A Riverbed Evolution Prediction Method Based on Multi-Source Data Fusion

Abstract

This invention provides a method for predicting riverbed evolution based on multi-source data fusion, belonging to the field of hydrological resource monitoring technology. It utilizes different devices to collect and fuse multi-source data, which is then preprocessed to form a unified three-dimensional grid data. A multi-scale feature map is generated based on a designed feature extraction network. Subsequently, a designed multi-channel convolutional neural network is used to simultaneously learn and extract features across multiple channels, performing hierarchical representation and fusion of different information sources. Finally, a designed time-series regression network is used to extract complex dependencies in the time series data through a two-layer LSTM structure, and an adaptive residual correction module learns residual information to improve prediction accuracy, obtaining the final prediction result. This invention demonstrates strong applicability and robustness in complex river environments. Through multi-source data fusion and innovative feature extraction and regression methods, it can address various riverbed morphological changes caused by natural factors and human activities.

Description

Technical Field

[0001] This invention belongs to the field of hydrological resource monitoring technology, and in particular relates to a method for predicting riverbed evolution based on multi-source data fusion. Background Technology

[0002] Globally, riverbed morphology evolution is influenced by multiple factors, including natural elements and human activities. Accurate prediction of riverbed morphology is crucial for water resource management, ecological protection, and disaster prevention and mitigation. In recent years, with intensified human activities and the impact of climate change, many rivers have experienced significant changes in their bed morphology, including channel deepening, riverbed erosion, and sedimentary alterations. These changes can have profound impacts on water quality, the ecological environment, riverbank stability, and flood risk.

[0003] Riverbed morphology evolution is a dynamic process encompassing phenomena such as erosion and sedimentation, driven by natural factors including water flow velocity, channel curvature, and topographic changes. Simultaneously, artificial structures such as dams, dikes, sand mining, and dredging significantly alter flow patterns and sediment transport behavior. For example, dam construction not only changes the speed and direction of water flow but also reduces downstream sediment, leading to riverbed erosion and increased channel depth. These complex interactions between human activities and natural factors influence the dynamic evolution of riverbed morphology, exacerbating channel degradation. However, current research on riverbed morphology evolution still faces many challenges. On the one hand, traditional research methods, such as field measurements and laboratory simulations, cannot fully cover the complex environment of rivers, limiting the spatial and temporal scales of research. On the other hand, numerical models provide effective tools for studying riverbed morphology evolution, especially in large river systems. However, the complexity of numerical simulations makes the process of building and calibrating models highly technical.

[0004] Against this backdrop, the introduction of machine learning techniques enables efficient processing and analysis of large-scale data, thereby allowing for more accurate predictions of riverbed evolution. However, existing technologies have the following drawbacks:

[0005] 1. Data Acquisition and Quality Control Issues: Current riverbed morphology research relies on multiple data sources, such as lidar, remote sensing imagery, sonar bathymetry, and hydrological station data. However, due to the different sources, their resolution, accuracy, and temporal consistency can vary significantly, directly impacting the effectiveness of data fusion. Data collected at different time periods may exhibit inconsistencies in accuracy, while the temporal lag of the data can limit the monitoring of rapidly changing riverbed morphology.

[0006] 2. High model coupling complexity: Traditional numerical models (such as TELEMAC and HEC-RAS) are mainly used for single-dimensional hydrodynamic simulations, while riverbed evolution prediction often requires integrated hydrodynamic and sediment transport models. The coupling of multi-dimensional models involves handling a large number of nonlinear equations and parameter adjustments, resulting in enormous computational costs and high parameter sensitivity, thus increasing the difficulty of model calibration.

[0007] 3. Selection and Application of Data Fusion Algorithms: In the process of multi-source data fusion, different algorithms have varying adaptability and processing capabilities for data features. For example, commonly used machine learning algorithms can effectively extract important features from data, but their ability to capture dynamic changes in time series data is limited. For the task of predicting riverbed morphology, conventional machine learning methods are insufficient in their ability to explain nonlinear changes and their model generalization, and cannot fully capture the combined influence of natural and anthropogenic factors on riverbed morphology.

[0008] 4. Model Uncertainty and Prediction Error: Riverbed evolution models contain various uncertainties, such as data errors, model assumptions, and environmental changes. These uncertainties directly affect the reliability of the prediction results. Although some methods (such as uncertainty analysis and error propagation) can be used to assess prediction errors, their effectiveness is limited for complex systems involving multi-source data fusion. Summary of the Invention

[0009] To address the above problems, this invention proposes a method for predicting riverbed evolution based on multi-source data fusion, comprising the following steps:

[0010] S1 collects multi-source riverbed data, preprocesses the data, and merges it to form a unified three-dimensional grid data.

[0011] S2, Based on the feature extraction network, a multi-scale feature map is generated; the feature extraction network includes a multi-scale convolutional neural network module and a mutual information feature selection module. The multi-scale convolutional neural network module extracts riverbed morphological features at different scales and combines them with a channel attention mechanism for weighted fusion; the mutual information feature selection module eliminates redundant features and obtains key features that cover the dynamic changes of the riverbed.

[0012] S3, based on a multi-channel convolutional neural network, simultaneously learns and extracts features across multiple channels, performing hierarchical representation and fusion of different information sources; the multi-channel convolutional neural network includes a multi-scale convolutional fusion module and a feature compression module; the multi-scale convolutional fusion module uses convolutional kernels of different sizes to extract detailed and large-scale features, and focuses on the most relevant features through a channel attention mechanism, fusing multi-channel information; the feature compression module further compresses the feature map through pooling layers, and when processing time-series data, it captures time dependencies through LSTM layers to generate serialized spatiotemporal features;

[0013] S4, based on a time series regression network, extracts complex dependencies in the time series through a two-layer LSTM structure, and improves prediction accuracy by learning residual information through an adaptive residual correction module, thus obtaining the final prediction result.

[0014] Preferably, the acquisition of multi-source riverbed data includes topographic data, hydrodynamic data, sediment data, and meteorological data. The topographic data includes three-dimensional point cloud coordinate data of topographic relief information acquired by lidar, two-dimensional image data of river channel morphology, surrounding vegetation cover, water body expansion, and shoreline retreat changes acquired by remote sensing imagery, and three-dimensional point cloud data of underwater riverbed morphology and depth acquired by sonar. The hydrodynamic data includes flow rate data, flow velocity data, and water level data. The sediment data includes suspended solids concentration data and sediment particle size distribution data. The meteorological data includes precipitation data and evaporation data.

[0015] Preferably, the process of fusing the data to form a unified three-dimensional mesh is as follows:

[0016] Define a unified grid: Based on the scope and resolution requirements of the riverbed area under study, define a unified three-dimensional spatial grid, determine the minimum and maximum x, y, z coordinates, denoted as (x... min y min , z min ) and (x max y max , z max ), used to determine the spatial extent of the grid, where the position of each grid point is (x... i y i .z i ), where i represents the index of the grid point; the grid resolution is based on the average resolution of the LiDAR, which is 1m;

[0017] Unified Data Conversion: The data collected by the lidar is 3D point cloud data, given in the form p1 = (x1, y1, z1), representing the 3D coordinates of the ground surface. p1 is then converted to a projected coordinate system.

[0018]

[0019] Z Radar =h

[0020] in, γ and γ are the original geographic coordinates, γ0 represents the center longitude of the projected coordinate system, N is the radius of curvature, and h is the elevation data, i.e. the elevation value measured by the lidar.

