Coastline change dynamic monitoring method driven by deep learning

Abstract

The invention provides a deep learning-driven coastline change dynamic monitoring method, and belongs to the technical field of image understanding and recognizing.The deep learning-driven coastline change dynamic monitoring method comprises the steps that a time sequence attention matrix is formed by constructing a multi-source satellite data fusion network, and coastline form change features are extracted by applying a fourth-order time sequence convolutional neural network to construct a first fluctuation matrix; and performing dimension reduction processing by using a space-time attention mechanism to generate a second fluctuation matrix, and converting the second fluctuation matrix into a quantitative index through a fluctuation conversion matrix. The method comprises the steps of inputting data into a pre-trained CoastDynNet model to analyze a physical mechanism, performing fine adjustment by applying a coastline evolution optimization function, performing multi-scale decomposition by using a WaveletTempNet network to separate long and short term changes, and finally constructing a multi-model integration framework by using a Bayesian model averaging method to generate a comprehensive coastline dynamic evaluation report.

Description

Technical Field

[0001] The present invention belongs to the technical field of image understanding and recognition, and more specifically, relates to a method for dynamically monitoring coastline changes driven by deep learning. Background Art

[0002] As the boundary line between land and sea, the dynamic change monitoring of the coastline is of great significance for coastal zone management, disaster prevention and mitigation, and ecological protection. Traditional coastline monitoring mainly relies on single-source satellite remote sensing data, extracts the coastline through methods such as image segmentation and feature extraction, and then analyzes its change law through multi-temporal comparison. These methods can achieve basic monitoring functions under ideal conditions, but their accuracy is limited in complex environments. Some studies have attempted to combine numerical simulation methods to establish physically-driven coastline evolution models, such as GENESIS, Delft3D, etc., to improve the prediction ability.

[0003] However, traditional technologies have obvious defects: First, the time resolution of single-source satellite data is limited, making it difficult to capture short-term fluctuations; second, pure image processing methods are difficult to accurately distinguish natural fluctuations from long-term evolution trends; third, pure physical models are computationally complex and parameter-sensitive, making it difficult to quickly apply them to large-area monitoring. In addition, existing methods often separate data-driven and physical mechanisms, lacking an effective fusion mechanism, resulting in insufficient model accuracy.

[0004] Currently, the dynamic monitoring of coastline changes urgently needs to solve the technical challenges of efficient fusion of multi-source data, accurate decomposition of long-term and short-term changes, and collaborative optimization of physical constraints and data-driven. Especially in the context of increasingly intensified climate change and human activities, how to construct a comprehensive monitoring method for coastline changes that combines physical rationality and prediction accuracy has become a difficult problem jointly faced by the academic and engineering communities, and there is an urgent need to propose new technical solutions that can improve the accuracy of dynamic monitoring of coastline changes. That is to say, there is a technical problem of insufficient accuracy in the dynamic monitoring of coastline changes in the existing technology. Summary of the Invention

[0005] In view of this, the present invention provides a method for dynamically monitoring coastline changes driven by deep learning, which can solve the technical problem of insufficient accuracy in the dynamic monitoring of coastline changes in the existing technology.

[0006] The present invention is implemented as follows: The present invention provides a method for dynamically monitoring coastline changes driven by deep learning, including: constructing a multi-source satellite data fusion network, combining time-series images obtained from satellites with different orbital periods to form a temporal attention matrix; applying a fourth-order temporal convolutional neural network to extract the morphological change features of the coastline, and establishing a first fluctuation matrix in a recursive manner; using a spatio-temporal attention mechanism to perform dimensionality reduction on the first fluctuation matrix, generating a second fluctuation matrix and converting it into a quantitative coastline change index; inputting it into the pre-trained coastal dynamics model CoastDynNet, and using the sediment transport equation to analyze the physical mechanism of coastline changes; applying a coastline evolution optimization function to finely adjust the time interpolation results; using the pre-trained temporal decomposition network WaveletTempNet to perform multi-scale decomposition on the coastline change sequence; designing a spectral analysis module, and calculating the coastline stability index and change frequency matrix based on the separation results; using the Bayesian model averaging method to construct a multi-model integration framework.

[0007] Among them, the temporal attention matrix is a two-dimensional feature representation matrix formed by weighted fusion of satellite image data at different time points through a multi-head attention mechanism, and is used to capture key change points in the time series.

[0008] Among them, the first fluctuation matrix is a primary feature representation of the morphological changes of the coastline extracted from the original satellite images, and is a feature matrix composed of multi-dimensional indicators including position offset, area change rate, and morphological complexity.

[0009] Among them, the second fluctuation matrix is a low-dimensional and high-information-density feature representation formed after dimensionality reduction, removing redundant information and noise and retaining the core change patterns.

[0010] Among them, the sediment transport equation is a physical mechanism equation describing the movement law of sediments around the coastline. The inputs include wave energy density, incident angle, sediment particle size distribution, tidal amplitude, and alongshore current velocity, and the outputs are the net sediment migration amount per unit time and the coastline position change rate.

[0011] Among them, the coastline evolution optimization function is used to constrain and optimize the deep learning prediction results based on physical constraints. The inputs include the initial coastline morphology, predicted coastline morphology, terrain slope, sediment characteristics, and extreme event intensity, and the output is the corrected coastline morphology that meets the physical constraint conditions.

[0012] Among them, the structure of the CoastDynNet model is a multi-layer encoding and decoding architecture, including a sediment dynamics physical constraint layer, a tidal influence attention module, a wave dynamics calculation unit, and a morphological evolution prediction layer; in the CoastDynNet model, the parameters of the sediment dynamics physical constraint layer are determined by the ratio of the longshore sediment transport rate to the vertical sediment transport rate, the parameters of the tidal influence attention module are determined by the ratio of the tidal period to the coastline response time, and the parameters of the wave dynamics calculation unit are determined by the ratio of the mean wave height to the critical erosion wave height.

[0013] Among them, the structure of the WaveletTempNet model is an end-to-end deep learning network composed of a multi-scale wavelet transform encoding layer, a time series feature extraction layer, a long short-term memory network layer, and a reconstruction decoding layer; in the WaveletTempNet model, the wavelet basis function parameters of the wavelet transform encoding layer are determined by the coastline change period, the number of attention heads of the time series feature extraction layer is determined by the input sequence length, and the hidden state dimension of the long short-term memory network layer is determined by the feature complexity.

[0014] Among them, the stability index is an indicator that quantifies the ability of the coastline to resist external disturbances and is calculated by the ratio of the long-term change trend to the short-term fluctuation amplitude; the change frequency matrix is a two-dimensional matrix that records the change frequency characteristics of different sections of the coastline. The rows represent the spatial positions, the columns represent the change frequency intervals, and the matrix element values represent the change intensity at the corresponding position in a certain frequency interval.

[0015] Among them, the Bayesian model averaging method is a statistical method that performs weighted averaging on the prediction results of multiple models based on the principle of Bayesian inference. By calculating the prediction error distribution of each model on the historical validation set, its credibility weight in the integration framework is determined, realizing the complementary advantages of the models and the quantification of prediction uncertainty.

[0016] The present invention constructs a temporal attention matrix through a multi-source satellite data fusion network, combines a fourth-order temporal convolutional neural network to extract the coastline morphological change features, and applies a spatio-temporal attention mechanism for feature dimensionality reduction and optimization, realizing high-precision detection of coastline changes. The method ingeniously integrates the pre-trained coastal dynamics model CoastDynNet and the temporal decomposition network WaveletTempNet, and at the same time introduces a Bayesian model averaging mechanism, forming a monitoring framework that deeply integrates data-driven and physical mechanisms.

