Riverbed evolution prediction method based on multi-source data fusion

Abstract

The invention provides a riverbed evolution prediction method based on multi-source data fusion, and belongs to the technical field of hydrological resource monitoring. The method comprises the following steps: acquiring and fusing multi-source data by using different devices, performing data preprocessing to form unified three-dimensional grid data, and then generating a multi-scale feature map based on a designed feature extraction network; based on the designed multi-channel convolutional neural network, learning and extracting features on multiple channels at the same time, and performing hierarchical expression and fusion on different information sources; and finally, based on the designed time sequence regression network, extracting a complex dependency relationship in the time sequence through a double-layer LSTM structure, learning residual error information through an adaptive residual error correction module to improve prediction precision, and obtaining a final prediction result. The method shows relatively high applicability and robustness in a complex river channel environment, and can cope with various river bed form changes caused by natural factors and human activities through multi-source data fusion and an innovative feature extraction and regression method.

Description

Technical Field

[0001] The invention belongs to the technical field of hydrological resource monitoring, and in particular relates to a riverbed evolution prediction method based on multi-source data fusion. Background Art

[0002] Globally, riverbed morphology is affected by multiple factors, including natural factors and human activities. Accurate prediction of riverbed morphology is of great significance for water resource management, ecological protection, and disaster prevention and mitigation. In recent years, with the intensification of human activities and the impact of climate change, the riverbed morphology of many rivers has changed significantly, including deepening of river channels, riverbed erosion, and sedimentation changes. These changes may have a profound impact on water quality, ecological environment, riverbank stability, and flood risks.

[0003] The evolution of riverbed morphology is a dynamic process that includes phenomena such as riverbed erosion and siltation, and is driven by natural factors such as water flow velocity, river channel curvature, and topographic changes. At the same time, artificial structures such as dams, dikes, sand mining, and channel dredging have also significantly changed water flow patterns and sediment transport behavior. For example, the construction of dams not only changes the speed and direction of water flow, but also causes a reduction in downstream sediments, leading to riverbed erosion and increasing the depth of the river channel. The complex interaction between these human activities and natural factors affects the dynamic evolution of riverbed morphology and aggravates the degradation of river channels. However, the 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 the river channel, limiting the spatial and temporal scale of the research. On the other hand, numerical models provide an effective tool for studying the evolution of riverbed morphology, especially in large river systems. However, the complexity of numerical simulations makes the process of building and calibrating models very technically demanding.

[0004] In this context, by introducing machine learning technology, large-scale data can be efficiently processed and analyzed, thereby making more accurate predictions of riverbed evolution. However, the existing technology has the following defects:

[0005] 1. Data acquisition and quality control issues: Current riverbed morphology research relies on multiple data sources, such as lidar, remote sensing images, sonar sounding, and hydrological station data. However, due to the different sources of these data, their resolution, accuracy, and time consistency may vary greatly, which directly affects the effect of data fusion. Data collected in different time periods may have inconsistent accuracy, and the time lag of the data will limit the monitoring of rapidly changing riverbed morphology.

[0006] 2. High model coupling complexity: Traditional numerical models (such as TELEMAC, HEC-RAS, etc.) are mainly used for single-dimensional hydrodynamic simulation, while riverbed evolution prediction often requires comprehensive hydrodynamic and sediment transport models. In the process of multi-dimensional model coupling, a large number of nonlinear equations and parameter adjustments need to be processed, the amount of calculation is huge and the parameter sensitivity is high, which increases 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 different adaptability and processing capabilities for data features. For example, commonly used machine learning algorithms can effectively extract important features from data, but have limited ability to capture dynamic changes in time series data. For the task of predicting riverbed morphology, conventional machine learning methods are insufficient in their ability to explain nonlinear changes and model generalization, and cannot fully capture the combined impact of natural and human factors on riverbed morphology.

[0008] 4. Model uncertainty and prediction error: There are many uncertainties in the riverbed evolution model, such as data errors, model assumptions, and environmental changes. These uncertainties will directly affect the reliability of the prediction results. Although some methods (such as uncertainty analysis, error propagation, etc.) can be used to evaluate the prediction error, these methods are limited in effectiveness for complex systems with multi-source data fusion. Summary of the invention

[0009] In view of the above problems, the present invention proposes a riverbed evolution prediction method based on multi-source data fusion, which comprises the following steps:

[0010] S1, collect and obtain multi-source riverbed data, pre-process the data, and fuse them into unified 3D grid data;

[0011] S2, based on the feature extraction network, generating a multi-scale feature map; 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 of different scales, and combines them with the channel attention mechanism for weighted fusion; the mutual information feature selection module eliminates redundant features and obtains key features covering the dynamic changes of the riverbed;

[0012] S3, based on a multi-channel convolutional neural network, simultaneously learns and extracts features on multiple channels, and hierarchically expresses and fuses 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 convolution kernels of different sizes to extract details and large-scale features, and focuses on the most relevant features through a channel attention mechanism to fuse multi-channel information; the feature compression module further compresses the feature map through a pooling layer, and when processing time series data, captures time dependencies through an LSTM layer to generate serialized spatiotemporal features;

[0013] S4, based on the time series regression network, extracts the complex dependencies in the time series through a double-layer LSTM structure, and learns the residual information through an adaptive residual correction module to improve the prediction accuracy and obtain the final prediction result.

[0014] Preferably, the acquisition obtains multi-source riverbed data, including topographic data, hydrodynamic data, sediment data and meteorological data; the topographic data includes three-dimensional point cloud coordinate data of terrain undulation information obtained by lidar, includes two-dimensional image data of river channel morphology, surrounding vegetation coverage, water body expansion and shoreline retreat changes obtained by remote sensing images, and also includes three-dimensional point cloud data of underwater morphology and depth of riverbed obtained by sonar; the hydrodynamic data includes flow data, flow velocity data and water level data; the sand data includes suspended matter concentration data and sediment particle size distribution data; the meteorological data includes precipitation data and evaporation data.