[0021] Remote sensing image data typically only contains planar coordinates p2 = (x2, y2), which are directly mapped onto a plane in a 3D grid, with the z value defaulting to 0;

[0022] The sonar-acquired data p3 = (x3, y3, z3) records the water depth values, which need to be converted to absolute elevation, i.e., the distance from the vertical line of the ground point to the geoid. Using a unified three-dimensional projected coordinate system, the absolute elevation z is calculated as follows:

[0023] z S =HD

[0024] Where D is the water depth data, which is the depth relative to the water surface, and H is the elevation of the water surface;

[0025] Hydrodynamic data p4 includes water level and velocity information; water level data is directly mapped onto the z-axis in the projected coordinate system as the z-value; velocity is decomposed into three-dimensional components (v... s v y v z Interpolate to each grid point on the corresponding plane of the 3D mesh;

[0026] The sediment data includes the suspended sediment concentration p5(x5,y5), which is directly recorded on the z-plane of the grid, i.e., z = z m The sediment concentration p5 of all grid points is obtained by interpolation using the IDW method. This involves interpolating the concentration data of the surrounding sampling points q5(xj, yj) and repeating the interpolation calculation for each grid point until the sediment concentration p5 of all grid points is obtained by IDW interpolation.

[0027] Meteorological data is in two-dimensional format, P6, including precipitation and evaporation, and is directly assigned to the surface layer of the grid.

[0028] After the above transformations, all data are standardized and converted to a unified three-dimensional coordinate system. LiDAR and sonar data provide z-coordinate values ​​for constructing three-dimensional topography of the surface and underwater areas. Remote sensing imagery data is mapped to the (x, y) plane, assigning attribute values ​​to the surface layer. Hydrodynamic and sediment data are spatially interpolated to different depth layers of the grid. Meteorological data is mapped to the surface grid, ultimately yielding data P represented by a three-dimensional grid. grid .

[0029] Preferably, the multi-scale convolutional neural network module in the feature extraction network specifically comprises:

[0030] For 3D mesh data P grid We designed multi-scale convolutional kernels, setting different kernel sizes (k*k) to extract multi-level spatial information. We also added a temporal dimension to construct a three-dimensional convolutional kernel. The formula for calculating the three-dimensional convolution is as follows:

[0031] F k，t =σ(P grid *W k，t +b k，t )

[0032] Among them, W k，t ∈R k×k×T×D The kernel is a three-dimensional convolution, including spatial dimension D and temporal dimension T; b k，t For the corresponding bias term, F k，t This represents the feature mapping at scale k and time step T;

[0033] Secondly, in order to capture data P more comprehensively grid The multi-scale features are fused by fusing the feature maps {F3, F5, F7} at each scale. The feature fusion is performed using a weighted summation method.

[0034] F multi =α·F3+β·F5+γ·F7

[0035] Here, α, β, and γ represent the fusion weights, which are adaptively learned through model training. The mean squared error is used as the loss function to measure the difference between the predicted and actual values, thereby optimizing the model's prediction accuracy for riverbed evolution characteristics.

[0036] Preferably, the adaptive feature selection module in the feature extraction network performs the following specific processing steps:

[0037] First, feature importance is measured. This is done by calculating the mutual information (MI) between each feature channel and the target variable to determine the importance of the feature. MI measures the degree of information correlation between the feature and the target variable, and its calculation formula is as follows:

[0038]

[0039] Among them, F i Represents the fusion feature F multi The i-th feature channel in the equation; Y is the target variable, i.e., the change in riverbed height; p(f i ,y) is the feature channel F i The joint probability distribution of the target variable Y and p(f i p(y) and p(y) are marginal probability distributions;

[0040] Then calculate the feature pair F. i Compared with other features F j The correlation coefficient p(F) between them i F j Pearson correlation coefficient was used:

[0041]

[0042] Subsequently, the effective mutual information MI is defined. eff (F i ;Y) to correct redundancy between features:

[0043]

[0044] Where δ represents the weighting coefficient, which balances the correlation between features and the target and the redundancy between features;

[0045] Finally, according to MI eff (F i ;Y) Sort the features and select the top N features F with the highest effective mutual information. selected This serves as the input for subsequent multi-channel convolutional neural network models.

[0046] F selected ={F k，t |MI eff (F i ;Y)≥ε}

[0047] Among them, F selected It is the feature set after mutual information filtering, where ε is the threshold of mutual information, that is, only feature maps with mutual information greater than or equal to the threshold are selected;

[0048] The feature matrix F after mutual information feature selection selected This will be used as input to the model for spatiotemporal prediction.

[0049] Preferably, the multi-scale convolutional fusion module in S3 includes multi-scale convolutional units and fusion units;

[0050] The multi-scale convolutional unit is designed with multiple convolutional layers of different kernel sizes to extract feature information at different spatial scales. For each kernel size, a convolution operation is performed on each channel to obtain the corresponding feature map. The convolution calculation formula is as follows:

[0051]

[0052] in, Let X represent the feature map of the i-th layer at scale k. (l-1) This is the output feature map of the previous layer. This represents the convolution kernel of the l-th layer at scale k. Here, σ is the bias term, and σ is the activation function ReLU.

[0053] The fusion unit introduces a channel attention mechanism, enabling the network to adaptively allocate channel weights based on current feature information; this applies to the multi-scale feature maps extracted through convolution. The features are fused to obtain the feature fusion results for each layer; the weight of each channel is defined as α.k The channel weighted fusion result is:

[0054]

[0055] in, g(·) is the channel weight learning function, which adaptively weights the feature maps of each channel.

[0056] Preferably, the feature compression module in S3 specifically comprises:

[0057] The feature maps after multi-channel fusion are pooled through a pooling layer to further compress the feature space and retain key features:

[0058]

[0059] Where pooling() represents the average pooling operation, and the compressed feature map As input for the next layer;

[0060] Simultaneously, a recurrent neural network (LSTM) layer is added after the output of the convolutional layer to capture dynamic changes in the time dimension, namely, the temporal changes of features such as riverbed height and sedimentation rate. The output of the pooling layer is input into the LSTM layer to obtain the serialized spatiotemporal features y. t .

[0061] Preferably, the data processing procedure for the time series regression network is as follows:

[0062] Based on the constructed time series data, firstly, a two-layer LSTM structure is used to extract the short-term and long-term dependencies of the time series. The first layer of LSTM focuses on capturing short-term change features, while the second layer of LSTM further analyzes long-term trends to generate preliminary prediction results. Then, an attention mechanism is added to adaptively allocate the weights of each time step, making the model more focused on the time step information that has a key impact on future evolution. Finally, an adaptive residual correction module is used to further optimize the prediction results to dynamically compensate for biases and improve the overall prediction accuracy and the model's temporal consistency.

[0063] Preferably, the dual-layer LSTM structure is as follows:

[0064] In the first layer of LSTM, the input data y is received at each time step t. t And the hidden state of the previous time step. and memory unit The calculation formula for the first LSTM layer is:

[0065]

[0066] in, These represent the hidden state and memory unit of the first LSTM layer at time step t. As the output of the first LSTM layer;

[0067] An attention mechanism is introduced, generating a set of attention weights at each time step t to help the model adaptively allocate attention at different time steps in the time series. The attention weights μ t The calculation formula is:

[0068]

[0069] Among them, ω, W a b a μ is a learnable parameter. t It is the attention weight at time step t, representing the importance of the feature at the current time step;

[0070] The output of the first LSTM layer after incorporating the attention mechanism represents the hidden states of all time steps. We obtain a global weighted feature representation s by summing the values ​​according to the attention weights. (1) :

[0071]

[0072] Weighted s (1) Indicates the short-term characteristics of riverbed evolution;

[0073] In the second LSTM layer, the weighted output s of the first LSTM layer is... (1) As input to the second LSTM layer, long-term dependency features are further extracted; the calculation formula for the second LSTM layer is:

[0074]

[0075] in, These represent the hidden state and memory unit of the second-layer LSTM at time step t; As the output of the second LSTM layer;

[0076] Hidden state generated by two-layer LSTM Used to generate preliminary prediction results at time step t+1

[0077] Preferably, the adaptive residual correction module specifically comprises:

[0078] Residual calculation: Calculate the true value Y at each time step t. t and predicted value The difference between them, i.e., the residual:

[0079]

[0080] Where, r t This represents the prediction error at time step t;

[0081] Residual Correction Model: Using linear regression to fit the residuals, the residual correction model learns the model bias based on past predictions and generates a residual correction value for each time step.

