Channel area erosion and deposition evolution quantitative evaluation system based on convolutional neural network and graph convolutional network

Abstract

The invention discloses a channel area erosion and deposition evolution quantitative evaluation system based on a convolutional neural network and a graph convolutional network, and belongs to the technical field of marine surveying and mapping, underwater terrain dynamic monitoring and artificial intelligence crossing. The system comprises a data processing module used for calculating the uncertainty of sounding points and generating a comparison point set with optimal water depth estimation; the feature engineering module is used for fusing multi-source heterogeneous data to construct an enhanced feature vector; a built-in multi-scale spatial-temporal feature fusion network of the terrain erosion and deposition evolution intelligent analysis module synchronously outputs the erosion and deposition type, the evolution rate and the uncertainty of each point according to the feature vectors; and the verification output module is used for verifying a result and generating a thematic map and a statistical report. According to the method, full-process automation and quantification of channel erosion and deposition evolution from data processing to intelligent evaluation are realized, and the analysis efficiency, precision and decision support capability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of marine surveying, underwater topographic dynamic monitoring and artificial intelligence, specifically to a quantitative assessment system for the scour and sedimentation evolution of waterways based on convolutional neural networks and graph convolutional networks. Background Technology

[0002] The waterway area is subject to multiple influences from runoff, tides, storms, and human engineering activities, resulting in rapid seabed topographic evolution and strong spatial heterogeneity, posing a continuous challenge to navigation safety and maintenance dredging projects. Single-beam echo sounding remains the primary method for periodic topographic monitoring in this area due to its ease of operation and low cost. However, the dynamic marine environment and the difficulty in fully replicating ship tracks lead to inherent differences in the locations of echo sounding points acquired at different times, resulting in low accuracy and insufficient reliability of traditional comparison methods based on fixed points or simple spatial interpolation.

[0003] Existing technologies, such as "Multi-Period Water Depth Profiling Analysis Method and Application with Joint Uncertainty," improve the reliability of multi-period data comparison by introducing measurement uncertainty propagation and Kalman filtering. However, such methods still have significant limitations: First, the determination of "significant changes" in terrain relies on globally unified statistical hypothesis testing, which cannot adapt to local complex terrain and data quality changes; second, for detected areas of change, manual interpretation of scour and deposition types and rate estimation are still required, resulting in low levels of automation and intelligence, making it difficult to meet the rapid analysis needs of large-scale, high-frequency monitoring; third, there is a lack of a model framework capable of comprehensively considering water depth values, measurement errors, spatial correlations, temporal effects, and external environmental driving factors for end-to-end intelligent decision-making.

[0004] To address the aforementioned issues, there is an urgent need for a quantitative assessment system for the scouring and sedimentation evolution of waterways based on convolutional neural networks and graph convolutional networks, which can solve the problems existing in traditional methods. Summary of the Invention

[0005] The purpose of this invention is to provide a quantitative assessment system for the evolution of scour and sedimentation in waterways based on convolutional neural networks and graph convolutional networks. This system automates and quantifies the entire process of data processing and intelligent assessment of waterway scour and sedimentation evolution, significantly improving analysis efficiency, accuracy, and decision support capabilities.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A quantitative assessment system for the scouring and sedimentation evolution of waterways based on convolutional neural networks and graph convolutional networks includes: The data processing module is used to receive raw data from multiple single-beam bathymetry measurements and auxiliary sensor data, calculate the initial uncertainty of each bathymetry point through the uncertainty propagation model, generate a set of comparison points along the preset measurement line, and use the Kalman filter algorithm to fuse information from neighboring bathymetry points, outputting the optimal water depth estimate and the corresponding water depth estimate uncertainty of each comparison point in each measurement period. The feature engineering module, connected to the data processing module, is used to fuse multi-source heterogeneous data for comparison points to construct enhanced feature vectors. The multi-source heterogeneous data includes at least: water depth and uncertainty information output by the data processing module, channel dredging engineering vector data obtained from the public government platform, hydrodynamic and suspended sediment concentration data obtained from the marine reanalysis database and remote sensing products, and spatial and geomorphological context features calculated based on basic geographic information data. The intelligent analysis module for topographic erosion and deposition evolution is connected to the feature engineering module. It has a built-in multi-scale spatiotemporal feature fusion network to receive enhanced feature vectors and simultaneously output the erosion and deposition type classification results, erosion and deposition evolution rate estimates, and corresponding rate estimation uncertainties for each comparison point. The verification output module is connected to the intelligent analysis module for terrain erosion and sedimentation evolution, and is used to verify the accuracy and reliability of the model output results.

[0007] Furthermore, the data processing module includes: The uncertainty calculation unit is used to calculate the initial horizontal and vertical uncertainties of each sounding point based on the standards of the International Hydrographic Organization and the error propagation formula, taking into account GNSS positioning error, attitude measurement error, sound velocity profile error, tide level error and the depth sounder's own error. The comparison point generation unit is used to generate a series of comparison points along the planned survey line at preset intervals, and calculate the influence radius based on the uncertainty and distance of the sounding points, and associate the sounding points within the influence radius with the corresponding comparison points; The data fusion unit uses a Kalman filter to perform recursive optimal estimation on multiple sounding point sequences associated with the same comparison point, and updates the optimal water depth estimate and final estimation uncertainty of the comparison point in the current period.

