Water level monitoring method of image segmentation model

Abstract

This application provides a water level monitoring method using an image segmentation model, comprising: acquiring multimodal flow data of a target water area; extracting features from the multimodal flow data to obtain a turbulence feature map, wherein the multimodal flow data includes flow motion data and water body image data; performing physical constraint segmentation processing on the turbulence feature map and the water body image data to obtain a water body region segmentation mask and a corresponding pixel-level confidence map; performing water level gauge geometric correction processing on the water body region segmentation mask to obtain a corrected water level line; and performing water level value conversion processing on the corrected water level line and the pixel-level confidence map to obtain the water level measurement value and uncertainty data. This method can effectively distinguish between the true water level line and dynamic flow artifacts, ensuring monitoring accuracy under complex hydrological conditions.

Description

Technical Field

[0001] This invention belongs to the field of water level monitoring technology, and in particular relates to a water level monitoring method using an image segmentation model. Background Technology

[0002] With the widespread application of image processing technology in water conservancy monitoring, computer vision-based water level monitoring methods are gradually replacing traditional contact sensor technologies. These non-contact solutions capture water surface images using cameras and automatically identify water level lines by combining edge detection or semantic segmentation algorithms, significantly reducing equipment deployment and maintenance costs. In traditional technologies, for highly dynamic water flow scenarios (such as spillways and turbulent rivers), classic edge detection algorithms such as the Canny operator or basic segmentation models are mainly used to process image data.

[0003] However, existing technologies face significant challenges: splashes and foam from high-speed flowing water create numerous false edge interferences in images. These dynamic interference features are highly similar in texture and shape to the real water level, making it difficult for algorithms to accurately distinguish them, ultimately resulting in the output of multiple broken or offset false water level lines. Especially under extreme hydrological conditions such as torrential rains and flood discharges, the misclassification rate of traditional segmentation models increases significantly, seriously affecting the reliability of monitoring data. Summary of the Invention

[0004] Therefore, it is necessary to provide a water level monitoring method based on an image segmentation model to address the aforementioned technical problems. This method can effectively distinguish between real water level lines and dynamic water flow artifacts through cross-domain technology fusion, ensuring monitoring accuracy under complex hydrological conditions.

[0005] In a first aspect, this application provides a water level monitoring method for an image segmentation model, comprising:

[0006] Multimodal flow data of the target water area is acquired, and features are extracted from the multimodal flow data to obtain a turbulence feature map. The multimodal flow data includes flow motion data and water body image data.

[0007] Physical constraint segmentation is performed on turbulence feature maps and water body image data to obtain water body region segmentation masks and corresponding pixel-level confidence maps.

[0008] The water body region segmentation mask is subjected to water level geometric correction processing to obtain the corrected water level line;

[0009] The water level values ​​are converted from the calibration water level line and the pixel-level confidence map to obtain the water level measurement value and uncertainty data.

[0010] In one embodiment, multimodal flow data of the target water area is acquired, and feature extraction is performed on the multimodal flow data to obtain a turbulence feature map, including:

[0011] The vorticity distribution characteristic map is obtained by performing curl calculation on the velocity vector field of the water flow data.

[0012] Gradient filtering is performed on the polarization gradient magnitude of the water body image data to obtain a polarization gradient map.

[0013] The vorticity distribution characteristic map is processed by exponential decay transformation to obtain the turbulence intensity coefficient map;

[0014] The turbulence intensity coefficient map and polarization gradient map are subjected to Hadamard product operation to obtain the turbulence characteristic map.

[0015] In one embodiment, physical constraint segmentation is performed on the turbulence feature map and water image data to obtain a water region segmentation mask and a corresponding pixel-level confidence map, including:

[0016] Deformable convolution feature extraction is performed on water body image data to obtain multi-scale texture feature maps;

[0017] The turbulence feature map is processed by a spatiotemporally gated recurrent network to output a motion feature map, which is used to characterize the motion pattern of dynamic artifacts.

[0018] Feature fusion processing under fluid continuity constraints is performed on multi-scale texture feature maps and motion feature maps to generate a fused feature field;

[0019] The fused feature field is subjected to confidence-weighted segmentation processing to output a water body region segmentation mask and a pixel-level confidence map.

[0020] In one embodiment, the water body region segmentation mask is subjected to water level geometric correction processing to obtain a corrected water level line, including:

[0021] The water body region segmentation mask is processed by maximum connected component contour extraction to obtain the initial boundary point set;

[0022] Based on the turbulence feature map, the initial boundary point set is subjected to motion stability screening to obtain the steady-state feature point set;

[0023] The steady-state feature point set is fitted with a non-uniform rational B-spline curve to generate a water level model.

[0024] The homography projection matrix is ​​calculated based on the world coordinates of the preset water level gauge point. The coordinate system of the water level model is then aligned using the homography projection matrix, and the corrected water level line is output.

[0025] In one embodiment, the calibration water level line and pixel-level confidence map are subjected to water level value conversion processing to obtain water level measurement values ​​and uncertainty data, including:

[0026] Extract the ordinate position data from the calibration water level line;

[0027] Extract the confidence distribution histogram of the target water area based on the pixel-level confidence map;

[0028] Construct an adaptive interpolation weight matrix based on the confidence distribution histogram;

[0029] Based on the adaptive interpolation weight matrix, weighted least squares fitting is performed on the height values ​​and ordinate position data of adjacent water gauge calibration points to generate water level measurement values ​​and uncertainty data.

[0030] In one embodiment, a feature fusion process under fluid continuity constraints is performed on the multi-scale texture feature map and motion feature map to generate a fused feature field, including:

[0031] Channel splicing is performed on multi-scale texture feature maps and motion feature maps to generate initial fused features;

[0032] The initial fused features are processed by divergence field calculation to output a feature divergence distribution map.

[0033] When the absolute mean of the feature divergence distribution exceeds a preset threshold, the gating weight iterative adjustment process is triggered.

[0034] Repeated divergence field calculations are performed until the preset fluid continuity condition is met, and the fused characteristic field is output.

[0035] In one embodiment, the initial boundary point set is subjected to motion stability screening based on the turbulence feature map to obtain a steady-state feature point set, including:

[0036] Curvature extremum detection is performed on the initial boundary point set to extract candidate key points;

[0037] The following formula is used to score the stability of the amplitude of the turbulence feature map at the candidate key point location, thus obtaining the motion stability score:

[0038]

[0039] Among them, s i Let T(x) be the motion stability score of the i-th candidate keypoint. i ,y i ) is a turbulence characteristic map at point (x i ,y i The eigenvectors of ) and τ is the turbulence intensity normalization coefficient;

[0040] Adaptive threshold segmentation is performed based on motion stability scores to remove outliers whose motion stability scores are below a preset stability threshold.