[0082]

[0083] Where g(·) represents the residual correction model, and m is the number of time steps of the input;

[0084] Final prediction output: The original prediction results of the two-layer LSTM model. and residual correction value Add them together to obtain the final prediction result:

[0085]

[0086] The output of the time series regression network is the final predicted sequence. It is a predicted value of the riverbed state at a future time step t+1.

[0087] Preferably, the feature extraction network, the multi-channel convolutional neural network, and the time series regression network all need to be trained, and the specific training strategy is as follows:

[0088] 3D mesh data P grid The dataset is divided into training and test sets in an 8:2 ratio, and the division method is random to ensure that the model has enough training samples and generalization ability.

[0089] The average of the sum of five indicators—the mean square error between the true and predicted values ​​of water depth, the mean square error between the true and predicted values ​​of flow velocity, the mean square error between the true and predicted values ​​of water level, the mean square error between the true and predicted values ​​of precipitation evaporation, and the mean square error between the true and predicted values ​​of sediment concentration—was used as the loss function for training the feature extraction network, the multi-channel convolutional neural network, and the time series regression network.

[0090] When the loss function value calculated using the validation set fails to decrease for 10 consecutive epochs, the training of the feature extraction network, multi-channel convolutional neural network, and time series regression network is complete.

[0091] Compared with the prior art, the present invention has the following beneficial effects:

[0092] 1. Improved Data Quality and Consistency: By fusing multi-source data, this method significantly improves the accuracy and consistency of input data. The multi-source data acquisition and processing module utilizes data from lidar, remote sensing imagery, sonar depth sounding, etc., and ensures temporal and spatial consistency of data from various sensors through data anomaly detection and time synchronization processing.

[0093] The improved data quality ensures the stability of the model input, especially in the monitoring of dynamic changes in the riverbed. Even if there are differences in the frequency and resolution of data collection from different data sources, the data fusion step can still provide complete and accurate three-dimensional gridded input data, which effectively improves the reliability and accuracy of subsequent predictions.

[0094] 2. Accurate Extraction of Multi-Scale Riverbed Features: Based on the designed feature extraction network, a multi-scale convolutional neural network and adaptive feature selection method are employed to accurately extract riverbed morphological change features at different scales, avoiding redundant information and improving feature recognition efficiency. Multi-scale convolutional kernels can fully extract local details and overall change trends of the riverbed, while the channel attention mechanism, through adaptive weighting, allows the model to focus on the most critical features. The feature selection module further filters out features highly correlated with riverbed evolution, reducing redundant information input and improving the model's learning efficiency and prediction accuracy. These designs ensure the comprehensiveness and effectiveness of feature extraction, enabling the model to better capture the complex characteristics of riverbed evolution.

[0095] 3. Multi-channel feature fusion enhances prediction performance: A multi-channel convolutional neural network design enables effective fusion of multi-source features, allowing simultaneous learning of dynamic features related to riverbed morphological evolution across multiple feature dimensions. The multi-scale convolutional fusion module structure allows different feature maps to be input via independent channels, thereby enhancing the hierarchical representation of feature fusion. The weighted fusion of multi-scale convolution and channel attention mechanisms effectively enhances the model's adaptability across different spatial scales, while the pooling layer in the feature compression module reduces redundant information, enabling the model to focus on more significant riverbed change trends. These designs significantly improve the model's ability to capture complex dynamic changes in the riverbed.

[0096] 4. Enhancing the Time-Series Prediction Accuracy of Riverbed Morphology: The time-series regression network, combining an attention mechanism with a two-layer LSTM structure and an adaptive residual correction module, enables the model to achieve high time-series prediction accuracy, effectively capturing both short-term and long-term trends in riverbed morphology. By integrating the attention mechanism with a two-layer LSTM, the model can adaptively allocate weights at each time step of the time series, paying particular attention to changes in riverbed morphology at key moments, capturing both short-term fluctuations and long-term trends. The adaptive residual correction module further improves prediction accuracy and stability by adjusting biases in the time-series prediction, ensuring the model provides predictions of future riverbed morphology that more closely reflect reality. This design allows the model to effectively address the uncertainties of riverbed evolution at different time scales, supporting high-precision prediction of dynamic riverbed processes.

[0097] 5. Enhanced Model Applicability and Robustness: This invention demonstrates strong applicability and robustness in complex river environments. Through multi-source data fusion and innovative feature extraction and regression methods, it can address various riverbed morphological changes caused by natural factors and human activities. The model can accurately identify local erosion and deposition in the riverbed and analyze large-scale river changes, making it suitable for diverse river environments and different riverbed evolution scenarios. Simultaneously, multi-channel and adaptive feature selection methods ensure the model can filter noisy data, and adaptive residual correction further improves prediction robustness. This allows the model to maintain high prediction accuracy and stability even under conditions of uneven data quality and complex environmental changes. Attached Figure Description

[0098] Figure 1 This is the overall logical block diagram of the riverbed evolution prediction method based on multi-source data fusion of the present invention.

[0099] Figure 2 This is a schematic diagram of the overall process for riverbed-related data acquisition and processing in this invention.

[0100] Figure 3 This is a schematic diagram of the feature extraction network processing flow of the present invention.

[0101] Figure 4 This is a schematic diagram of the time series regression network architecture that incorporates the attention mechanism of this invention.

[0102] Figure 5 This is a comparison chart of the predictions of the two-layer LSTM model combined with the attention mechanism in the embodiment. Detailed Implementation

[0103] The invention will be further described below with reference to specific embodiments.

[0104] This invention proposes a riverbed evolution prediction method based on multi-source data fusion. The prediction method involves processing different data source sets using multiple data processing modules combined with a prediction model. The overall flowchart of the multi-source data fusion-based riverbed evolution prediction is shown below. Figure 1 As shown.

[0105] I. Riverbed-related data collection and processing

[0106] In riverbed evolution prediction, data acquisition and collection are fundamental to achieving high-precision analysis and reliable prediction. The collection and integration of multi-source data helps to accurately capture the complexity of riverbed changes, providing crucial support for subsequent model construction and algorithm application.

[0107] 1. Data Collection

[0108] In this invention, to realize a riverbed evolution prediction method based on multi-source data fusion, the first step is to collect and organize relevant riverbed hydrological resources. This allows for the capture of spatial and temporal variation characteristics of the riverbed, improving the accuracy of riverbed morphology and flow field analysis. It provides information on hydrodynamics, topography, and sediment, supporting the analysis of sediment transport, riverbed erosion, and sedimentation.

[0109] Topographic data: This section integrates lidar data, remote sensing imagery, and sonar bathymetry data. Among them:

[0110] (1) LiDAR data was collected by an aircraft-mounted LiDAR sensor flying along the river area to obtain high-precision terrain undulation information, including terrain data of the riverbank and shallow water areas. The LiDAR sensor used had a scanning frequency of 300kHz, a point density of 5 points / square meter, and vertical and horizontal accuracies of ±5-15 cm and ±10-30 cm, respectively. The sampling period was once per quarter. The final data obtained was three-dimensional point cloud coordinate data P1=(x1, y1, z1), which was used to represent the terrain model.

[0111] (2) Remote sensing imagery is captured by drones to monitor changes in river morphology and surrounding vegetation cover, water expansion, and shoreline retreat. Sampling is conducted monthly. The final result is two-dimensional image data P2 = (x2, y2) containing spatial location information, which is used for monitoring changes in water body, vegetation cover, and riverbank, and can also extract river features.

[0112] (3) Sonar bathymetry data is obtained by using sonar equipment and surveying the riverbed from a vessel. This data includes underwater morphological data of the riverbed, determining the river depth, riverbed elevation, and sediment distribution. The spatial resolution is 26 meters, the vertical accuracy is ±10-30 centimeters, and the sampling period is once per season. Finally, three-dimensional point cloud data of the covered area is obtained, which can generate complete underwater topography and riverbed depth variation P3=(X3,y3,z3).