[0017] This method effectively solves multiple defects existing in traditional technologies: on the one hand, multi-source data fusion and temporal attention mechanism significantly improve the temporal resolution, making short-term change monitoring possible; on the other hand, the temporal decomposition network successfully distinguishes long-term trends from short-term fluctuations, realizing multi-scale change feature extraction; more importantly, by introducing physical constraints and coastal dynamics models, the physical rationality of the monitoring results is ensured, and the problem of insufficient accuracy of pure data-driven methods is solved.

[0018] Through the organic combination of deep learning and coastal dynamics, the present invention constructs a complete dynamic monitoring method for coastline changes, which not only realizes high-precision and multi-scale change monitoring, but also reveals the physical mechanism behind the changes, provides a scientific basis for coastal zone management and protection, solves the technical problem of insufficient accuracy in dynamic monitoring of coastline changes, and significantly improves the accuracy and reliability of dynamic monitoring of coastline changes. Brief Description of the Drawings

[0019] Figure 1 It is a flowchart of the method of the present invention.

[0020] Figure 2 It is a diagram of the temporal changes of the coastline monitored by multi-source satellites and the response diagram of typhoon events in Example 2.

[0021] Figure 3 It is a diagram of the multi-scale decomposition result of coastline changes in Example 2.

[0022] Figure 4 It is a diagram of the spectrum analysis result of coastline changes in Example 2. Detailed Embodiments

[0023] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] As Figure 1 shown, it is a flowchart of a dynamic monitoring method for coastline changes driven by deep learning provided by the present invention. This method includes the following steps: S01. Construct a multi-source satellite data fusion network, combine time series images obtained by satellites with different orbital periods, and form a temporal attention matrix; S02. Apply a fourth-order temporal convolutional neural network to extract the morphological change features of the coastline, and establish a first fluctuation matrix in a recursive manner; S03. Use the spatio-temporal attention mechanism to perform dimensionality reduction processing on the first fluctuation matrix to generate a second fluctuation matrix; S04. Establish a fluctuation conversion matrix to convert the second fluctuation matrix into a quantitative coastline change index; S05. Input the pre-trained coastal dynamics model CoastDynNet, and use the sediment transport equation to analyze the physical mechanism of coastline change; S06. Apply the coastline evolution optimization function to refine the time interpolation result to improve the short-term prediction accuracy; S07. Use the pre-trained time series decomposition network WaveletTempNet to perform multi-scale decomposition on the coastline change sequence to separate the long-term trend and short-term fluctuations; S08. Design a spectrum analysis module, and calculate the coastline stability index and change frequency matrix based on the separation result; S09. Use the Bayesian model averaging method to construct a multi-model integration framework, and weighted integrate the outputs of the physical model and the deep learning model according to the prediction error to generate a comprehensive coastline dynamic assessment report; Among them, the time series attention matrix specifically refers to a two-dimensional feature representation matrix formed by weighted fusion of satellite image data at different time points through the multi-head attention mechanism, which is used to capture key change points in the time series.

[0025] Among them, the first fluctuation matrix specifically refers to the primary feature representation of the coastline morphological change extracted from the original satellite image, which is a feature matrix composed of multi-dimensional indicators such as position offset, area change rate, and morphological complexity.

[0026] Among them, the second fluctuation matrix specifically refers to the low-dimensional high-information density feature representation formed after dimensionality reduction processing, which removes redundant information and noise and retains the core change pattern.

[0027] Among them, the fluctuation transformation matrix specifically refers to the transformation matrix that maps the abstract feature space to the coastline change index space with clear physical meaning, and realizes the mapping relationship from features to actual change amounts.

[0028] Among them, the sediment transport equation specifically refers to the physical mechanism equation that describes the movement law of sediments around the coastline. The inputs include wave energy density, incident angle, sediment particle size distribution, tidal amplitude, and alongshore current velocity, and the outputs are the net sediment migration amount per unit time and the coastline position change rate; Among them, the coastline evolution optimization function is used to constrain and optimize the deep learning prediction result based on physical constraints. The inputs include the initial coastline morphology, predicted coastline morphology, terrain slope, sediment characteristics, and extreme event intensity, and the output is the corrected coastline morphology that meets the physical constraint conditions; Among them, the specific structure of the pre-trained CoastDynNet model is a multi-layer encoding and decoding architecture, including a sediment dynamics physical constraint layer, a tidal influence attention module, a wave dynamics calculation unit, and a morphological evolution prediction layer. The parameters of the sediment dynamics physical constraint layer are determined by the ratio of the longshore sediment transport rate to the vertical sediment transport rate. The parameters of the tidal influence attention module are determined by the ratio of the tidal period to the coastline response time. The parameters of the wave dynamics calculation unit are determined by the ratio of the mean wave height to the critical erosion wave height. The steps for establishing the training dataset during the pre-training process of the CoastDynNet model specifically include collecting satellite observation data of different types of coastlines around the world for many years, extracting hydrometeorological and ocean dynamics parameters in the corresponding regions, constructing physical simulation data of the combined action of tides and waves, marking the occurrence time periods of extreme events and extracting the morphological change characteristics before and after, and classifying and sorting according to different coast types and dynamic conditions to form a training sample library. The steps for pre-training the CoastDynNet model specifically include first pre-training on the physical simulation data to learn the basic dynamics laws, then fine-tuning on the global satellite observation data to learn the actual coast response patterns, then performing transfer learning for coast types similar or close to the region and the sea area to be measured to optimize the local adaptability, and finally optimizing the model parameters through error backpropagation by comparing the morphological differences between the real observation sequence and the model prediction sequence. Among them, the specific structure of the pre-trained WaveletTempNet model is an end-to-end deep learning network composed of a multi-scale wavelet transform encoding layer, a time series feature extraction layer, a long short-term memory network layer, and a reconstruction decoding layer. The wavelet basis function parameters of the wavelet transform encoding layer are determined by the coastline change period. The number of attention heads of the time series feature extraction layer is determined by the input sequence length. The hidden state dimension of the long short-term memory network layer is determined by the feature complexity. The steps for establishing the training dataset during the pre-training process of the WaveletTempNet model specifically include collecting different time-scale change sequences of typical coastlines around the world, performing standardized preprocessing on the sequences to eliminate the influence of dimensions, manually annotating different scale change components as training targets, constructing a mixed sequence containing long-term trends and short-term fluctuations as the input, and performing stratified sampling according to different change patterns and frequency characteristics to form a training set and a validation set. The steps for pre-training the WaveletTempNet model specifically include first performing initial training on the synthetic data to learn the basic decomposition ability, then performing supervised learning on the real coastline sequences to improve the generalization performance, then optimizing the parameters by comparing the consistency between the outputs of different decomposition levels and the theoretical decomposition results, and finally ensuring the lossless retention of information in the decomposition and reconstruction processes through reconstruction accuracy verification. Among them, the multi-layer wavelet transform network specifically refers to a neural network structure designed based on the wavelet transform principle, which realizes the multi-resolution analysis of signals in the time and frequency domains by learning wavelet basis functions of different scales.

[0029] Among them, the stability index specifically refers to an indicator that quantifies the ability of the coastline to resist external disturbances, and is calculated by the ratio of the long-term change trend to the short-term fluctuation amplitude.