[0015] Preferably, the fusion forms unified three-dimensional grid data, and the specific process is:

[0016] Define a unified grid: According to the scope and resolution requirements of the riverbed area to be studied, define a unified three-dimensional space grid and determine the minimum and maximum x, y, and z coordinates, denoted as (x min ,y min , z min ) and (x max ,y max , z max ), which is 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 uses the average resolution of the laser radar, 1m;

[0017] Unified data conversion: The data collected by the lidar is three-dimensional point cloud data, given in the form of p1 = (x1, y1, z1), representing the three-dimensional coordinates of the surface. p1 is converted into a projection coordinate system:

[0018]

[0019] Z Radar =h

[0020] in, and γ are the original geographic coordinates, γ0 represents the central longitude of the projection 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 usually only contains plane coordinates p2 = (x2, y2), which are directly mapped to the plane in the three-dimensional grid, and the z value defaults to 0;

[0022] The data p3 = (x3, y3, z3) collected by the sonar records the water depth value and needs to be converted into absolute elevation, that is, the distance from the vertical line of the ground point to the geoid. In order to conform to the unified three-dimensional projection coordinate system, the absolute elevation z is calculated as follows:

[0023] z S =HD

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

[0025] The hydrodynamic data p4 includes water level and flow velocity information; the water level data is directly mapped to the z-axis in the projection coordinates as the z value; the flow velocity is decomposed into three-dimensional components (v s , v y , v z ), interpolate to each grid point on the corresponding plane of the three-dimensional grid;

[0026] The sediment data includes the suspended sediment concentration p5(x5,y5). The suspended sediment concentration data is recorded directly on the z plane of the grid, i.e. z = z m , and interpolated to the grid by the IDW method, that is, interpolation calculation is performed based on the concentration data of the surrounding sampling points q5 (xj, yj), and the above interpolation calculation is repeated for each grid point until the sediment concentration values ​​p5 of all grid points are obtained by IDW interpolation;

[0027] Meteorological data are two-dimensional data P6, including precipitation and evaporation, which are directly assigned to the surface layer of the grid;

[0028] After the above transformation, all data are standardized and converted to a unified three-dimensional coordinate system; the lidar data and sonar data provide z coordinate values ​​for constructing the three-dimensional terrain of the surface and underwater; the remote sensing image data are mapped to the (x, y) plane, and the attribute values ​​of the surface layer are assigned; the hydrodynamic data and sediment data are distributed to different depth layers of the grid through spatial interpolation; the meteorological data are mapped to the surface grid, and finally the data P represented by the three-dimensional grid is obtained. grid .

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

[0030] For 3D grid data P grid , design a multi-scale convolution kernel, set the convolution kernel size k*k of different scales respectively, to extract multi-level spatial information, and add the time dimension at the same time to construct a three-dimensional convolution kernel. The three-dimensional convolution calculation formula 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 is a three-dimensional convolution kernel, including the spatial dimension D and the temporal dimension T; b k，t is the corresponding bias term, F k，t represents the feature map at scale k and time step T;

[0033] Secondly, to capture data more comprehensively grid The multi-scale features of the feature map {F3, F5, F7} at each scale are fused, and the feature fusion is in the form of weighted summation:

[0034] F multi =α·F3+β·F5+γ·F7

[0035] Among them, α, β, and γ represent fusion weights, which are adaptively learned through model training. The mean square error is used as the loss function to measure the gap between the predicted value and the true value, and to optimize the prediction accuracy of the model for riverbed evolution characteristics.

[0036] Preferably, the specific processing process of the adaptive feature selection module in the feature extraction network is as follows:

[0037] First, the feature importance is measured. The importance of the feature is determined by calculating the mutual information between each feature channel and the target variable. The mutual information MI is used to measure the information correlation between the feature and the target variable. The calculation formula is:

[0038]

[0039] Among them, F i Represents the fusion feature F multi The i-th characteristic channel in; 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, p(f i ) and p(y) are marginal probability distributions;

[0040] Then calculate the feature pair F i With other features F j The correlation coefficient p(F i , F j ), using the Pearson correlation coefficient:

[0041]

[0042] Then define the effective mutual information MI eff (F i ; Y) to correct the redundancy between features:

[0043]

[0044] Among them, δ represents the weight coefficient, balancing the correlation between features and targets and the redundancy between features;

[0045] Finally, according to MI eff (F i ; Y) Sort the features and select the first N features F with the highest effective mutual information selected , as the input of the subsequent multi-channel convolutional neural network model:

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

[0047] Among them, F selected is the feature set filtered by mutual information, ε is the threshold of mutual information, that is, only feature graphs with mutual information greater than or equal to the threshold are selected;

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

[0049] Preferably, the multi-scale convolution fusion module in S3 includes a multi-scale convolution unit and a fusion unit;

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

[0051]

[0052] in, represents the feature map of the first layer at scale k, X (l-1) is the output feature map of the previous layer, represents the convolution kernel of the lth layer at scale k, is the bias term, σ is the activation function ReLU;

[0053] The fusion unit introduces a channel attention mechanism, which enables the network to adaptively assign channel weights according to the current feature information; Fusion is performed to obtain the feature fusion result of each layer; the weight of each channel is defined as αk , then the channel weighted fusion result is:

[0054]

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

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

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

[0058]

[0059] Among them, pooling() represents the average pooling operation, and the compressed feature map is the input of the next layer;

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

[0061] Preferably, the data processing process of the time series regression network is:

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

[0063] Preferably, the double-layer LSTM structure is specifically:

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

[0065]

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

[0067] The attention mechanism is introduced to generate a set of attention weights at each time step t, which helps the model to adaptively allocate attention to different steps of the time series. The attention weight μ t The calculation formula is:

[0068]

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

[0070] Combined with the first layer LSTM output after the attention mechanism, the hidden states of all time steps According to the weighted sum of attention weights, a global weighted feature representation s is obtained (1) :

[0071]

[0072] Weighted s (1) Represents short-term features of riverbed evolution;

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

[0074]

[0075] in, They are the hidden state and memory unit of the second layer LSTM at time step t respectively; As the output of the second layer LSTM;

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

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

[0078] Residual calculation: At each time step t, calculate the true value Y t and predicted values The difference between them, that is, the residual:

[0079]

[0080] Among them, r t 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 residual correction values ​​for each time step:

[0082]

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

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

[0085]

[0086] The output of the time series regression network is the final prediction sequence It is the predicted value of the riverbed state at the 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:

[0088] The three-dimensional grid data P grid The training set and the test set are divided into two sets in a ratio of 8:2. The division method is random division to ensure that the model has sufficient training samples and generalization ability.

