw Initial elevation value Z ini The elevation values ​​Z of the DSM are generated using a radar-constrained moving surface fitting function, along with a terrain slope correction factor θ (generally 0.8 when the slope is >30°, and 1 otherwise). Z = a0 + a1x + a2y + a3x 2 +a4xy+a5y 2 +ω×(Z_rad qw -Z ini )×θ, a0~a5 are surface fitting coefficients (solved by the least squares method), ω is the radar data weight; then calculate the DSM elevation Z. DSM Measured elevation Z of ground control point GCP The absolute difference ΔZ, based on the DSM elevation values ​​corresponding to the ground control points, must satisfy ΔZ ≤ the corresponding threshold (adjusted according to the terrain, e.g., 50cm in mountainous areas) for accuracy verification. The elevation values ​​Z calculated from all windows are then stitched together (using overlapping windows for smooth transition), ultimately resulting in a DSM elevation grid covering the entire area. This outputs a satellite and UAV digital land model (DSM) that meets the required accuracy. sat and DSM uav By combining the optical image pixel coordinates (x, y) and the DSM elevation Z... DSM A second orthorectification is performed to obtain the optimized orthorectified images I_DOM of the satellite and UAV. sat,opt and I_DOM uav,opt ; B5: First, call the projection transformation function (X,Y). trans =proj_trans((X,Y) orig The system transforms the original coordinate system src_crs (such as WGS84 latitude and longitude and local plane coordinates) of satellite and UAV imagery into the target coordinate system tgt_crs; then, the SIFT algorithm is used to extract the data from I_DOM. sat,opt and I_DOMuav,opt Feature points for registration are extracted, and the coordinates and descriptors of these feature points (e.g., shorelines, bridges, fixed structures) are output. Descriptors are matched using Euclidean distance, and mismatched pairs are eliminated using the RANSAC algorithm to obtain valid matching pairs (Match_pairs). The affine transformation matrix M is then calculated based on Match_pairs, and I_DOM is used to... uav,opt DSM uav The aligned I_DOM is obtained through the M-transformation. uav,al DSM uav,al To ensure a complete match with satellite data in spatially overlapping areas; for spatially non-overlapping areas: directly stitch together satellite high-precision product sets and UAV high-precision product sets to retain single-source data (satellite data covers a large area of ​​non-key regions, while UAV data covers key areas such as river cross-sections). For spatially overlapping areas, based on the DSM elevation accuracy ΔZ of satellite and UAV data... sat and ΔZ uav Calculate the fusion weights w for satellites and drones sat and w uav The DOM and DSM fusion functions are used respectively. fu and DSM fu The global orthophoto I_DOM and the global digital surface model DSM are obtained. This embodiment requires specific explanation that the high-precision product set of satellites and drones includes orthophotos and digital terrain models; w sat =ΔZ uav / (ΔZ sat +ΔZ uav ), w uav =ΔZ sat / (ΔZ sat +ΔZ uav ); DOM fu (x,y)=w sat ×I_DOM sat,opt (x,y)+w uav ×I_DOM uav,al (x,y), DSM fu (x,y)=w sat ×DSM sat (x,y)+w uav ×DSM uav,al .

[0026] Please see Figure 3 As shown, the intelligent water-blocking obstacle recognition module includes a water datum surface generation unit, an obstacle candidate area detection unit, and a three-dimensional feature fusion extraction unit. Based on orthophotos and digital surface models, it identifies water-blocking obstacles through a deep learning segmentation model and outputs the intelligent recognition results of water-blocking obstacles. The water datum generation unit collects the average value μ_Z of historical water level data from multiple time periods (e.g., based on hydrological data statistics from the past 5 years) and the standard deviation σ_Z of historical water level data. It also combines the confidence coefficient α (generally taken as 1.28 to 1.96, corresponding to a 90% to 97.5% confidence interval to ensure that the datum covers normal water level fluctuations) and the seasonal correction function f_sea(t), where t is the monitoring time (month), to generate the water datum model Z_ba(x,y), which represents the digital water level datum elevation at the target water area (x,y), and characterizes the normal water level elevation under unobstructed conditions. Z_ba(x,y) = μ_Z + α × σ_Z × f_sea(t). This embodiment requires a detailed explanation of the construction method of the seasonal correction function f_season(t): First, collect monthly average water level data for at least 5 years for the target water area; second, calculate the proportional coefficient of each month's water level relative to the annual average water level; finally, fit these proportional coefficients into a smooth periodic function (e.g., using a third-order Fourier series) or directly store them as a monthly lookup table to obtain f_season(t), for example, f_season(t)=1.0+0.1×sin(2π×(t-3) / 12), where the amplitude of 0.1 and the phase of 3 months can be adjusted by fitting actual data.