[0030] Among them, the change frequency matrix specifically refers to a two-dimensional matrix that records the change frequency characteristics of different sections of the coastline. The rows represent the spatial positions, the columns represent the change frequency intervals, and the matrix element values represent the change intensity at the corresponding position in a certain frequency interval.

[0031] Among them, the Bayesian model averaging method specifically refers to a statistical method that performs weighted averaging on the prediction results of multiple models based on the principle of Bayesian inference. By calculating the prediction error distribution of each model on the historical validation set, the credibility weight in the integrated framework is determined, so as to achieve complementary advantages of models and quantification of prediction uncertainty.

[0032] The specific implementation manners of the above steps are described in detail below.

[0033] The specific implementation manner of step S01 is to effectively integrate temporal information by establishing a multi-source satellite data fusion network. First, multi-temporal images of satellites with different orbital periods (including polar-orbiting satellites and geostationary satellites) are collected, and each image is radiometrically corrected and geometrically corrected to ensure data consistency. Then, a multi-head self-attention mechanism is used to construct a temporal attention matrix, and each attention head is responsible for learning the change characteristics at different time scales. Specifically, for the image feature at the th time point and the image feature at the th time point, the attention weight is calculated to form a -dimensional temporal attention matrix A, where is the length of the time series. This matrix can adaptively highlight the image features at key change moments and reduce the influence of noise. When the attention weight is greater than 0.7, it indicates that there is a significant correlation between the two time points and needs to be focused on. The role of this step is to establish a high-quality time series feature representation by fusing multi-source remote sensing data, laying a data foundation for subsequent coastline change analysis.

[0034] The specific implementation of step S02 is to apply a fourth-order temporal convolutional neural network to extract the morphological change features of the coastline. First, four layers of temporal convolutional layers are designed, and each layer uses different-sized temporal convolutional kernels (3, 5, 7, and 9 respectively) to capture multi-scale temporal features. Then, the convolutional results are processed by a recurrent neural network, specifically using a long short-term memory network (LSTM) structure, with the hidden layer dimension set to 256 and the time step set to the length of the image sequence. The output of the recurrent network is mapped through a fully connected layer to generate the first fluctuation matrix F containing multi-dimensional indicators such as position offset, area change rate, and morphological complexity. Among them, the position offset is obtained by calculating the Euclidean distance between corresponding points of the coastline at adjacent time points; the area change rate represents the percentage change in the land-sea boundary area per unit time; the morphological complexity is calculated through the fractal dimension, and the value range is 1-2, and the larger the value, the more complex the coastline. When the position offset exceeds 10 meters per year or the area change rate exceeds 5% per year, it is determined as a significant change area. The purpose of this step is to extract the primary feature representation of the coastline morphological change from the original images and capture the key change information.

[0035] The specific implementation of step S03 is to use a spatio-temporal attention mechanism to reduce the dimension of the first fluctuation matrix. First, a spatio-temporal attention module is constructed, which contains two attention sub-modules in the temporal dimension and the spatial dimension. The temporal dimension attention sub-module identifies key time points by calculating the correlation matrix of features at different time points; the spatial dimension attention sub-module identifies key spatial regions by calculating the correlation matrix of features at different spatial positions. Then, the outputs of the two attention sub-modules are weighted and fused, and the weight coefficients are automatically optimized through backpropagation. Finally, the principal component analysis algorithm is applied to reduce the dimension of the fusion result, and the principal components with an explained variance ratio reaching 95% are retained to form the second fluctuation matrix G with low dimension and high information density. When the contribution rate of the largest principal component exceeds 60%, it indicates the existence of a dominant change pattern. The role of this step is to remove redundant information and noise in the first fluctuation matrix, extract the core change pattern, and provide a more concise feature representation for subsequent quantitative analysis.

[0036] The specific implementation of step S04 is to establish a fluctuation transformation matrix, which maps the abstract feature space to the coastline change index space with clear physical meanings. First, a three-layer fully connected neural network is designed. The number of nodes in the input layer is equal to the dimension of the second fluctuation matrix, the number of nodes in the hidden layer is 128, and the number of nodes in the output layer is equal to the number of target change indexes. Then, the known ground measured data is used as supervision information, and the network parameters are optimized by the least squares method to learn the mapping relationship from the feature space to the change index space. The finally obtained fluctuation transformation matrix T can convert the second fluctuation matrix G into a series of quantitative coastline change indexes, including the annual average retreat rate, the seasonal fluctuation amplitude, the extreme event response intensity, etc. When the root mean square error between the predicted value and the measured value is less than 0.15, it is considered that the transformation matrix reaches the acceptable accuracy. The purpose of this step is to convert the abstract features extracted by the deep learning model into change indexes with actual physical meanings, which is convenient for professionals to understand and for decision-making support.

[0037] The specific implementation of step S05 is to input the quantitative indexes into the pre-trained coastal dynamics model CoastDynNet, and analyze the physical mechanism of coastline change through the sediment transport equation. First, the change indexes obtained in step S04 are input into the CoastDynNet model together with the auxiliary data (such as wave parameters, tidal data, meteorological data). Then, the sediment dynamics physical constraint layer inside the model calculates the longshore sediment transport rate and the vertical sediment transport rate . When the ratio is greater than 3, it indicates that the longshore sediment transport is the dominant process. Then, the wave dynamics calculation unit evaluates the wave erosion ability according to the ratio of the average wave height to the critical erosion wave height. When this ratio is greater than 1.2, the coastline is in the erosion state. Finally, the morphological evolution prediction layer integrates the outputs of each module to generate a physical interpretation report of the coastline change, including the dominant dynamic mechanism, the sediment budget status, and the change trend prediction. The role of this step is to combine the data-driven statistical analysis with the physical mechanism model, reveal the essential reasons behind the coastline change, and improve the interpretability of the monitoring results.

[0038] The specific implementation of step S06 is to refine the time interpolation result using the coastline evolution optimization function. First, an optimization objective function is constructed based on physical constraint conditions, including terrain slope limitations, sediment characteristics constraints, and extreme event impact factors. Then, the Lagrange multiplier method is used to transform the constraint conditions into penalty terms, which together with the prediction error term form a complete optimization function. Next, the gradient descent algorithm is used for iterative optimization, and the iteration stops when the maximum difference between the results of two adjacent iterations is less than 0.5 meters. Finally, the corrected coastline morphology that meets the physical constraint conditions is obtained. For areas significantly affected by tides, additional correction is required according to the tidal level change, and the tidal level correction coefficient ranges from 0.8 to 1.2. The purpose of this step is to correct the prediction results driven by pure data by introducing physical constraint conditions, making the prediction results more in line with the actual coastal dynamics process and improving the short-term prediction accuracy.

[0039] The specific implementation of step S07 is to perform multi-scale decomposition on the coastline change sequence using the pre-trained time series decomposition network WaveletTempNet. First, the corrected coastline change sequence is input into the wavelet transform encoding layer of the WaveletTempNet model. This layer, based on the principle of discrete wavelet transform, uses wavelet basis functions of different scales (such as Haar wavelet, Daubechies wavelet) to perform multi-resolution analysis on the input signal. Then, the time series feature extraction layer extracts the time features of each decomposition level through the multi-head self-attention mechanism. When the length of the input sequence is n, the number of attention heads is set to min(8, n / 4). Next, the long short-term memory network layer further processes the extracted features to capture long-term dependencies. Finally, the reconstruction decoding layer reorganizes the processing results to separate the long-term trend component and the short-term fluctuation component. When the reconstruction error is less than 5% of the standard deviation of the original sequence, the decomposition quality is considered acceptable. The role of this step is to decompose the complex coastline change sequence into components of different time scales, facilitating the separate analysis of long-term evolution trends and short-term dynamic changes.