[0008] Furthermore, the enhanced feature vector constructed by the feature engineering module includes: The core measurement features include the water depth estimates, water depth estimation uncertainties, and water depth changes of the comparison point itself in the two measurement periods; Engineering activity characteristics, including whether the comparison point is located within the known waterway dredging project area and its closest distance to the project boundary; Environmental dynamic characteristics, including the average flow velocity, dominant flow direction and average surface suspended sediment concentration at the comparison point location obtained by interpolation during the corresponding measurement period; Spatial contextual features include the coordinates of the comparison point, its distance from the shoreline, its distance from the river mouth, and the coding of the geomorphic unit type to which it belongs.

[0009] Furthermore, the classification of the geomorphic unit types is based on official nautical charts, waterway improvement planning maps, or historical research results, and the types include at least the main channel, waterway slope, side beach, and deep channel.

[0010] Furthermore, the multi-scale spatiotemporal feature fusion network includes: The local feature extraction branch, composed of a one-dimensional convolutional neural network, is used to process the sequential features centered on the target comparison point and composed of its spatial nearest neighbors to extract micro-topography spatial patterns. The global spatial relationship branch consists of two layers of graph convolutional networks. It transmits information based on the Delaunay triangulation constructed from all comparison points and aggregates multi-order neighbor information of nodes to capture regional spatial dependencies. The temporal feature extraction branch, including multi-head self-attention mechanism and feedforward network, is used to perform deep modeling of time span and temporal context encoding to capture temporal nonlinear effects; The feature fusion and interaction module is used to concatenate the output features of the three branches, learn higher-order feature interactions using an explicit feature cross-network, and finally generate high-level comprehensive features through a fully connected layer. The multi-task output head includes a classification head for outputting the probability of scour or siltation type and a regression head for simultaneously outputting the mean of the evolution rate and its log-variance.

[0011] Furthermore, the local feature extraction branch specifically includes: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, and a fully connected layer.

[0012] Furthermore, the intelligent analysis module for terrain erosion and sedimentation evolution employs a weighted multi-task loss function L during training. total =λ*L cls +(1-λ)*L reg In the formula, L cls For the focus loss function optimized for class imbalance, L reg Let λ be the loss function for heteroscedastic regression based on uncertainty, and λ be the weighting coefficient.

[0013] Furthermore, the verification output module includes: The confidence screening unit is used to estimate uncertainty and classification probability based on the rate output of the model, set a threshold, and automatically identify evaluation results with low confidence. The external validation unit is used to compare the scouring and deposition distribution and rate output by the model with the real changes derived from the high-precision multibeam terrain data of the same period, and to calculate the classification accuracy, root mean square error and coefficient of determination to quantify the accuracy. The visualization mapping unit is used to automatically generate spatial distribution maps of scour and sedimentation types, contour maps of evolution rates, uncertainty distribution maps, and comprehensive analysis maps along key profiles.

[0014] In summary, the present invention has at least one of the following beneficial technical effects: 1. Achieved full-process automation and intelligence: Through the pipeline design of data processing - feature construction - model evaluation - visualization, it has completely changed the traditional manual interpretation and inefficient mode, realized the batch and automated processing of waterway scour and siltation assessment, and significantly improved the efficiency of operation.

[0015] 2. Improved accuracy and reliability of evaluation results: This is based on the rigorous quantification and propagation of single-beam measurement uncertainty, ensuring the reliability of data input. The innovative multi-scale spatiotemporal feature fusion network can simultaneously capture local details, spatial correlations, and temporal effects, and introduces uncertainty estimation, enabling the model to self-assess prediction reliability, resulting in more reliable and interpretable output results.

[0016] 3. Deep integration of multi-source information and decision support: The system creatively integrates multi-source information such as measurement data, engineering records, hydrodynamic environment, and remote sensing products to construct features, enabling the artificial intelligence model to have physical cognitive ability close to expert experience and strong decision support.

[0017] 4. Significant practical value and promising prospects for application: The system output includes quantitative rates, siltation and sedimentation volumes, thematic maps, and uncertainty indicators, which can be directly used for channel dredging planning, maintenance effectiveness assessment, and navigation safety early warning, providing a powerful technical tool for port and waterway management departments. This methodology can be extended to the monitoring and research of other areas with rapidly changing underwater topography, such as estuaries, coastlines, and reservoirs. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the branch structure for local feature extraction; Figure 3 This is a schematic diagram of the global spatial relationship branch structure; Figure 4 This is a schematic diagram of the branch structure for temporal feature extraction; Figure 5 This is a schematic diagram of the feature fusion and interaction module structure; Figure 6 This is a cross-sectional view of water depth changes; Figure 7 This is a map showing the distribution of scour and sedimentation rates along the course of the river. Figure 8 A graph showing the distribution of uncertainty in model predictions along the path; Figure 9 A distribution map of the probability along the erosion and siltation type classification. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] like Figure 1 As shown, this invention provides a quantitative assessment system for the evolution of scour and sedimentation in waterway areas based on convolutional neural networks and graph convolutional networks, comprising: The data processing module is used to receive raw data from multiple single-beam bathymetry measurements and auxiliary sensor data, calculate the initial uncertainty of each bathymetry point through the uncertainty propagation model, generate a set of comparison points along the preset measurement line, and use the Kalman filter algorithm to fuse information from neighboring bathymetry points, outputting the optimal water depth estimate and the corresponding water depth estimate uncertainty of each comparison point in each measurement period. The feature engineering module, connected to the data processing module, is used to fuse multi-source heterogeneous data for comparison points to construct enhanced feature vectors. The multi-source heterogeneous data includes at least: water depth and uncertainty information output by the data processing module, channel dredging engineering vector data obtained from the public government platform, hydrodynamic and suspended sediment concentration data obtained from the marine reanalysis database and remote sensing products, and spatial and geomorphological context features calculated based on basic geographic information data. The intelligent analysis module for topographic erosion and deposition evolution is connected to the feature engineering module. It has a built-in multi-scale spatiotemporal feature fusion network to receive enhanced feature vectors and simultaneously output the erosion and deposition type classification results, erosion and deposition evolution rate estimates, and corresponding rate estimation uncertainties for each comparison point. The verification output module connects to the intelligent analysis module for topographic erosion and sedimentation evolution. It is used to verify the accuracy and reliability of the model output results and generate thematic maps and quantitative statistical reports on erosion and sedimentation evolution.