[0113] (4) Hydrodynamic Data: This section includes flow rate, velocity, and water level data. Hydrodynamic data is acquired through velocity sensors and water level gauges deployed at fixed hydrological stations. To determine the impact of flow rate on riverbed evolution and sediment movement, flow rate data is collected at a frequency of approximately 1000 kHz with an accuracy of ±1-2%, and the sampling period is once daily. Simultaneously, changes in water level in the river channel are determined to monitor erosion and sedimentation. The water level gauges have a range of 0-30 m with an accuracy of ±1 mm, and the sampling period is once daily. This data is a time-series information recorded as attributes, stored in CSV format, and represented as P4, to represent the impact of water flow on riverbed morphology and water level fluctuations.

[0114] (5) Sediment Data: This section includes suspended solids concentration data and sediment particle size distribution data. The data is obtained through continuous monitoring at fixed points using water quality sensors. Turbidity measurement range is 0-1000 NTU, with an accuracy of ±5 NTU, and sampling is performed daily. This data is a time-series record stored as an attribute, formatted as CSV. It is represented as P5 to indicate sediment transport and to assess riverbed erosion and deposition.

[0115] (6) Meteorological Data: This section includes precipitation and evaporation data. Continuous measurements at meteorological stations reveal the impact of precipitation on riverbed erosion and flood transport, as well as the impact of evaporation on river balance. Rainfall is measured using a rain gauge ranging from 0-500 mm / h, with a daily sampling period. Evaporation is measured using an evaporation sensor, ranging from 0-20 mm / day, with a daily sampling period. This data is a time-series record stored as an attribute, denoted as P6, to represent the impact of precipitation on river flow and evaporation on water level changes and the watershed's water balance.

[0116] 2. Data anomaly detection and handling

[0117] In the scenario of riverbed evolution prediction, the collected data (P1-P6) are first subjected to anomaly detection. This is mainly used to identify and remove data points with large noise or errors, thereby improving the reliability and accuracy of the data. For different types of data (such as lidar, remote sensing imagery, sonar depth sensing, hydrodynamics, sediment, and meteorological data), this invention employs different anomaly detection methods, the specific process of which includes:

[0118] Calculate basic statistics: Calculate basic statistics for each data type, including the mean. Standard deviation σ, first quartile Q1, third quartile Q3, and interquartile range IQR.

[0119] Select a detection method: Select an appropriate detection method based on the distribution characteristics of the data: If the data is close to a normal distribution, use the three-sigma method. If the data deviates from a normal distribution, use the IQR method.

[0120] Apply the anomaly detection formula: For each data point x in the multi-source data P1 to P6 i Apply the corresponding anomaly detection formula and mark the points that meet the conditions as outliers:

[0121] Three-sigma method formula:

[0122] IQR method formula: x i <Q1 - 1.5×IQR or x i >Q3 + 1.5×IQR

[0123] Record and process outliers: Record the data marked as outliers and decide whether to delete, replace, or conduct further inspections as needed.

[0124] The data after anomaly detection processing is (p1 to p6).

[0125] 3. Time synchronization processing

[0126] Since the acquisition frequencies of the data of each part p1 to p6 are different, in order to uniformly process and analyze the data, time synchronization processing is required. To unify the time, the present invention performs interpolation processing on the missing data in units of days.

[0127] The present invention uses the method of linear interpolation for interpolation to complete time synchronization. It is expressed as follows:

[0128]

[0129] Among them, p day (t) is the daily data after interpolation, t0 and t1 are the time points of two adjacent months respectively, p(t1) and p(t0) are the data values of the corresponding months respectively. T is the specific date for which interpolation is required. After time synchronization, a new data set p1 to p6 is obtained.

[0130] 4. Data fusion

[0131] It can be understood that in the prediction of riverbed evolution, the collected topographic, hydrodynamic, sediment, and meteorological data all come from different sensors and detection devices, and the data types collected are different. When analyzing the data, it is necessary to fuse the data to construct a comprehensive model reflecting the evolution of the riverbed morphology.

[0132] In this invention, the lidar generates approximately 7 million data points quarterly, with a single data collection of 100-500 MB / square kilometer; remote sensing impact data generates approximately 10 MB of data and 48 images over a one-square-kilometer area; sonar depth sounding generates approximately 40,000 records per square kilometer; hydrodynamic data includes approximately 8,760 flow records; sediment data includes approximately 8,760 records; and meteorological data also includes approximately 8,760 records.

[0133] To integrate different data types collected from multiple sensors, data fusion is necessary. After anomaly detection, the data (p1 to p6) are fused. This invention employs a weighted inverse distance interpolation algorithm to integrate terrain, hydrodynamic, and attribute data of different spatial resolutions and data types into a unified three-dimensional mesh, facilitating further analysis and prediction.

[0134] Specifically, weighted inverse distance interpolation (IDW) estimates the attribute values ​​of unsampled locations by averaging the sampled point data according to distance. This is used to fuse data from lidar, sonar bathymetry, hydrodynamics, and sediment properties at different resolutions and sampling frequencies, constructing a unified spatial grid to support subsequent analysis and modeling of riverbed morphology evolution. The following is the specific process for achieving multi-source data fusion.

[0135] Define a unified grid: Based on the scope and resolution requirements of the riverbed area under study, define a unified three-dimensional spatial grid. Determine the minimum and maximum x, y, z coordinates, denoted as (x...). min y min , z min ) and (x max y max , z max ), used to determine the spatial extent of the grid. The position of each grid point is (x... i y i .z i ), where i represents the index of the grid point. The grid resolution should take into account both the original data resolution and the target application accuracy; in this invention, an average resolution of 1m for the lidar is adopted.

[0136] Unified data conversion:

[0137] (1) The data collected by the lidar is three-dimensional point cloud data, given in the form p1 = (x1, y1, z1), representing the three-dimensional coordinates of the Earth's surface. Converting p1 to a projected coordinate system:

[0138]

[0139] Z Radar =h

[0140] in, γ represents the original geographic coordinates, y0 represents the central longitude of the projected coordinate system, N is the radius of curvature, and h is the elevation data, i.e., the elevation value measured by the lidar.

[0141] (2) The data acquired by remote sensing images are image data. Since they are two-dimensional image data, remote sensing image data usually only contains planar coordinates p2 = (x2, y2), which can be directly mapped onto the plane in the three-dimensional grid. The z value is 0 by default.

[0142]

[0143] z Ro =0

[0144] (3) The sonar depth data p3 = (x3, y3, z3) records the water depth values ​​(relative to the water surface reference), which need to be converted to absolute elevation, i.e., the distance from the vertical line of the ground point to the geoid, to conform to a unified three-dimensional projected coordinate system. The absolute elevation z is calculated as follows:

[0145] z S =HD

[0146] Where D represents the water depth relative to the water surface, and H represents the elevation of the water surface.

[0147] (4) Hydrodynamic data p4 includes water level and velocity information, typically time-series data, and is collected at fixed points. Water level data can be directly mapped onto the z-axis in the projected coordinate system as z-values. Velocity can be decomposed into three-dimensional components (v... s v y v z The data is interpolated to each grid point on the corresponding plane of the 3D mesh. Let the coordinates of the hydrodynamic sampling point be (x...). h y h The water level is z. h Then directly use the water level data z h Assign the value to the z-coordinate of the grid point. That is: z = z h .

[0148] (5) The sediment data includes suspended sediment concentration and is time-series data. Since sediment data is mostly collected at specific cross-sections, it needs to be spatially interpolated into a three-dimensional grid. Suspended sediment concentration data can be directly recorded on the z-plane of the grid, i.e., z = z0. m The sediment concentration is then interpolated onto the grid using the IDW method. For any target point Q(x, y) on the grid plane, its sediment concentration is represented as p5(x5, y5), which can be calculated by interpolation based on the concentration data p5 of the surrounding sampling points q5(xj, yj). The interpolated value p5(Q) is expressed as follows:

[0149]

[0150] Where p5(q5) represents the sediment concentration value at sampling point q5. This represents the distance between grid point Q and sampling point q5. A is the attenuation exponent, with a value of 2, and n is the number of sampling points involved in the interpolation. The above interpolation calculation is repeated for each grid point until the sediment concentration value p5 of all grid points is obtained through IDW interpolation.