[0040] The specific implementation of step S08 is to design a spectrum analysis module to further process the separated time series. First, apply the fast Fourier transform to the short-term fluctuation component to obtain the frequency-domain representation and identify the main periodic components. When the frequency with an energy ratio exceeding 15% is identified as a significant period. Then calculate the ratio of the long-term trend amplitude to the standard deviation of the short-term fluctuations as the stability index SI. When SI is greater than 2, it indicates that the coastline in this section is relatively stable, and when SI is less than 0.5, it indicates instability. Next, construct a change frequency matrix, where the rows represent the spatial sampling points and the columns represent the predefined frequency intervals (such as interannual changes, seasonal changes, tidal cycle changes, etc.). The matrix element values represent the change intensity at the corresponding position in a certain frequency interval. Finally, based on the spectrum analysis results, identify potential periodic change patterns and abnormal change events. The purpose of this step is to quantitatively evaluate the dynamic characteristics of the coastline and provide a scientific basis for management decisions.

[0041] The specific implementation of step S09 is to construct a multi-model integration framework using the Bayesian model averaging method. First, collect the prediction results of physical models (such as CoastDynNet) and deep learning models (such as WaveletTempNet) on the historical validation set. Then, based on the principle of Bayesian inference, calculate the prediction error distribution of each model. When the error follows a normal distribution, use the mean squared error, and when it follows a long-tailed distribution, use the mean absolute percentage error. Next, determine the credibility weight of each model according to the error distribution. The weight calculation formula is an exponential function of the model accuracy, ensuring that models with higher accuracy obtain greater weights. Finally, perform a weighted average of the outputs of each model to generate a comprehensive coastline dynamic assessment report, including change trend prediction, uncertainty estimation, and risk level assessment. When the prediction difference between models exceeds 20%, expert review is required. The role of this step is to make full use of the complementary advantages of different types of models to improve the accuracy and reliability of predictions, and at the same time quantify the uncertainty of the prediction results.

[0042] Furthermore, the specific structure of the CoastDynNet model is a multi-layer encoding and decoding architecture, which consists of four core modules. The first module is the sedimentation dynamics physical constraint layer, which uses the improved CERC formula to calculate the coastal sediment transport rate and establishes a constraint relationship in combination with the vertical sediment transport rate. Its parameters are determined by the ratio of the coastal sediment transport rate to the vertical sediment transport rate. When the ratio is greater than 3, a strong coastal sediment transport parameter configuration is used, and when the ratio is less than 0.5, a strong vertical sediment transport parameter configuration is used. The second module is the tidal impact attention module, which uses a gated recurrent unit network to process tidal time series data. Its parameters are determined by the ratio of the tidal cycle to the coastline response time. When the ratio is less than 0.1, the tidal impact can be ignored. The third module is the wave dynamics calculation unit, which calculates the wave erosion effect based on the improved Battjes-Janssen wave energy dissipation model. Its parameters are determined by the ratio of the average wave height to the critical erosion wave height. When the ratio is greater than 1.5, the strong erosion mode is activated. The fourth module is the morphological evolution prediction layer, which integrates the outputs of the first three modules and applies the improved one-dimensional ecological cellular automaton algorithm to predict the morphological evolution of the coastline. The steps for establishing the pre-training dataset of the CoastDynNet model include: first, collecting multi-year satellite observation data of different types of coastlines around the world (including sandy coasts, reef coasts, muddy coasts, etc.), with a time span of no less than 10 years; then extracting the hydrological, meteorological and ocean dynamic parameters of the corresponding area, including wave spectra, tidal curves, wind field data, sediment particle size distribution, etc.; then constructing physical simulation data of the combined action of tides and waves, with a simulation time step of no more than 1 hour; then marking the time periods of extreme events such as typhoons and storm surges, and extracting the characteristics of coastal morphological changes before and after; finally, classifying and sorting according to different coastal types and dynamic conditions to form a training sample library of no less than 1,000 samples. The pre-training steps of the CoastDynNet model include: first pre-training on physical simulation data to learn basic sedimentary dynamics laws, with no less than 100 pre-training rounds; then fine-tuning on global satellite observation data to learn actual coastal response patterns, with no less than 50 fine-tuning rounds; then transfer learning is carried out for regions with similar or similar coastal types to optimize local adaptability, with no less than 30 transfer learning rounds; finally, by comparing the morphological differences between the real observation sequence and the model prediction sequence, the root mean square error is used as the loss function, and error back propagation is performed to optimize the model parameters, with no less than 20 optimization rounds.

[0043] Furthermore, the specific structure of the WaveletTempNet model is an end-to-end deep learning network composed of four key layers. The first layer is the multi-scale wavelet transform encoding layer, which uses the discrete wavelet transform principle to perform multi-scale decomposition on the input time series. The wavelet basis function can select the Haar, Daubechies or Symlet series, and its parameters are determined by the coastline change cycle. When the cycle is less than 1 month, the Haar wavelet is selected, when the cycle is 1 to 12 months, the Daubechies wavelet is selected, and when the cycle is greater than 12 months, the Symlet wavelet is selected. The second layer is the time series feature extraction layer, which uses a multi-head self-attention mechanism to capture the time dependency. The number of attention heads is determined by the length of the input sequence, which is generally set to 1 / 4 to 1 / 8 of the sequence length. The third layer is the long short-term memory network layer, which uses a bidirectional LSTM structure to process sequence information. Its hidden state dimension is determined by the feature complexity, with a simple change mode set to 64 and a complex change mode set to 256. The fourth layer is the reconstruction decoding layer, which recombines the processed features into a time series through an inverse wavelet transform and separates the long-term trend and short-term fluctuation components. The steps for establishing the pre-training dataset of the WaveletTempNet model include: first, collecting the change sequences of typical coastlines around the world at different time scales, including daily, monthly, annual and decadal scale data; then, standardizing and preprocessing the sequences, and using the Z-score method to eliminate the dimension effect; then, experts manually annotate the change components at different scales as training targets; then, construct a mixed sequence containing long-term trends and short-term fluctuations as input, with a mixing ratio ranging from 0.1 to 10; finally, stratified sampling is performed according to different change patterns and frequency characteristics to form training sets and validation sets, with a ratio of 8:2 between the training set and the validation set. The pre-training steps of the WaveletTempNet model include: first, initial training on synthetic data to learn basic decomposition capabilities, with no less than 200 initial training rounds; then, supervised learning on real coastline sequences to improve generalization performance, with no less than 100 supervised learning rounds; then, by comparing the consistency of the outputs of different decomposition levels with the theoretical decomposition results, parameter optimization is performed using the root mean square error and structural similarity index as loss functions, with no less than 50 rounds of parameter optimization; finally, reconstruction accuracy verification is performed to ensure that the information of the decomposition and reconstruction processes is losslessly retained, and the training is considered to have met the standard when the reconstruction error is less than 3% of the standard deviation of the original signal.