[0027] The obstacle candidate region detection unit performs pixel grayscale value standardization processing on the input orthophoto I_DOM (normalizing the pixel grayscale value to the [0,1] interval), inputs the standardized full I_DOM image into the pre-trained deep learning segmentation model, outputs a pixel-level prediction map, sets a confidence threshold (e.g., 0.7, or 0.65 if water surface reflection is severely reduced), marks the pixel regions with prediction probabilities ≥ the threshold as suspected water-blocking obstacle candidate regions, generates a raster format candidate region map, and determines the position coordinates and confidence of suspected water-blocking obstacles on the DOM; In this embodiment, it is important to note that the obstacle candidate region detection module employs an improved U-Net neural network model. This model embeds a squeeze-excitement (SE) attention module into the encoder part of the standard U-Net to enhance the ability to extract obstacle texture features. The model uses labeled samples containing various typical water-blocking obstacles such as construction waste and sand deposits. By learning from these multi-class samples, the model can extract common texture and contour features of various obstacles that are significantly different from the background water area / shoreline, thereby training the model to identify suspected areas. During training, a weighted sum of cross-entropy loss and Dice loss is used as the objective function. After training, the model can perform pixel-level segmentation of the input DOM and output a binary prediction map of the suspected obstacle region (1 represents the obstacle region, and 0 represents the background region).

[0028] This embodiment uses a medium-sized river as the monitoring object. The intelligent extraction module for water area and obstacles works as follows: 1. Based on hydrological data from the past 5 years (μ_Z=120.5m, σ_Z=0.8m), α=1.645 (95% confidence interval) is set, and the monitoring time is t=July (summer, f_season=1.1), generating Z_base(x,y)=120.5+1.645×0.8×1.1≈121.9m; ​​2. An improved U-Net model is used to detect potential obstacle candidate areas on the DOM, with a recall rate of 96.2%; 3. ΔZ=0.4m and S_min=1m are set. 2 AR_obj∈[0.2~5.0], fused with DSM and reference plane, 17 fixed obstacles were screened out (excluding 3 boats and 12 floating objects), and their coordinates, volume and other parameters were extracted; according to actual measurement, the obstacle extraction false detection rate of this system is only 2.8%, and a total of 17 obstacles were accurately identified with no missed detection, effectively improving the river flood control safety guarantee capability.