[0041] Output a subset of feature points that conform to the steady-state characteristics of the fluid, and determine it as the steady-state feature point set.

[0042] Secondly, this application also provides a water level monitoring device based on an image segmentation model, comprising:

[0043] The water body perception module is used to acquire multimodal water flow data of the target water area, extract features from the multimodal water flow data, and obtain turbulence feature maps. The multimodal water flow data includes water flow motion data and water body image data.

[0044] The physical constraint segmentation module is used to perform physical constraint segmentation on turbulence feature maps and water body image data to obtain water body region segmentation masks and corresponding pixel-level confidence maps.

[0045] The geometric correction module is used to perform water level geometric correction processing on the water body area segmentation mask to obtain the corrected water level line;

[0046] The water level monitoring module is used to perform water level value conversion processing on the calibration water level line and pixel-level confidence map to obtain water level measurement values ​​and uncertainty data.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the water level monitoring method of the above-mentioned image segmentation model.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the water level monitoring method of the above-described image segmentation model.

[0049] The aforementioned image segmentation model for water level monitoring acquires and fuses multimodal flow data composed of water flow motion data and water image data to achieve cross-domain feature extraction of water dynamics, generating a turbulence feature map containing turbulent motion laws. Physical constraint segmentation processing is used to jointly analyze the turbulence feature map and water image data. By introducing fluid dynamics laws as constraints for the segmentation algorithm, pseudo-edge interference from splashing water and foam is effectively suppressed, outputting a water region segmentation mask with pixel-level confidence assessment. Geometric correction of the water level gauge eliminates segmentation deviations caused by shooting angle or water flow morphology, generating an accurately corrected water level line. Water level value conversion is performed using the pixel-level confidence map, simultaneously outputting the water level measurement value and its uncertainty data. This technical approach, through the core design of multimodal data fusion and embedding physical laws, directly distinguishes between the real water level line and dynamic artifacts during the segmentation stage, ensuring the accuracy and reliability of monitoring under complex hydrological conditions. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating a water level monitoring method using an image segmentation model provided in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the structure of a water level monitoring device based on an image segmentation model provided in an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] First, a brief introduction to the terms used in the embodiments of this application will be given.

[0055] Image segmentation models are deep learning-based computer vision processing architectures. Their core function is to perform refined classification and recognition of each pixel in a digital image according to semantic categories, automatically interpreting image content through an end-to-end feature learning mechanism. This model achieves multi-level feature extraction and spatial information reconstruction through an encoder-decoder structure, ultimately outputting a pixel-level segmentation mask with clear semantic boundaries, providing a structured region perception foundation for water level monitoring.

[0056] Turbulence feature maps are digital representations of the nonlinear motion patterns of water bodies, generated through spatiotemporal feature encoding of multimodal water flow data. They not only include visual features such as texture and edges found in traditional images, but also incorporate fluid dynamic parameters (such as Reynolds stress and vortex intensity) to form tensor data reflecting water flow energy dissipation and turbulent structure. This feature map transforms the physical laws of water motion into a feature space recognizable by computational models, providing a physical basis for suppressing transient disturbances such as splashing water.

[0057] Water gauge geometric correction is a spatial transformation method designed for water conservancy monitoring scenarios. By establishing a projection mapping relationship between the camera coordinate system and the water gauge scale plane, it eliminates water level distortion caused by tilted shooting angles, water surface curvature, or wave undulations. The core principle is to convert the pixel-level water level lines obtained from image segmentation into an absolute elevation position perpendicular to the direction of gravity, based on perspective geometry principles, thus solving the parallax error inherent in non-contact measurements.

[0058] Uncertainty data is a quantitative indicator of the accuracy of water level measurements. It can be derived from pixel-level confidence maps through a Monte Carlo propagation model. It includes standard uncertainty (standard deviation of the measured value) and expanded uncertainty (confidence interval), comprehensively reflecting the cumulative impact of factors such as segmentation error, sensor noise, and geometric correction deviation on the final water level value. This data enables the monitoring system to have error perception capabilities and can still assess the reliability of the results under extreme hydrological conditions.

[0059] Based on the above definitions, the implementation environment of the water level monitoring method using the image segmentation model provided in this application embodiment will be described. Indicatively, this implementation environment includes: a sensor array, a processor, and a storage device. The sensor array includes, but is not limited to, visible light cameras, infrared thermal imagers, millimeter-wave radars, lidar, ultrasonic current meters, Doppler current meters, inertial measurement units (IMUs), etc.; the processor can be a central processing unit, graph processor, multi-core processor, or artificial intelligence chip, etc.; the storage device can be a distributed storage device or centralized storage, without limitation here.

[0060] Based on the above explanations of terms and implementation environments, the application scenarios of the embodiments of this application are described. The water level monitoring method using the image segmentation model provided in the embodiments of this application can be applied to scenarios including but not limited to the following:

[0061] In the scenario of flood discharge monitoring in water conservancy projects, especially in high-dynamic water flow areas such as flood discharge gates and spillways of water conservancy hubs, this technical solution can effectively overcome the interference of water splash and foam accumulation on water level identification during flood discharge by integrating water flow motion data (such as particle image velocity data) with real-time monitoring images and combining them with a segmentation model constrained by fluid mechanics. It can accurately output the water level line and confidence assessment during the flood discharge process, providing key data support for gate opening control.

[0062] In urban flood control and drainage monitoring scenarios, for turbulent water flow scenarios such as urban rivers and drainage outlets during rainstorms, the solution uses multimodal data to extract turbulence features and uses physical constraint segmentation to suppress false edge noise generated by raindrop impact and water flow vortex, thereby achieving robust identification of water levels in flooded areas and assisting urban flooding early warning systems in quickly responding to the risk of sudden water level changes.

[0063] In natural river hydrological monitoring scenarios, such as mountain rivers and steep slopes where traditional sensors are difficult to deploy in the field, this method can use images and water flow data collected by drones or fixed cameras to eliminate water level distortion caused by shooting tilt angle through geometric correction, and output water level values ​​and uncertainties that conform to hydrological standards, supporting all-weather automated monitoring of hydrological stations in remote areas.

[0064] As an illustration, the water level monitoring method of the image segmentation model provided in this application embodiment can also be applied to other application scenarios. This is only an example and does not limit the specific application scenarios.