[0044] The primary difficulty faced by traditional coastline monitoring is the relatively long satellite revisit cycle. Taking the Landsat series of satellites as an example, its revisit cycle is 16 days, and this time interval far exceeds the characteristic time scales of many coastal dynamic changes. Due to this inherent time resolution limitation, when extreme meteorological events such as storm surges and storm tides occur, these events often have a significant impact on the coastline within a few hours to a few days, and the traditional single-satellite observation system cannot capture these short-term change processes in a timely manner. This results in the difficulty of completely recording the morphological changes of the coastline before and after extreme events, and key change points are missed, making there a "time blind spot" in coastline dynamic monitoring. In addition, even if data is obtained according to the established revisit cycle under ideal conditions, the actual available data is often further reduced due to factors such as cloud cover and sensor failures, and the effective observation interval may be extended to several months, making short-term change monitoring more difficult. This problem of insufficient time resolution severely restricts the accurate monitoring and mechanism understanding of the dynamic changes of the coastline, especially the assessment of the impact of catastrophic extreme events.

[0045] The present invention effectively solves the problem of insufficient time resolution caused by the long traditional satellite revisit cycle by constructing a multi-source satellite data fusion network (step S01). This method no longer relies on a single satellite data source, but integrates time series images obtained from multiple satellites with different orbital periods (such as optical satellites Landsat, Sentinel-2, and radar satellite Sentinel-1, etc.) to form a complementary coverage high-density time series. The temporal attention matrix constructed by the multi-head self-attention mechanism can adaptively fuse image data of different phases and different sensors, highlighting the characteristics of key change moments and effectively improving the time resolution. At the same time, the fourth-order temporal convolutional neural network (step S02) adopts a multi-scale temporal convolutional kernel design, which can capture the change characteristics of different time scales and has a high sensitivity to short-term mutation events. The coastline evolution optimization function (step S06) introduces physical constraint conditions, which can perform reasonable interpolation within the time intervals where observation data is missing to fill the "time blind spot". In addition, the multi-model integration framework constructed by the Bayesian model averaging method (step S09) can comprehensively utilize the advantages of physical models and deep learning models, and use the prior knowledge provided by physical models to improve the prediction reliability in data-sparse regions, effectively coping with the mutation situations caused by extreme events. The comprehensive application of these technical means enables the present invention to greatly improve the time resolution of coastline monitoring, thereby effectively capturing the short-term changes of the coastline caused by extreme events such as storm surges.

[0046] The second difficulty in monitoring coastline changes is the difficulty in effectively distinguishing the long-term trends caused by climate change and the short-term changes caused by seasonal fluctuations or extreme events. The dynamic changes of the coastline are the composite result of processes on multiple time scales, including the interdecadal long-term trends caused by factors such as sea-level rise and sediment supply changes, the periodic seasonal changes caused by factors such as monsoon alternation and tidal cycles, and the short-term violent fluctuations caused by extreme events such as typhoons and storm surges. These change processes on different time scales are superimposed and interact with each other in the observed data. If they cannot be effectively separated, it will lead to problems such as deviation in trend analysis, inaccurate assessment of the impact of extreme events, and reduced reliability of prediction results. Traditional analysis methods often use simple time series filtering or regression analysis, which are difficult to accurately separate complex multi-scale change components, especially when there are missing or non-uniformly sampled data, and the separation effect is even less satisfactory.

[0047] To address the problem of separating multi-scale signals of coastline changes, the present invention uses a pre-trained time series decomposition network WaveletTempNet (step S07) to effectively separate long-term trends and short-term fluctuations. This network is designed based on the principle of wavelet transform. Through the multi-scale wavelet transform coding layer, it performs multi-resolution analysis on the coastline change sequence. Wavelet basis functions of different scales can accurately extract change components with different periods. The time series feature extraction layer of the network captures the time features of each decomposition level through the multi-head self-attention mechanism, and the long short-term memory network layer further processes the sequence information to capture long-term dependence relationships. Finally, through the reconstruction decoding layer, the processing results are reorganized to achieve the accurate separation of long-term trends and short-term fluctuations. On this basis, the spectral analysis module (step S08) further processes the separated time series, identifies the main periodic components through the fast Fourier transform, calculates the stability index to quantitatively evaluate the coastline stability, and constructs a change frequency matrix to describe the change characteristics of different sections at different frequencies. At the same time, the CoastDynNet model (step S05) analyzes the physical mechanism of coastline changes through the sediment transport equation and can identify the driving factors of changes on different time scales. The collaborative application of these technologies enables the present invention to effectively distinguish the long-term trends caused by climate change and the short-term fluctuations caused by seasonal or extreme events, providing reliable technical support for dynamic assessment and prediction of the coastline.

[0048] The following details the mathematical models or calculation processes involved in the present invention.

[0049] In step S01, the calculation process of the multi-head self-attention mechanism for constructing the time series attention matrix is specifically expressed as follows: ; ; ; In the formula, is the attention weight of the image at the -th time point and the image at the -th time point; is the unnormalized attention score; and are the image feature vectors at the -th and -th time points respectively; and are the query transformation matrix and the key transformation matrix respectively, which are obtained through model training; is the dimension of the feature vector; is the length of the time series; is the -dimensional temporal attention matrix.

[0050] Among them, the parameter acquisition method is: and are obtained by feature extraction from the original satellite image, specifically including Step 1: performing radiometric correction and geometric correction on the satellite image; Step 2: using a convolutional neural network to extract the spatial features of the image to obtain the feature vector. and are optimized through the backpropagation algorithm during the model training process, and the training objective is to minimize the error between the predicted coastline and the actual coastline. usually has a value range of 64 to 256.

[0051] The calculation principle of this attention mechanism is based on dot product similarity. By calculating the inner product of the image features at different time points to measure the correlation between them, and dividing by for scaling to avoid the vanishing gradient problem. The exponential function and the normalization operation transform the attention score into a probability distribution to ensure that the sum of all weights is 1. This design can adaptively highlight the image features at key change moments and effectively improve the temporal modeling ability. Compared with the traditional linear weighting method, this mechanism can capture non-linear time dependencies and is more robust to outliers.

[0052] In step S02, the calculation processes of the position offset, the area change rate, and the morphological complexity are specifically expressed as follows: ; ; ; In the formula, is the position offset at the and are respectively at the The x - coordinate and y - coordinate of a sampling point; is the number of coastline sampling points; is The rate of change of area at time, in percentage; is The land - sea boundary area at time, in square meters; is The morphological complexity at time, dimensionless; is The coastline length at time, in meters.

[0053] Among them, the parameter acquisition method is as follows: and are obtained by extracting the coastline from satellite images, specifically including Step 1: Using threshold segmentation or deep learning methods to extract the water - land boundary; Step 2: Vectorize the boundary to obtain the coordinates of equally - spaced sampling points. is obtained by calculating the area of the water body region, specifically including Step 1: Counting the pixels of the segmentation result; Step 2: Multiplying the number of pixels by the square of the pixel resolution to obtain the actual area. is obtained by calculating the geometric length of the coastline, specifically including Step 1: Calculating the length of the vectorized coastline; Step 2: Multiplying the pixel length by the resolution to obtain the actual length.

[0054] The calculation principles of these indicators are based on Euclidean distance, percentage change rate, and fractal geometry respectively. The position offset reflects the degree of change in the coastline position, and the displacement of corresponding sampling points at adjacent time points is calculated through Euclidean distance; the area change rate reflects the relative change in the land - sea interface area, and using the percentage form is convenient for comparison of different regions; the morphological complexity is based on the fractal dimension theory and is calculated through the logarithmic ratio of the coastline length to the area, which can quantitatively describe the geometric complexity of the coastline. These three indicators together constitute a multi - dimensional feature representation of the coastline morphological change, and can depict the dynamic change characteristics of the coastline more comprehensively than traditional single indicators.

