Method for identifying river cross-sectional morphology by coupling artificial intelligence and riverbed geomorphology

Abstract

The application discloses a river cross section shape identification method coupling artificial intelligence and river bed geomorphology, and comprises the following steps: unmanned aerial vehicle complete river channel topography measurement; river cross section type knowledge graph construction; cross section shape characteristic parameter construction; double-encoder framework construction based on artificial intelligence; most similar cross section searching; satellite remote sensing river width and elevation acquisition and water surface elevation correction; unknown cross section geometric shape adjustment. The application couples artificial intelligence and river bed geomorphology, solves the problem that a purely data-driven artificial intelligence model lacks understanding of physical causality, and deduces unreasonable black box results in a physically unreasonable way when the result is generalized to rivers or extreme hydrological events outside the training set; and realizes river cross section parameter acquisition under non-contact measurement conditions. The application breaks through the problem of accurate measurement of river cross section parameters under severe conditions such as flood disaster water damage, major geological disasters, unreachable boundaries and overseas key rivers.

Description

Technical Field

[0001] This invention relates to the fields of water conservancy engineering technology, geological disaster monitoring and satellite remote sensing runoff monitoring, especially to the acquisition of river channel deformation after extreme floods and river runoff data that cannot be measured on-site. Specifically, it is a method for identifying river cross-section morphology by coupling artificial intelligence and riverbed geomorphology. Background Technology

[0002] River cross-sectional morphology is a key element in hydrological calculations of river flow and a crucial fundamental parameter in modern hydraulic engineering, hydrological modeling, and river geomorphology research. The cross-sectional morphology of a river channel directly constitutes the boundary conditions for water flow, and its accuracy directly determines the reliability of flood evolution simulations, sediment transport calculations, waterway improvement projects, and eco-hydrological model operation. Long-term, detailed monitoring and rapid inference of river cross-sectional morphology have become fundamental requirements in hydraulic engineering, river ecological restoration, and flood control. However, in many river sections, especially in remote, disaster-stricken, military, and ecologically restricted areas, implementing conventional contact-based cross-sectional measurements presents significant difficulties: personnel cannot reach these areas, measurement costs are high, timeliness is poor, and safety risks exist, resulting in a long-term lack of representative cross-sectional morphology data for many regions. Therefore, a non-contact method that can compensate for the scarcity of on-site data is urgently needed to derive accurate river channel morphology.

[0003] Currently, traditional methods for obtaining river cross-sectional morphology mainly rely on contact measurement techniques, such as underwater sonar bathymetry and the combination of global navigation satellite systems and water level gauges. These methods offer extremely high local accuracy, but are inherently labor-intensive and inefficient, providing only discrete point or line measurements, making it difficult to achieve systematic coverage of large-scale river networks and multi-temporal conditions. Especially in remote, topographically complex mountain rivers, on-site measurements are difficult to promote due to their high economic costs, significant safety risks, and long revisit cycles.

[0004] To acquire flood runoff data using remote sensing technologies and methods, various existing techniques utilize satellites, drones, and hydrological models to calculate and monitor flood inundation extent and runoff. Chinese invention patent application number 02245598.1 designs a simple measuring ruler for on-site measurement of river cross-sectional morphology. Chinese invention patent application number 201710069538.1 designs a method to reconstruct historical river channel morphology using riverbed sediments. Chinese invention patent application number 202311564799.2 uses drones and on-site measurement methods to identify river cross-sectional morphology. Chinese invention patent application number 202411731153.3 uses satellite remote sensing to describe the complex morphology of braided rivers. Chinese invention patent application number 202410674354.8 uses multi-source remote sensing data to simulate the migration of river cross-sectional morphology.

[0005] The aforementioned existing technologies have solved the problems of obtaining flood inundation range and monitoring river runoff during non-flood seasons to some extent, but there are still some technical problems, including: 1) It is difficult to obtain cross-sectional morphology data in inaccessible areas, and existing methods for monitoring river cross-sectional morphology require monitoring tools or soil samples; 2) There is no non-contact remote sensing method for obtaining river cross-sections, and existing methods mainly use remote sensing to obtain river morphology, focusing on surface river network distribution or longitudinal morphological characterization; 3) Existing methods cannot achieve large-scale acquisition of river channel morphology data in space. Using UAV equipment can only collect data in limited river cross-sections, and it is difficult to achieve large-scale cross-sectional morphology acquisition at the regional scale; 4) It is difficult to achieve high-precision cross-sectional morphology acquisition by only using remote sensing to obtain surface information; 5) The methods lack versatility and are difficult to apply on a large scale in operations. Summary of the Invention

[0006] The purpose of this invention is to overcome the technical problems in obtaining river cross-sectional morphology under non-contact measurement conditions, and to overcome the technical deficiencies in the existing technology. This invention proposes a method for river cross-sectional morphology recognition that couples artificial intelligence with riverbed geomorphology. This objective is achieved through the following technical solution:

[0007] A method for recognizing the cross-sectional morphology of rivers by coupling artificial intelligence with riverbed geomorphology includes the following steps:

[0008] Step 1: Complete UAV river topographic survey: Acquire complete UAV image data of the monitored river to obtain the elevation changes above the water surface, use sonar depth sounding to measure the underwater topography, fit the elevation changes above the water surface and the underwater topography, and merge them to generate a complete three-dimensional digital river channel, that is, obtain a digital river cross section; all the measured digital river cross sections form a cross section database.

[0009] Step 2: Construction of a knowledge graph of river cross-section types: Classify the morphology of the digitized river cross-sections measured in Step 1 to obtain river cross-section categories;

[0010] Step 3: Construction of cross-sectional morphological characteristic parameters: For each river cross section, determine its river development stage, river landform type, and floodplain type, and use satellite imagery and data to calculate embedding rate, curvature, gradient, and width-to-depth ratio to construct cross-sectional morphological characteristic parameters;

[0011] Step 4: Construction of an AI-based dual-encoder framework: Utilizing a contrastive learning method that combines physical information, the digitized river cross-sections obtained in Step 1 and the cross-sectional morphological feature parameters obtained in Step 3 are mapped together into the same embedding space. Furthermore, the river cross-section category from Step 2 is used to set a judgment threshold for the cross-sectional morphological feature parameters in Step 3. The digitized river cross-sections obtained in Step 1 are processed using a Transformer encoder to extract geometric features; the cross-sectional morphological feature parameters from Step 3 are processed using a multi-layer fully connected neural network to extract morphological features. Finally, a mapping relationship model between the digitized river cross-sections and morphological feature parameters is obtained.

[0012] Step 5: Finding the most similar cross section to the unknown cross section: Based on the morphological feature parameters of the unknown cross section obtained in Step 3, input the mapping relationship model obtained in Step 4, and search for the digital river cross section with the highest cosine similarity to the unknown cross section in the database of Step 1. Make a judgment: The judgment is to determine whether it meets the river cross section category in Step 2. If it does, then this cross section is the most similar cross section, and proceed to Step 6. If it does not meet the criteria, then search for the cross section with the highest cosine similarity among the remaining cross sections, and continue to make judgments until you enter Step 6.