[0029] The three-dimensional feature fusion extraction unit first extracts the location coordinates (x, y) of the suspected water-blocking obstacle and the elevation Z at the (x, y) location in the DSM. DSM Align (x,y) with the water level datum elevation Z_ba(x,y); then set the elevation threshold Z. th (Adjust according to the height of the monitoring area, for example, 0.3 in shallow water and 0.5 in deep water), Z DSM The difference between (x,y) and Z_ba(x,y) is ≥ Z th The target is marked as a water-blocking obstacle; then, a morphological rule constraint function SS is constructed using the planar area S_obj of the candidate water-blocking obstacle region, the minimum area threshold S_min, and the aspect ratio AR of the suspected region, to filter regions whose shape conforms to the fixed obstacle characteristics; the formula of SS is: S_obj≥S_min and AR belongs to the aspect ratio of the water-blocking obstacle (for example, the aspect ratio of blocky building debris is between 0.2 and 5, calculated based on the boundary of the suspected region on the DOM using historical samples), and calculated using the DOM pixel size output by S2 (the ground size corresponding to the DOM pixel is pixel_size, the number of DOM pixels in the suspected region is counted and multiplied by pixel_size). 2 S_obj is obtained; then, the difference between the DSM elevation and the datum elevation at each (x,y) position within the suspected area is accumulated, and then multiplied by the ground area pixel_size corresponding to the DOM pixels. 2 The actual volume V_ini is obtained; finally, the intelligent identification results of the water-blocking obstacle are obtained, including the location, boundary, elevation and volume of the water-blocking obstacle. Dynamic monitoring module for water-blocking obstacles: Based on the intelligent identification results of water-blocking obstacles, it performs 3D entity time-series data alignment, dynamically monitors and quantifies change types, and issues early warning level information to the management terminal for human-computer interaction based on the quantified anomaly results, including: C1: Collect the point cloud coordinates (x, y, z) of the 3D entities of obstacles at different time phases t1, t2, ..., tn, along with the corresponding DSM and monitoring area coordinates. Based on the point cloud registration method ICP, integrate the Euclidean distance attenuation weight and time interval weight w_t of the matching point pairs of N point clouds into the error function EE, EE=(1 / N)×∑||P_i-(R×Q_i+T)|| 2 ×exp(-d_i / d0)×w_t, P_i and Q_i are the point cloud coordinates of the obstacle's 3D entity in the target period t2 and the reference period t1 (usually the moment when the obstacle's 3D entity data is first acquired), respectively. R is a 3×3 rotation matrix, T is a 3×1 translation vector, d_i and d0 are the Euclidean distance and reference distance of the i-th point, respectively. w_t gradually decreases with the increase of the time interval, for example, 1.0 is taken for time interval ≤ 7 days, 0.95 is taken for interval > 7 days, and further decreases when the time interval is longer. Then, when EE is less than the set threshold (usually taken as 0.05~0.1m), the optimal rotation matrix R and translation vector T are obtained and applied to the point cloud P_i of the target period to obtain the aligned obstacle point cloud PQ_i of the target period. C2: Compare point cloud PQ_i with point cloud Q_i at the reference time, and calculate the volume change ΔV, area change ΔS, and spatial displacement change ΔL respectively; when S_in(S_t2,S_t1) / S_t2 < area first threshold S_th1 (e.g., 0.1) and the volume V_t2 at time t2 ≥ minimum effective volume threshold V_min, the obstacle change type is new; when S_in(S_t2,S_t1) / S_t1 < S_th1 and t2 When the volume V_t2 < V_min, the obstacle change type is demolition; when ΔV > 0 and S_in(S_t2,S_t1) / S_t1 ≥ the second area threshold S_th2 (e.g., 0.5), the obstacle change type is expansion; when ΔV < 0 and S_in(S_t2,S_t1) / S_t2 ≥ S_th2, the obstacle change type is contraction; S_in(S_t2,S_t1) is the intersection of the planar areas of the obstacles in two time phases. In this embodiment, it should be specifically explained that the volume change ΔV and the area change ΔS are obtained from the difference between the volume / area at time t2 and time t1, and the spatial displacement change ΔL is obtained from the distance formula between two points in three-dimensional space at time t2 and time t1.