[0055] In step S03, the calculation process of the spatio - temporal attention mechanism is specifically expressed as follows: ; ; ; In the formula, is the attention output in the time dimension; , , are the query matrix, key matrix, and value matrix in the time dimension respectively; is the attention output in the spatial dimension; , , are respectively the query matrix, key matrix, and value matrix in the spatial dimension; is the feature dimension; is the spatio-temporal attention fusion result; is the attention weight coefficient in the time dimension, and its value range is 0 to 1.

[0056] Among them, the parameter acquisition method is as follows: , , , , , are all obtained through linear transformation of the first fluctuation matrix , specifically including Step 1: Design six different linear transformation matrices , , , , , ; Step 2: Multiply the first fluctuation matrix with these transformation matrices respectively to obtain the corresponding query matrix, key matrix, and value matrix. is automatically optimized through the backpropagation algorithm, and the initial value is set to 0.5.

[0057] The design principle of the spatio-temporal attention mechanism is to decompose the attention calculation into two independent calculation processes in the time dimension and the spatial dimension, and then integrate the information of the two dimensions through weighted fusion. The time dimension attention focuses on the dependence between different time points, and the spatial dimension attention focuses on the correlation between different spatial positions. This decomposition design enables the model to capture the key information in both the time and space dimensions simultaneously, significantly improving the feature extraction ability. The weighting coefficient is determined through automatic optimization and can flexibly adjust the importance of the two dimensions according to the specific data characteristics. Compared with the traditional single-dimension attention mechanism, this method can more effectively handle the complex dependence of spatio-temporal data.

[0058] In step S04, the calculation process of the fluctuation transformation matrix is specifically expressed as follows: ; ; In the formula, is the fluctuation transformation matrix; , , are respectively the weight matrices of the three layers of the neural network; , , are respectively the bias vectors of the three layers of the neural network; For modifying the linear unit activation function; is the second fluctuation matrix; is the transformed quantitative index vector of coastline change, including multiple indexes such as annual average retreat rate, seasonal fluctuation amplitude, extreme event response intensity, etc.

[0059] Among them, the parameter acquisition method is: and and and and and Obtained through the supervised learning method, specifically including step 1: Collecting known ground measured data as training samples, including indexes such as actual measured annual average retreat rate, seasonal fluctuation amplitude, etc.; step 2: Using the second fluctuation matrix corresponding to the time period as the input and the measured indexes as the output target; step 3: Using the least squares method to optimize the neural network parameters, and the objective function is where is the measured index vector, is the regularization coefficient, and its value range is 0.001 to 0.1, is the Frobenius norm of the weight matrix.

[0060] The design principle of the fluctuation transformation matrix is to establish a mapping relationship from the abstract feature space to the index space with clear physical meaning through a deep neural network. The three-layer neural network structure can effectively fit complex non-linear mapping relationships, The activation function introduces non-linear transformation ability and solves the gradient vanishing problem. The least squares loss function measures the difference between the predicted index and the measured index, and the regularization term prevents overfitting. This design transforms the abstract features extracted by deep learning into change indexes with actual physical meaning, improving the interpretability and practicality of the model. Compared with the traditional linear transformation method, this method can capture the relationship between more complex features and indexes.

[0061] In step S05, the calculation process of the sediment transport equation is specifically expressed as follows: ; ; ; In the formula, is the longshore sediment transport rate, and the unit is ; is the vertical sediment transport rate, and the unit is ; and and are empirical coefficients; is the wave height in the wave breaking area, and the unit is ; is the wave incident angle in the wave breaking zone, with the unit of radian; is the beach slope, dimensionless; is the longshore gradient of wave height; is the rate of change of coastline position, with the unit of ; is the critical water depth of wave action, with the unit of ; is the longshore distance coordinate, with the unit of ; is the distance coordinate perpendicular to the reference line, with the unit of ; is the time, with the unit of .

[0062] Among them, the parameter acquisition methods are as follows: Obtained through wave monitoring data or wave model calculation, specifically including Step 1: Collecting inshore wave observation data or calculating wave propagation using the third-generation wave model (such as the SWAN model); Step 2: Calculating the evolution of waves in the inshore area according to the shallow water equation to determine the wave height in the wave breaking zone. Obtained through wave direction observation data, specifically including Step 1: Collecting deep-water wave direction data; Step 2: Calculating the incident angle of waves in the breaking zone after refraction according to Snell's law. Obtained through submarine topography survey, specifically including Step 1: Conducting submarine topography survey; Step 2: Calculating the average slope of the inshore area. , , are empirical coefficients that need to be calibrated through measured data. The value range of is 0.1 - 0.4, The value range of can be estimated through , where is the significant wave height.

[0063] Among them, considering that the longshore sediment transport rate is proportional to the 2.5th power of the wave height and the sine value of the incident angle, reflecting the influence of wave energy and incident angle on longshore sediment transport; the vertical sediment transport rate It consists of two parts. The first term represents the onshore sediment transport caused by waves, which is proportional to the square of the wave height and the slope. The second term represents the offshore sediment transport caused by the wave height gradient. The coastline position change rate equation is based on the principle of sediment mass conservation and calculates the change rate of the coastline position through the spatial gradient of the longshore sediment transport rate and the vertical sediment transport rate. This physical model can quantitatively describe the dynamic mechanism of coastline evolution and provide a physical explanation for coastline changes. Compared with pure data-driven methods, this physical model has better interpretability and extrapolation ability.

[0064] In step S06, the calculation process of the coastline evolution optimization function is specifically expressed as follows: ; ; ; In the formula, is the coastline evolution optimization function; is the optimized coastline position vector; is the coastline position vector preliminarily predicted by the model; is the spatial gradient of the coastline position; , , , are weight coefficients; is the physical constraint condition term; and are the maximum and minimum allowable slopes of the coastline respectively; is the optimized final coastline position vector; is the number of discrete sampling points of the coastline; is the th sampling point of the coastline position.

[0065] Among them, the parameter acquisition method is: is obtained by preliminary prediction through a deep learning model; , , , are weight coefficients, determined by the cross-validation method, has a value range of 0.5 - 2.0, has a value range of 0.1 - 0.5, has a value range of 0.05 - 0.2, has a value range of 1.0 - 5.0; and are determined according to the actual coastal terrain characteristics, usually has a value range of 30° - 45°, The value range of is usually from 0° to 5°. The optimization process uses the gradient descent algorithm, specifically including Step 1: Initialize ; Step 2: Calculate the gradient of the objective function with respect to ; Step 3: Update in the opposite direction of the gradient, with a step size of

[0066] (the value range is from 0.01 to 0.1); Step 4: Repeat Step 2 and Step 3 until convergence or the maximum number of iterations is reached.

[0067] The design principle of the coastline evolution optimization function is to introduce physical constraint conditions on the basis of data-driven prediction, and transform the constrained problem into an unconstrained optimization problem through the Lagrange multiplier method. The objective function consists of four parts: the first item is the data item, which ensures that the optimization result is close to the preliminary prediction; the second item is the smoothing item, which prevents unnatural jagged fluctuations of the coastline; the third item is the curvature item, which controls the curvature change of the coastline; the fourth item is the physical constraint item, which ensures that the coastline slope is within a reasonable range. This design combines the flexibility of deep learning with the constraint of physical models, and can generate coastline prediction results that not only conform to data laws but also meet physical constraints. Compared with pure data-driven methods, this method can effectively suppress overfitting and improve the physical rationality of prediction.