[0021] Next, we will provide a detailed introduction to the specific processing of each module: I. Data Processing Module The data processing module includes: The uncertainty calculation unit is used to calculate the initial horizontal and vertical uncertainties of each sounding point based on the standards of the International Hydrographic Organization and the error propagation formula, taking into account GNSS positioning error, attitude measurement error, sound velocity profile error, tide level error and the depth sounder's own error. The comparison point generation unit is used to generate a series of comparison points along the planned survey line at preset intervals, and calculate the influence radius based on the uncertainty and distance of the sounding points, and associate the sounding points within the influence radius with the corresponding comparison points; The data fusion unit uses a Kalman filter to perform recursive optimal estimation on multiple sounding point sequences associated with the same comparison point, and updates the optimal water depth estimate and final estimation uncertainty of the comparison point in the current period.

[0022] Since the above parts all use existing technology, they will not be described in detail.

[0023] II. Feature Engineering Module The enhanced feature vectors constructed by the feature engineering module include: 1. Core measurement features, which come from the output of the data processing module, include the optimal water depth estimate of the comparison point in the two periods (T1, T2), the corresponding water depth estimate uncertainty, the water depth change between the two periods, and the average and standard deviation of the water depth of all comparison points within a radius of 50 meters centered on the point, used to describe local topographic relief.

[0024] 2. Project activity characteristics are collected from publicly available announcements and completion acceptance maps of waterway maintenance and dredging projects on the official websites of local maritime bureaus, waterway administrations, or transportation departments. This includes binary indicators to determine whether a comparison point is located within a known dredging project area and its closest distance to the project boundary. The binary indicator is obtained by spatially overlaying the comparison point coordinates with the officially published project scope vector map to generate a binary feature indicating whether the point is located within the planned or implemented dredging area between the two measurement phases. The closest distance to the project boundary is obtained by calculating the shortest Euclidean distance between the comparison point and the boundary of the most recent dredging project. This data can be directly calculated from the project scope vector map using GIS software.

[0025] 3. Environmental dynamic characteristics, collected from publicly available global or regional reanalysis datasets and remote sensing inversion products, including average flow velocity, dominant flow direction, and average surface suspended sediment concentration at the corresponding measurement period obtained through interpolation at the comparison point locations.

[0026] 4. Spatial context features are collected from publicly available basic geographic information data and historical nautical charts, including the coordinates of the comparison points, their distance from the shoreline, their distance from the estuary, and the code of the geomorphic unit type to which they belong. The classification of the geomorphic unit type is based on official nautical charts, waterway improvement planning maps, or historical research results. The types include at least the main channel, waterway slope, side beach, and deep channel, and are represented in the form of unique thermal codes.

[0027] After obtaining the core measurement features, engineering activity features, environmental dynamic features, and spatial context features, these features are processed to construct a dataset usable by subsequent models, specifically: 1. Spatial and temporal alignment A core table is created based on the unique ID, coordinates, and measurement time of each comparison point.

[0028] Using Geographic Information System (GIS) tools, all external vector data (project scope, geomorphic zones, shoreline) are spatially joined with the comparison points, and corresponding attributes are assigned to them.

[0029] External grid or sequence data (flow velocity, sediment concentration) are bilinearly interpolated in time and space to ensure that each comparison point has a corresponding environmental parameter value at a specific measurement time.

[0030] 2. Feature vector normalization assembly For each comparison point, all its feature values ​​(numerical and categorical) are assembled into a fixed-dimensional array.

[0031] Standardize numerical features (such as water depth, distance, concentration, and flow velocity), for example, by using Z-score standardization to make the mean 0 and the standard deviation 1, in order to accelerate model training and convergence.

[0032] One-hot encoding is used for categorical features (such as geomorphic units and dredging markers).

[0033] 3. Expert Label Preparation This is crucial for building a supervised learning model, requiring the involvement of domain experts and utilizing more reliable data sources than single-beam scanning (such as centimeter-precision DEMs generated by simultaneous multi-beam scanning, sediment sampling analysis reports, and long-term fixed section monitoring data) as "ground truth".