[0030] C3: When the absolute value of the volume change |ΔV| ≥ the first volume warning threshold ΔV_th1 or the absolute value of the spatial displacement change |ΔL| ≥ the first area warning threshold ΔL_th1, a red Level 1 warning is issued. When the second volume warning threshold ΔV_th2 ≤ |ΔV| < ΔV_th1 or the second area warning threshold ΔL_th2 ≤ |ΔL| < ΔL_th1, an orange Level 2 warning is issued. When |ΔV| < V_th2 or |ΔL| < ΔL_th2, no warning is issued; ΔV_th1 > ΔV_th2, ΔL_th1 > ΔL_th2. The warning information includes the warning level, trigger threshold, change location, change type, quantitative indicators, three-dimensional change comparison chart, and handling measures for different warning levels. For example, the handling measures for a red warning are emergency on-site verification and increased monitoring frequency, while the handling measures for an orange warning are routine on-site verification and normalized frequency. The handling measure for no warning is to follow the original monitoring cycle without additional on-site verification. The dynamic monitoring risk assessment and feedback module combines obstacle extraction parameters with water datum data to calculate the water resistance coefficient and classify risk levels, and feeds back abnormal risk warning information to the management terminal for human-computer interaction; the obstacle extraction parameters include the volume V_t2 and height H1 of the target period, and the projected area A of the obstacle on the cross-section. pr The monitoring cross-sectional area S, the average water depth H10 of the cross-section, and the net height of the obstacle relative to the reference plane (Z) are all monitored. DSM -Z_ba), construct the water-blocking coefficient quantification formula KK. When KK ≥ the first risk threshold KK_th1 (e.g., 0.3), it is classified as an abnormal high-risk level. When the second risk threshold KK_th2 (e.g., 0.1) ≤ KK < KK_th1, it is classified as an abnormal medium-risk level. When KK < KK_th2, it is low-risk, with no warning. The abnormal risk warning information is transmitted to the management terminal for human-computer interaction. The management terminal takes disposal measures according to the risk level. For example, the disposal measures for high risk are to immediately start clearing obstacles and review flood discharge capacity, etc. The disposal measures for medium risk are to formulate a time-limited rectification plan and conduct rectification acceptance, etc. The disposal measures for low risk are to follow the original monitoring cycle without additional on-site verification. The KK = b1×(V_t2 / (S×H10)) + b2×(H1 / H10) + b3×(A pr / S)+b4×((Z DSM -Z_ba) / H10), b1, b2, b3 and b4 are the corresponding weighting coefficients, for example b1=0.4, b2=0.2, b3=0.2 and b4=0.2.

[0031] In this embodiment, it should be specifically explained that the cross-sectional area refers to the area of ​​the cross section of the water body perpendicular to the direction of water flow (which can be understood as "the cross section that cuts through the water body"). The larger S is, the wider the flood passage of the cross section is, and the smaller the proportion of obstacles that hinder flood flow; conversely, the smaller the S is, the larger the proportion of obstacles. The average water depth of the cross section is obtained by averaging the actual water depths of multiple measuring points within the cross section.

[0032] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging, characterized in that: include: Multi-source remote sensing data acquisition module: Through the remote sensing data acquisition module, multi-source stereo remote sensing data of the monitored water area are acquired according to the preset monitoring cycle to obtain the first multi-source remote sensing dataset; Remote sensing data adaptive preprocessing module: Through data preprocessing technology, the first multi-source remote sensing dataset is preprocessed to obtain the second multi-source remote sensing dataset after the data quality verification is passed; Remote sensing stereo imaging modeling module: Based on the second multi-source remote sensing dataset, the pixel displacement of corresponding points is obtained through an improved matching and fitting function, and the accuracy requirement is judged. Based on the judgment, the initial elevation is obtained through the result. Based on the generated orthophoto and the initial elevation, a digital surface model is generated, and finally, the global orthophoto and the global digital surface model are obtained. The intelligent water-blocking obstacle recognition module includes a water datum surface generation unit, an obstacle candidate area detection unit, and a three-dimensional feature fusion and extraction unit. Based on orthophotos and digital surface models, it identifies water-blocking obstacles through a deep learning segmentation model and outputs the intelligent recognition results of water-blocking obstacles. Dynamic monitoring module for water-blocking obstacles: Based on the intelligent identification results of water-blocking obstacles, it performs three-dimensional entity time-series data alignment, dynamically monitors and quantifies the change type, and issues early warning level information to the management terminal for human-computer interaction based on the quantified abnormal results; Dynamic monitoring risk assessment and feedback module: Combines obstacle extraction parameters and water datum data to calculate water resistance coefficient and classify risk levels, and feeds back abnormal risk warning information to the management terminal for human-computer interaction.