[0067] In step S08, the calculation process of the stability index and the change frequency matrix is specifically expressed as follows: ; ; where is the stability index, dimensionless; is the amplitude of the long-term trend, and the unit depends on the specific variable; is the standard deviation of the short-term fluctuation, and the unit is the same as ; is an element of the change frequency matrix, indicating the change intensity of the spatial position in the frequency interval ; is the energy of the -th frequency component in the spectral analysis; is the frequency of the -th frequency component; is the indicator function, which takes 1 when the condition holds and 0 otherwise; is the Dirac function, which is used for spatial positioning; is the total number of frequency components; is the position of the -th spatial sampling point.

[0068] Among them, the parameter acquisition method is: Obtained by dividing the difference between the start and end points of the long-term trend component by the time interval, specifically including Step 1: Obtain the long-term trend component separated by the WaveletTempNet model; Step 2: Calculate the change rate of the trend component over the entire time span. Obtained by calculating the standard deviation of the short-term fluctuation component, specifically including Step 1: Obtain the short-term fluctuation component separated by the WaveletTempNet model; Step 2: Calculate the standard deviation of this component. and Obtained by performing a fast Fourier transform on the short-term fluctuation component, specifically including Step 1: Apply the FFT algorithm to the short-term fluctuation component; Step 2: Calculate the energy spectrum and corresponding frequencies of each frequency component. and are the predefined frequency interval boundaries. Common interval settings include interannual variation (1 - 5 years), seasonal variation (0.25 - 1 year), tidal cycle variation (0.5 - 1 day), etc.

[0069] The calculation principle of the stability index is based on the concept of signal-to-noise ratio. Taking the long-term trend amplitude as the "signal" and the short-term fluctuation standard deviation as the "noise", the ratio of the two reflects the stability of the coastline. The higher the stability index, the more the long-term trend dominates and the more stable the coastline change; the lower the index, the more the short-term fluctuation dominates and the more unstable the coastline change. The calculation principle of the change frequency matrix is based on spectral analysis and spatial discretization. By calculating the energy distribution at different spatial positions in different frequency intervals, a two-dimensional matrix is constructed to describe the spatio-temporal change characteristics of the coastline. These indicators provide a quantitative assessment of the dynamic characteristics of the coastline, helping to identify high-risk areas and key change patterns. Compared with the traditional single-index evaluation method, this method provides a more comprehensive and detailed description of the change characteristics.

[0070] In step S09, the calculation process of the Bayesian model averaging method is specifically expressed as follows: ; ; ; where, is the weight of the th model; is the mean squared error of the th model on the validation set; is the temperature parameter, controlling the concentration of the weight distribution; is the total number of models; is the prediction result of the th model; is the integrated prediction result; is the prediction variance of the th model; It is the variance of the integrated prediction, representing the uncertainty of the prediction.

[0071] Among them, the parameter acquisition method is as follows: It is obtained by evaluating the performance of each model on the historical validation set, specifically including Step 1: dividing the historical data into a training set and a validation set; Step 2: calculating the mean square error between the predicted values and the actual values of each model on the validation set. It is determined by the cross-validation method, and the value range is usually 0.1 - 10. A larger value makes the weight distribution more concentrated on the model with the best performance. For deep learning models, it can be estimated by the Monte Carlo dropout method, specifically including Step 1: keeping the dropout activation during the prediction stage; Step 2: performing multiple forward propagations and calculating the variance of the prediction results. For physical models, it can be estimated by the parameter perturbation method, specifically including Step 1: making small random perturbations to the parameters of the physical model; Step 2: statistically analyzing the variances of the prediction results under different parameter combinations.

[0072] The design principle of the Bayesian model averaging method is based on the Bayesian statistical framework, regarding the model weights as a function of the model conditional probability. The weight calculation formula adopts a negative exponential form, enabling models with better performance (smaller errors) to obtain higher weights. The temperature parameter controls the "hardness" of weight allocation. The integrated prediction adopts a weighted average method, considering the prediction results and corresponding weights of each model. The prediction variance calculation formula consists of two parts: the first part reflects the uncertainty within each model, and the second part reflects the uncertainty between models. This design makes full use of the complementary advantages of different types of models, taking into account both the flexibility of data-driven models and the theoretical basis of physical models, thereby improving the accuracy and reliability of the prediction. Compared with the simple average or single model method, this method can provide more accurate prediction results and more reasonable uncertainty estimates.

[0073] Specifically, the principle of the present invention is as follows: The technical principle of the present invention is based on the deep integration of deep learning and coastal dynamics, and achieves a technical breakthrough by constructing a multi-level linkage coastline change monitoring framework of "data - feature - physics". At the data level, the multi-source satellite data fusion network breaks through the time resolution limitation of a single data source, effectively integrates the observation advantages of satellites with different orbital periods through the temporal attention matrix mechanism, and enhances the information richness of the original data. The fourth-order temporal convolutional neural network can efficiently capture multi-scale features in temporal data, and its recursive structure design enables the model to have the ability to learn long-distance dependencies, effectively extracting the internal laws of coastline morphological changes.

[0074] The introduction of the spatio-temporal attention mechanism and the wave transformation matrix solves the mapping problem from the high-dimensional feature space to the change index space with clear physical meaning, ensuring the efficiency and interpretability of feature extraction. The CoastDynNet model is designed based on the principles of sediment dynamics. Its multi-layer encoding and decoding architecture closely combines coastline change prediction with physical process simulation. Through the tidal influence attention module and the wave dynamics calculation unit, it realizes the understanding of the coastline evolution process from the physical mechanism. The WaveletTempNet model is based on the wavelet transform principle and can achieve multi-scale decomposition of the coastline change sequence, providing a mathematical basis for distinguishing long-term trends and short-term fluctuations.

[0075] The Bayesian model averaging method solves the multi-model integration problem from a statistical perspective. By determining the credibility weights of different models through the prediction error distribution, it realizes the complementary advantages of physical models and deep learning models. This integration strategy not only improves the prediction accuracy but also quantifies the prediction uncertainty, providing more comprehensive support for decision-making. The reason why the method of the present invention can effectively solve the technical problem of insufficient accuracy in dynamic monitoring of coastline changes lies in that it breaks the barrier between data analysis and physical simulation in traditional methods, constructs a complete knowledge chain from data to physics, enables the monitoring results to conform to both data laws and physical constraints, and realizes the organic unity of multi-source data fusion, multi-scale feature extraction, and physical mechanism analysis, thus significantly improving the accuracy of dynamic monitoring of coastline changes.

[0076] The following provides a specific embodiment 1 of the present invention. The specific implementation manners of each step in this embodiment 1 are described in detail as follows.

[0077] In the specific implementation manner of step S01, the calculation of the attention weight is achieved through the following formula: ; ; ; In the formula, is the attention weight between the image at the th time point and the image at the th time point; is the unnormalized attention score; and are the image feature vectors at the th and the th time points respectively; and are the query transformation matrix and the key transformation matrix respectively, which are obtained through model training; is the dimension of the feature vector, and its value range is usually 64 - 256; is the length of the time series; is Dimensional temporal attention matrix. When the attention weight is greater than 0.7, it indicates a significant correlation between two time points and requires key attention. The role of this step is to establish a high-quality time series feature representation by fusing multi-source remote sensing data, laying a data foundation for subsequent coastline change analysis. This mechanism can adaptively highlight the image features at key change moments, reduce the influence of noise, and effectively improve the temporal modeling ability.