[0013] Step 6: Satellite remote sensing width and elevation acquisition and water surface elevation correction; Based on the satellite remote sensing data source, water body classification is obtained. Combined with public DEM data, sampling is performed at 10 m intervals. According to physical laws, the elevations at both ends of the water surface should be consistent. Therefore, the lower value of the water body elevations at both ends of the multi-period images is taken as the corrected DEM data.

[0014] Step 7: Adjustment of the geometry of the unknown cross section: Extract the most similar cross section obtained in Step 5 into a pure shape between 0 and 1, and combine it with the corrected elevation in Step 6 to obtain the basic geometry of the unknown cross section; adjust the water depth, smooth and densify the basic geometry to obtain the river cross section result of the unknown cross section.

[0015] Further optimization, in step four, involves the following specific steps for constructing the AI-based dual encoder framework:

[0016] 1) Construction of a digital river cross-section geometric encoder based on the Transformer architecture:

[0017] For the digital river cross-section sequence obtained in step one p i For the i-th distance-elevation point pair in a cross-section sequence, its spatial topological features are extracted using a Transformer encoder. Position encoding is used to enhance the model's ability to handle the spatial order of cross-section points. The input coordinates are embedded with positions, and the position encoding is performed according to the following formula:

[0018] (5)

[0019] (6)

[0020] In the formula, Indicates the index position of the cross-section point in the sequence; Index representing the embedded dimension; This represents the total dimension of the encoder's feature embeddings; This represents the generated position encoding vector;

[0021] A digital river cross-section geometry encoder captures key geometric features of the cross-section using a multi-head self-attention mechanism. The calculation formula is as follows:

[0022] (7)

[0023] In the formula, This represents the geometric feature matrix after processing by the self-attention layer. Represents the input cross-sectional coordinate sequence matrix; These represent the weight matrices for the query, key, and value learned by the model, respectively. The dimension of the key vector is used to scale the dot product result to prevent gradient vanishing; T represents the transpose matrix.

[0024] 2) Construction of a morphological feature parameter encoder based on a multi-layer fully connected neural network:

[0025] Regarding the cross-sectional morphology feature parameter vector obtained in step three Construct an encoder based on a multilayer fully connected neural network (MLP). That is, to obtain the morphological feature parameter encoder based on a multi-layer fully connected neural network. Project it into an embedding space consistent with the geometric features;

[0026] Morphological feature parameter encoder of multilayer fully connected neural network The specific calculation formula is as follows:

[0027] (8)

[0028] In the formula, This represents the embedding vector generated after encoding the morphological feature parameters; This represents the input vector of morphological feature parameters output from step three; This represents the weight matrix of each layer's connections; Represents the bias term vector for each layer; ReLU represents the nonlinear activation function;

[0029] 3) Setting hard judgment thresholds based on physical information classification:

[0030] This step introduces the river cross-section category from step two. We set a judgment threshold based on physical meaning for morphological features, and construct a physical constraint loss function for the model based on this threshold to ensure that the mapping relationship generated by the model does not violate the basic principles of geomorphology.

[0031] The physical constraint loss function is calculated using the following formula:

[0032] (9)

[0033] In the formula, This represents the physical constraint loss value; This represents the model's prediction or mapping of the first... Each morphological feature parameter value; These represent the river categories respectively. The Hard upper and lower thresholds for each feature parameter; Indicates the total number of feature parameters;

[0034] 4) Joint training of the contrastive learning mapping model is achieved by constructing:

[0035] Comparison loss functions, forced digitization of river cross-section geometric embedding vectors With morphological feature embedding vector Align within the embedded space;

[0036] The joint objective function is calculated using the following formula:

[0037] (10)

[0038] In the formula, Represents the total training loss function; This represents the contrastive learning loss, used to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs; The weighting coefficients of the physical constraint terms are used to adjust the degree of influence of physical laws on the model learning process;

[0039] When the mapping relationship is established, When convergence is achieved, the nonlinear mapping function between the geometric and morphological features of the digital river cross-section is established as follows:

[0040] (11)

[0041] In the formula, This is the final mapping relationship model.

[0042] Further optimization is achieved by using the following formula for cosine similarity in step five:

[0043] (12)

[0044] (13)

[0045] In the formula, This represents the projection vector of the cross section to be investigated in the feature embedding space; This represents the input vector of the morphological feature parameters of the cross-section to be investigated extracted in step three; Represents the vector of the section to be investigated With the database Geometric feature vectors of each cross section Cosine similarity between them.

[0046] Furthermore, the specific formulas for obtaining satellite remote sensing elevation and correcting water surface elevation in step six are as follows:

[0047] The improved Normalized Differential Water Index (MNDWI) is used to identify the extent of water bodies in river channels and generate a water body mask matrix. :

[0048] (14)

[0049] In the formula, For the first Periodic images in coordinates Water body signage at the location; For adaptive threshold extraction;

[0050] Using ALOS 12.5m resolution DEM data, discrete sampling was performed on the extracted water body boundaries with a spatial step size of 10 m to obtain the initial elevation values ​​of the boundary points on both sides of the water body; for any cross section Identify its left bank water boundary point Water boundary point on the right bank For water boundary elevations acquired from multiple image periods, the lower of the observations at both ends is taken as the baseline elevation of the water surface in that period to eliminate the problem of artificial elevation in the DEM. The elevation is calculated using the following formula:

[0051] (15)

[0052] In the formula, This represents the corrected water surface reference elevation of the k-th cross section under the image of period t; , These represent the sampling positions of the left and right bank waterline of the k-th section in the t-th image, respectively; This represents the elevation function of the corresponding coordinate points obtained from the original ALOS DEM.

[0053] Corrected boundary elevation Map back to the river channel topography model and update the original DEM feature points.

[0054] Step six involves obtaining the river width using existing technology. SAR data is obtained from satellite remote sensing data sources, and the river surface width is retrieved using a hybrid pixel decomposition algorithm based on the SAR data. The calculation formula is as follows:

[0055] (16)

[0056] In the formula, W is the width of the water surface, in meters; x is the total number of pixels in the region of interest of the river valley. Let j be the SAR value of pixel j; and These are the thresholds for land and water SAR, respectively; m is the actual area of ​​a pixel. 2 ; Let be the length of the selected valley region of interest, in meters.

[0057] Furthermore, the specific operation for obtaining the basic geometric shape of the unknown cross-section in step seven is as follows: The most similar cross-section sequence obtained in step five... Dimensionless processing is performed to extract the pure shape reflecting the riverbed evolution characteristics, which is a pure shape between 0 and 1. This shape is then mapped to the water surface width determined in step six. and elevation datum In space, obtain the basic geometric shape

[0058] (17)

[0059] (18)

[0060] In the formula, This represents the horizontal position vector after scaling. Represents the vertical position vector after scaling; Indicates by Extract the i-th value of the normalized horizontal and vertical axes; , This represents the maximum and minimum horizontal distances of the original sequence of similar cross sections; , This represents the maximum and minimum elevation values ​​of the original sequence of similar cross sections; This represents the river width obtained in step six; Step 6 outputs the corrected river boundary reference elevation; This indicates the preset river depth at the target cross-section.