2. The dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging according to claim 1, characterized in that: The initial elevation includes: B1: Inputting the epipolar image I_epi and grayscale correction coefficient γ(x,y) from the second multi-source remote sensing dataset into the grayscale correction type covariance matching function, where (p,q) are the horizontal and vertical displacement parameters of the image, and obtaining the matching value R(p,q) corresponding to the image displacement; B2: Within the preset search interval, select discrete combinations of (p, q) one by one according to the preset pixel step size to obtain corresponding R(p, q). If the largest R(p0, q0) is less than the set threshold, the region is determined to be a match failure, and the elevation mean interpolation of adjacent windows is used to fill the gap. At the same time, if the pixel step size is greater than or equal to the required accuracy, the combination (p0, q0) with the largest R(p0, q0) and greater than or equal to the corresponding threshold is the pixel displacement of the corresponding point. If the pixel step size is less than the required accuracy, sub-pixel interpolation is performed. Taking the horizontal disparity p as the center, one point is taken on both the left and right sides according to the pixel step size, resulting in three sets of points p and corresponding R(p). Then, based on these three sets of data, a quadratic interpolation function R(p) = ap is constructed. 2 +bp+c, and solve for the coefficients a, b, and c. Differentiate R(p) with respect to 0 to obtain the extreme point of R(p), which is the optimal disparity p0. Then, convert the horizontal disparity p0 of the corresponding points into the initial elevation Z. ini ; I_epi imagery of satellite and UAV epipolar lines respectively sat and I_epi uav Perform the B1-B2 matching operation to obtain the initial elevation Z of the satellite and the UAV. ini,sat and Z ini,uav .

3. The dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging according to claim 2, characterized in that: The orthophoto and digital land model: B3: Convert the optical image pixel coordinates (x,y) into geodetic coordinates (X,Y,Z) including elevation using an RPC function; then convert the geodetic coordinates (X,Y,Z) into the orthophoto pixel coordinates (x1,y1) using the projection function proj(); finally, iterate through each pixel coordinate using the resampling function Resample() to obtain the initial satellite and UAV orthophoto I_DOM respectively. sat,in and I_DOM uav,in ; B4: Based on the planar pixel coordinates (x, y) of the initial orthophoto, according to the pre-divided windows, the overlap rate of the windows must meet the corresponding threshold. For the position coordinates (x, y) of all pixels within each window, and combined with radar data I_rad qw The corresponding elevation value Z_rad qw Initial elevation value Z ini The elevation values ​​Z of the DSM (Digital Surface Model) are generated using a radar-constrained moving surface fitting function, along with a terrain slope correction factor θ. Accuracy verification is then performed based on the requirement that the absolute difference ΔZ ≤ the corresponding threshold for all DSM elevation values ​​corresponding to ground control points. The elevation values ​​Z obtained from all windows are then stitched together to obtain a DSM elevation grid covering the entire area, thus outputting a satellite and UAV digital surface model (DSM) that meets the required accuracy. sat and DSM uav By combining the optical image pixel coordinates (x, y) and the DSM elevation Z... DSM A second orthorectification is performed to obtain the optimized orthorectified images I_DOM of the satellite and UAV. sat,opt and I_DOM uav,opt .

4. The dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging according to claim 3, characterized in that: The global orthophoto image and global digital surface model: B5: First, the projection transformation function is called to convert the original coordinate system src_crs of the satellite and UAV imagery into the target coordinate system tgt_crs; then, the SIFT algorithm is used to extract the imagery from I_DOM. sat,opt and I_DOM uav,opt Feature points for registration are extracted, and their coordinates and descriptors are output. Descriptors are matched using Euclidean distance, and mismatched pairs are eliminated using the RANSAC algorithm to obtain valid matching pairs (Match_pairs). The affine transformation matrix M is then calculated based on Match_pairs, and I_DOM is... uav,opt DSM uav The aligned I_DOM is obtained through the M-transformation. uav,al DSM uav,al For non-overlapping spatial regions: directly stitch together satellite high-precision product sets and UAV high-precision product sets to retain single-source data; for spatially overlapping regions, use satellite and UAV DSM elevation accuracy ΔZ. sat and ΔZ uav Calculate the fusion weights w for satellites and drones sat and w uav The DOM and DSM fusion functions are used respectively. fu and DSM fu This yields the global orthophoto image I_DOM and the global digital surface model DSM.