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

[0090] When the loss function value calculated using the validation set cannot be reduced for 10 consecutive epochs, the training of the feature extraction network, multi-channel convolutional neural network, and time series regression network is completed.

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

[0092] 1. Improve data quality and consistency: Through the fusion of multi-source data, this method greatly improves the accuracy and consistency of input data. The multi-source data acquisition and processing module uses data such as lidar, remote sensing images, and sonar depth sounding, and ensures the temporal and spatial consistency of multiple sensor data through data anomaly detection and time synchronization processing.

[0093] The improvement in data quality ensures the stability of model input, especially in the monitoring of dynamic changes in riverbeds. 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 grid input data, effectively improving the reliability and accuracy of subsequent predictions.

[0094] 2 Accurately extract multi-scale riverbed features: Based on the designed feature extraction network, the multi-scale convolutional neural network and adaptive feature selection method can accurately extract the morphological change characteristics of the riverbed at different scales, avoid redundant information, and improve the efficiency of feature recognition. The multi-scale convolution kernel can fully extract the local details and overall change trends of the riverbed, and the channel attention mechanism uses adaptive weighting to enable the model to focus on the most critical features. The feature selection module further screens out features that are highly relevant to the evolution of the riverbed, reduces redundant information input, and improves the learning efficiency and prediction accuracy of the model. These designs ensure the comprehensiveness and effectiveness of feature extraction, allowing the model to better capture the complex characteristics of riverbed evolution.

[0095] 3. Multi-channel feature fusion improves prediction performance: The multi-channel convolutional neural network design realizes the effective fusion of multi-source features, and can simultaneously learn the dynamic characteristics of riverbed morphological evolution in multiple feature dimensions. The multi-scale convolution fusion module structure allows different feature maps to be input in independent channels, thereby improving the hierarchical expression of feature fusion. The weighted fusion of multi-scale convolution and channel attention mechanism effectively enhances the adaptability of the model at different spatial scales. At the same time, the pooling layer of the feature compression module reduces redundant information, allowing 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. Improve the time series prediction accuracy of riverbed morphology: The double-layer LSTM structure of the time series regression network combined with the attention mechanism and the adaptive residual correction module enable the model to have a high time series prediction accuracy and can effectively capture the short-term and long-term change trends of the riverbed morphology. By combining the attention mechanism with the double-layer LSTM, the model can adaptively assign weights in each time step of the time series, especially paying more attention to the changes in the riverbed morphology at key moments, capturing the short-term fluctuations and long-term trends of the riverbed. The adaptive residual correction module further improves the prediction accuracy and stability by adjusting the deviation in the time series prediction, ensuring that the model can give a more realistic prediction result of the future riverbed morphology. This design enables the model to effectively cope with the uncertainty of riverbed evolution at different time scales and support high-precision prediction of riverbed dynamic processes.

[0097] 5. Improve the applicability and robustness of the model: The present invention shows strong applicability and robustness in complex river environments. Through multi-source data fusion and innovative feature extraction and regression methods, it can cope with various riverbed morphological changes caused by natural factors and human activities. The model can not only accurately identify local erosion and sedimentation of the riverbed, but also analyze large-scale river changes. It is suitable for a variety of river environments and different riverbed evolution scenarios. At the same time, multi-channel and adaptive feature selection methods ensure that the model can filter noise data, and adaptive residual correction further improves the robustness of the prediction. This enables the model to maintain high prediction accuracy and stability even when data quality is uneven and environmental changes are complex. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0099] Figure 2 It is a schematic diagram of the overall process of riverbed related data collection and processing of the present invention.

[0100] Figure 3 Schematic diagram of the feature extraction network processing flow of the present invention.

[0101] Figure 4 Schematic diagram of the time series regression network architecture combined with the attention mechanism of the present invention.

[0102] Figure 5 This is a comparison chart of model predictions of a double-layer LSTM combined with an attention mechanism in the embodiment. DETAILED DESCRIPTION

[0103] The invention will be further described below in conjunction with specific embodiments.

[0104] The present invention proposes a riverbed evolution prediction method based on multi-source data fusion. The prediction method is obtained by processing different data source sets by multiple data processing modules in combination with prediction models. The overall flow chart of riverbed evolution prediction based on multi-source data fusion is as follows: Figure 1 shown.

[0105] 1. Riverbed related data collection and processing

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

[0107] 1. Data Collection

[0108] In the present invention, to realize the riverbed evolution prediction method of multi-source data fusion, the collection and organization of riverbed hydrological resources must be completed first. The spatial and temporal variation characteristics of the riverbed can be captured, and the accuracy of riverbed morphology and flow field analysis can be improved. It provides information on hydrodynamics, topography, sediment, etc., and supports the analysis of sediment transport, riverbed erosion and siltation.

[0109] Topographic data: This part includes the integration of lidar data, remote sensing images and sonar depth data. Among them:

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

[0111] (2) Remote sensing images are taken by drones to monitor changes in river morphology and surrounding vegetation coverage, water body expansion, and shoreline retreat. The sampling period is once a month. Finally, two-dimensional image data P2 = (x2, y2) containing spatial position information is obtained, which is used to monitor changes in water bodies, vegetation coverage, and river banks, and to extract river characteristics.

[0112] (3) Sonar sounding data uses sonar equipment and ships sailing along the river channel to obtain underwater riverbed morphology data, determine river channel depth, riverbed elevation and sediment distribution, with a spatial resolution of 26 meters, a vertical accuracy of ±10-30 cm, and a sampling period of once a quarter. Finally, the three-dimensional point cloud data of the covered area is obtained, which can generate a complete underwater topography and riverbed depth change P3 = (X3, y3, z3).