[0061] Furthermore, step seven involves adjusting the water depth based on the basic geometry: using gradient calculations to find the flattest central region of the riverbed, creating a Gaussian weighted distribution in the central region, and multiplying the water depth by the Gaussian weights. The specific operation is as follows:

[0062] To simulate the real deep-channel scouring characteristics of natural rivers, the flattest areas of the riverbed were identified, and nonlinear depth adjustment was applied to these areas using a Gaussian function. The local slope of the basic geometric morphology sequence was calculated using first-order difference. Identify flat areas at the bottom of the riverbed and determine the lateral center location. , The calculation formula is as follows:

[0063] (19)

[0064] In the formula, This represents the local geometric gradient at the i-th sampling point; , This represents the i-th horizontal position and the base elevation value obtained after scaling transformation; , This represents the (i+1)th horizontal position and the base elevation value obtained after scaling.

[0065] Identify gradient values Less than the preset threshold A continuous region, and determine the lateral center coordinates of that region. The Gaussian weighted water depth superposition is performed according to the following formula:

[0066] (20)

[0067] In the formula, This represents the final corrected elevation vector after completing the simulation of the underwater morphology. This represents the base elevation value obtained after scaling. This indicates the reference average water depth obtained from hydrological data; The span parameter represents the Gaussian distribution, used to control the attenuation range of the influence of water depth correction on both banks of the riverbed; the final corrected elevation vector represents the completed simulation of the underwater morphology. This indicates the horizontal position obtained after scaling.

[0068] In step seven: smoothing and refining the unknown cross-sectional shape after water depth adjustment specifically involves: to ensure that the generated cross-section has continuous physical curvature and meets the accuracy requirements of engineering calculations, a cubic spline interpolation algorithm is used to fit the discrete point set, ensuring that the second derivative is continuous at all sampling nodes and within any adjacent sampling point interval. Inside, elevation value The calculation formula is as follows:

[0069] (twenty one)

[0070] In the formula, express The i-th elevation value; , express The horizontal distance between the i-th and i+1-th sampling points; The horizontal distance is the independent variable; The coefficients are polynomials, determined by the elevation at the nodes, the first derivative, the second derivative, and the third derivative.

[0071] According to the preset encryption factor Equal-interval encrypted sampling is performed to generate the final river cross-section result sequence. The calculation formula is as follows:

[0072] (twenty two)

[0073] In the formula, This represents the final three-dimensional digital river channel result; This represents the j-th value of the lateral distance after resampling; This represents the j-th value of the longitudinal elevation obtained from the mapping; , Indicates the horizontal distance between the start and end points of the cross-section; express The total number of sampling points; This represents the cubic spline interpolation function; Indicates the encryption factor.

[0074] The advantages and beneficial effects of this invention are:

[0075] 1) This invention revolutionizes the traditional river cross-section monitoring paradigm by innovatively coupling artificial intelligence with riverbed geomorphology, hydrodynamics, and remote sensing to accurately acquire river cross-section information. It constructs a physically-driven, AI-powered intelligent inference framework for river cross-section morphology, utilizing existing river cross-section data for model training, calibration, and generalization testing. This enables high-precision inference and visualization of river cross-section morphology inaccessible areas. While purely data-driven AI models possess powerful nonlinear fitting capabilities, their lack of understanding of these physical causal relationships can easily lead to physically unreasonable "black box" results when generalized to rivers outside the training set or to extreme hydrological events. This invention transforms geomorphological knowledge into internal constraints for the model, which is crucial for ensuring the scientific rationality of the inference results.

[0076] 2) The method of this invention enables the acquisition of river cross-sectional parameters under non-contact measurement conditions. It solves the problem of accurately measuring river cross-sectional parameters under harsh conditions such as flood damage, major geological disasters, inaccessible borders, and critical rivers outside of China. Accurate acquisition of river cross-sections provides the most fundamental data for calculating river runoff, water resource assessment, and flood and drought prevention in inaccessible areas. Simultaneously, it provides essential basic data for remote sensing runoff monitoring, addressing the challenge of accurately monitoring water resources in overseas regions, particularly the inflow from upstream countries. This provides crucial data for mitigating flood disasters caused by extreme precipitation and irregular flooding in upstream countries. Attached Figure Description

[0077] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0078] Figure 1 This is a schematic diagram illustrating the operation of an embodiment of the present invention;

[0079] Figure 2 This invention provides a river type classification for the river development knowledge graph.

[0080] Figure 3 The selected river section is used for the example study;

[0081] Figure 4 The digital river channel morphology is shown in the measured cross-section of the embodiment;

[0082] Figure 5 The cross-sectional shape is simulated in the embodiment;

[0083] Figure 6 The results show the accuracy verification of the measured and simulated cross-sections in the embodiments.

[0084] Figure 7 Schematic diagram of floodplain width and main channel width;

[0085] Figure 8 A diagram illustrating the aspect ratio calculation. Detailed Implementation

[0086] Example 1:

[0087] A method for river cross-section morphology recognition that couples artificial intelligence with riverbed geomorphology is presented, with an implementation example deployed on the Changhua River cross-section in Hainan Province. See the operation diagram below. Figure 1 Specifically, it includes the following steps:

[0088] Step 1: Complete UAV River Topographic Survey: Acquire complete UAV image data of the monitored river to obtain the elevation changes above the water surface. The UAV image data includes image data of the river shape, elevation, and nearby land features with the target river cross-section; use sonar depth sounding to measure the underwater topography; fit the elevation changes above the water surface and the underwater topography, and merge them to generate a complete three-dimensional digital river channel; that is, obtain a digital river cross-section.

[0089] The monitoring river section should be located in a gentle section with stable slopes on both banks, or a standard cross section should be artificially constructed. The length between the upstream and downstream monitoring sections should not be less than 100 meters.

[0090] The drone's flight altitude is controlled between 70-120m based on actual flight conditions. The flight path is set with the river as the center. For example, the flight range is set to 700×350m, where 700m is the length of the flight path in the direction of the river and 350m is the total width on both sides of the river.

[0091] Step Two: Construction of a Knowledge Graph of River Cross-Section Types: Based on the fundamental principles of riverbed geomorphology and hydrodynamics, and combined with the morphological development process of river cross-sections, the river cross-section morphologies measured in Step One are classified to obtain river cross-section categories, such as... Figure 2 As shown, river cross-sections are mainly classified into three categories: V-shaped: including five specific types such as flat-bottomed narrow canyons, flat-bottomed widening, multi-point grooves, connected grooves, and lateral V-shaped; Parabolic: including seven specific types such as symmetrical parabolic, mid-channel parabolic, lateral channel parabolic, lateral parabolic, strongly lateral parabolic, inner shoal parabolic, and central bulge; Trapezoidal: including six specific types such as symmetrical trapezoidal, mid-channel trapezoidal, lateral channel trapezoidal, lateral trapezoidal, steep and gentle slopes, and rectangular.

[0092] Step 3: Construction of Cross-Sectional Morphological Feature Parameters: For each river cross-section, determine its river development stage, landform type, and floodplain type. Calculate the embedding rate and tortuosity using Sentinel-2 imagery, and calculate the gradient and width-to-depth ratio using an ALOS 12.5 m DEM. Construct the following seven morphological feature parameters for the cross-section. Some of the calculated results are shown in Table 1. The calculation rules are as follows:

[0093] 1) Valley development stage

[0094] The developmental stages of river valleys divide river evolution into three phases: juvenile, mature, and senile, based on Davis's theory of geomorphic evolution cycles. This parameter is significant in revealing the dynamic balance between river erosion and deposition. In the juvenile phase, downward erosion dominates, resulting in narrow, deep valleys, often V-shaped, with steep cross-sections and rapid currents. In the mature phase, erosion and deposition tend to be balanced, leading to wider valleys and relatively stable cross-sections. In the senile phase, lateral erosion dominates, resulting in broad, flat valleys with shallow cross-sections, easily forming terraces.