[0034] Based on the "ground true value", experts determined whether the changes at each comparison point during the period from T1 to T2 were "scour", "siltation" or "stable" (classification label), and calculated the exact annual average scour and siltation thickness (regression label).

[0035] 4. Dataset partitioning and serialization The final generated complete sample set, which includes feature vectors and ground truth labels, is randomly divided according to a preset ratio (e.g., 70% training set, 15% validation set, and 15% test set) to ensure consistent distribution.

[0036] The partitioned dataset is serialized into a standard format file (such as .npz, .h5, or .tfrecord) for efficient reading by subsequent models.

[0037] III. Intelligent Analysis Module for Topographic Erosion and Sedimentation This paper primarily introduces the multi-scale spatiotemporal feature fusion network in the intelligent analysis module for topographic erosion and deposition evolution. This network includes a local feature extraction branch, a global spatial relationship branch, a temporal feature extraction branch, a feature fusion and interaction module, and a multi-task output head. The local feature extraction branch employs an improved convolutional neural network, and the global spatial relationship branch employs an improved graph convolutional network. A detailed description follows: 1. Input of the multi-scale spatiotemporal feature fusion network Input: The enhanced feature vector set {F} output by the feature engineering module for N comparison points. j |j=1,...,N}.

[0038] Each feature vector F j The model entry point is clearly divided into four logical groups to correspond to different processing branches: (1) Point-like features (P) j ): Dimension D p =9. Includes the core attribute of the comparison point j itself: Z T1 Z T2 , σ T1 , σ T2 ΔZ represents the flow velocity, sediment concentration, distance from the shoreline, and distance from the estuary. Where Z... T1 Z T2 These are the optimal water depth estimates for the two periods (T1 and T2), respectively, σ T1 σ T2 The values ​​are the water depth estimation uncertainties for the two periods (T1 and T2), respectively, and ΔZ is the water depth change between the two periods.

[0039] (2) Local neighborhood statistical characteristics (L j ): Dimension D l =4. Z contains all comparison points within a circular neighborhood centered at point j with a radius R = 50 meters. T1 and Z T2 The mean and standard deviation are four statistical measures.

[0040] (3) Spatial topology and category characteristics (S j ): Dimension Ds =6+M. Includes: x-coordinate, y-coordinate, dredging identifier (0 / 1), distance from the dredging area, and an M-dimensional one-hot encoded vector, where M is the total number of geomorphic unit types. For example, if M=4, it specifically includes the main channel, slope, beach, and others.

[0041] (4) Time series characteristics (T) j ): Scalar Δt, i.e., time span, in years.

[0042] 2. Local Feature Extraction Branch (1) The input of the branch is: for each comparison point j, find its K=8 nearest neighbor comparison points in space. P of these K+1 points (including itself) j and L j By performing feature concatenation, a shape of (K+1, D) is obtained. p +D l The matrix is ​​(9, 13). This matrix is ​​considered as a one-dimensional pseudo-image where the spatial order is sorted by the distance of the points from the center point j.

[0043] (2) such as Figure 2 As shown, the branch structure specifically includes: The first convolutional layer uses four kernels of size 3 with a stride of 1 and same padding. It is followed by a ReLU activation function. This layer extracts primary local patterns from a 9×13 input and outputs a 9×4 feature map.

[0044] First pooling layer: Max pooling is used, with a pooling window size of 2 and a stride of 2. The output feature map size is 4×4. This operation downsamples to enhance feature invariance.

[0045] The second convolutional layer uses eight kernels of size 3 with a stride of 1 and same padding. It is followed by ReLU activation. The output feature map size is 4×8.

[0046] Second pooling layer: Max pooling is performed again, with a window size of 2 and a stride of 2. The output feature map size is 2×8.

[0047] Flattening layer: Flattens the 2×8 feature map into a 16-dimensional vector.

[0048] Fully connected layer: Pass the 16-dimensional vector through a fully connected layer (weight matrix dimension 16×16, bias dimension 16), apply ReLU activation, and output d. local =16-dimensional local eigenvectors V local j .

[0049] (3) The specific principle of the branch is as follows: the one-dimensional convolution slides along the direction of the spatial neighbor sequence, and the weights of its convolution kernel are learned during training, which can effectively capture the continuous patterns of local micro-topographic changes, such as the undulation of sand ripples and the edge gradient of scour pits. For example, one convolution kernel may learn to recognize the transition pattern from deep to shallow (representing a slope), while another may recognize the deep-shallow-deep depression pattern (representing a scour pit). The pooling layer enhances the robustness to small spatial displacements and disturbances by taking local maxima, making the features more stable. The fully connected layer compresses the extracted local spatial patterns into a compact representation.

[0050] 3. Global Spatial Relationship Branches (1) The input to the branch is: based on the planar coordinates (x, y) of all N comparison points, construct an undirected graph G = (V, E) using the Delaunay triangulation algorithm. Each node v j The characteristic of ∈V is u j =concat(P j S j (), that is, the splicing of point-like and spatial topological features, with a dimension of D. p +D s -M (coordinates are already included in S) j (In the middle). Edge E represents the geographic proximity relationship. The adjacency matrix A of the graph is obtained for GCN after self-loop addition and degree matrix normalization.