[0113] (4) Hydrodynamic data: This part includes flow rate data and water level data. Hydrodynamic data is obtained through flow velocity sensors and water level meters installed at fixed hydrological stations. To determine the impact of flow on riverbed evolution and sediment movement, the flow meter at the hydrological station collects flow data at a frequency of about 1000kHz, with an accuracy of ±1-2%, and a sampling period of once a day; at the same time, to determine the changes in water levels in the river channel and monitor river channel erosion and siltation, the water level meter has a range of 0-30m, an accuracy of ±1mm, and a sampling period of once a day. This part of the data is a time series information recorded in the form of attributes, stored in CSV format, represented as P4, to represent the impact of water flow on riverbed morphology and water level fluctuations.

[0114] (5) Sediment data: This part includes suspended matter concentration data and sediment particle size distribution data. The data is obtained by continuous monitoring of water quality sensors at fixed points. The turbidity measurement range is 0-1000NTU, the accuracy is ±5NTU, and the sampling period is once a day. This part of the data is a time series information recorded in the form of attributes and stored in CSV format. It is represented as P5 to indicate the sediment transportation situation and judge the riverbed erosion and deposition.

[0115] (6) Meteorological data: This part includes precipitation data and evaporation data. Through continuous measurement at the meteorological station, we can obtain the impact of precipitation on the riverbed erosion and flood transport in the basin confluence process, as well as the impact of water evaporation loss on the river balance. The precipitation is measured by a rain gauge with a range of 0-500mm / h and a sampling period of once a day; the evaporation sensor measures evaporation with a range of 0-20mm / day and a sampling period of once a day. This part of the data is a time series information recorded in the form of attributes and stored in CSV format. It is represented as P6 to represent the impact of precipitation on river flow and the impact of evaporation on water level changes and basin water balance.

[0116] 2. Data anomaly detection and processing

[0117] In the riverbed evolution prediction scenario, the collected data (P1-P6) are firstly subjected to data anomaly detection, which is mainly used to identify and remove data points with large noise or errors to improve the reliability and accuracy of the data. For different types of data (such as lidar, remote sensing images, sonar depth, hydrodynamics, sediment and meteorological data), the present invention adopts different anomaly detection methods, and the specific process 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 according to 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 parts 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 on a daily basis.

[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 the present invention, the laser radar generates about 7 million points on a quarterly basis, and a single collection is 100-500MB / square kilometer per month; the remote sensing impact data generates about 10MB of data and 48 images in an area of ​​one square kilometer; the sonar sounding generates about 40,000 records in one square kilometer; the hydrodynamic data has a total of about 8760 flow records; the sediment data has about 8760 records; and the meteorological data also has about 8760 records.

[0133] In order to integrate different data types collected by multi-source sensors, data fusion is required. After data anomaly detection, the data (p1~p6) are fused. The present invention adopts a weighted inverse distance interpolation algorithm to integrate terrain, hydrodynamics and attribute data of different spatial resolutions and data types into a unified three-dimensional grid for further analysis and prediction.

[0134] Specifically, the weighted inverse distance interpolation (IDW) method estimates the attribute values ​​of unsampled locations by weighted averaging the sampling point data according to the distance, so as to fuse the lidar, sonar bathymetry, hydrodynamics and sediment attribute data of different resolutions and sampling frequencies, and construct a unified spatial grid to support the subsequent analysis and modeling of riverbed morphological evolution. The following is the specific process of realizing multi-source data fusion.

[0135] Define a unified grid: Define a unified three-dimensional space grid based on the scope and resolution requirements of the riverbed area to be studied. Determine the minimum and maximum x, y, and z coordinates, denoted as (x min ,y min , z min ) and (x max ,y max , z max ), which is used to determine the spatial range 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 resolution of the grid should comprehensively consider the original data resolution and the target application accuracy. In the present invention, the average resolution of the laser radar is 1m.

[0136] Unified data conversion:

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

[0138]

[0139] Z Radar =h

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

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

[0142]

[0143] z Ro =0

[0144] (3) The data p3 = (x3, y3, z3) collected by sonar sounding records the water depth value (relative to the water surface reference) and needs to be converted into absolute elevation, that is, the distance from the ground point vertical line to the geoid. In order to conform to the unified three-dimensional projection coordinate system. The absolute elevation z is calculated as follows:

[0145] z S =HD

[0146] Among them, D is the depth of water relative to the water surface, and the elevation of the water surface is H.

[0147] (4) Hydrodynamic data p4 includes water level and flow velocity information, which is usually time series data and is collected at a fixed point. Water level data can be directly mapped to the z-axis in the projection coordinates as z values. Flow velocity can be decomposed into three-dimensional components (v s , v y , v z ), interpolate to each grid point on the corresponding plane of the three-dimensional grid. Assume that the coordinates of the hydrodynamic sampling point are (x h ,y h ), the water level is z h , then directly convert the water level data z h Assign the value to the z coordinate of the grid point. That is: z = z h .

[0148] (5) Sediment data includes suspended sediment concentration, which is time series data. Since sediment data are mostly collected at specific sections, they need to be interpolated into a three-dimensional grid in space. Suspended sediment concentration data can be directly recorded on the z plane of the grid, that is, z = z m , and interpolated to the grid using the IDW method. For any target point Q(x, y) on the grid plane, its sediment concentration is expressed as p5(x5, y5), which can be interpolated based on the concentration data p5 of the surrounding sampling point q5(xj, yj). The interpolated value p5(Q) is expressed as follows:

[0149]

[0150] Among them, p5(q5) represents the sediment concentration value at sampling point q5. Represents the distance between grid point Q and sampling point q5. A is the attenuation exponent, which is 2, and n is the number of sampling points involved in the interpolation. Repeat the above interpolation calculation for each grid point until the sediment concentration values ​​p5 of all grid points are obtained through IDW interpolation.

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

[0152] After the above conversion steps, all data are standardized and converted to a unified three-dimensional coordinate system. LiDAR data and sonar depth data: provide z coordinate values ​​for constructing three-dimensional terrain on the surface and underwater. Remote sensing image data: mapped to the (x, y) plane, and attribute values ​​of the surface layer. Hydrodynamic data and sediment data: distributed to different depth layers of the grid through spatial interpolation. Meteorological data: mapped to the surface grid. Finally, the data P represented by the three-dimensional grid is obtained. grid .