[0065] In one exemplary embodiment, such as Figure 1 As shown, a water level monitoring method based on an image segmentation model is provided. This embodiment illustrates the application of this method to a terminal in the aforementioned implementation environment. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 104:

[0066] Step 101: Obtain multimodal water flow data of the target water area, extract features from the multimodal water flow data to obtain a turbulence feature map, wherein the multimodal water flow data includes water flow motion data and water body image data.

[0067] Specifically, multimodal water flow data of the target water area can be acquired through synchronous acquisition equipment. Water flow motion data is captured in real-time by a radar current meter or a particle image velocimetry (PIV) system to capture the surface velocity distribution of the water body, while water image data is acquired using a fixed, waterproof industrial camera at a frame rate of ≥25fps using RGB or near-infrared spectral images. Specifically, feature extraction of the multimodal water flow data is achieved using a cross-modal fusion convolutional neural network: firstly, spatiotemporal texture features are extracted from consecutive frame images through a 3D convolutional layer, and simultaneously, the velocity data is encoded into a motion vector field using an optical flow algorithm; further, the two types of data are stitched together in the latent space through a feature alignment module, and then the salient features of the turbulent region are dynamically weighted through a gated attention mechanism, ultimately outputting a turbulent feature map containing water flow direction, vortex intensity, and surface spectral characteristics. For example, an improved RAFT optical flow algorithm can be combined with a ResNet-50 backbone network to explicitly mark highly turbulent regions (such as the vortex region downstream of a spillway) in the feature map, providing kinematic prior knowledge for subsequent segmentation.

[0068] Step 102: Perform physical constraint segmentation on the turbulence feature map and water body image data to obtain the water body region segmentation mask and the corresponding pixel-level confidence map.

[0069] Specifically, when performing physically constrained segmentation on turbulence feature maps and water body image data, a dual-branch feature fusion network can be constructed: the water body image branch uses deformable convolutional kernels to extract multi-scale texture features, while the turbulence branch uses temporally gated recurrent units to encode dynamic motion patterns. Furthermore, fluid continuity constraints are enforced during the feature fusion stage: the fusion feature divergence field is calculated, and the feature weights are iteratively adjusted until the divergence converges to zero, ensuring that the segmentation results conform to the law of conservation of mass. The output water body region segmentation mask eliminates wave-induced hole artifacts, and the synchronously generated pixel-level confidence map quantifies the reliability of the segmentation results at each location, where low-confidence regions are strongly correlated with high-amplitude regions in the turbulence feature map.

[0070] Step 103: Perform water level geometric correction on the water body region segmentation mask to obtain the corrected water level line.

[0071] Specifically, water level gauge geometric correction is performed based on a water body region segmentation mask. Candidate water level lines in the mask are detected using Hough transform, and then spatial mapping is performed using a calibration water level gauge installed on the shore. Specifically, a perspective projection model of the water level gauge control points is established: a homography matrix is ​​constructed using pre-calibrated key points of the water level gauge scale (such as 0cm, 50cm, and 100cm markers), and the projection transformation from the image plane to the plane perpendicular to the water level gauge is achieved by solving a least-squares problem. For example, when an angle is detected between the water level line and the water level gauge scale, an affine transformation is used to correct the tilt error of the water level line caused by the camera's pitch angle, outputting a sub-pixel-level corrected water level line parallel to the actual water level plane.

[0072] Step 104: Perform water level value conversion processing on the calibration water level line and pixel-level confidence map to obtain water level measurement values ​​and uncertainty data.

[0073] Specifically, water level value conversion is performed based on the calibration water level line and pixel-level confidence map. The water level measurement value is calculated according to the linear mapping relationship between the water gauge scale and pixel coordinates. Taking the zero point of the water gauge as the reference, the physical scale value corresponding to the water level line is determined by bicubic interpolation. Furthermore, an uncertainty propagation model is constructed in conjunction with the confidence map: pixel areas with confidence <0.5 are marked as high uncertainty areas. Based on the area proportion and spatial distribution of these areas (e.g., the dispersion of artifact clusters), Bayesian inference is used to calculate the standard deviation of the water level value. The final output includes the water level measurement value (unit: cm) and its uncertainty data (95% confidence interval), where the uncertainty data directly reflects the quantitative impact of dynamic water flow interference on measurement accuracy.

[0074] The aforementioned image segmentation model for water level monitoring acquires and fuses multimodal flow data composed of water flow motion data and water image data to achieve cross-domain feature extraction of water dynamics, generating a turbulence feature map containing turbulent motion laws. Physical constraint segmentation processing is used to jointly analyze the turbulence feature map and water image data. By introducing fluid dynamics laws as constraints for the segmentation algorithm, pseudo-edge interference from splashing water and foam is effectively suppressed, outputting a water region segmentation mask with pixel-level confidence assessment. Geometric correction of the water level gauge eliminates segmentation deviations caused by shooting angle or water flow morphology, generating an accurately corrected water level line. Water level value conversion is performed using the pixel-level confidence map, simultaneously outputting the water level measurement value and its uncertainty data. This technical approach, through the core design of multimodal data fusion and embedding physical laws, directly distinguishes between the real water level line and dynamic artifacts during the segmentation stage, ensuring the accuracy and reliability of monitoring under complex hydrological conditions.

[0075] In one embodiment, multimodal flow data of the target water area is acquired, and feature extraction is performed on the multimodal flow data to obtain a turbulence feature map, including:

[0076] The vorticity distribution characteristic map is obtained by performing curl calculation on the velocity vector field of the water flow data.

[0077] Specifically, the water flow data is processed. This data is a three-dimensional velocity vector field acquired by an acoustic Doppler current meter, containing the velocity vectors of the water flow in the X, Y, and Z directions in the Cartesian coordinate system. Based on these velocity vectors, the vorticity value of each pixel is calculated. For the two-dimensional velocity components in the image plane, the velocity gradient of adjacent grid points is calculated using the central difference method, thereby obtaining the vorticity distribution perpendicular to the imaging plane. This transforms the invisible fluid rotational motion into a visualized vorticity map, where high vorticity regions correspond to vortices or turbulent core areas in the water flow, providing a physical basis for subsequent identification of dynamic interference sources.

[0078] The polarization gradient magnitude of the water body image data is subjected to gradient filtering to obtain a polarization gradient map.

[0079] Specifically, the synchronously acquired water image data is obtained by a polarization camera in the visible light band. The camera is rotated in front of the lens and linear polarizers are inserted in four directions: 0°, 45°, 90°, and 135°, to obtain four subframes with different polarization angles. For each pixel position, the Stokes parameters I, Q, and U are calculated and the degree of polarization and polarization angle are further derived. Then, the two-dimensional gradient magnitude of the polarization degree image is obtained. After suppressing high-frequency noise through anisotropic diffusion filtering, a polarization gradient map with edge preservation characteristics is generated. This map enhances the polarization difference between the water surface mirror reflection area and the wave foam area.