[0051] (2) such as Figure 3 As shown, the branch structure is as follows: a two-layer graph convolutional network (GCN) is used. The operation of the graph convolutional layer is defined as: H {（l+1）} =σ(ÃH {（l）} W {（l）} ), where H {（l）} W is the feature matrix of the nodes in the l-th layer. {（l）} It is a learnable weight matrix, σ is the activation function, and Ã=D {-1 / 2}（A+I） D {-1 / 2} It is a normalized adjacency matrix with self-loops.

[0052] ① First convolutional layer: Input: Node feature matrix H {（0）} =U∈R {N×Fin} F in =dim(u j )))

[0053] Specific instructions: H {（1）} =ReLU(ÃH {（0）} W {（0）} W {（0）} ∈R {Fin×16}These are learnable weights.

[0054] Output: Node intermediate features H {（1）} ∈R {N×16} .

[0055] ② Second convolutional layer: Operation: H {（2）} =ÃH {（1）} W {（1）} Among them, W {（1）} ∈R {16×dglobal} d global =8. This layer does not have an activation function to preserve linear features.

[0056] Output: Node final feature H {（2）} ∈R {N×8} Node v j The corresponding row vector is the global spatial feature V. global j .

[0057] (3) The principle of the branch is as follows: GCN is implemented through the local first-order approximation of spectral graph convolution. The first layer aggregates the information of each node and its first-order neighbors, and introduces nonlinearity through ReLU activation. The second layer further aggregates, and since the square approximation of à is equivalent to considering the second-order neighbors, the final feature V of each node is... global j It integrates information from all nodes within its two-hop neighborhood. This allows the model to learn spatial dependency patterns; for example, node characteristics in upstream scour zones may influence the deposition probability of downstream nodes. The weight matrix W adjusts the intensity and pattern of information transmission during the learning process.

[0058] 4. Temporal Feature Extraction Branch (1) The input to the branch is: a scalar time series feature Δt and a learnable context code (dimensional ctx) representing the measurement year / season. dim =8). Context encoding is a lookup table that maps discrete temporal contexts (such as summer 2019) to continuous vectors.

[0059] (2) such as Figure 4 As shown, the branch structure is as follows: ① Embedding and Mapping Layer: Δt and its corresponding context encoding vector are concatenated and mapped to d through a fully connected layer (9 input dimensions, 8 output dimensions, ReLU activation). temp =8-dimensional initial time series vector.

[0060] ② Multi-head self-attention layer: Set the number of heads h=2. The initial time series vector is passed through three independent linear layers (each with an 8×4 weight matrix) to generate Query(Q), Key(K), and Value(V), with each head dimension d... k =d v =4.

[0061] Calculate scaled dot product attention: Attention(Q, K, V) = softmax(QK) T / sqrt(d k ))V.

[0062] The outputs of the two heads (each head outputs 4 dimensions) are concatenated to obtain an 8-dimensional vector, which is then fused through an output linear layer (8×8).

[0063] ③ Normalization and feedforward networks: First, perform residual connections and layer normalization: Z = LayerNorm (Attention) Output +Initial Vector ).

[0064] Feedforward network: Two fully connected layers. The first layer maps 8 dimensions to 16 dimensions (ReLU activation), and the second layer maps back to 8 dimensions.

[0065] After further residual connections and layer normalization, the final d is output. temp =8-dimensional time series feature vector V temp j .

[0066] (3) The principle of the branch is as follows: The self-attention mechanism allows the model to dynamically weigh different influence patterns at different time spans by calculating the correlation (attention score) between all positions in the input sequence. Here, the input sequence actually has only one position (the current time span and context), but through the multi-head mechanism, the model can learn information from different representation subspaces in parallel. The feedforward network provides additional nonlinear transformation capabilities. This structure helps to distinguish different feature patterns of short-term (e.g., after dredging) and long-term (e.g., natural evolution) processes.

[0067] 5. Feature Fusion and Interaction Module (1) The input of the module is: the feature vectors V of the three branches. local j (16-dimensional), V global j (8-dimensional), V temp j (8-dimensional).

[0068] (2) such as Figure 5 As shown, the module's structure is as follows: ①Splicing layer: Vfused j =concat(V local j V global j V temp j This yields a 32-dimensional fusion feature.

[0069] ② Feature Cross Network: Employing the cross layer in Deep & Cross Network (DCN-V2), high-order interactions between features are explicitly learned: Let x0 = V fused j .

[0070] Level l crossover operation: x {l+1} =x0⊙(W l x l +b l )+x l Where ⊙ denotes element-wise multiplication, W l and b l These are the weights and biases of the l-th layer. Here, we use L=2 layers of cross-interaction. The first cross-interaction layer outputs a dimension of 32, and the second layer also outputs a dimension of 32.

[0071] The final output cross feature V cross j (32 dimensions).

[0072] ③ Deep Neural Networks: V cross j Through a 2-layer fully connected neural network: First fully connected layer: 32-dimensional input, 32-dimensional output, ReLU activation, and Dropout applied (rate=0.3).

[0073] The second fully connected layer has a 32-dimensional input and a 16-dimensional output, with ReLU activation.

[0074] The output is a high-level comprehensive feature V. final j (16 dimensions).

[0075] (3) The principle of the module is that simple feature splicing may not be able to fully capture the complex interaction relationship between features of different modalities. The cross layer can effectively learn feature combinations (such as the combination of high flow velocity and main channel may strongly indicate scouring) through explicit cross-product. The deep fully connected layer further performs nonlinear transformation and feature compression to extract the information that is most discriminative for the final task.