[0151] (6) Meteorological data is usually two-dimensional (such as precipitation and evaporation), and is generally directly assigned to the surface layer of the grid. The mapping of the precipitation and evaporation data P6 at each grid point is: z = zW e .

[0152] After the above transformation steps, all data are standardized and transformed into a unified three-dimensional coordinate system. LiDAR and sonar depth data provide z-coordinate values ​​for constructing three-dimensional topography of the surface and underwater. Remote sensing imagery data is mapped to the (x, y) plane and assigned attribute values ​​to the surface layer. Hydrodynamic and sediment data are distributed to different depth layers of the grid through spatial interpolation. Meteorological data is mapped to the surface grid. The final result is data P represented by a three-dimensional grid. grid .

[0153] This standardization and transformation method ensures the spatial and temporal consistency of multi-source data, providing unified data information for riverbed evolution prediction. The overall process of riverbed-related data acquisition and processing is as follows: Figure 2 As shown.

[0154] II. Feature Extraction Network Design

[0155] In riverbed evolution prediction, multi-source data fusion provides the model with rich spatial, temporal, and topographical features. The fused 3D grid data P is obtained by uniformly transforming and mapping the data through weighted inverse distance interpolation to a 3D grid. grid This includes data on riverbed depth, flow velocity, water level, precipitation evaporation, and sediment concentration. Because these features vary significantly in spatial scale and scope, directly inputting them into the model may lead to information redundancy and noise accumulation. To improve the model's learning efficiency and prediction accuracy, it is necessary to extract key information from multi-scale features and reduce redundancy through effective feature selection techniques.

[0156] Therefore, this invention proposes a multi-scale convolutional neural network and an adaptive feature selection algorithm based on mutual information to solve the above problems. Multi-scale convolution can optimize dataset P. grid The model is decomposed layer by layer at different spatial scales to extract local and global features; while feature selection based on mutual information can screen out features that are highly correlated with riverbed evolution but are independent of each other, and further optimize the model.

[0157] Riverbed environmental information P based on multi-source data fusion grid It is a three-dimensional grid data with dimensions P. grid ∈R H ×W×D Where H and W correspond to the spatial dimensions of the riverbed area, respectively. D represents the number of data channels, which in this invention include information from different data sources, such as water depth, flow velocity, water level, precipitation evaporation, and sediment concentration.

[0158] 1. Multi-scale convolutional neural network module

[0159] In riverbed evolution prediction, multi-scale feature extraction methods help models identify complex riverbed evolution patterns by capturing feature changes at different spatial scales. Multi-scale feature extraction is a technique that extracts information from different scales and resolutions to capture multi-level features of data, and is particularly suitable for spatially complex or multi-pattern tasks.

[0160] Specifically, firstly, multi-scale convolutional kernel design is performed. This is for 3D mesh data P. grid The training and test set ratio was set to 8:2 to ensure the model had sufficient training samples and generalization ability. Multi-scale convolutional kernels were designed, with different kernel sizes (k*k, k = 3, 5, 7) to extract multi-level spatial information. ReLU was used as the activation function for the output activation of the convolutional layers. A temporal dimension was also added to construct a three-dimensional convolutional kernel. The formula for calculating the three-dimensional convolution is as follows:

[0161] F k =σ(P grid *W k +b k )

[0162] Among them, W k ∈R k×k×D b is a convolution kernel matrix of scale k, used to extract features at a specific scale; k For the corresponding bias term, σ is the activation function, and F is the activation function. k This represents the feature map at scale k, which contains the dataset P. grid Local and global information at this scale.

[0163] Secondly, feature fusion is performed. This is to capture data P more comprehensively. grid The multi-scale features are fused by fusing the feature maps {F3, F5, F7} at each scale. Feature fusion is performed using a weighted summation method.

[0164] F multi =α·F3+β·F5+γ·F7

[0165] Where α, β, and γ represent the fusion weights, which are adaptively learned through model training. The mean squared error is used as the loss function to measure the difference between the predicted and actual values, optimizing the model's prediction accuracy for riverbed evolution characteristics. The fused feature map F... multi It includes spatial features at different scales, which are used as inputs for subsequent feature selection and modeling.

[0166] This invention introduces a temporal dimension to the existing multi-scale convolutional feature extraction, increasing sensitivity to temporal evolution. Specifically, this invention proposes a three-dimensional convolutional kernel: that is, performing convolution operations in the H×W×T space, where T represents the time step.

[0167] This method allows for the selection of recent data segments for convolution processing in the time dimension to capture short-term trends in riverbed changes, while combining historical data to obtain long-term deposition and erosion trends.

[0168] Furthermore, adding a time dimension, the formula for calculating 3D convolution is as follows:

[0169] F k，t =σ(P grid *W k，t +b k，t )

[0170] Among them, W k，t ∈R k×k×T×D The kernel is a three-dimensional convolution, including spatial dimension D and temporal dimension T; b k，t For the corresponding bias term, F k，t This represents the feature map at scale k and time step T.

[0171] By adding a time dimension, we can not only extract features at different scales in space, but also extract multi-scale evolution information in time, providing richer feature support for predicting the future trend of riverbeds.

[0172] 2. Mutual Information Feature Selection Module

[0173] To further extract effective features related to riverbed evolution, the fused feature map F... multi Feature selection is performed. Feature selection is based on an improved mutual information algorithm to ensure that the model only receives features that significantly contribute to the target prediction.

[0174] Specifically, the first step is to measure feature importance. The importance of a feature is determined by calculating the mutual information (MI) between each feature channel and the target variable (future riverbed height change). MI measures the degree of informational correlation between a feature and the target variable, and its calculation formula is as follows:

[0175]

[0176] Among them, F i Represents the fusion feature F multi The i-th feature channel in the equation; Y is the target variable, i.e., the change in riverbed height; p(f i ,y) is the feature channel F i The joint probability distribution of the target variable Y and p(f i p(y) and p(y) are marginal probability distributions.

[0177] Secondly, improved mutual information feature selection. It's understandable that traditional mutual information only measures the correlation between features and the target, without considering feature redundancy. The improved algorithm avoids redundant features entering the model by introducing correlation between features. The specific steps are as follows:

[0178] (1) Calculate F for each feature i The corresponding mutual information MI(F) i ;Y).

[0179] (2) Calculate the feature pair F i Compared with other features F j The correlation coefficient ρ(F) between them i F j In this invention, the Pearson correlation coefficient is used:

[0180]

[0181] (3) Define effective mutual information (MI) eff (F i ;Y) to correct redundancy between features:

[0182]

[0183] Where δ represents the weighting coefficient, which balances the correlation between features and the target and the redundancy between features.

[0184] (4) According to MI eff (F i ;Y) Sort the features and select the top N features F with the highest effective mutual information. selected This serves as the input for subsequent multi-channel convolutional neural network models.

[0185] F selected =(F k，t |MI eff (F i ;Y)≥ε}

[0186] Among them, F selected It is the feature set filtered by mutual information. ε is the threshold for mutual information, that is, only feature maps with mutual information greater than or equal to the threshold are selected.

[0187] The feature matrix F after mutual information feature selection selected This feature matrix will be used as input to the model for spatiotemporal prediction. It not only contains salient features at different scales but also filters out redundant information to improve the model's prediction efficiency and accuracy. The feature extraction network flowchart is shown below. Figure 3 As shown.