[0095] 2) Landform type

[0096] Landform types are categorized into mountains, hills, and plains, representing the slope, flow velocity, and hydrodynamic effects of rivers. Mountain rivers have narrow cross-sections and steep slopes, resulting in intense erosion and strong hydrodynamics. Hilly rivers are transitional, with moderately wide cross-sections, balancing erosion and deposition. Plain rivers have wide, shallow cross-sections and gentle slopes, with deposition as the dominant process and low flow velocities. The elevation standard deviation within a 1 km radius is calculated from the DEM, along with the overall slope. The elevation standard deviation varies significantly in mountains, with an overall slope > 15°. The elevation standard deviation varies moderately in hills, with an overall slope between 5° and 15°. The elevation standard deviation varies less in plains, with an overall slope < 5° and minimal local undulation.

[0097] 3) Floodplain type

[0098] Floodplain types are categorized into no floodplain, bilateral floodplain, left-side floodplain, and right-side floodplain. A no-floodplain refers to a river channel without continuous low-lying alluvial platforms on either side, with no obvious floodplain formation, and no obvious stagnant depressions or alluvial deposits near the shore in satellite imagery. A bilateral floodplain refers to a river channel with distinct floodplain zones on both sides, which can be submerged during floods. From upstream to downstream, a left-side floodplain refers to a floodplain zone primarily located on the left side of the river channel, while the right side is lacking or significantly narrower; a right-side floodplain refers to a floodplain primarily located on the right side of the river channel, while the left side is lacking or significantly narrower.

[0099] 4) Embedding rate

[0100] The entrenchment ratio (ER) is the ratio of the width of the floodplain to the width of the main channel. In riverbed geomorphology, it is a key indicator used to quantify the degree of vertical constraint on the river channel, reflecting the depth of incision in the valley and the connectivity of the floodplain. A lower ratio indicates deeper channel incision, greater vertical restriction on flood discharge, and poorer floodplain connectivity; conversely, a higher ratio indicates a well-developed floodplain and high hydrological connectivity between the river and the floodplain. The calculation formula is as follows:

[0101] (1)

[0102] In the formula, ER represents the embedding rate, and W floodprone W refers to the width of the river floodplain. bankfull Refers to the width of the main channel of a river.

[0103] River width was extracted using Sentinel-2 remote sensing imagery. The 75th percentile of the river width from multi-year imagery represents the floodplain width, while the median river width is the full channel width. A schematic diagram illustrating their positional relationships is shown below. Figure 7 .

[0104] 5) Ratio

[0105] Slope (S) represents the gradient of a river section and is significant in driving water flow dynamics. A high slope increases flow velocity and promotes erosion; a low slope slows flow and facilitates deposition. Using Sentinel-2 remote sensing imagery, the river centerline is extracted. The elevation difference along the river centerline is extracted using a DEM (Digital Elevation Model). An elevation sequence is obtained through sampling. The elevation difference and corresponding distance are calculated to derive the slope. The calculation formula is as follows:

[0106] (2)

[0107] In the formula, S represents the gradient, Δh is the elevation difference, and L is the length of the river segment.

[0108] 6) Curvature

[0109] Sinuosity (SI) quantifies the degree of river channel tortuosity, reflecting energy dissipation and lateral erosion intensity. A sinuosity >1.5 indicates a meandering river, increasing friction and slowing flow velocity; 1.1-1.5 indicates a meandering river; and <1.1 indicates a straight river. In geomorphology, this parameter indicates the tortuous changes a river undergoes to adapt to geological conditions. The river centerline is extracted using Sentinel-2 remote sensing imagery, and the total length of the river centerline is calculated. The distance between the starting and ending points is then calculated as the straight-line distance to the valley floor. Sinuosity is the ratio of the actual river channel length to the straight-line distance to the valley floor, calculated using the following formula:

[0110] (3)

[0111] In the formula, SI represents the curvature, and L... channel L is the total length of the river's centerline. valley This is the straight-line distance to the valley floor.

[0112] 7) Width-to-depth ratio

[0113] The width-to-depth ratio (W / D) is the ratio of the width of the main channel to the average depth of the main channel. It is a key dimensionless parameter in riverbed geomorphology that characterizes the cross-sectional morphology of a river. The calculation formula is as follows:

[0114] (4)

[0115] In the formula, W / D is the width-to-depth ratio, W bankfull D is the width of the main channel of the river. bankfull The average depth of the main channel.

[0116] River width was extracted using Sentinel-2 remote sensing imagery. The median river width from multi-year imagery was used as the main channel width. Water depth was measured using DEM or altimeter satellite data sources as the average depth. A schematic diagram showing the locations of the main channel width and average depth is shown below. Figure 8 .

[0117] Step 4: Construction of an AI-based dual-encoder framework: Utilizing a contrastive learning method that combines physical information, the digitized river cross-sections obtained in Step 1 and the cross-sectional morphological feature parameters obtained in Step 3 are mapped to the same embedding space. Furthermore, a hard threshold is set for the cross-sectional morphological features in Step 3, based on the river cross-section category from Step 2. Specifically, the digitized river cross-sections obtained in Step 1 are processed using a Transformer encoder to extract geometric features, while the cross-sectional morphological feature parameters in Step 3 are processed using a multi-layer fully connected neural network. The final result is a mapping model between the digitized river cross-sections and their morphological feature parameters.

[0118] The specific construction steps are as follows:

[0119] 1) Construction of a digital river cross-section geometric encoder based on Transformer architecture

[0120] For the digital river cross-section sequence obtained in step one The spatial topological features are extracted using a Transformer encoder. To enable the model to handle the spatial order of cross-sectional points, positional encoding is used to embed the input coordinates. The positional encoding is performed according to the following formula:

[0121] (5)

[0122] (6)

[0123] In the formula, Indicates the index position of the cross-section point in the sequence; Index representing the embedded dimension; This represents the total dimension of the encoder's feature embeddings; This represents the generated position encoding vector;

[0124] A digital river cross-section geometry encoder that captures key geometric features of the cross-section using a multi-head self-attention mechanism. The calculation formula is as follows:

[0125] (7)

[0126] In the formula, This represents the geometric feature matrix after processing by the self-attention layer. Represents the input cross-sectional coordinate sequence matrix; These represent the weight matrices for the query, key, and value learned by the model, respectively. This represents the dimension of the key vector, used to scale the dot product result to prevent gradient vanishing.

[0127] 2) Construction of a morphological feature parameter encoder based on a multi-layer fully connected neural network

[0128] A morphological feature parameter encoder based on a multilayer fully connected neural network is used to construct the cross-sectional morphological feature parameter vector obtained in step three. Construct an encoder based on a multilayer fully connected neural network (MLP). Project it into an embedding space consistent with the geometric features.