[0076] 6. Multi-task output head (1) Input: Advanced comprehensive feature V finalj .

[0077] (2) The siltation and sedimentation sorting head includes: The third fully connected layer has a 16-dimensional input and a 16-dimensional output, and uses the ReLU activation function.

[0078] Fourth fully connected layer: 16-dimensional input, 2-dimensional output.

[0079] Softmax layer: Apply the Softmax function to the 2D output to obtain the class probability distribution P. j =[p 冲刷 p 淤积 The final predicted category is the category L with the highest probability. j' =argmax(P j ).

[0080] (3) Evolution rate regression head: Fifth fully connected layer: 16-dimensional input, 16-dimensional output, using the ReLU activation function.

[0081] The sixth fully connected layer has a 16-dimensional input and a 2-dimensional output. These two outputs represent the mean velocity μ. j and logarithmic variance log(σ) j2 ).

[0082] Final prediction rate R j' =μ j The prediction uncertainty (standard deviation) is σ. j =exp(0.5*log(σ) j2 )).

[0083] (4) Its specific principle is as follows: multi-task learning is adopted, and classification and regression share most of the feature extraction networks, which can promote each other and improve generalization ability. The uncertainty output of the regression head enables the model to self-evaluate the reliability of the prediction results, which is very important in practical applications and can screen out high uncertainty prediction points that need to be manually checked.

[0084] 7. Next, we will introduce the loss function of the multi-scale spatiotemporal feature fusion network: Total loss L total It is a weighted multi-tasking loss: L total =λ*L cls +(1-λ)*L reg In the formula, L cls For the focus loss function optimized for class imbalance, L reg Here, λ is the heteroscedastic regression loss function based on uncertainty, and λ is the weighting coefficient. In this invention, λ = 0.7.

[0085] Classification loss L cls The focus loss with class weights is used to handle potential class imbalance and focus on hard-to-classify samples.

[0086] Regression loss L reg Using heteroscedasticity regression loss, the model simultaneously predicts the mean μ and variance σ. 2 .

[0087] 8. The training steps for the multi-scale spatiotemporal feature fusion network will be introduced next: (1) Initialization: All neural network weights are initialized using the He normal distribution, and the bias is initialized to 0.

[0088] (2) Optimizer: The AdamW optimizer was used with an initial learning rate of lr=0.001, weight decay of weight_decay=0.01, beta1=0.9, and beta2=0.999.

[0089] (3) Training cycle: Set the maximum number of training cycles Epochs=200 and the batch size BatchSize=64.

[0090] Forward propagation: Data is input into the network in batches to calculate the predicted P. j μ j ,log(σ j2 ).

[0091] Loss calculation: based on the true label L j (Category Index) and y j Calculate L (actual speed) respectively cls and L reg Then calculate the weighted sum L. total .

[0092] Backpropagation and Optimization: Calculating L total Perform gradient clipping (gradient norm threshold = 1.0) on the gradient of the model parameters, and call the step() method of the AdamW optimizer to update the parameters.

[0093] (4) Learning rate scheduling: The ReduceLROnPlateau scheduler is used to monitor the validation set loss. When the loss stops decreasing after 10 consecutive periods, the learning rate is multiplied by a factor of 0.5. The minimum learning rate is set to 1e-6.

[0094] (5) Early stopping: When the validation set loss no longer decreases after 20 consecutive pauses, stop training and load the model weights that perform best on the validation set.

[0095] (6) Hyperparameter tuning: Using a Bayesian optimization framework (such as Optuna), the key hyperparameters are searched on the validation set for approximately 50 iterations. The search space includes: λ∈[0.5, 0.8], d local Given d_global ∈ [8, 32], d_global ∈ [4, 16], learning rate ∈ [1e-4, 1e-2] (logscale), and dropout rate ∈ [0.1, 0.5]. The optimization objective is to maximize the overall metric Score = 2 * (1 - NRMSE) * F1. score / (1-NRMSE)+F1 score ), where NRMSE is the normalized root mean square error, F1 score The macro F1 score for classification.

[0096] 9. The multi-scale spatiotemporal feature fusion network ultimately outputs a structured result dictionary, which includes: (1) Scour / Sedimentation type: The predicted category (scour / sedimentation) and the corresponding confidence probability for each comparison point. For example, {'type': 'sedimentation', 'prob': 0.92}.

[0097] (2) Evolution rate: the predicted annual average siltation thickness (m / year) at each comparison point, with positive values ​​indicating siltation and negative values ​​indicating scouring.

[0098] (3) Prediction uncertainty: The standard deviation of the rate prediction for each point, σ (meters / year). The larger the value of σ, the more uncertain the model's rate estimate for that point is.

[0099] IV. Verification Output Module This module is the final application layer, and the verification output module includes: The confidence screening unit is used to estimate uncertainty and classification probability based on the rate output of the model, set a threshold, and automatically identify evaluation results with low confidence. The external validation unit is used to compare the scouring and deposition distribution and rate output by the model with the real changes derived from the high-precision multibeam terrain data of the same period, and to calculate the classification accuracy, root mean square error and coefficient of determination to quantify the accuracy. The visualization mapping unit is used to automatically generate spatial distribution maps of scour and sedimentation types, contour maps of evolution rates, uncertainty distribution maps, and comprehensive analysis maps along key profiles.