[0080] An exponential decay transformation is performed on the vorticity distribution characteristic map to obtain the turbulence intensity coefficient map.

[0081] Specifically, the vorticity distribution feature map is input into the exponential decay transformation function, which takes the absolute value of vorticity as the independent variable and controls the output dynamic range by a preset decay coefficient, so that the response in the high vorticity region saturates quickly while the low vorticity region still maintains a distinguishable gradient, and finally forms a turbulence intensity coefficient map. This coefficient map is numerically normalized to the interval [0,1], which can directly characterize the contribution weight of local turbulence energy to the subsequent segmentation task.

[0082] The turbulence intensity coefficient map and polarization gradient map are subjected to Hadamard product operation to obtain the turbulence characteristic map.

[0083] Specifically, the turbulence intensity coefficient map and the polarization gradient map are multiplied in the same pixel coordinates using a Hadamard product operation, i.e., multiplied pixel by pixel. The resulting product retains polarization edge information while being adaptively weighted according to turbulence energy: when a pixel has both high polarization gradient and high turbulence intensity, its output value approaches 1, marking it as a suspected splash edge; when the polarization gradient is high and the turbulence intensity is low, the output value is suppressed, marking it as a static reflection or a false edge of a still water surface. The tensor obtained through this operation is the turbulence feature map. This cross-modal feature fusion mechanism has physical significance; only when a region simultaneously has strong water rotation motion and significant polarization changes will the product operation produce a high response value; while single sensor noise or isolated artifacts will be suppressed due to the low response of another mode. The generated turbulence feature map not only clearly marks the spatiotemporal distribution of interference sources such as eddies and splashes, but also provides prior weights for subsequent segmentation algorithms by quantifying their intensity, enabling the model to actively focus on high-turbulence regions and implement targeted suppression. This feature fusion method based on physical laws significantly improves the algorithm's adaptability in complex hydrological scenarios compared to image segmentation that relies solely on data-driven approaches.

[0084] In one embodiment, physical constraint segmentation is performed on the turbulence feature map and water image data to obtain a water region segmentation mask and a corresponding pixel-level confidence map, including:

[0085] Deformable convolution feature extraction is performed on water body image data to obtain multi-scale texture feature maps.

[0086] For example, deformable convolutional feature extraction can be performed on water image data by using convolutional kernels with learnable offsets that slide in the image space. By predicting the dynamic positional offset of each sampling point, the local deformation features of the water flow texture are adaptively matched. For example, a four-level feature pyramid structure is constructed. The first level uses a 7×7 standard convolution to capture large-scale water surface structures; the second to fourth levels use 3×3 deformable convolutional layers, whose offsets are generated by shallow features through 1×1 convolutional regression, allowing the convolutional kernels to bend and cover the non-rigid edges of wave-breaking areas. Furthermore, a feature alignment module can be introduced into the skip connections to eliminate multi-scale feature misalignment caused by camera shake, outputting a multi-scale texture feature map containing both macroscopic water level contours and microscopic turbulence details.

[0087] The turbulence feature map is processed by a spatiotemporally gated recurrent network to output a motion feature map, which is used to characterize the motion pattern of dynamic artifacts.

[0088] For example, the turbulence feature map is input into a spatiotemporally gated recurrent network (ST-GRU) to obtain a motion feature map. Specifically, the network is designed with a two-branch architecture: the spatial branch extracts the geometric patterns of vortex regions through gated convolution; the temporal branch uses causal convolution to model the motion trajectory of artifacts. For example, a forget-gate priority mechanism is set in the gating unit: when a conflict between the feature vector and the fluid continuity equation is detected (such as an isolated splash point), the weight of the historical state is automatically reduced. During processing, the network takes five consecutive frames of turbulence feature maps as input, filters the temporal features that conform to the motion law of viscous fluid through the memory gate, and outputs a motion feature map representing the foam dissipation path and the direction of water splash.

[0089] Feature fusion processing under fluid continuity constraints is performed on multi-scale texture feature maps and motion feature maps to generate a fused feature field.

[0090] For example, this method fuses multi-scale texture feature maps and motion feature maps under physical constraints, splicing the two types of features along the channel dimension, generating initial fused features through convolutional layers, and introducing a fluid continuity constraint: based on the law of conservation of mass in incompressible fluids, the divergence tensor of the fused feature field is calculated, and an optimization objective function is constructed to force the divergence to approach zero. Specifically, an iterative gradient descent algorithm can also be used to adjust the feature weights until the feature field satisfies the fluid continuity condition (i.e., no source / sink anomalies). This constraint eliminates non-physical holes or isolated regions in the segmentation mask caused by optical interference, ensuring that the segmentation boundary conforms to fluid kinematics characteristics.

[0091] The fused feature field is subjected to confidence-weighted segmentation processing to output a water body region segmentation mask and a pixel-level confidence map.

[0092] For example, pixel-level segmentation results and confidence assessments are generated based on the optimized fused feature field. Specifically, a dual-channel decoder architecture can be adopted: the main branch generates an initial segmentation mask by progressively upsampling through transposed convolution; the confidence branch outputs pixel-level uncertainty estimates in parallel. For example, feature confidence gating is set at skip connections: when the cosine similarity between the fused feature vector and the prior turbulence physics is lower than a threshold, the decoding weight of the feature vector is automatically reduced. In the final segmentation stage, the uncertainty estimation map is transformed into a confidence weight matrix, which is applied to the initial mask through Hadamard product to suppress erroneous predictions in high-uncertainty regions, and the water region segmentation mask and pixel-level confidence map are output simultaneously. This embodiment effectively suppresses splash interference in high-speed water flow by using deformable convolution to capture texture, spatiotemporally gated network to model dynamic artifacts, fluid continuity constraint fusion, and confidence-weighted segmentation, and still outputs high-confidence water region segmentation results under extreme conditions such as rainstorms or flood discharges.

[0093] In one embodiment, the water body region segmentation mask is subjected to water level geometric correction processing to obtain a corrected water level line, including:

[0094] The maximum connected component contour is extracted from the water body region segmentation mask to obtain the initial boundary point set.