[0129] Morphological feature parameter encoder of multilayer fully connected neural network The specific calculation formula is as follows:

[0130] (8)

[0131] In the formula, This represents the embedding vector generated after encoding the morphological feature parameters; This represents the input vector of morphological feature parameters output from step three; This represents the weight matrix of each layer's connections; Represents the bias term vector for each layer; Represents the nonlinear activation function ReLU.

[0132] 3) Setting hard judgment thresholds based on physical information classification

[0133] This step introduces the river cross-section category from step two. We set a hard threshold penalty term based on physical meaning for morphological features to ensure that the mapping relationship generated by the model does not violate the basic principles of geomorphology.

[0134] The physical constraint loss function is calculated using the following formula:

[0135] (9)

[0136] In the formula, This represents the physical constraint loss value; This represents the model's prediction or mapping of the first... Each morphological feature parameter value; These represent the river categories respectively. The Hard upper and lower thresholds for each feature parameter; This indicates the total number of feature parameters.

[0137] 4) Joint training of contrastive learning mapping models by constructing

[0138] Comparison of loss functions, forced digitization of river cross-section geometric embedding vectors With morphological feature embedding vector Align within the embedded space.

[0139] The joint objective function is calculated using the following formula:

[0140] (10)

[0141] In the formula, Represents the total training loss function; This represents the contrastive learning loss, used to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs; The weight coefficients of the physical constraint terms are used to adjust the degree of influence of physical laws on the model learning process.

[0142] When the mapping relationship is established, When convergence is achieved, the nonlinear mapping function between the geometric and morphological features of the digital river cross-section is established as follows:

[0143] (11)

[0144] In the formula, This is the final constructed mapping model, through which the mapping relationship can be realized when the morphological parameters are known. Under the given conditions, retrieve highly reliable digital river cross-sectional sequences.

[0145] Step 5: Find the most similar cross section: Based on the morphological feature parameters of the unknown cross section obtained in Step 3, input the mapping relationship model obtained in Step 4, find the digital river cross section with the highest cosine similarity in Step 1, and determine whether it matches the river cross section type in Step 2. If it matches, it is the most similar cross section, stop the search, and proceed to Step 6. If it does not match, find the second most similar cross section (the cross section with the highest cosine similarity among the remaining cross sections) and continue to make judgments until you enter Step 6.

[0146] The vector of unknown cross-sectional morphological feature parameters obtained in step three Input the morphological feature parameters into the encoder trained in step four. In this process, the mathematical correlation between the cross-section and the geometric feature vectors of each candidate cross-section in the database is calculated:

[0147] (12)

[0148] (13)

[0149] In the formula, This represents the projection vector of the cross section to be investigated in the feature embedding space; This represents the pre-trained parameter encoder mapping operator in step four; This represents the input vector of the morphological feature parameters of the cross-section to be investigated extracted in step three; Represents the vector of the section to be investigated With the database Geometric feature vectors of each cross section Cosine similarity between them.

[0150] Step 6: Obtaining river width and elevation from satellite remote sensing and correcting water surface elevation: Based on the satellite remote sensing data source, water body classification is obtained. Combined with ALOS 12.5m DEM data, samples are taken at 10 m intervals, as shown in Table 3. According to physical laws, the elevations at both ends of the water surface should be consistent. Therefore, the lower value of the water body elevations at both ends of the multi-period images is taken as the corrected DEM data.

[0151] The improved Normalized Difference Water Index (MNDWI) is used to identify the extent of water bodies in the river channel and generate a water body mask matrix. :

[0152] (14)

[0153] In the formula, For the first Periodic images in coordinates Water body signage at the location; For adaptive threshold extraction.

[0154] Using ALOS 12.5m resolution DEM data, discrete sampling was performed at 10m spatial steps along the extracted water body boundaries to obtain the initial elevation values ​​of the boundary points on both sides of the water body. For any cross-section... Identify its left bank water boundary point Water boundary point on the right bank For water boundary elevations acquired from multiple image periods, the lower of the observations at both ends is taken as the baseline elevation for that period to eliminate the issue of artificial elevations in the DEM. The elevation is calculated using the following formula:

[0155] (15)

[0156] Corrected boundary elevation Map back to the river channel topography model and update the original DEM feature points.

[0157] SAR data is obtained from satellite remote sensing data sources. The width of the river surface is inverted using a hybrid pixel decomposition algorithm based on the SAR data, using the algorithm of formula (16).

[0158] Step 7: Adjusting the Geometry of the Unknown Cross-Section: Extract the most similar cross-section obtained in Step 5 into a pure shape between 0 and 1. Combine this with the corrected DEM from Step 6, and take the elevation values ​​according to weights to obtain the basic geometry of the unknown cross-section. Use gradient calculation to find the flattest central region of the riverbed. Create a Gaussian weighted distribution in the river center region. Adjust the basic geometry by multiplying the water depth by the Gaussian weight. Finally, use cubic spline interpolation to smooth and refine the unknown cross-section shape after water depth adjustment, obtaining the river cross-section result of the unknown cross-section, such as... Figure 5 As shown.

[0159] The specific process is as follows:

[0160] The specific steps for adjusting the geometry of the unknown cross-section in step seven are as follows:

[0161] (1) The most similar digital river cross-section sequence obtained in step five Dimensionless processing is performed to extract the pure shape that reflects the riverbed evolution characteristics, and this shape is mapped to the water surface width determined in step six. and elevation datum In space, obtain the basic geometric shape

[0162] (17)

[0163] (18)

[0164] In the formula, This represents the horizontal position vector after scaling. Represents the vertical position vector after scaling; Indicates by Extract the i-th value of the normalized horizontal and vertical axes; , This represents the maximum and minimum horizontal distances of the original sequence of similar cross sections; , This represents the maximum and minimum elevation values ​​of the original sequence of similar cross sections; This represents the river width obtained in step six; Step 6 outputs the corrected river boundary reference elevation; This indicates the preset river depth at the target cross-section.

[0165] (2) To simulate the real deep channel scouring characteristics of natural rivers, the flattest area of ​​the riverbed was identified, and a Gaussian function was used to nonlinearly adjust the depth of this area. This was achieved by... Perform first-order gradient calculation Identify gradient values ​​less than a preset threshold A continuous region, and determine the lateral center coordinates of that region. The Gaussian weighted water depth superposition is performed according to the following formula:

[0166] (19)

[0167] In the formula, This represents the local geometric gradient at the i-th sampling point; , This represents the i-th horizontal position and the base elevation value obtained after scaling transformation; , This represents the (i+1)th horizontal position and the base elevation value obtained after scaling.

[0168] Identify gradient values Less than the preset threshold A continuous region, and determine the lateral center coordinates of that region. The Gaussian weighted water depth superposition is performed according to the following formula:

[0169] (20)

[0170] In the formula, This represents the final corrected elevation vector after completing the simulation of the underwater morphology. This represents the base elevation value obtained after scaling. This indicates the reference average water depth obtained from hydrological data; The span parameter represents the Gaussian distribution, used to control the attenuation range of the influence of water depth correction on both banks of the riverbed; the final corrected elevation vector represents the completed simulation of the underwater morphology. This indicates the horizontal position obtained after scaling.