[0100] In addition, the verification output module can also be configured with other functions according to specific needs, and this application does not impose any restrictions on it.

[0101] To make the technical solution and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment takes a key section of a waterway in my country with a length of 2 kilometers (from the starting point S to the ending point E) as the application object, and the monitoring period is from January 2023 (T1 period) to January 2024 (T2 period), with a time span of Δt = 1 year.

[0102] 1. Data Preparation and Input Collect raw data and auxiliary information from two phases of single-beam bathymetry for this channel section. Simultaneously, acquire the following multi-source data: Engineering data: According to the waterway management department, a maintenance dredging was carried out in June 2023 in the section 500 to 800 meters from the starting point.

[0103] Environmental data: Grid data such as the average flow velocity and suspended sediment concentration in the region during the same period were obtained by interpolation from public databases.

[0104] Basic geographic data: Obtain shoreline vectors and the division of geomorphic units in the "main channel" of the waterway.

[0105] 2. System Implementation Process and Data Output The above data is input into the system of the present invention, and each module runs in sequence.

[0106] (1) Data processing module output The module performs uncertainty propagation calculations and Kalman filtering on the raw depth sounding data to generate a series of 41 comparison points spaced 50 meters apart.

[0107] (2) Output of feature engineering module The module constructs an enhanced feature vector for each comparison point.

[0108] (3) Output of the intelligent analysis module for topographic erosion and sedimentation evolution The aforementioned features are input into a pre-trained multi-scale spatiotemporal feature fusion network. The model predicts scour and sedimentation at 41 comparison points across the entire cross-section, outputting results including scour and sedimentation type, evolution rate, and uncertainty. The model performs excellently on the reserved validation set in this embodiment: scour and sedimentation classification accuracy reaches 95.6%, the root mean square error (RMSE) of evolution rate prediction is 0.09 m / year, and the coefficient of determination R0 is [missing information]. 2 It is 0.91.

[0109] (4) Verify the output of the output module The module validates and visualizes the model results, generating a series of thematic maps, specifically: ① Generate a water depth variation profile, specifically as follows: Figure 6As shown, with distance on the horizontal axis and water depth on the vertical axis, water depth profile curves for periods T1 and T2 are plotted. The solid blue line represents the water depth in period T1, and the dashed red line represents the water depth in period T2. The graph clearly shows that in the dredged section between 500 and 800 meters, the profile line for period T2 is significantly lower than that for period T1, forming a "pit," which directly reflects the scouring effect of the project. In other sections, the profile line for period T2 is mostly higher than that for period T1, indicating widespread siltation.

[0110] ② Generate a distribution map of scour and sedimentation evolution rates, specifically as follows: Figure 7 As shown, the data is presented using a dual-axis format of "bar chart + line chart". The bar chart (blue represents positive sedimentation, orange represents negative scour) shows drastic local changes, especially the dense orange high bars in the dredged section; the black cumulative curve shows an overall sedimentation trend starting from the starting point, which drops sharply in the dredged section due to strong scour, and then recovers and rises slowly.

[0111] ③ Generate a map showing the distribution of uncertainty in the model predictions, specifically as follows: Figure 8 As shown in the figure, the line graph shows the variation of uncertainty along the process, with a significant peak in the dredging section (500-800 meters), reaching a maximum of 0.22 m / yr. Based on this, the system automatically identifies this area as a "high uncertainty area that needs to be reviewed".

[0112] ④ Generate a probability map for classifying scour and sedimentation, as shown in the following example. Figure 9 As shown in the figure, the line graph shows the change in the siltation probability along the process. The probability value drops below the 0.5 threshold in the dredging section, which corresponds exactly to the scour area judged by the model.

[0113] 3. This embodiment demonstrates the entire process of the system of the present invention, from multi-source data input and intelligent processing and analysis to the visualization output of results, using a complete set of simulated data. The results show that: The system has successfully achieved automated and quantitative assessment of scour and sedimentation evolution, eliminating the reliance on manual interpretation.

[0114] The proposed multi-scale spatiotemporal feature fusion network has high prediction accuracy and can provide prediction uncertainty, with reliable and interpretable results.

[0115] The system outputs rich and intuitive results, including various thematic maps, which can be directly used for scientific decision-making in waterway maintenance and management. For example, it can accurately plan the dredging scope and volume based on "cumulative siltation" and "scour hotspots".

[0116] This embodiment fully verifies the effectiveness, advancement, and practical value of the system of the present invention for quantitative assessment of scour and sedimentation evolution in waterway areas.