[0153] This standardization and conversion method can ensure the consistency of multi-source data in space and time, and provide unified data information for riverbed evolution prediction. Figure 2 shown.

[0154] 2. Feature Extraction Network Design

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

[0156] Therefore, the present 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 grid The different spatial scales in the model are decomposed layer by layer to extract local and global features; and the feature selection based on mutual information can screen out features that are highly correlated with riverbed evolution but independent of each other, further optimizing the model.

[0157] Riverbed environmental information based on multi-source data fusion grid It is a three-dimensional grid data with dimension 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 the present invention includes 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 the prediction of riverbed evolution, the multi-scale feature extraction method helps the model identify the complex evolution pattern of the riverbed by capturing the feature changes at different spatial scales. Multi-scale feature extraction is a technology that extracts information from different scales and resolutions to capture the multi-level features of data, and is particularly suitable for task scenarios with complex space or multiple patterns.

[0160] Specifically, first, a multi-scale convolution kernel is designed. grid The ratio of training set to test set is set to 8:2 to ensure that the model has sufficient training samples and generalization ability. Multi-scale convolution kernels are designed, and the convolution kernel sizes of different scales are set to k*k (k=3, 5, 7) to extract multi-level spatial information. ReLU is used as the activation function for the output activation of the convolution layer. At the same time, the time dimension is added to construct a three-dimensional convolution kernel. The three-dimensional convolution calculation formula is as follows:

[0161] F k =σ(P grid *W k +b k )

[0162] Among them, W k ∈R k×k×D is the convolution kernel matrix of scale k, which is used to extract features of a specific scale; b k is the corresponding bias term, σ is the activation function, F k 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 to capture data P more comprehensively. grid The multi-scale features of the feature map {F3, F5, F7} at each scale are fused. Feature fusion adopts the form of weighted summation:

[0164] F multi =α·F3+β·F5+γ·F7

[0165] Among them, α, β, and γ are represented as fusion weights, which are adaptively learned through model training. The mean square error is used as the loss function to measure the gap between the predicted value and the true value, and to optimize the prediction accuracy of the model for the riverbed evolution characteristics. The fused feature map F multi Contains spatial features at different scales, which are used as input for subsequent feature selection and modeling.

[0166] The time dimension is introduced on the basis of the original multi-scale convolution feature extraction to increase the sensitivity to time evolution. Specifically, a three-dimensional convolution kernel is proposed in the present invention: that is, the convolution operation is performed in the H×W×T space, where T represents the time step.

[0167] In this way, data fragments from recent years can be selected for convolution processing in the time dimension to capture short-term trends in riverbed changes, while historical data can be combined to obtain long-term sedimentation and erosion trends.

[0168] Furthermore, adding the time dimension, the three-dimensional convolution calculation formula 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 is a three-dimensional convolution kernel, including the spatial dimension D and the temporal dimension T; b k，t is the corresponding bias term, F k，t represents the feature map at scale k and time step T.

[0171] By adding the time dimension, not only can features at different scales be extracted spatially, but also multi-scale evolution information can be extracted temporally, providing richer feature support for future riverbed trend prediction.

[0172] 2. Mutual Information Feature Selection Module

[0173] In order to further extract effective features related to riverbed evolution, the fusion feature map F multi Perform feature selection. Feature selection is based on a modified mutual information algorithm to ensure that the model only receives features that contribute significantly to the target prediction.

[0174] Specifically, the feature importance is measured first. The importance of the feature is determined by calculating the mutual information between each feature channel and the target variable (future riverbed height change). Mutual information MI is used to measure the information correlation between the feature and the target variable, and its calculation formula is:

[0175]

[0176] Among them, F i Represents the fusion feature F multi The i-th characteristic channel in; 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, p(f i ) and p(y) are marginal probability distributions.

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

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

[0179] (2) Calculate feature pair F i With other features F j The correlation coefficient ρ(F i , F j ). In the present invention, the Pearson correlation coefficient is used:

[0180]

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

[0182]

[0183] Among them, δ represents the weight coefficient, which balances the correlation between features and targets and the redundancy between features.

[0184] (4) According to MI eff (F i ; Y) Sort the features and select the first N features F with the highest effective mutual information selected , as the input of the subsequent multi-channel convolutional neural network model:

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

[0186] Among them, F selected is the feature set filtered by mutual information. ε is the threshold of 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 , which will be used as the input of the model for spatiotemporal prediction. This feature matrix not only contains the significant features at different scales, but also filters out redundant information to improve the prediction efficiency and accuracy of the model. The feature extraction network process structure diagram is shown in Figure 3 shown.

[0188] By extracting features at different spatial and temporal scales, it helps to 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] 3. Multi-channel Convolutional Neural Network

[0190] Multi-channel convolutional neural network (MC-CNN) is a convolutional neural network structure that processes multi-dimensional and multi-feature data. The key to MC-CNN is to input multiple input feature maps into the model at the same time for convolution operation, so that the network can learn and extract features on multiple channels at the same time, thereby improving the ability to capture and understand complex information. It efficiently processes multi-source data through modules such as multi-channel input, feature extraction and fusion. First, the multi-scale convolution fusion module receives multiple features, uses convolution kernels of different sizes to extract details and large-scale features, and focuses on the most relevant features through the channel attention mechanism to fuse multi-channel information; the feature compression module further compresses the feature map through the pooling layer to improve the training efficiency of the model. And when processing time series data, the LSTM layer captures time dependencies and generates dynamic change predictions. The synergy of each module enables MC-CNN to perform well in multi-dimensional data processing tasks.

[0191] In the riverbed evolution prediction task, a multi-scale feature map F is generated through the feature extraction network. selected ,These feature maps cover the local and global information of the dynamic changes of the riverbed. In order to make full use of these multi-scale features and further improve the model's ability to capture complex riverbed evolution patterns, the present invention introduces a multi-channel convolutional neural network. Through multi-channel design, these feature maps are used as multiple input channels of the network, so that the model can process data at multiple scales simultaneously, thereby avoiding information loss and better capturing the complex changes in riverbed morphology.

[0192] Since the feature map contains not only riverbed morphology information, but also a variety of 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 can design convolution operations for each feature channel separately, so that the model can analyze each feature map independently and then combine the results to achieve hierarchical expression and fusion of different information sources.