[0095] Specifically, based on the water body region segmentation mask, an improved boundary tracing algorithm can be used to extract the contour of the largest connected region. The segmentation mask is a binary image, where white pixels represent regions identified as water bodies. By scanning pixel value abrupt changes along the image row direction, all potential connected region boundaries are identified. The area ratio of each connected region is calculated, and dominant connected regions with an area ratio exceeding a preset threshold (e.g., ≥85%) are selected as target regions. Clockwise boundary tracing is performed on the selected connected regions: the contour coordinates are recorded pixel by pixel along the 8-neighborhood search rule, generating an initial boundary point set composed of continuous pixels. This step effectively eliminates isolated regions caused by noise or small floating objects, ensuring that subsequent processing focuses on the main water body boundary.

[0096] Based on the turbulence feature map, the initial boundary point set is subjected to motion stability screening to obtain the steady-state feature point set.

[0097] Specifically, by combining the physical characteristics of the turbulence feature map with historical time-series data to filter motion artifacts from the initial boundary points, the intensity value of each boundary point in the current frame's turbulence feature map and the variance of its intensity distribution in the most recent time-series frames can be extracted. This includes: removing boundary points whose turbulence feature values ​​exceed the dynamic threshold (corresponding to non-steady-state regions such as splashing water and vortex disturbances); and excluding boundary points with significant fluctuations in turbulence intensity in consecutive frames (corresponding to transient artifacts or regions of sudden changes in local turbulence). Through this dual physical constraint mechanism, this method can retain a set of steady-state feature points that conform to the dynamic equilibrium characteristics of the still water surface, forming a denoised candidate water level boundary.

[0098] The steady-state feature point set is fitted with a non-uniform rational B-spline curve to generate a water level model.

[0099] Specifically, the Non-Uniform Rational B-Spline (NURBS) algorithm is used to fit continuous curves to discrete steady-state feature points. Node vectors are adaptively generated based on the spatial distribution density of boundary points: the node spacing is shortened in densely populated areas such as nearshore regions to improve local fitting accuracy, while the spacing is expanded in sparse areas such as open water to maintain overall curve smoothness. The weight factor for each boundary point is dynamically allocated based on the confidence level of turbulence characteristics; high-confidence points have stronger control over the curve shape. During the fitting process, the objective can be optimized by minimizing the combination of the weighted residual sum of squares and the curvature energy function, ensuring that the curve closely fits the actual boundary while avoiding overfitting local noise. The resulting NURBS parametric model accurately represents the macroscopic geometry of the water level line and eliminates the sawtooth distortion caused by local turbulence.

[0100] The homography projection matrix is ​​calculated based on the world coordinates of the preset water level gauge point. The coordinate system of the water level model is then aligned using the homography projection matrix, and the corrected water level line is output.

[0101] Specifically, water level spatial mapping can be achieved based on the preset physical coordinates of the calibration plate. By identifying the feature markers (such as checkerboard corners) of the water gauge calibration plate in the image, a three-dimensional correspondence between the image pixel coordinates and the calibration plate's world coordinate system is established. The homography matrix is ​​calculated using a direct linear transformation algorithm to establish the perspective projection mapping relationship from the image plane to the water gauge elevation, eliminating geometric distortion caused by camera tilt or lens distortion. The NURBS curve control points are substituted into the homography matrix for coordinate transformation, converting the curve from the image pixel coordinate system to a physical coordinate system strictly aligned with the water gauge scale reference. The output corrected water level line eliminates perspective projection errors and can be directly converted to the actual water level elevation value based on the water gauge scale. This embodiment eliminates non-steady-state boundary interference through a dynamic stability screening mechanism, accurately represents the macroscopic contour of the water body using NURBS parametric modeling, and eliminates geometric distortion by combining the mapping of the water gauge calibration coordinate system. This reduces water level positioning errors in wave disturbance scenarios, and the output results have a clear physical spatial reference, meeting the accuracy requirements of hydrological monitoring.

[0102] In one embodiment, the calibration water level line and pixel-level confidence map are subjected to water level value conversion processing to obtain water level measurement values ​​and uncertainty data, including:

[0103] Extract the ordinate position data from the calibration water level line;

[0104] Extract the confidence distribution histogram of the target water area based on the pixel-level confidence map;

[0105] Construct an adaptive interpolation weight matrix based on the confidence distribution histogram;

[0106] Based on the adaptive interpolation weight matrix, weighted least squares fitting is performed on the height values ​​and ordinate position data of adjacent water gauge calibration points to generate water level measurement values ​​and uncertainty data.

[0107] Specifically, this technical solution, based on a geometrically corrected parametric water level model, extracts discrete elevation sampling points along the water gauge calibration area and generates a confidence distribution histogram of the target water area based on a pixel-level confidence map. Specifically, the confidence values ​​of each sampling point are extracted along the corrected water level trajectory, and a statistical frequency distribution is divided into preset intervals. For example, the confidence value range [0,1] is divided into 20 intervals, and the probability of the trajectory points within each interval is calculated. Further, the histogram is Gaussian smoothed to eliminate statistical noise caused by sudden changes in local turbulence, outputting a probability density distribution model characterizing the reliability of the water level measurement. An adaptive interpolation weight matrix is ​​constructed based on the confidence distribution histogram. Specifically, the weight allocation function expression is as follows:

[0108]

[0109] Among them, wi p(c) represents the normalized weight coefficient for the i-th sampling point. i ) represents sampling point c i The histogram probability density corresponding to the confidence level, max(p) is the global peak value of the probability density function p(c), μ is the mean of the high confidence region (typically >0.7), σ is the standard deviation of the distribution, and Gaussian decay term. To enhance the contribution of high-confidence points close to the steady-state baseline μ, for example, sampling points located in the peak probability density range are assigned the highest weight, while weights in the low probability density range are attenuated, forming a weight matrix that suppresses low-confidence measurement points. For example, the adaptive interpolation weight matrix and the height values ​​of adjacent water gauge calibration points are input into a weighted least squares fitting framework, where the calibration point height is a known physical quantity, the ordinate position data is the observed value, and the weight matrix serves as a residual penalty factor in solving the normal equation. The resulting water level measurement is the optimal height estimate from the fitted output. The fitting residuals are weighted by the weight matrix, and the standard deviation is calculated, which is output as uncertainty data synchronously with the water level measurement. This embodiment, through a confidence-driven adaptive weighting mechanism, directly integrates segmentation uncertainty into the water level conversion process, ensuring that robust water level measurements and their quantified uncertainties are still output even in areas with significant splash disturbances.

[0110] In one embodiment, a feature fusion process under fluid continuity constraints is performed on the multi-scale texture feature map and motion feature map to generate a fused feature field, including:

[0111] Channel splicing is performed on multi-scale texture feature maps and motion feature maps to generate initial fused features.