[0117] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0118] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0121] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A quantitative assessment system for the evolution of scour and sedimentation in waterway areas based on convolutional neural networks and graph convolutional networks, characterized in that, include: The data processing module is used to receive raw data from multiple single-beam bathymetry measurements and auxiliary sensor data, calculate the initial uncertainty of each bathymetry point through the uncertainty propagation model, generate a set of comparison points along the preset measurement line, and use the Kalman filter algorithm to fuse information from neighboring bathymetry points, outputting the optimal water depth estimate and the corresponding water depth estimate uncertainty of each comparison point in each measurement period. The feature engineering module, connected to the data processing module, is used to fuse multi-source heterogeneous data for comparison points to construct enhanced feature vectors. The multi-source heterogeneous data includes at least: water depth and uncertainty information output by the data processing module, channel dredging engineering vector data obtained from the public government platform, hydrodynamic and suspended sediment concentration data obtained from the marine reanalysis database and remote sensing products, and spatial and geomorphological context features calculated based on basic geographic information data. The intelligent analysis module for topographic erosion and deposition evolution is connected to the feature engineering module. It has a built-in multi-scale spatiotemporal feature fusion network to receive enhanced feature vectors and simultaneously output the erosion and deposition type classification results, erosion and deposition evolution rate estimates, and corresponding rate estimation uncertainties for each comparison point. The verification output module is connected to the intelligent analysis module for terrain erosion and sedimentation evolution, and is used to verify the accuracy and reliability of the model output results.

2. The quantitative assessment system for waterway scouring and sedimentation evolution based on convolutional neural networks and graph convolutional networks according to claim 1, characterized in that, The data processing module includes: The uncertainty calculation unit is used to calculate the initial horizontal and vertical uncertainties of each sounding point based on the standards of the International Hydrographic Organization and the error propagation formula, taking into account GNSS positioning error, attitude measurement error, sound velocity profile error, tide level error and the depth sounder's own error. The comparison point generation unit is used to generate a series of comparison points along the planned survey line at preset intervals, and calculate the influence radius based on the uncertainty and distance of the sounding points, and associate the sounding points within the influence radius with the corresponding comparison points; The data fusion unit uses a Kalman filter to perform recursive optimal estimation on multiple sounding point sequences associated with the same comparison point, and updates the optimal water depth estimate and final estimation uncertainty of the comparison point in the current period.

3. The quantitative assessment system for waterway scouring and sedimentation evolution based on convolutional neural networks and graph convolutional networks according to claim 2, characterized in that, The enhanced feature vectors constructed by the feature engineering module include: The core measurement features include the water depth estimates, water depth estimation uncertainties, and water depth changes of the comparison point itself in the two measurement periods; Engineering activity characteristics, including whether the comparison point is located within the known waterway dredging project area and its closest distance to the project boundary; Environmental dynamic characteristics, including the average flow velocity, dominant flow direction and average surface suspended sediment concentration at the comparison point location obtained by interpolation during the corresponding measurement period; Spatial contextual features include the coordinates of the comparison point, its distance from the shoreline, its distance from the river mouth, and the coding of the geomorphic unit type to which it belongs.

4. The quantitative assessment system for waterway scouring and sedimentation evolution based on convolutional neural networks and graph convolutional networks according to claim 3, characterized in that, The classification of the geomorphic unit types is based on official nautical charts, waterway improvement planning maps, or historical research results, and the types include at least the main channel, waterway slope, side beach, and deep channel.

5. The quantitative assessment system for waterway scouring and sedimentation evolution based on convolutional neural networks and graph convolutional networks according to claim 4, characterized in that, The multi-scale spatiotemporal feature fusion network includes: The local feature extraction branch, composed of a one-dimensional convolutional neural network, is used to process the sequential features centered on the target comparison point and composed of its spatial nearest neighbors to extract micro-topography spatial patterns. The global spatial relationship branch consists of two layers of graph convolutional networks. It transmits information based on the Delaunay triangulation constructed from all comparison points and aggregates multi-order neighbor information of nodes to capture regional spatial dependencies. The temporal feature extraction branch, including multi-head self-attention mechanism and feedforward network, is used to perform deep modeling of time span and temporal context encoding to capture temporal nonlinear effects; The feature fusion and interaction module is used to concatenate the output features of the three branches, learn higher-order feature interactions using an explicit feature cross-network, and finally generate high-level comprehensive features through a fully connected layer. The multi-task output head includes a classification head for outputting the probability of scour or siltation type and a regression head for simultaneously outputting the mean of the evolution rate and its log-variance.

6. The quantitative assessment system for waterway scouring and sedimentation evolution based on convolutional neural networks and graph convolutional networks according to claim 5, characterized in that, The local feature extraction branch specifically includes: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, and a fully connected layer.

7. The quantitative assessment system for waterway scouring and sedimentation evolution based on convolutional neural networks and graph convolutional networks according to claim 6, characterized in that, The intelligent analysis module for terrain erosion and sedimentation evolution uses a weighted multi-task loss function L during training. total =λ*L cls +(1-λ)*L reg In the formula, L cls For the focus loss function optimized for class imbalance, L reg Let λ be the loss function for heteroscedastic regression based on uncertainty, and λ be the weighting coefficient.

8. The quantitative assessment system for waterway scouring and sedimentation evolution based on convolutional neural networks and graph convolutional networks according to claim 7, characterized in that, The verification output module includes: The confidence screening unit is used to estimate uncertainty and classification probability based on the rate output of the model, set a threshold, and automatically identify evaluation results with low confidence. The external validation unit is used to compare the scouring and deposition distribution and rate output by the model with the real changes derived from the high-precision multibeam terrain data of the same period, and to calculate the classification accuracy, root mean square error and coefficient of determination to quantify the accuracy. The visualization mapping unit is used to automatically generate spatial distribution maps of scour and sedimentation types, contour maps of evolution rates, uncertainty distribution maps, and comprehensive analysis maps along key profiles.

Images