[0193] 1. Multi-scale convolution fusion module

[0194] According to 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 It represents a characteristic map in the spatial dimension (including water depth, flow velocity, water level, precipitation evaporation, and sediment concentration), and N represents the number of characteristic channels.

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

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

[0198] Convolutional layers with multiple convolution kernel sizes are designed to extract feature information of different spatial scales. For each convolution kernel size, a convolution operation is performed on each channel to obtain the corresponding feature map. The multi-channel convolutional neural network combined with the channel attention mechanism also uses the ReLU function as the activation function in the convolution layer and the mean square error MSE as the loss function to optimize the results of multi-channel feature fusion and make the prediction more accurate.

[0199] The convolution calculation formula is as follows:

[0200]

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

[0202] Secondly, the multi-scale feature map extracted by the fusion unit through convolution The fusion is performed to obtain the feature fusion result of each layer. In this process, the channel attention mechanism is introduced to enable the network to adaptively allocate channel weights according to the current feature information.

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

[0204]

[0205] in:

[0206]

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

[0208] 2. Feature compression module:

[0209] The pooling layer is used to compress the feature map of the convolution output in space, retain the main features and reduce redundant information, avoiding the overfitting problem caused by redundant features. For the riverbed evolution prediction model, the pooling layer helps to highlight the large-scale evolution trend and weaken the short-term or small-scale random changes, so as to pay more attention to the stable evolution signal.

[0210] The feature map after multi-channel fusion is pooled to further compress the feature space and retain key features. The present invention uses average pooling to reduce the data dimension, thereby improving computational efficiency and reducing the impact of redundant information on the model:

[0211]

[0212] Among them, pooling() represents the average pooling operation, and the compressed feature map is the input for the next layer.

[0213] Since riverbed evolution has time series characteristics, a recurrent neural network LSTM layer is added after the output of the convolution layer to capture the dynamic changes in the time dimension, that is, the time changes of riverbed height sedimentation rate and other characteristics. The output of the pooling layer is input into the LSTM unit to obtain serialized spatiotemporal features:

[0214]

[0215] y t It represents the cumulative effect of riverbed change characteristics over time. Through time series modeling, MC-CNN can predict the possible change pattern of the riverbed in the future time period.

[0216] 4. Time Series Regression Network Design

[0217] The morphological changes of the riverbed are not only affected by local factors in space (topography and flow velocity), but also show dynamic evolution characteristics over time. Therefore, when predicting the future changes of the riverbed, the present invention considers the time series characteristics. Designing a time series regression network can generate more time-consistent future prediction results based on the riverbed characteristics at current and historical moments, thereby improving the prediction accuracy and stability.

[0218] Specifically, after the multi-channel convolutional neural network extracts multi-scale features and obtains the preliminary prediction output result y, in order to more accurately predict the future evolution trend of the riverbed, the present invention designs and introduces a time series regression network. The design goal of the network is to combine the spatial features of MC-CNN with the temporal dynamics of the riverbed, capture the pattern and trend of the riverbed changing over time, and optimize the prediction accuracy of the future riverbed morphology. The introduction of a double-layer LSTM combined with an attention mechanism aims 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 double-layer LSTM structure. The first layer of LSTM focuses on capturing short-term change features, and the second layer of LSTM further analyzes long-term trends to generate preliminary prediction results. The attention mechanism is added to adaptively allocate the weights of each time step, so that the model focuses more on the time step information that has a key impact on future evolution. Then, the adaptive residual correction module is used to further optimize the prediction results to dynamically compensate for deviations and improve the overall prediction accuracy and the temporal consistency of the model. Through this module design, the model can achieve high-precision prediction of riverbed morphology and provide scientific support for river management and water resources regulation.

[0219] 1. Time series data construction

[0220] The prediction result y output by MC-CNN t represents the spatiotemporal characteristics of the riverbed at time step t. In order to capture its trend over time, y t In the specific time series construction, the y of multiple consecutive time steps is converted into time series input. t The values ​​are used as the input sequence of LSTM.

[0221] Specifically, assuming that the total time step is T, construct the time series input Y = [y1, y2, ..., y T ], where each y t Represents the predicted state of the riverbed at a certain point in the past. This time series is input into the regression module to analyze the temporal trend of the characteristics and predict the future evolution of the riverbed.

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

[0223] In the time series regression module, the present invention designs a two-layer LSTM architecture combined with an attention mechanism to extract complex dependencies in time series. The two-layer LSTM structure is deeper and more flexible than the single-layer LSTM, and can better capture short-term and long-term time dependencies. The introduction of the attention mechanism enables the model to automatically select and give higher weights to key time steps to highlight features that have a greater impact on prediction. The working principle of the two-layer LSTM combined with the attention mechanism is as follows:

[0224] The activation function is set to use the Tanh function, which is used as the activation function of LSTM. Tanh is suitable for processing time series data. Similarly, the mean square error is used as the loss function. The continuous prediction value of LSTM output is optimized through the loss function. The test samples are time series riverbed data at different times.

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

[0226]

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

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

[0229] The attention mechanism is introduced to generate a set of attention weights at each time step t, which helps the model to adaptively allocate attention at different steps of the time series. The calculation formula of the attention weight μt is:

[0230]

[0231] Among them, ω, W a , b a is a learnable parameter. t is the attention weight at time step t, indicating the importance of the feature at the current time step. Through the weight calculation of the attention mechanism, the model can assign different weights to each time step, so that the key time step obtains a higher weight in the model learning process, thereby effectively capturing the key dynamic changes of the riverbed.

[0232] Combined with the first layer LSTM output after the attention mechanism, the hidden states of all time steps According to the weighted sum of attention weights, a global weighted feature representation s is obtained (1) :

[0233]

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

[0235] In the second LSTM layer, the weighted first layer LSTM output s (1) As the input of the second layer LSTM, long-term dependency features are further extracted. The calculation formula of the second layer LSTM is:

[0236]

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

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

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

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

[0241] 3. Adaptive residual correction module

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

[0243] Residual calculation: At each time step t, calculate the true value Y t and predicted values The difference between them, that is, the residual:

[0244]

[0245] Among them, r t represents the prediction error at time step t.