5. The dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging according to claim 1, characterized in that: The water datum generation unit collects the average value μ_Z and the standard deviation σ_Z of historical water level data from multiple time periods. It also combines the confidence coefficient α and the seasonal correction function f_sea(t), where t is the monitoring time, to generate a water datum model Z_ba(x,y), which represents the digital water level datum elevation of the target water area at position (x,y).

6. The dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging according to claim 1, characterized in that: The obstacle candidate region detection unit performs pixel grayscale value standardization processing on the input orthophoto I_DOM, inputs the standardized full I_DOM image into the pre-trained deep learning segmentation model, outputs a pixel-level prediction map, sets a confidence threshold, marks the pixel regions with prediction probabilities ≥ the threshold as suspected water-blocking obstacle candidate regions, generates a raster format candidate region map, and determines the position coordinates and confidence of suspected water-blocking obstacles on the DOM.

7. A dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging according to claim 6, characterized in that: The three-dimensional feature fusion extraction unit first extracts the location coordinates (x, y) of the suspected water-blocking obstacle and the elevation Z at the (x, y) location in the DSM. DSM Align (x,y) with the water level datum elevation Z_ba(x,y); then set the elevation threshold Z. th , will Z DSM The difference between (x,y) and Z_ba(x,y) is ≥ Z th The target is marked as a water-blocking obstacle; then, a morphological rule constraint function SS is constructed using the planar area S_obj of the candidate water-blocking obstacle region, the minimum area threshold S_min, and the aspect ratio AR of the suspected region, to filter regions whose shapes conform to the fixed obstacle characteristics; then, the difference between the DSM elevation and the reference elevation at each (x,y) position in the suspected region is accumulated, and then multiplied by the ground area pixel_size corresponding to the DOM pixel. 2 The actual volume V_ini is obtained; finally, the intelligent identification results of the water-blocking obstacle are obtained, including the location, boundary, elevation and volume of the water-blocking obstacle.

8. The dynamic monitoring system for water-blocking obstacles based on remote sensing stereo imaging according to claim 1, characterized in that: The dynamic monitoring module for water-blocking obstacles includes: C1: collecting the point cloud coordinates (x, y, z) of the three-dimensional entities of obstacles at different time phases t1, t2, ..., tn, the corresponding DSM and monitoring area coordinates; based on the point cloud registration method ICP, integrating the Euclidean distance attenuation weight and time interval weight w_t of the matching point pairs of N point clouds into the error function EE; then, when EE is less than a set threshold, obtaining the optimal rotation matrix R and translation vector T, and applying it to the point cloud P_i of the target period to obtain the aligned target period obstacle point cloud PQ_i; C2: Compare point cloud PQ_i with reference period point cloud Q_i, and calculate volume change ΔV, area change ΔS, and spatial displacement change ΔL respectively; when S_in(S_t2,S_t1) / S_t2 < area first threshold S_th1 and volume V_t2 at period t2 ≥ minimum effective volume threshold V_min, the obstacle change type is new; when S_in(S_t2,S_t1) / S_t1 < S_th1 and volume V_t2 at period t2 < V_min, the obstacle change type is demolition; when ΔV > 0 and S_in(S_t2,S_t1) / S_t1 ≥ area second threshold S_th2, the obstacle change type is expansion; when ΔV < 0 and S_in(S_t2,S_t1) / S_t2 ≥ S_th2, the obstacle change type is contraction; S_in(S_t2,S_t1) is the intersection of the planar areas of the two time phase obstacles; C3: When the absolute value of the volume change |ΔV| ≥ the first volume warning threshold ΔV_th1 or the absolute value of the spatial displacement change |ΔL| ≥ the first area warning threshold ΔL_th1, a red level one warning message is issued. When the second volume warning threshold ΔV_th2 ≤ |ΔV| < ΔV_th1 or the second area warning threshold ΔL_th2 ≤ |ΔL| < ΔL_th1, an orange level two warning message is issued. When |ΔV| < V_th2 or |ΔL| < ΔL_th2, no warning is issued, ΔV_th1 > ΔV_th2, ΔL_th1 > ΔL_th2. The warning information includes the warning level, trigger threshold, change location, change type, quantitative indicators, three-dimensional change comparison chart, and handling measures for different warning levels.

Images