[0188] By extracting features at different spatial and temporal scales, it helps identify topographic changes and dynamic patterns at different scales during riverbed evolution. By using convolution kernels of different scales, multi-level feature information from local to global can be obtained, covering subtle local evolution, sedimentary features, and overall topographic trends at larger scales.

[0189] III. Multichannel Convolutional Neural Networks

[0190] Multichannel Convolutional Neural Networks (MC-CNN) are convolutional neural network architectures for processing multidimensional, multi-feature data. The key to MC-CNN lies in simultaneously inputting multiple input feature maps into the model for convolution operations, enabling the network to learn and extract features across multiple channels simultaneously, thereby enhancing its ability to capture and understand complex information. It efficiently processes multi-source data through modules such as multichannel input, feature extraction, and fusion. First, the multi-scale convolutional fusion module receives multiple features, uses convolutional kernels of different sizes to extract detailed and large-scale features, and focuses on the most relevant features through a channel attention mechanism, fusing multi-channel information. The feature compression module further compresses feature maps through pooling layers, improving the model's training efficiency. When processing time-series data, it captures temporal dependencies through LSTM layers, generating dynamic change predictions. The synergistic effect of these modules makes MC-CNN outstanding in multidimensional data processing tasks.

[0191] In the task of predicting riverbed evolution, a multi-scale feature map F was generated using a feature extraction network. selected These feature maps encompass both local and global information about the dynamic changes in the riverbed. To fully utilize these multi-scale features and further improve the model's ability to capture complex riverbed evolution patterns, this invention introduces a multi-channel convolutional neural network. Through multi-channel design, these feature maps are used as multiple input channels of the network, enabling the model to process data simultaneously at multiple scales, thereby avoiding information loss and better capturing the complex changes in riverbed morphology.

[0192] Because feature maps contain not only riverbed morphology information but also various data such as hydrodynamics and sediment concentration, these features reflect the hierarchical characteristics of riverbed morphology at different scales. The multi-channel structure of MC-CNN allows for the design of separate convolution operations for each feature channel, enabling the model to analyze each feature map independently and then combine the results to achieve hierarchical representation and fusion of different information sources.

[0193] 1. Multi-scale convolutional fusion module

[0194] Based on the obtained multi-scale feature set F selected This feature set is composed of features from different scales and different data sources, and contains rich information on riverbed changes.

[0195] Specifically, the multi-source feature set is F selected = [F1, F2, ..., F N ], where each F i ∈R H×W This represents a feature map in the spatial dimension (including water depth, flow velocity, water level, precipitation evaporation, and sediment concentration), where N represents the number of feature channels.

[0196] Furthermore, the feature set is used as the multi-channel input to MC-CNN to construct the network's input matrix X = [F1, F2, ..., F...]. N Where X∈R H×W×N This represents an input tensor with N channels.

[0197] First, the multi-scale convolutional unit uses multiple convolutional kernels to extract features from the input multi-channel data. Each convolutional kernel spatially identifies riverbed change information at different scales, revealing the local evolution, sediment accumulation, and overall trend of the riverbed.

[0198] We design convolutional layers with multiple kernel sizes to extract features at different spatial scales. For each kernel size, we perform a convolution operation on each channel to obtain the corresponding feature map. The multi-channel convolutional neural network, incorporating a channel attention mechanism, also uses ReLU as the activation function and Mean Squared Error (MSE) as the loss function in the convolutional layers to optimize the multi-channel feature fusion results, making the prediction more accurate.

[0199] The formula for calculating convolution is as follows:

[0200]

[0201] in, Let X represent the feature map of the i-th layer at scale k. (l-1) This is the output feature map of the previous layer. This represents the convolution kernel of the l-th layer at scale k. σ is the bias term. σ is the ReLU activation function.

[0202] Secondly, the fusion unit extracts multi-scale feature maps through convolution. The features of each layer are fused together to obtain the fusion results. A channel attention mechanism is introduced during this process, enabling the network to adaptively allocate channel weights based on the current feature information.

[0203] Specifically, the weight of each channel is defined as α. k The channel weighted fusion result is:

[0204]

[0205] in:

[0206]

[0207] g(·) is the channel weight learning function. By adaptively weighting the feature mapping of each channel, it ensures that the MC-CNN model can adaptively focus on different riverbed features in different regions, thereby improving its adaptability to the spatial characteristics of the riverbed.

[0208] 2. Feature compression module:

[0209] Pooling layers spatially compress the feature maps output by convolution, preserving key features and reducing redundant information, thus avoiding overfitting caused by redundant features. In riverbed evolution prediction models, pooling layers help highlight large-scale evolution trends while mitigating short-term or small-scale random changes, thereby focusing more on stable evolution signals.

[0210] The feature maps fused from multiple channels are pooled to further compress the feature space while retaining key features. This invention employs average pooling to reduce data dimensionality, thereby improving computational efficiency and reducing the impact of redundant information on the model.

[0211]

[0212] Where pooling() represents the average pooling operation, and the compressed feature map This serves as the input for the next layer.

[0213] Since riverbed evolution exhibits time-series characteristics, a recurrent neural network (LSTM) layer is added after the output of the convolutional layer to capture dynamic changes in the time dimension, such as the temporal variations in features like riverbed height and sedimentation rate. The output of the pooling layer is then input into the LSTM unit to obtain the sequential spatiotemporal features.

[0214]

[0215] y t This demonstrates the cumulative effect of riverbed change characteristics over time. Through time series modeling, MC-CNN can predict possible riverbed change patterns in future time periods.

[0216] IV. Time Series Regression Network Design

[0217] Riverbed morphology changes are influenced not only by local spatial factors (topography and flow velocity) but also exhibit dynamic evolution over time. Therefore, this invention considers time-series characteristics when predicting future riverbed changes. Designing a time-series regression network can generate more temporally consistent future predictions based on current and historical riverbed characteristics, improving prediction accuracy and stability.

[0218] Specifically, after extracting multi-scale features and obtaining preliminary prediction output y from a multi-channel convolutional neural network, this invention introduces a time-series regression network to more accurately predict the future evolution trend of the riverbed. The design goal of this network is to combine the spatial features of MC-CNN with the temporal dynamics of the riverbed, capturing the patterns and trends of riverbed changes over time to optimize the prediction accuracy of future riverbed morphology. A two-layer LSTM combined with an attention mechanism is introduced to more accurately predict the dynamic trend of riverbed evolution. Based on the constructed time-series data, the short-term and long-term dependencies of the time series are first extracted through a two-layer LSTM structure. The first LSTM focuses on capturing short-term change features, while the second LSTM further analyzes long-term trends to generate preliminary prediction results. An attention mechanism is added to adaptively allocate weights to each time step, making the model more focused on time step information that has a key impact on future evolution. Then, an adaptive residual correction module is used to further optimize the prediction results to dynamically compensate for biases and improve the overall prediction accuracy and the model's temporal consistency. Through this module design, the model can achieve high-precision prediction of riverbed morphology, providing scientific support for river management and water resource regulation.

[0219] 1. Time series data construction

[0220] The predicted result y output by MC-CNN t This represents the spatiotemporal characteristics of the riverbed sequence at time step t. To capture its trend over time, y needs to be... t Transform it into a time series input. In the specific time series construction, the y-values ​​of multiple consecutive time steps are... t The values ​​are used as the input sequence for the LSTM.

[0221] Specifically, assuming the total time step is T, construct the time series input Y = [y1, y2, ..., y3]. T ], where each y t This represents the predicted state of the riverbed at a specific point in the past. This time series is then input into a regression module to analyze the temporal trends of these features and predict future riverbed evolution.

[0222] 2. Design of a two-layer LSTM structure incorporating attention mechanism

[0223] In the time series regression module, this invention designs a two-layer LSTM architecture incorporating an attention mechanism to extract complex dependencies in time series data. The two-layer LSTM structure is more deep and flexible than a single-layer LSTM, better capturing both short-term and long-term temporal dependencies. The introduction of the attention mechanism allows the model to automatically select and assign higher weights to key time steps, highlighting features that have a greater impact on prediction. The working principle of the two-layer LSTM with attention mechanism is as follows:

[0224] The Tanh function, suitable for LSTM activation, is used as the activation function and is well-suited for time series data processing. Similarly, the mean squared error is used as the loss function. The loss function is used to optimize the continuous predicted values ​​output by the LSTM. The test samples are time series riverbed data at different time points.