[0112] Specifically, multi-scale texture feature maps extracted by deformable convolutions are fused with motion feature maps output by a spatiotemporally gated network to form multimodal data. The two feature tensors are concatenated along the channel dimension to form channel-expanded fusion precursor data. Channel compression and cross-modal interaction learning are performed on the concatenated features through a 1×1 convolutional layer to output an initial fusion feature field containing texture and motion attributes. This process preserves the detail resolution of the original features while establishing a low-level correlation mapping between texture and motion.

[0113] The initial fused features are processed by divergence field calculation to output the feature divergence distribution map.

[0114] Specifically, physical constraints are imposed on the initial fusion feature field based on the principle of fluid dynamics continuity. For the vector features at each spatial location in the feature field, the two-dimensional planar divergence value is calculated: the Hamiltonian operator is simulated by a convolution kernel to perform differential operations on the feature vector field, and a scalar divergence distribution map representing the inflow / outflow state of the local region is output. The mathematical and physical meaning of the divergence value is the source and sink intensity distribution of the feature vector. When there is a non-zero extreme value region in the divergence field, it indicates that the feature at that location violates the law of conservation of fluid mass (such as the abnormal phenomenon of water flow being generated or disappearing out of thin air).

[0115] When the absolute mean of the feature divergence distribution exceeds a preset threshold, the gating weight iterative adjustment process is triggered.

[0116] For example, when the global absolute mean of the divergence distribution map exceeds a preset reliability threshold, a gating weight adjustment mechanism is triggered. Specifically, this involves: identifying the spatial distribution range of high divergence regions and locating the set of abnormal coordinate positions in the feature field that violate the continuity principle; constructing a differentiable feature weight adjustment matrix to reduce the contribution weight of the corresponding feature channels in the fusion process in the abnormal regions; updating the feature fusion coefficients using a gradient descent algorithm to minimize the divergence field energy function while controlling the feature reconstruction error; recalculating the divergence distribution map of the optimized feature field and verifying whether its global mean has dropped below the threshold. This iterative process continues until the following convergence conditions are met: the global mean of the divergence distribution map is within an acceptable error tolerance range, and the divergence fluctuation amount of three consecutive iterations is less than the stability judgment threshold.

[0117] Repeated divergence field calculations are performed until the preset fluid continuity condition is met, and the fused characteristic field is output.

[0118] Specifically, when the iterative optimization satisfies the fluid continuity constraint, the final fused feature field is output. This feature field, while preserving the original texture and motion information, is forced to satisfy the continuity differential constraint of incompressible fluid. This manifests as a continuous and smooth streamline shape in the water region of the feature space, and there are no abnormal flow sources or sinks outside of physical laws. This fused feature field serves as the input to the subsequent segmentation module, ensuring the physical rationality of the segmentation results in a hydrodynamic sense.

[0119] In one embodiment, the initial boundary point set is subjected to motion stability screening based on the turbulence feature map to obtain a steady-state feature point set, including:

[0120] Curvature extremum detection is performed on the initial boundary point set to extract candidate key points;

[0121] The following formula is used to score the stability of the amplitude of the turbulence feature map at the candidate key point location, thus obtaining the motion stability score:

[0122]

[0123] Among them, s i Let T(x) be the motion stability score of the i-th candidate keypoint. i ,y i ) is a turbulence characteristic map at point (x i ,y i The eigenvectors of ) and τ is the turbulence intensity normalization coefficient;

[0124] Adaptive threshold segmentation is performed based on motion stability scores to remove outliers whose motion stability scores are below a preset stability threshold.

[0125] Output a subset of feature points that conform to the steady-state characteristics of the fluid, and determine it as the steady-state feature point set.

[0126] Specifically, this technical solution calculates the local curvature change rate point by point along the boundary contour: the curvature radius is estimated by the small arc segment formed by three adjacent points, and abrupt changes in curvature exceeding the neighborhood average are marked as candidate key points. These locations typically correspond to the edge of broken waves or the center of turbulent vortices, exhibiting stronger motion instability. The set of candidate key points selected in this step serves as the focus area for subsequent stability assessment. Based on the above formula, the motion stability of the candidate key points is quantified by combining the vector field data of the turbulence feature map. For the coordinates of each candidate key point, the feature vector magnitude of the corresponding position in the turbulence feature map is extracted. Based on the physical law that turbulence energy intensity is negatively correlated with motion stability, an exponential stability scoring function is constructed: the feature vector magnitude is input into the negative exponential calculation unit for normalization, and the score decreases nonlinearly with increasing turbulence intensity. This scoring function maps the original turbulence features to a standardized stability measurement interval, with high-score regions representing low-disturbance states conforming to hydrostatic equilibrium. Furthermore, an adaptive screening based on fluid dynamics characteristics is performed using stability score data. A dynamic stability threshold, adaptively generated based on historical turbulence statistics, is set. Candidate key points are iterated one by one, and their stability scores are compared with the current threshold. Points with scores below the threshold are identified as motion outliers, their physical nature belonging to unsteady regions caused by local eddies or transient wave disturbances. The remaining points meeting the threshold are classified as valid points conforming to macroscopic fluid steady-state characteristics. This screening process simultaneously references the characteristic parameter range of steady-state flow in hydraulic theory to ensure that the retained point set satisfies the slowly varying flow characteristics with low energy dissipation. The final steady-state feature point set is output based on the comprehensive screening results. Topological continuity verification is performed on the retained point set: the spatial distribution integrity of the screened points within the original boundary is checked, and local interpolation is performed if necessary to avoid missing key sections. The determined steady-state feature point set includes two typical locations: the mainstream stable zone in open water far from shore disturbances, and the low-disturbance transition zone in near-shore slow-wave areas conforming to the surface gravity wave propagation law. This set accurately represents the true water level boundary with fluid equilibrium characteristics. This embodiment uses curvature extremum detection to locate boundary disturbance-sensitive areas and combines it with an exponential stability scoring model of turbulence feature vectors to achieve a physically driven dynamic screening mechanism, effectively overcoming the pseudo-boundary interference generated by traditional edge detection methods in wave-breaking zones. The adaptive threshold segmentation algorithm integrates fluid dynamics steady-state theory, accurately eliminating unsteady outliers such as eddy disturbances while preserving the macroscopic topology of the water level line, significantly improving the physical reliability of the steady-state feature point set under complex turbulent scenarios. The constructed screening system synergistically analyzes the kinematic characteristics (turbulence intensity) and geometric characteristics (curvature abrupt changes) of the water level boundary, which not only reduces the false screening rate under strong interference environments but also ensures that the final generated steady-state feature point set conforms to the dynamic propagation law of surface gravity waves, providing a data foundation with strict fluid dynamic consistency for subsequent water level line fitting.