[0246] Residual correction model: Use 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:

[0247]

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

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

[0250]

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

[0252] Through residual correction, the deviation of the model at different time steps can be dynamically adjusted, thereby improving the accuracy and robustness of the time series regression model. The short-term and long-term dynamic changes of the riverbed morphology are captured by the double-layer LSTM structure to ensure that the model can effectively identify seasonal or interannual riverbed evolution. By comparing the prediction errors of LSTM models with different time series lengths, the accuracy of the model in the time dimension is evaluated. The flowchart of the dual time series regression module combined with the attention mechanism is shown in the figure below. Figure 4 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 grid data P grid The training set and the test set are divided into training set and test set in a ratio of 8:2 by random division to ensure that the model has sufficient training samples and generalization ability; secondly, the average value of the sum of the mean square error between the true value and the predicted value of water depth, the mean square error between the true value and the predicted value of flow velocity, the mean square error between the true value and the predicted value of water level, the mean square error between the true value and the predicted value of precipitation evaporation, and the mean square error between the true value and the predicted value of sediment concentration is used as the loss function during the training of feature extraction network, multi-channel convolutional neural network and time series regression network; finally, when the loss function value calculated using the validation set cannot be reduced for 10 consecutive epochs, the training of feature extraction network, multi-channel convolutional neural network and time series regression network is completed.

[0254] To verify the effectiveness of the double-layer LSTM model combined with the attention mechanism proposed in the present invention, a simulation comparison is performed with the benchmark LSTM at a time step of 100. By comparing the prediction results of the benchmark LSTM and the double-layer LSTM combined with the attention mechanism, it can be seen that the introduction of the attention mechanism significantly improves the prediction accuracy of the model, especially when processing complex time series data. When combining multi-source data to predict riverbed evolution, the double-layer LSTM combined with the attention mechanism can effectively fuse multi-source information, thereby improving prediction performance. The comparison diagram is shown in Figure 2. Figure 5 shown.

[0255] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0256] Although the above describes the specific implementation methods of the present invention, it is 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 on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A riverbed evolution prediction method based on multi-source data fusion, characterized in that: The following steps are involved: S1, collect and obtain multi-source riverbed data, pre-process the data, and fuse them into unified 3D grid data; S2, based on the feature extraction network, generating a multi-scale feature map; 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 of different scales, and combines them with the channel attention mechanism for weighted fusion; the mutual information feature selection module eliminates redundant features and obtains key features covering the dynamic changes of the riverbed; S3, based on a multi-channel convolutional neural network, simultaneously learns and extracts features on multiple channels, and hierarchically expresses and fuses 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 convolution kernels of different sizes to extract details and large-scale features, and focuses on the most relevant features through a channel attention mechanism to fuse multi-channel information; the feature compression module further compresses the feature map through a pooling layer, and when processing time series data, captures time dependencies through an LSTM layer to generate serialized spatiotemporal features; S4, based on the time series regression network, extracts the complex dependencies in the time series through a double-layer LSTM structure, and learns the residual information through an adaptive residual correction module to improve the prediction accuracy and obtain the final prediction result.

2. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: The multi-source riverbed data collected includes topographic data, hydrodynamic data, sediment data and meteorological data; the topographic data includes three-dimensional point cloud coordinate data of terrain undulation information obtained by lidar, two-dimensional image data of river channel morphology, surrounding vegetation coverage, water body expansion and shoreline retreat changes obtained by remote sensing images, and three-dimensional point cloud data of underwater morphology and depth of the riverbed obtained by sonar; the hydrodynamic data includes flow data, flow velocity data and water level data; the sand data includes suspended matter concentration data and sediment particle size distribution data; the meteorological data includes precipitation data and evaporation data.

3. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 2, characterized in that: The fusion forms unified three-dimensional grid data, and the specific process is as follows: Define a unified grid: According to the scope and resolution requirements of the riverbed area to be studied, define a unified three-dimensional space grid and determine the minimum and maximum x, y, and z coordinates, denoted as (x min ,y min , z min ) and (x max ,y max , z max ), which is 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 uses the average resolution of the laser radar, 1m; Unified data conversion: The data collected by the lidar is three-dimensional point cloud data, given in the form of p1 = (x1, y1, z1), representing the three-dimensional coordinates of the surface. p1 is converted into a projection coordinate system: Z Radar =h in, and γ are the original geographic coordinates, γ0 represents the central longitude of the projection coordinate system, N is the radius of curvature, and h is the elevation data, i.e., the elevation value measured by the lidar; Remote sensing image data only contains plane coordinates p2 = (x2, y2), which are directly mapped to the plane in the three-dimensional grid, and the z value defaults to 0; The data p3 = (x3, y3, z3) collected by the sonar records the water depth value, which needs to be converted into absolute elevation, that is, the distance from the vertical line direction of the ground point to the geoid, in order to conform to the unified three-dimensional projection coordinate system. The absolute elevation z is calculated as follows: z S =H-D Among them, D is the depth of water data relative to the water surface, and the elevation of the water surface is H; The hydrodynamic data p4 includes water level and flow velocity information; the water level data is directly mapped to the z-axis in the projection coordinates as the z value; the flow velocity is decomposed into three-dimensional components (v s , v y , v z ), interpolate to each grid point on the corresponding plane of the three-dimensional grid; The sediment data includes the suspended sediment concentration p5(x5,y5). The suspended sediment concentration data is recorded directly on the z plane of the grid, i.e. z = Z m , and interpolated to the grid by the IDW method, that is, interpolation calculation is performed based on the concentration data of the surrounding sampling points q5 (xj, yj), and the above interpolation calculation is repeated for each grid point until the sediment concentration values ​​p5 of all grid points are obtained by IDW interpolation; Meteorological data are two-dimensional data P6, including precipitation and evaporation, which are directly assigned to the surface layer of the grid; After the above transformation, all data are standardized and converted to a unified three-dimensional coordinate system; the lidar data and sonar data provide z coordinate values ​​for constructing the three-dimensional terrain of the surface and underwater; the remote sensing image data are mapped to the (x, y) plane, and the attribute values ​​of the surface layer are assigned; the hydrodynamic data and sediment data are distributed to different depth layers of the grid through spatial interpolation; the meteorological data are mapped to the surface grid, and finally the data P represented by the three-dimensional grid is obtained. grid .

4. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: The multi-scale convolutional neural network module in the feature extraction network is specifically: For 3D grid data P grid , design a multi-scale convolution kernel, set the convolution kernel size k*k of different scales respectively, to extract multi-level spatial information, and add the time dimension at the same time to construct a three-dimensional convolution kernel. The three-dimensional convolution calculation formula is as follows: F k，t =σ(P grid *W k，t +b k,t ) Among them, W k，t ∈R k×k×T×D is a three-dimensional convolution kernel, including the spatial dimension D and the temporal dimension T; b k，t is the corresponding bias term, F k，t represents the feature map at scale k and time step T; Secondly, to capture data more comprehensively grid The multi-scale features of the feature map {F3, F5, F7} at each scale are fused, and the feature fusion is in the form of weighted summation: F multi =α·F3+β·F5+γ·F7 Among them, α, β, and γ represent fusion weights, which are adaptively learned through model training.

5. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 4, characterized in that: The mutual information feature selection module in the feature extraction network has the following specific processing steps: First, the feature importance is measured. The importance of the feature is determined by calculating the mutual information between each feature channel and the target variable. The mutual information MI is used to measure the information correlation between the feature and the target variable. The calculation formula is: Among them, F i Represents the fusion feature F multi The i-th characteristic channel in; 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, p(f i ) and p(y) are marginal probability distributions; Then calculate the feature pair F i With other features F j The correlation coefficient ρ(F i , F j ), using the Pearson correlation coefficient: Then define the effective mutual information MI eff (F i ; Y) to correct the redundancy between features: Among them, δ represents the weight coefficient, balancing the correlation between features and targets and the redundancy between features; Finally, according to MI eff (F i ; Y) Sort the features and select the first N features F with the highest effective mutual information selected , as the input of the subsequent multi-channel convolutional neural network model: F selected ={F k，t |MI eff (F i ;Y)≥ε} Among them, F selected is the feature set filtered by mutual information, ε is the threshold of mutual information, that is, only feature graphs with mutual information greater than or equal to the threshold are selected; The feature matrix F after mutual information feature selection selected , which will be used as the input of the model for spatiotemporal prediction.

6. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: The multi-scale convolution fusion module in S3 includes a multi-scale convolution unit and a fusion unit; The multi-scale convolution unit designs convolution layers with multiple convolution kernel sizes to extract feature information of different spatial scales. For each convolution kernel size, a convolution operation is performed on each channel to obtain a corresponding feature map. The convolution calculation formula is as follows: in, represents the feature map of the first layer at scale k, X (l-11) is the output feature map of the previous layer, represents the convolution kernel of the lth layer at scale k, is the bias term, σ is the activation function ReLU; The fusion unit introduces a channel attention mechanism, which enables the network to adaptively assign channel weights according to the current feature information; Fusion is performed to obtain the feature fusion result of each layer; the weight of each channel is defined as α k , then the channel weighted fusion result is: in, g(·) is the channel weight learning function, which adaptively weights the feature map of each channel.

7. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 6, characterized in that: The feature compression module in S3 is specifically: The feature map after multi-channel fusion is pooled through the pooling layer to further compress the feature space and retain the key features: Among them, pooling() represents the average pooling operation, and the compressed feature map is the input of the next layer; At the same time, a recurrent neural network LSTM layer is added after the output of the convolutional layer to capture the dynamic changes in the time dimension, that is, the temporal changes of features such as riverbed height and sedimentation rate. The output of the pooling layer is input to the LSTM layer to obtain the serialized spatiotemporal features y t .

8. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: The data processing process of the time series regression network is as follows: Based on the constructed time series data, firstly, the short-term and long-term dependencies of the time series are extracted through a two-layer LSTM structure. The first layer of LSTM focuses on capturing short-term change characteristics, and the second layer of LSTM further analyzes long-term trends to generate preliminary prediction results. Then, the attention mechanism is added to adaptively allocate the weights of each time step, so that the model can focus more on the time step information that has a key impact on future evolution. Finally, the adaptive residual correction module is used to further optimize the prediction results to dynamically compensate for deviations and improve the overall prediction accuracy and the temporal consistency of the model.

9. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 8, characterized in that: The double-layer LSTM structure is specifically as follows: In the first LSTM layer, the input data y is received at each time step t t and the hidden state at the previous time step and memory unit The calculation formula of the first layer of LSTM is: in, are the hidden state and memory unit of the first layer LSTM at time step t, As the output of the first layer of LSTM; The attention mechanism is introduced to generate a set of attention weights at each time step t, which helps the model to adaptively allocate attention to different steps of the time series. The attention weight μ t The calculation formula is: Among them, ω, W a , b a is a learnable parameter, μ t is the attention weight at time step t, indicating the importance of the feature at the current time step; Combined with the first layer LSTM output after the attention mechanism, the hidden states of all time steps According to the weighted sum of attention weights, a global weighted feature representation s is obtained (1) : Weighted s (1) Represents short-term features of riverbed evolution; In the second LSTM layer, the weighted first layer LSTM output s (1) As the input of the second layer LSTM, long-term dependency features are further extracted; the calculation formula of the second layer LSTM is: in, They are the hidden state and memory unit of the second layer LSTM at time step t respectively; As the output of the second layer LSTM; Hidden states generated by a two-layer LSTM Used to generate preliminary prediction results for time step t+1 10. A riverbed evolution prediction method based on multi-source data fusion as claimed in claim 8, characterized in that: The adaptive residual correction module is specifically: Residual calculation: At each time step t, calculate the true value Y t and predicted values The difference between them, that is, the residual: Among them, r t 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 residual correction values ​​for each time step: Among them, g(·) represents the residual correction model, and m is the number of input time steps; Final prediction output: The original prediction results of the two-layer LSTM model and residual correction value Add them together to get the final prediction result: The output of the time series regression network is the final prediction sequence It is the predicted value of the riverbed state at the future time step t+1.

Images