[0225] In the first layer of LSTM, the input data y is received at each time step t. t And the hidden state of the previous time step. and memory unit The calculation formula for the first LSTM layer is:

[0226]

[0227] in, These represent the hidden state and memory unit of the first LSTM layer at time step t. As the output of the first layer of LSTM.

[0228] The first layer of LSTM is mainly used to capture short-term dynamic features of riverbed changes, such as changes in erosion and deposition rates over a short period of time.

[0229] An attention mechanism is introduced, generating a set of attention weights at each time step t to help the model adaptively allocate attention across different time steps. The formula for calculating the attention weights μt is:

[0230]

[0231] Among them, ω, W a b a μ is a learnable parameter. t The attention weight at time step t represents the importance of the features at that time step. Through the weight calculation of the attention mechanism, the model can assign different weights to each time step, so that key time steps receive higher weights during the model learning process, thereby effectively capturing the key dynamic changes of the riverbed.

[0232] The output of the first LSTM layer after incorporating the attention mechanism represents the hidden states of all time steps. We obtain a global weighted feature representation s by summing the values ​​according to the attention weights. (1) :

[0233]

[0234] Weighted s (1) This indicates the short-term characteristics of riverbed evolution.

[0235] In the second LSTM layer, the weighted output s of the first LSTM layer is... (1) The second LSTM layer is used as input to further extract long-term dependency features. The calculation formula for the second LSTM layer is:

[0236]

[0237] in, These represent the hidden state and memory unit of the second-layer LSTM at time step t. As the output of the second LSTM layer.

[0238] The second layer of LSTM focuses primarily on long-term trends in riverbed morphology, such as seasonal or interannual deposition and erosion patterns.

[0239] Furthermore, the hidden states generated by the two-layer LSTM Used to generate preliminary prediction results at time step t+1

[0240] The introduction of a two-layer LSTM structure with an attention mechanism can effectively extract the dynamic features of riverbed evolution in the time dimension. By extracting short-term and long-term features in stages through the two-layer structure, the prediction results of the model are more accurate and stable.

[0241] 3. Adaptive Residual Correction Module

[0242] In time series regression, predictions often exhibit some bias. To address this, an adaptive residual correction module is designed to further improve prediction accuracy by learning from residual information. The design process of the residual correction module is as follows:

[0243] Residual calculation: Calculate the true value Y at each time step t. t and predicted value The difference between them, i.e., the residual:

[0244]

[0245] Where, r t This represents the prediction error at time step t.

[0246] Residual Correction Model: This model uses linear regression to fit the residuals. It learns the model bias based on past predictions and generates a residual correction value for each time step.

[0247]

[0248] Where g(·) represents the residual correction model, and m is the number of time steps of the input.

[0249] Final prediction output: The original prediction results of the two-layer LSTM model. and residual correction value Add them together to obtain the final prediction result:

[0250]

[0251] The output of the time series regression network is the final predicted sequence. It is a prediction of the riverbed state at a future time step t+1, including predictions of future water depth, flow velocity, water level, precipitation evaporation, and sediment concentration.

[0252] Residual correction dynamically adjusts the model's bias at different time steps, thereby improving the accuracy and robustness of the time series regression model. A two-layer LSTM structure captures short-term and long-term dynamic changes in riverbed morphology, ensuring the model can effectively identify seasonal or interannual riverbed evolution. The model's accuracy in the time dimension is evaluated by comparing the prediction errors of LSTM models with different time series lengths. The flowchart of the dual-time series regression module incorporating the attention mechanism is shown below. Figure 4 As shown.

[0253] The feature extraction network, multi-channel convolutional neural network, and time series regression network are trained using the following steps: First, the three-dimensional mesh data P is... grid The dataset was divided into training and testing sets in an 8:2 ratio using random partitioning to ensure sufficient training samples and generalization ability. Secondly, the average of the sums of five indicators—the mean square error between the true and predicted values ​​of water depth, flow velocity, water level, precipitation evaporation, and sediment concentration—was used as the loss function for training the feature extraction network, multi-channel convolutional neural network, and time series regression network. Finally, training of the feature extraction network, multi-channel convolutional neural network, and time series regression network was considered complete when the loss function value calculated using the validation set failed to decrease for 10 consecutive epochs.

[0254] To verify the effectiveness of the proposed dual-layer LSTM model incorporating an attention mechanism, simulations were conducted comparing it with a baseline LSTM at 100 time steps. By comparing the prediction results of the baseline LSTM and the dual-layer LSTM incorporating the attention mechanism, it can be seen that the introduction of the attention mechanism significantly improves the model's prediction accuracy, especially when dealing with complex time-series data. When combining multi-source data for riverbed evolution prediction, the dual-layer LSTM incorporating the attention mechanism can effectively fuse multi-source information, thereby improving prediction performance. The comparison figures are shown below. Figure 5 As shown.

[0255] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0256] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting riverbed evolution based on multi-source data fusion, characterized in that, Includes the following steps: S1 collects multi-source riverbed data, preprocesses the data, and merges it to form a unified three-dimensional grid data. The acquisition of multi-source riverbed data includes topographic data, hydrodynamic data, sediment data, and meteorological data. The topographic data includes three-dimensional point cloud coordinates of topographic relief information acquired via lidar, two-dimensional image data of river channel morphology, surrounding vegetation cover, water body expansion, and shoreline retreat changes acquired via remote sensing imagery, and three-dimensional point cloud data of underwater riverbed morphology and depth acquired via sonar. The hydrodynamic data includes flow rate data, velocity data, and water level data. The sediment data includes suspended solids concentration data and sediment particle size distribution data. The meteorological data includes precipitation data and evaporation data. The process of fusing the data to form a unified three-dimensional mesh is as follows: Define a uniform grid: Based on the scope and resolution requirements of the riverbed area under study, define a uniform three-dimensional spatial grid, and determine the minimum and maximum values. x, y, z Coordinates, denoted as ( )and( ), used to determine the spatial extent of the grid, where the position of each grid point is ( ),in, Indicates the index of the grid points; the grid resolution is based on the average resolution of the LiDAR, which is 1m. Unified Data Conversion: The data collected by the lidar is 3D point cloud data, given in the form p1=(x1,y1,z1), representing the 3D coordinates of the ground surface. p1 is then converted to a projected coordinate system. ； ； ； in, and Original geographic coordinates The longitude of the center of the projected coordinate system is represented by N, the radius of curvature is N, and h is the elevation data, i.e. the elevation value measured by the lidar. The remote sensing image data only contains planar coordinates p2=(x2,y2), which are directly mapped onto the plane in the 3D grid, and the z value is 0 by default; The sonar data p3=(x3,y3,z3) records the water depth value, which needs to be converted into absolute elevation, i.e., the distance from the vertical line of the ground point to the geoid, in order to conform to a unified three-dimensional projected coordinate system. The absolute elevation z is calculated as follows: ； Where D is the water depth data, which is the depth relative to the water surface, and H is the elevation of the water surface; Hydrodynamic data p4 includes water level and flow velocity information; water level data is directly mapped onto the z-axis in the projected coordinate system as the z-value; flow velocity is decomposed into three-dimensional components ( Interpolate to each grid point on the corresponding plane of the 3D mesh; The sediment data includes the suspended sediment concentration p5(x5,y5), which is directly recorded on the z-plane of the grid. The sediment concentration p5 of all grid points is obtained by interpolation using the IDW method. This involves interpolating the concentration data of the surrounding sampling points q5(xj,yj) and repeating the interpolation calculation for each grid point until the sediment concentration p5 of all grid points is obtained by IDW interpolation. Meteorological data is in two-dimensional format, P6, including precipitation and evaporation, and is directly assigned to the surface layer of the grid. After the above transformations, all data are standardized and converted to a unified three-dimensional coordinate system. LiDAR and sonar data provide z-coordinate values ​​for constructing three-dimensional topography of the surface and underwater areas. Remote sensing imagery data is mapped to the (x,y) plane, assigning attribute values ​​to the surface layer. Hydrodynamic and sediment data are spatially interpolated and distributed to different depth layers of the grid. Meteorological data is mapped to the surface grid, ultimately yielding data represented by a three-dimensional grid. ; S2, Based on the feature extraction network, a multi-scale feature map is generated; the feature extraction network includes a multi-scale convolutional neural network module and a mutual information feature selection module. The multi-scale convolutional neural network module extracts riverbed morphological features at different scales and combines them with a channel attention mechanism for weighted fusion; the mutual information feature selection module eliminates redundant features and obtains key features that cover the dynamic changes of the riverbed. S3, based on a multi-channel convolutional neural network, simultaneously learns and extracts features across multiple channels, performing hierarchical representation and fusion of different information sources; the multi-channel convolutional neural network includes a multi-scale convolutional fusion module and a feature compression module; the multi-scale convolutional fusion module uses convolutional kernels of different sizes to extract detailed and large-scale features, and focuses on the most relevant features through a channel attention mechanism, fusing multi-channel information; the feature compression module further compresses the feature map through pooling layers, and when processing time-series data, it captures time dependencies through LSTM layers to generate serialized spatiotemporal features; S4, based on a time series regression network, extracts complex dependencies in the time series through a two-layer LSTM structure, and improves prediction accuracy by learning residual information through an adaptive residual correction module, thus obtaining the final prediction result.