[0127] In summary, the water level monitoring method based on the image segmentation model provided in this application constructs a closed-loop water level monitoring technology chain through a core design of multimodal data fusion and physical law embedding: It simultaneously acquires water flow motion data (such as radar velocity fields) and water image data (such as polarization spectral images); it fuses fluid dynamics curl calculation and optical polarization gradient analysis through a turbulence feature extraction module to generate a turbulence feature map that quantitatively characterizes the dynamics of water flow; it introduces a physical constraint mechanism in the segmentation stage, combining deformable convolution to extract multi-scale texture features, and uses a temporal gating network to model artifact motion patterns; it generates a high-confidence segmentation mask through feature fusion constrained by the fluid continuity equation; it achieves geometric correction of the water level line based on motion stability screening and NURBS curve fitting, eliminating interference from shooting angle and water flow morphology; and it simultaneously outputs water level measurements and their uncertainty data by combining confidence-driven adaptive weight allocation and weighted least squares fitting. This scheme integrates cross-disciplinary technologies of fluid mechanics, geometric modeling and computer vision to construct a full-link constraint system from pixel-level feature selection to physical space quantization. It can effectively distinguish between real water level boundaries and dynamic water flow artifacts, and achieve highly robust water level monitoring under complex hydrological conditions such as strong turbulence and multiple disturbances.

[0128] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0129] Based on the same inventive concept, this application also provides a water level monitoring device 10 for implementing the water level monitoring method of the image segmentation model involved above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more water level monitoring devices 10 for image segmentation models provided below can be found in the limitations of the water level monitoring method for image segmentation models above, and will not be repeated here.

[0130] In one exemplary embodiment, such as Figure 2 As shown, a water level monitoring device 10 based on an image segmentation model is provided, comprising:

[0131] The water body perception module 11 is used to acquire multimodal water flow data of the target water area, extract features from the multimodal water flow data, and obtain a turbulence feature map. The multimodal water flow data includes water flow motion data and water body image data.

[0132] The physical constraint segmentation module 12 is used to perform physical constraint segmentation processing on turbulence feature map and water body image data to obtain water body region segmentation mask and corresponding pixel-level confidence map.

[0133] The geometric correction module 13 is used to perform water level geometric correction processing on the water body area segmentation mask to obtain the corrected water level line;

[0134] The water level monitoring module 14 is used to perform water level value conversion processing on the calibration water level line and the pixel-level confidence map to obtain the water level measurement value and uncertainty data.

[0135] In one embodiment, the water sensing module 11 includes:

[0136] The curl calculation unit is used to perform curl calculation on the velocity vector field of water flow data to obtain a vorticity distribution characteristic map.

[0137] The gradient filtering unit is used to perform gradient filtering on the polarization gradient magnitude of water body image data to obtain a polarization gradient map.

[0138] The exponential transformation unit is used to perform exponential decay transformation on the vorticity distribution characteristic map to obtain the turbulence intensity coefficient map.

[0139] The Hadamard product operation unit is used to perform Hadamard product operations on the turbulence intensity coefficient map and polarization gradient map to obtain the turbulence characteristic map.

[0140] In one embodiment, the physical constraint segmentation module 12 includes:

[0141] Deformable convolutional units are used to perform deformable convolutional feature extraction on water image data to obtain multi-scale texture feature maps.

[0142] The spatiotemporal gating unit is used to perform spatiotemporal gating recurrent network processing on the turbulence feature map and output a motion feature map, which is used to characterize the motion pattern of dynamic artifacts.

[0143] The fluid-constrained fusion unit is used to perform feature fusion processing on multi-scale texture feature maps and motion feature maps under fluid continuity constraints to generate a fused feature field.

[0144] The confidence-weighted unit is used to perform confidence-weighted segmentation on the fused feature field and output a water region segmentation mask and a pixel-level confidence map.

[0145] In one embodiment, the geometric correction module 13 includes:

[0146] The contour extraction unit is used to perform maximum connected component contour extraction on the water body region segmentation mask to obtain the initial boundary point set.

[0147] The motion filtering unit is used to perform motion stability filtering on the initial boundary point set based on the turbulence feature map to obtain the steady-state feature point set.

[0148] The curve fitting unit is used to perform non-uniform rational B-spline curve fitting on the steady-state feature point set to generate a water level model.

[0149] The coordinate alignment unit is used to calculate the homography projection matrix based on the preset world coordinates of the water level gauge point, perform coordinate system alignment processing on the water level model through the homography projection matrix, and output the corrected water level line.

[0150] In one embodiment, the water level monitoring module 14 includes:

[0151] The coordinate extraction unit is used to extract the ordinate position data from the calibration water level line;

[0152] Histogram construction unit, used to extract the confidence distribution histogram of the target water area based on the pixel-level confidence map;

[0153] The weight generation unit is used to construct an adaptive interpolation weight matrix based on the confidence distribution histogram.

[0154] The water level fitting unit is used to perform weighted least squares fitting on the height values ​​and ordinate position data of adjacent water gauge calibration points based on the adaptive interpolation weight matrix, and generate water level measurement values ​​and uncertainty data.

[0155] In one embodiment, the fluid constraint fusion unit includes:

[0156] The channel stitching subunit is used to perform channel stitching processing on multi-scale texture feature maps and motion feature maps to generate initial fused features;

[0157] The divergence calculation subunit is used to perform divergence field calculation on the initial fused features and output the feature divergence distribution map.

[0158] The iterative triggering subunit is used to trigger the gating weight iterative adjustment process when the absolute mean of the feature divergence distribution map exceeds a preset threshold.

[0159] The iterative optimization sub-unit is used to repeatedly calculate and process the divergence field until the preset fluid continuity condition is met, and outputs the fused feature field.

[0160] In one embodiment, the motion screening unit includes:

[0161] The curvature detection unit is used to perform curvature extremum detection on the initial boundary point set and extract candidate key points.

[0162] The stability scoring unit is used to score the stability of the amplitude of the turbulence feature map at the candidate key point location using the following formula, thus obtaining a motion stability score:

[0163]

[0164] Among them, s i Let T(x) be the motion stability score of the i-th candidate keypoint. i ,y i ) is a turbulence characteristic map at point (x i ,y i The eigenvectors of ) and τ is the turbulence intensity normalization coefficient;

[0165] The threshold segmentation unit is used to perform adaptive threshold segmentation based on motion stability score, and remove outliers whose motion stability score is lower than the preset stability threshold.