[0171] (3) To ensure that the generated cross-section has continuous physical curvature and meets the accuracy requirements of engineering calculations, a cubic spline interpolation algorithm is used to fit the discrete point set to ensure that the second derivative is continuous at all sampling nodes and in any adjacent sampling point interval. Inside, elevation value The calculation formula is as follows:

[0172] (twenty one)

[0173] In the formula, express The i-th elevation value; , express The horizontal distance between the i-th and i+1-th sampling points; The horizontal distance is the independent variable; The coefficients are polynomials, determined by the elevation at the nodes, the first derivative, the second derivative, and the third derivative.

[0174] According to the preset encryption factor Equal-interval encrypted sampling is performed to generate the final river cross-section result sequence. The calculation formula is as follows:

[0175] (twenty two)

[0176] In the formula, This represents the final three-dimensional digital river channel result; This represents the horizontal distance value after resampling; This represents the longitudinal elevation value obtained from the mapping; , This represents the horizontal distance between the start and end points of the cross-section; express The total number of sampling points; This represents the cubic spline interpolation function; Indicates the encryption factor.

[0177] The following is the relevant data for this embodiment:

[0178] Table 1 shows a partial database of cross-sectional parameters.

[0179] Table 2 shows the cross-sectional parameters of the Changhua River in Hainan Province and the most similar cross-sectional information obtained from the query.

[0180] Table 3 shows the DEM and river binary classification results obtained at different distances along the Changhua River section in Hainan Province.

[0181] Table 1

[0182] 2023092101 102.7995 35.3471 Gansu Daxia River prime of life Plains No floodplain Symmetrical parabola 2023092102 103.2588 35.632 Gansu Daxia River prime of life Plains right floodplain Steep and gentle slopes 2023092103 108.594539 34.635958 Shaanxi Jinghe River prime of life Plains No floodplain Offset trapezoid 2023102801 115.895617 40.033392 Beijing Yongding River prime of life hills No floodplain Offset trapezoid 2024071401 116.034867 39.975328 Beijing Yongding River prime of life hills No floodplain lateral parabola 2024082201 110.3722247 19.74271216 Hainan Nandujiang prime of life Plains No floodplain Steep and gentle slopes 2024082202 110.3407954 19.16290795 Hainan Wanquan River prime of life Plains No floodplain Steep and gentle slopes 2024082302 109.1921561 18.39984851 Hainan Ningyuan River prime of life Plains No floodplain rectangle 2024082401 108.9075516 19.22187937 Hainan Changhua River prime of life Plains No floodplain lateral parabola 2024082402 108.9818359 19.46934932 Hainan Zhubijiang prime of life Plains No floodplain lateral parabola 2024082403 109.6905212 19.88074086 Hainan Wenlan River prime of life Plains No floodplain Steep and gentle slopes 2024112001 94.41226662 29.78802212 Tibet Eight and Eight Melodies childhood mountain No floodplain Multi-point groove 2024112002 94.40615804 29.76386778 Tibet Eight and Eight Melodies childhood mountain No floodplain rectangle 2024112003 94.40301777 29.75659667 Tibet Eight and Eight Melodies childhood mountain No floodplain Flat bottom narrow canyon 2024112101 94.46336999 29.4513744 Tibet Niyangqu old age mountain No floodplain Trapezoidal center 2024112103 94.44886829 29.45924413 Tibet Niyangqu old age mountain No floodplain Trapezoidal center 2025032702 94.37121087 29.59522898 Tibet Yinjiulongbaqu childhood mountain No floodplain Parabolic trough 2025062104 108.98 19.16 Hainan Changhua River prime of life Plains No floodplain Symmetrical trapezoid 2025062102 108.98 19.17 Hainan Gonglao River prime of life Plains No floodplain Strong side projectile 2025062101 109.05 19.25 Hainan Shilu River prime of life Plains No floodplain Symmetrical parabola 2025062105 109.16224 18.7719 Hainan Changhua River prime of life Plains No floodplain Lateral trapezoid 2025062202 109.89 19.1 Hainan Ding'an River prime of life Plains right floodplain Symmetrical trapezoid 2025062107 109.82 18.61 Hainan Lingshui River prime of life Plains No floodplain Trapezoidal center 2025062108 109.94 18.62 Hainan Duzong River prime of life hills No floodplain Lateral trapezoid 2025061801 110.39 19.69 Hainan Xunya River prime of life hills No floodplain Symmetrical trapezoid 2025061802 110.34 19.7 Hainan Wencun Water prime of life Plains Left floodplain Offset trapezoid 2025062003 110.198 19.74 Hainan Nandujiang old age Plains No floodplain Trapezoidal center 2025062002 110.22 19.63 Hainan Xinwuxi prime of life hills right floodplain Symmetrical trapezoid 2025062106 109.16 18.75 Hainan Da'an River prime of life hills right floodplain Multi-point groove 2025061001 88.036 31.303 Tibet Gerengcuo prime of life mountain No floodplain lateral parabola 2025062604 88.395 31.805 Tibet Rooted in Tibet prime of life hills No floodplain Lateral trapezoid 2025062501 88.225 30.9775 Tibet Baru Tsangpo prime of life hills Left floodplain Lateral V 2024101901A 118.76887 39.757896 Hebei Luanhe River old age Plains Left floodplain Symmetrical parabola 2025061501 118.850377 39.893036 Hebei Qinglong River prime of life Plains No floodplain lateral parabola 2025061502 118.313221 40.179521 Hebei Luanhe River prime of life mountain No floodplain Symmetrical trapezoid 2025061503 118.357387 40.585498 Hebei Waterfall River prime of life mountain right floodplain Mid-channel parabolic 2025061504 118.599531 40.730934 Hebei Waterfall River prime of life mountain right floodplain lateral parabola 2024102001A 118.14134 40.620907 Hebei Luanhe River prime of life mountain No floodplain Symmetrical parabola 2025061601 118.130166 40.645159 Hebei Liuhe prime of life mountain No floodplain Lateral trapezoid 2025061602 118.063001 40.846651 Hebei Luanhe River prime of life mountain right floodplain Mid-channel parabolic 2025061603 117.944474 41.02899 Hebei Wulie River prime of life mountain Left floodplain Parabolic trough 2025061604 117.76996 40.990648 Hebei Eson River prime of life mountain No floodplain Mid-channel parabolic 2025061701 118.372003 40.55313 Hebei Eson River prime of life mountain No floodplain Mid-channel parabolic 2025061702 117.098256 41.581552 Hebei Luanhe River prime of life mountain right floodplain lateral parabola 2025061901 110.06 19.11 Hainan Ding'an River prime of life Plains No floodplain Strong side projectile 2025061903 110.07 19.085 Hainan Ding'an River prime of life Plains Left floodplain Mid-channel parabolic 2025062008 109.04 19.43 Hainan Zhubijiang prime of life Plains No floodplain Inner Beach Parabolic 2025062007 109.09 19.43 Hainan Daling River prime of life Plains No floodplain Lateral V 2025061101 88.86355876 30.5724688 Tibet Rooted in Tibet prime of life Plains Left floodplain Parabolic trough

[0183] Results of cross-sectional parameters for the Changhua River in Hainan Province and information on the most similar cross-section:

[0184] Table 2

[0185] Changhua River section 108.98 19.16 Hainan Changhua River prime of life Plains No floodplain Symmetrical trapezoid Most similar cross section 1 110.22 19.63 Hainan Xinwuxi prime of life Plains right floodplain Symmetrical trapezoid Most similar cross section 2 110.39 19.69 Hainan Xunya River prime of life Plains No floodplain Symmetrical trapezoid Most similar cross section 3 110.26 19.14 Hainan Wanquan River prime of life hills No floodplain Steep and gentle slopes