2. The method for predicting riverbed evolution based on multi-source data fusion as described in claim 1, characterized in that: The multi-scale convolutional neural network module in the feature extraction network is specifically as follows: For 3D mesh data We designed multi-scale convolutional kernels, setting different kernel sizes (k*k) to extract multi-level spatial information. We also added a temporal dimension to construct a three-dimensional convolutional kernel. The formula for calculating the three-dimensional convolution is as follows: ； in, It is a three-dimensional convolution kernel, including spatial dimensions. and time dimension ; For the corresponding bias term, This represents the feature mapping at scale k and time step T; Secondly, in order to capture data more comprehensively Multi-scale features, mapping features at each scale { The features are fused using a weighted summation method: ； in, , , This is represented as fusion weights, which are used for adaptive learning during model training.

3. The method for predicting riverbed evolution based on multi-source data fusion as described in claim 2, characterized in that: The mutual information feature selection module in the feature extraction network performs the following specific process: First, feature importance is measured. This is done by calculating the mutual information (MI) between each feature channel and the target variable to determine the importance of the feature. MI measures the degree of information correlation between the feature and the target variable, and its calculation formula is as follows: ； in, Indicates fusion features The first in One feature channel; The target variable is the change in riverbed height. For feature channels The joint probability distribution of the target variable Y. and It represents a marginal probability distribution; Then calculate the feature pairs. Compared with other features Correlation coefficient between Pearson correlation coefficient was used: ； Then, effective mutual information is defined. To correct redundancy between features: ； in, This represents the weighting coefficient, which balances the correlation between features and the target, as well as the redundancy among features. Finally, based on Sort the features and select the top N features with the highest effective mutual information. This serves as the input for subsequent multi-channel convolutional neural network models. ； in, It is the feature set filtered through mutual information. The threshold for mutual information is used to select only feature maps whose mutual information is greater than or equal to the threshold. Feature matrix after mutual information feature selection This will be used as input to the model for spatiotemporal prediction.

4. The method for predicting riverbed evolution based on multi-source data fusion as described in claim 1, characterized in that: The multi-scale convolutional fusion module in S3 includes multi-scale convolutional units and fusion units; The multi-scale convolutional unit is designed with multiple convolutional layers of different kernel sizes to extract feature information at different spatial scales. For each kernel size, a convolution operation is performed on each channel to obtain the corresponding feature map. The convolution calculation formula is as follows: ； in, Indicates the first l Feature mapping of the layer at scale k This is the output feature map of the previous layer. Indicates the first l The convolution kernel of the layer at scale k, For bias terms, The activation function is ReLU; The fusion unit introduces a channel attention mechanism, enabling the network to adaptively allocate channel weights based on current feature information; this applies to the multi-scale feature maps extracted through convolution. Perform fusion to obtain the feature fusion results of each layer; define the weight of each channel as follows. The channel weighted fusion result is: ； in, ; This is the channel weight learning function, which adaptively weights the feature maps of each channel.

5. The method for predicting riverbed evolution based on multi-source data fusion as described in claim 4, characterized in that: The feature compression module in S3 is specifically as follows: The feature maps after multi-channel fusion are pooled through a pooling layer to further compress the feature space and retain key features: ； in, This represents the average pooling operation, and the compressed feature map. As input for the next layer; Simultaneously, a recurrent neural network (LSTM) layer is added after the output of the convolutional layer to capture dynamic changes in the time dimension, such as the temporal variations of features like riverbed height and sedimentation rate. The output of the pooling layer is then input into the LSTM layer to obtain serialized spatiotemporal features. .

6. The method for predicting riverbed evolution based on multi-source data fusion as described in claim 1, characterized in that, The data processing procedure for the time series regression network is as follows: Based on the constructed time series data, firstly, a two-layer LSTM structure is used to extract the short-term and long-term dependencies of the time series. The first layer of LSTM focuses on capturing short-term change features, while the second layer of LSTM further analyzes long-term trends to generate preliminary prediction results. Then, an attention mechanism is added to adaptively allocate the weights of each time step, making the model more focused on the time step information that has a key impact on future evolution. Finally, an adaptive residual correction module is used to further optimize the prediction results to dynamically compensate for biases and improve the overall prediction accuracy and the model's temporal consistency.

7. The method for predicting riverbed evolution based on multi-source data fusion as described in claim 6, characterized in that, The dual-layer LSTM structure is specifically as follows: In the first layer of LSTM, the input data at each time step t is received. And the hidden state of the previous time step. and memory unit The calculation formula for the first LSTM layer is: ； in, These represent the hidden state and memory unit of the first LSTM layer at time step t. As the output of the first LSTM layer; An attention mechanism is introduced, generating a set of attention weights at each time step t to help the model adaptively allocate attention across different time series steps. The calculation formula is: ； in, , , For learnable parameters, It is the attention weight at time step t, representing the importance of the feature at the current time step; The output of the first LSTM layer after incorporating the attention mechanism represents the hidden states of all time steps. We obtain a global weighted feature representation by summing the values ​​according to the attention weights. : ； Weighted Indicates the short-term characteristics of riverbed evolution; In the second LSTM layer, the weighted output of the first LSTM layer is... As input to the second LSTM layer, long-term dependency features are further extracted; the calculation formula for the second LSTM layer is: ； in, These represent the hidden state and memory unit of the second-layer LSTM at time step t; As the output of the second LSTM layer; Hidden state generated by two-layer LSTM Used to generate preliminary prediction results at time step t+1 .

8. The method for predicting riverbed evolution based on multi-source data fusion as described in claim 6, characterized in that, The adaptive residual correction module is specifically as follows: Residual calculation: Calculate the true value at each time step t. and predicted value The difference between them, i.e., the residual: ； in, This represents the prediction error at time step t; Residual Correction Model: Using linear regression to fit the residuals, the residual correction model learns the model bias based on past predictions and generates a residual correction value for each time step. ； in, This represents the residual correction model, where m is the number of time steps in the input. Final prediction output: The original prediction results of the two-layer LSTM model. and residual correction value Add them together to obtain the final prediction result: ； The output of the time series regression network is the final predicted sequence. ; It is a predicted value of the riverbed state at a future time step t+1.

Images