[0166] The steady-state output unit is used to output a subset of feature points that conform to the steady-state characteristics of the fluid, and to determine it as the steady-state feature point set.

[0167] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a water level monitoring method based on an image segmentation model as described above.

[0168] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0169] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0170] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A water level monitoring method using an image segmentation model, characterized in that, The method includes: Multimodal water flow data of the target water area is acquired, and features are extracted from the multimodal water flow data to obtain a turbulence feature map. The multimodal water flow data includes water flow motion data and water body image data. Physical constraint segmentation processing is performed on the turbulence feature map and the water body image data to obtain a water body region segmentation mask and the corresponding pixel-level confidence map. The water body region segmentation mask is subjected to water level geometric correction processing to obtain the corrected water level line; The corrected water level line and the pixel-level confidence map are subjected to water level value conversion processing to obtain water level measurement values ​​and uncertainty data; The step of acquiring multimodal flow data of the target water area and extracting features from the multimodal flow data to obtain a turbulence feature map includes: The velocity vector field of the water flow data is processed by curl calculation to obtain a vorticity distribution characteristic map; The polarization gradient magnitude of the water body image data is subjected to gradient filtering to obtain a polarization gradient map; The vorticity distribution characteristic map is subjected to exponential decay transformation to obtain the turbulence intensity coefficient map; The turbulence intensity coefficient map and the polarization gradient map are subjected to Hadamard product operation to obtain the turbulence feature map; The step of performing physical constraint segmentation on the turbulence feature map and the water body image data to obtain a water body region segmentation mask and a corresponding pixel-level confidence map includes: The water body image data is subjected to deformable convolution feature extraction processing to obtain a multi-scale texture feature map; The turbulence feature map is processed by a spatiotemporally gated recurrent network to output a motion feature map, which is used to characterize the motion pattern of dynamic artifacts. The multi-scale texture feature map and motion feature map are subjected to feature fusion processing under fluid continuity constraints to generate a fused feature field; The fused feature field is subjected to confidence-weighted segmentation processing to output the water region segmentation mask and the pixel-level confidence map.

2. The method according to claim 1, characterized in that, The step of performing water level geometric correction processing on the water body region segmentation mask to obtain the corrected water level line includes: The water body region segmentation mask is subjected to maximum connected component contour extraction processing to obtain an initial boundary point set; Based on the turbulence feature map, the initial boundary point set is subjected to motion stability screening to obtain a steady-state feature point set; The steady-state feature point set is fitted with a non-uniform rational B-spline curve to generate a water level model. The homography projection matrix is ​​calculated based on the preset world coordinates of the water level gauge point. The water level model is then aligned with the coordinate system using the homography projection matrix, and the corrected water level line is output.

3. The method according to claim 1, characterized in that, The step of performing water level value conversion processing on the corrected water level line and the pixel-level confidence map to obtain water level measurement values ​​and uncertainty data includes: Extract the ordinate position data from the corrected water level line; Based on the pixel-level confidence map, extract the confidence distribution histogram of the target water area; Construct an adaptive interpolation weight matrix based on the confidence distribution histogram; Based on the adaptive interpolation weight matrix, the height values ​​of adjacent water gauge calibration points are fitted with the ordinate position data using weighted least squares to generate the water level measurement value and the uncertainty data.

4. The method according to claim 1, characterized in that, The step of performing feature fusion processing on the multi-scale texture feature map and motion feature map under fluid continuity constraints to generate a fused feature field includes: The multi-scale texture feature map and the motion feature map are subjected to channel splicing to generate an initial fused feature; The initial fusion features are processed by divergence field calculation to output a feature divergence distribution map; When the absolute mean of the feature divergence distribution map exceeds a preset threshold, the gating weight iterative adjustment process is triggered. Repeat the divergence field calculation process until the preset fluid continuity condition is met, and output the fused feature field.

5. The method according to claim 2, characterized in that, The step of performing motion stability screening on the initial boundary point set based on the turbulence feature map to obtain a steady-state feature point set includes: The initial boundary point set is subjected to curvature extremum detection processing to extract candidate key points; The motion stability score is obtained by scoring the amplitude of the turbulence feature map at the candidate key point location using the following formula: ； in, For the first The motion stability score of each of the candidate keypoints. The turbulence feature map at point eigenvectors, This is the normalization coefficient for turbulence intensity; An adaptive threshold segmentation process is performed based on the motion stability score to remove outliers whose motion stability scores are lower than a preset stability threshold. Output a subset of feature points that conform to the steady-state characteristics of the fluid, and determine it as the steady-state feature point set.

6. A water level monitoring device based on an image segmentation model, used to implement the method as described in any one of claims 1 to 5, characterized in that, The device includes: The water body sensing module is used to acquire multimodal water flow data of the target water area, extract features from the multimodal water flow data to obtain a turbulence feature map, wherein the multimodal water flow data includes water flow motion data and water body image data; The physical constraint segmentation module is used to perform physical constraint segmentation processing on the turbulence feature map and the water body image data to obtain a water body region segmentation mask and a corresponding pixel-level confidence map. The geometric correction module is used to perform water level geometric correction processing on the water body area segmentation mask to obtain the corrected water level line; The water level monitoring module is used to perform water level value conversion processing on the calibration water level line and the pixel-level confidence map to obtain water level measurement values ​​and uncertainty data; The water sensing module includes: The curl calculation unit is used to perform curl calculation on the velocity vector field of the water flow motion data to obtain a vorticity distribution characteristic map. The gradient filtering unit is used to perform gradient filtering on the polarization gradient magnitude of the water body image data to obtain a polarization gradient map. An exponential transformation unit is used to perform an exponential decay transformation on the vorticity distribution characteristic map to obtain a turbulence intensity coefficient map. The Hadamard product operation unit is used to perform Hadamard product operation on the turbulence intensity coefficient map and the polarization gradient map to obtain the turbulence feature map. The physical constraint segmentation module includes: Deformable convolutional units are used to perform deformable convolutional feature extraction processing on the water body image data to obtain multi-scale texture feature maps. A spatiotemporal gating unit is used to perform spatiotemporal gating cyclic network processing on the turbulence feature map and output a motion feature map, which is used to characterize the motion pattern of dynamic artifacts. The fluid constraint fusion unit is used to perform feature fusion processing on the multi-scale texture feature map and motion feature map under fluid continuity constraints to generate a fused feature field. The confidence weighting unit is used to perform confidence weighting segmentation processing on the fused feature field and output the water body region segmentation mask and the pixel-level confidence map.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images