[0186] Table 3

[0187] 0 108.9774 19.1578 84 0 0 0 0 0 10 108.9775 19.1579 84 0 0 0 0 0 20 108.9775 19.1579 80 0 0 0 0 0 30 108.9776 19.1580 80 0 0 0 0 0 40 108.9777 19.1581 80 0 0 0 0 0 50 108.9777 19.1581 74 0 0 0 0 0 60 108.9778 19.1582 70 0 0 0 0 0 70 108.9779 19.1582 70 0 0 0 0 0 80 108.9780 19.1583 70 0 0 0 0 0 90 108.9780 19.1584 62 0 0 0 1 0 100 108.9781 19.1584 55 1 1 1 1 1 110 108.9782 19.1585 55 1 1 1 1 1 120 108.9782 19.1586 55 1 1 1 1 1 130 108.9783 19.1586 44 1 1 1 1 1 140 108.9784 19.1587 44 1 1 1 1 1 150 108.9785 19.1587 34 1 1 1 1 1 160 108.9785 19.1588 34 1 1 1 1 1 170 108.9786 19.1589 34 1 1 1 1 1 180 108.9787 19.1589 34 1 1 1 1 1 190 108.9787 19.1590 34 1 1 1 1 1 200 108.9788 19.1591 34 1 1 1 1 1 210 108.9789 19.1591 34 1 1 1 1 1 220 108.9790 19.1592 34 1 1 1 1 1 230 108.9790 19.1592 34 1 1 1 1 1 240 108.9791 19.1593 34 1 1 1 1 1 250 108.9792 19.1594 34 1 1 1 1 1 260 108.9792 19.1594 34 1 1 1 1 1 270 108.9793 19.1595 34 1 1 1 1 1 280 108.9794 19.1596 34 1 1 1 1 1 290 108.9795 19.1596 34 1 1 1 1 1 300 108.9795 19.1597 34 1 1 1 1 1 310 108.9796 19.1597 34 1 1 1 1 1 320 108.9797 19.1598 47 1 1 1 1 1 330 108.9797 19.1599 47 1 1 1 1 1 340 108.9798 19.1599 47 1 1 1 1 1 350 108.9799 19.1600 47 1 1 1 1 1 360 108.9800 19.1601 53 1 1 1 1 1 370 108.9800 19.1601 53 1 1 1 1 1 380 108.9801 19.1602 53 0 0 1 0 1 390 108.9802 19.1602 66 0 0 0 0 0 400 108.9802 19.1603 78 0 0 0 0 0 410 108.9803 19.1604 78 0 0 0 0 0 420 108.9804 19.1604 78 0 0 0 0 0 430 108.9805 19.1605 88 0 0 0 0 0

[0188] In this calculation case, the three most similar cross sections to the Changhua River in Hainan Province are all in Hainan Province, which is consistent with the relevant theory of hydrogeographic similarity in riverbed geomorphology. At the same time, they have the same or similar cross section types as the Changhua River in Hainan Province, all of which are trapezoidal cross sections. This is consistent with the river development knowledge graph constructed based on the principles of riverbed geomorphology and hydrodynamics, where the most similar cross section and the simulated cross section have the same river cross section type.

[0189] The coefficient of determination R is obtained by calculating the accuracy of the scatter plots of the fitted river cross section and the field verification section. 2 The mean square error (RMSE) is 0.944, the Nash efficiency coefficient (NSE) is 3.732 m, and the mean square error (MAE) is 2.706 m. These indicators show that the simulated river cross-sections obtained by this method are close to the measured river cross-sections. The river cross-section morphology recognition method coupled with artificial intelligence and riverbed geomorphology can effectively identify river cross-section types and simulate the basic morphology of river cross-sections.

[0190] Finally, it should be noted that the above description is only used to illustrate the technical solutions of the present invention and is not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention.

Claims

1. A method for recognizing river cross-section morphology by coupling artificial intelligence with riverbed geomorphology, characterized in that: include: Step 1: Complete Riverbed Topographic Survey by UAV: ​​Acquire UAV image data of the monitored riverbed to obtain the elevation changes above the water surface, use sonar depth sounding to measure the underwater topography, fit the elevation changes above the water surface and the underwater topography to obtain a digital river cross section; all measured digital river cross sections form a cross section database. Step 2: Construction of a knowledge graph of river cross-section types: Classify the morphology of the digitized river cross-sections measured in Step 1 to obtain river cross-section categories; Step 3: Construction of cross-sectional morphological characteristic parameters: For each river cross section, construct cross-sectional morphological characteristic parameters; Step 4: Construction of a dual encoder framework based on artificial intelligence: The digital river cross section obtained in Step 1 and the cross section morphology feature parameters obtained in Step 3 are mapped together into the same embedding space. Combined with the river cross section category in Step 2, a judgment threshold is set for the cross section morphology feature parameters; thus, a mapping relationship model between the digital river cross section and the morphology feature parameters is obtained. Step 5: Find the most similar cross section to the unknown cross section: Obtain the morphological feature parameters of the unknown cross section, input the mapping relationship model obtained in Step 4, and search for the digital river cross section with the highest cosine similarity to the unknown cross section based on the database in Step 1, and make a judgment: The judgment is to determine whether it meets the river cross section category in Step 2. If it does, it is the most similar cross section, the search stops, and Step 6 is performed. If it does not meet, the cross section with the highest cosine similarity among the remaining cross sections is searched, and the judgment continues. Step Six: Obtaining River Width and Elevation via Satellite Remote Sensing and Correcting Water Surface Elevation; Step 7: Adjustment of the geometry of the unknown cross section: Based on the most similar cross section obtained in Step 5 and the elevation corrected in Step 6, the basic geometry of the unknown cross section is obtained, and the water depth is adjusted, smoothed and densified to obtain the river cross section result of the unknown cross section.

2. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 1, characterized in that: The UAV imagery data in step one includes image data showing the shape and elevation of the target river channel and information on nearby features.

3. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 1, characterized in that: In step two, the river cross-section categories include: V-shaped, parabolic, and trapezoidal. The V-shaped category includes flat-bottomed narrow canyons, flat-bottomed widening, multi-point grooves, connected grooves, and lateral V-shaped sections. The parabolic category includes symmetrical parabolic sections, central channel parabolic sections, lateral channel parabolic sections, lateral parabolic sections, strongly lateral parabolic sections, inner shoal parabolic sections, and central bulges. The trapezoidal category includes symmetrical trapezoidal sections, central channel trapezoidal sections, lateral channel trapezoidal sections, lateral trapezoidal sections, steep and gentle slopes, and rectangles.

4. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 1, characterized in that: In step three, the cross-sectional morphological characteristic parameters include: the river development stage, landform type, floodplain type, embedment rate, curvature, gradient, and width-to-depth ratio.

5. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 4, characterized in that: Valley development stages: dividing river evolution into juvenile, mature, and senile stages to reveal the dynamic balance between river erosion and deposition; Landform types: divided into mountains, hills and plains, representing the slope, flow velocity and hydrodynamic effects of rivers; Floodplain types: divided into no floodplain, double-sided floodplain, left-side floodplain, and right-side floodplain; Embedding rate: is the ratio of the width of the floodplain to the width of the main channel, used to quantify the degree of vertical constraint of the river channel; Gradient: Indicates the slope of a river section, which is the ratio of the elevation difference to the length of the river section; Sincerity: The ratio of the actual length of the river channel to the straight-line distance to the valley floor; Width-to-depth ratio: This is the ratio of the width of the main channel of a river to its average depth.

6. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 1, characterized in that: In step four, the specific steps for constructing the AI-based dual encoder framework are as follows: 1) Construction of a digital river cross-section geometric encoder based on the Transformer architecture: For the digital river cross-section sequence obtained in step one p i For the i-th distance-elevation point pair in a cross-section sequence, extract its spatial topological features using a Transformer encoder; To enhance the model's ability to handle the spatial order of cross-section points, position encoding is performed on the input coordinates using the following formula: （5） （6） In the formula, Indicates the index position of the cross-section point in the sequence; Index representing the embedded dimension; This represents the total dimension of the encoder's feature embeddings; This represents the generated position encoding vector; A digital river cross-section geometry encoder captures key geometric features of the cross-section using a multi-head self-attention mechanism. The calculation formula is as follows: （7） In the formula, This represents the geometric feature matrix after processing by the self-attention layer. Represents the input cross-sectional coordinate sequence matrix; These represent the weight matrices for the query, key, and value learned by the model, respectively. The dimension of the key vector is used to scale the dot product result to prevent gradient vanishing; T represents the transpose matrix. 2) Construction of a morphological feature parameter encoder based on a multi-layer fully connected neural network: Regarding the cross-sectional morphology feature parameter vector obtained in step three Construct an encoder based on a multilayer fully connected neural network (MLP). That is, to obtain the morphological feature parameter encoder based on a multi-layer fully connected neural network. Project it into an embedding space consistent with the geometric features; Morphological feature parameter encoder of multilayer fully connected neural network The specific calculation formula is as follows: （8） In the formula, This represents the embedding vector generated after encoding the morphological feature parameters; This represents the input vector of morphological feature parameters output from step three; This represents the weight matrix of each layer's connections; Represents the bias term vector for each layer; Represents the nonlinear activation function ReLU; 3) Setting hard judgment thresholds based on physical information classification: This step introduces the river cross-section category from step two. We set a judgment threshold based on physical meaning for morphological features, and construct a physical constraint loss function for the model based on this threshold. The physical constraint loss function is calculated using the following formula: （9） In the formula, This represents the physical constraint loss value; This represents the model's prediction or mapping of the first... Each morphological feature parameter value; These represent the river categories respectively. The Hard upper and lower thresholds for each feature parameter; Indicates the total number of feature parameters; 4) Joint training of the contrastive learning mapping model is achieved by constructing: Comparison loss functions, forced digitization of river cross-section geometric embedding vectors With morphological feature embedding vector Align within the embedded space; The joint objective function is calculated using the following formula: （10） In the formula, Represents the total training loss function; This represents the contrastive learning loss, used to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs; The weighting coefficients of the physical constraint terms are used to adjust the degree of influence of physical laws on the model learning process; When the mapping relationship is established, When convergence is achieved, the nonlinear mapping function between the geometric and morphological features of the digital river cross-section is established as follows: （11） In the formula, This is the final mapping relationship model that is constructed.

7. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 6, characterized in that: The specific formula for cosine similarity mentioned in step five is as follows: （12） （13） In the formula, This represents the projection vector of the cross section to be investigated in the feature embedding space; The input vector represents the extracted morphological feature parameters of the cross-section to be investigated. Represents the vector of the section to be investigated With the database Geometric feature vectors of each cross section Cosine similarity between them.

8. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 1, characterized in that: The specific formulas for satellite remote sensing elevation acquisition and water surface elevation correction mentioned in step six are as follows: The improved Normalized Differential Water Index (MNDWI) is used to identify the extent of water bodies in river channels and generate a water body mask matrix. : （14） In the formula, For the first Periodic images in coordinates Water body signage at the location; For adaptive threshold extraction; Using ALOS 12.5m resolution DEM data, discrete sampling was performed on the extracted water body boundaries with a spatial step size of 10 m to obtain the initial elevation values ​​of the boundary points on both sides of the water body; for any cross section Identify its left bank water boundary point Water boundary point on the right bank For water boundary elevations acquired from multiple image periods, the lower of the observations at both ends is taken as the baseline elevation of the water surface in that period to eliminate the problem of artificial elevation in the DEM. The elevation is calculated using the following formula: （15） In the formula, This represents the corrected water surface reference elevation of the k-th cross section under the image of period t; , These represent the sampling positions of the left and right bank waterline of the k-th section in the t-th image, respectively; This represents the elevation function of the corresponding coordinate points obtained from the original ALOS DEM. Corrected boundary elevation Map back to the river channel topography model and update the original DEM feature points.

9. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 1, characterized in that: The specific steps for obtaining the basic geometric shape of the unknown cross-section in step seven are as follows: The sequence of the most similar cross-sections obtained in step five... Dimensionless processing is performed to extract the pure shape reflecting the riverbed evolution characteristics, which is a pure shape between 0 and 1. This shape is then mapped to the water surface width determined in step six. and elevation datum In space, obtain the basic geometric shape （17） （18） In the formula, Represents the horizontal position vector after scaling; Represents the vertical position vector after scaling; Indicates by Extract the i-th value of the normalized horizontal and vertical axes; , This represents the maximum and minimum horizontal distances of the original sequence of similar cross sections; , This represents the maximum and minimum elevation values ​​of the original sequence of similar cross sections; This represents the river width obtained in step six; Step 6 outputs the corrected river boundary reference elevation; This indicates the preset river depth of the target cross-section.

10. The method for recognizing river cross-section morphology by coupling artificial intelligence and riverbed geomorphology according to claim 1, characterized in that: Step seven describes adjusting the water depth based on the basic geometry: using gradient calculations to find the flattest central region of the riverbed, creating a Gaussian weighted distribution in the central region, and multiplying the water depth by the Gaussian weights. The specific operation is as follows: To simulate the real deep-channel scouring characteristics of natural rivers, the flattest areas of the riverbed were identified, and nonlinear depth adjustment was applied to these areas using a Gaussian function. The local slope of the basic geometric morphology sequence was calculated using first-order difference. Identify flat areas at the bottom of the riverbed and determine the lateral center location. , The calculation formula is as follows: （19） In the formula, This represents the local geometric gradient at the i-th sampling point; , This represents the i-th horizontal position and the base elevation value obtained after scaling transformation; , This represents the (i+1)th horizontal position and the base elevation value obtained after scaling. Identify gradient values Less than the preset threshold A continuous region, and determine the lateral center coordinates of that region. The Gaussian weighted water depth superposition is performed according to the following formula: （20） In the formula, This represents the final corrected elevation vector after completing the simulation of the underwater morphology. This represents the base elevation value obtained after scaling. This indicates the reference average water depth obtained from hydrological data; The span parameter represents the Gaussian distribution and is used to control the attenuation range of the effect of water depth correction on both banks of the riverbed. This represents the final corrected elevation vector after completing the simulation of the underwater morphology. This indicates the horizontal position obtained after scaling.

Images