[0078] The specific implementation of step S02 is to apply a fourth-order temporal convolutional neural network to extract the coastline morphological change features. First, four temporal convolutional layers are designed, and each layer uses different-sized temporal convolutional kernels (3, 5, 7, 9 respectively) to capture multi-scale temporal features. Then, the convolutional results are processed by a recurrent neural network, specifically using a long short-term memory network structure, with the hidden layer dimension set to 256 and the time step set to the length of the image sequence. The output of the recurrent network is mapped through a fully connected layer to generate the first fluctuation matrix F containing multi-dimensional indicators such as position offset, area change rate, and morphological complexity. The calculation formulas for these indicators are as follows: ; ; ; In the formula, is the position offset at time and are the x coordinate and y coordinate of the th sampling point at time is the number of coastline sampling points; is the area change rate at time is the land-sea boundary area at time is the morphological complexity at time is the coastline length at time

[0079] When the position offset exceeds 10 m / year or the area change rate exceeds 5% / year, it is determined as a significant change area. The purpose of this step is to extract the primary feature representation of the coastline morphological change from the original image and capture the key change information. These three indicators together constitute the multi-dimensional feature representation of the coastline morphological change, which can depict the dynamic change characteristics of the coastline more comprehensively than traditional single indicators.

[0079] The specific implementation of step S03 is the same as the foregoing and will not be elaborated in detail here.

[0080] The specific implementation of step S04 is to establish a fluctuation transformation matrix, mapping the abstract feature space to the coastline change index space with clear physical meanings. First, a three-layer fully connected neural network is designed. The number of nodes in the input layer is equal to the dimension of the second fluctuation matrix, the number of nodes in the hidden layer is 128, and the number of nodes in the output layer is equal to the number of target change indices. The transformation process is achieved through the following formula: ; ; In the formula, is the fluctuation transformation matrix; , , are the weight matrices of the three layers of the neural network respectively; , , are the bias vectors of the three layers of the neural network respectively; is the rectified linear unit activation function; is the second fluctuation matrix; is the quantitative index vector of the coastline change after transformation, including multiple indices such as the annual average retreat rate, seasonal fluctuation amplitude, and extreme event response intensity. Then, the known ground measured data is used as the supervision information, and the network parameters are optimized through the least squares method to learn the mapping relationship from the feature space to the change index space. The finally obtained fluctuation transformation matrix T can convert the second fluctuation matrix G into a series of quantitative coastline change indices. When the root mean square error between the predicted value and the measured value is less than 0.15, it is considered that the transformation matrix reaches the acceptable accuracy. The purpose of this step is to convert the abstract features extracted by the deep learning model into change indices with actual physical meanings, facilitating the understanding of professionals and decision-making support. The design of the fluctuation transformation matrix establishes a mapping relationship from the abstract feature space to the index space with clear physical meanings through a deep neural network, improving the interpretability and practicability of the model.

[0081] The specific implementation of step S05 is to input the quantitative indices into the pre-trained coastal dynamics model CoastDynNet, and analyze the physical mechanism of the coastline change through the sediment transport equation. The sediment transport calculation formula is as follows: ; ; ; In the formula, is the longshore sediment transport rate, with the unit of ; is the vertical sediment transport rate, with the unit of ; , , are empirical coefficients, ranges from 0.1 to 0.4, The value range of is from 0.01 to 0.05, the value range of is from 0.1 to 0.3; is the wave height in the wave breaking zone, with the unit of ; is the wave incident angle in the wave breaking zone, with the unit of radian; is the beach slope, dimensionless; is the longshore gradient of wave height; is the change rate of coastline position, with the unit of ; is the critical water depth of wave action, with the unit of ; is the longshore distance coordinate, with the unit of ; is the distance coordinate perpendicular to the reference line, with the unit of ; is the time, with the unit of . First, input the change index obtained in step S04 and auxiliary data (such as wave parameters, tidal data, meteorological data) into the CoastDynNet model together. Then, the sediment dynamics physical constraint layer inside the model calculates the longshore sediment transport rate and the vertical sediment transport rate . When the ratio is greater than 3, it indicates that the longshore sediment transport is the dominant process. Then, the wave dynamics calculation unit evaluates the wave erosion ability according to the ratio of the average wave height to the critical erosion wave height. When this ratio is greater than 1.2, the coastline is in an erosion state. Finally, the morphological evolution prediction layer integrates the outputs of each module to generate a physical explanation report of the coastline change. The function of this step is to combine data-driven statistical analysis with a physical mechanism model to reveal the essential reasons behind the coastline change and improve the interpretability of the monitoring results. The sediment transport equation is based on classical coastal dynamics theory, combines the CERC formula and the vertical sediment transport model, and can quantitatively describe the dynamic mechanism of coastline evolution, providing a physical explanation for the coastline change.

[0082] The specific implementation method of step S06 is to apply the coastline evolution optimization function to finely adjust the time interpolation result. The mathematical expression of the optimization function is: ; ; ; In the formula, is the coastline evolution optimization function; is the optimized coastline position vector; is the coastline position vector preliminarily predicted by the model; is the spatial gradient of the coastline position; , , , are weight coefficients, ranges from 0.5 to 2.0, ranges from 0.1 to 0.5, ranges from 0.05 to 0.2, ranges from 1.0 to 5.0; is a physical constraint term; and are the maximum and minimum allowable slopes of the coastline respectively, usually ranges from 30° to 45°, usually ranges from 0° to 5°; is the final coastline position vector after optimization; is the number of discrete sampling points of the coastline; is the coastline position of the

[0083] The specific implementation of step S07 is the same as the foregoing and will not be elaborated herein.

[0084] The specific implementation of step S08 is to design a spectrum analysis module to further process the separated time series. The calculation of the stability index and the change frequency matrix is as follows: ; ; In the formula, is the stability index, dimensionless; is the amplitude of the long-term trend, and the unit depends on the specific variable; is the standard deviation of the short-term fluctuation, and the unit is the same as ; represents an element of the change frequency matrix, indicating the spatial position in the frequency interval the change intensity; is the energy of the th frequency component in the spectral analysis; is the th frequency of the th frequency component; is the indicator function, taking 1 when the condition holds and 0 otherwise; is the Dirac function for spatial positioning; is the th position of the spatial sampling point. First, apply the fast Fourier transform to the short-term fluctuation component to obtain the frequency-domain representation and identify the main periodic components. Frequencies with an energy proportion exceeding 15% are identified as significant periods. Then, calculate the ratio of the long-term trend amplitude to the standard deviation of the short-term fluctuations as the stability index SI. When SI is greater than 2, it indicates that the coastline in this section is relatively stable, and when SI is less than 0.5, it indicates instability. Next, construct the change frequency matrix, where the rows represent the spatial sampling points and the columns represent predefined frequency intervals (such as interannual changes, seasonal changes, tidal cycle changes, etc.). The matrix element values represent the change intensity at the corresponding position in a certain frequency interval. Finally, based on the spectral analysis results, identify potential periodic change patterns and abnormal change events. The purpose of this step is to quantitatively evaluate the dynamic characteristics of the coastline and provide a scientific basis for management decisions. The calculation of the stability index is based on the concept of signal-to-noise ratio, taking the long-term trend amplitude as the "signal" and the standard deviation of the short-term fluctuations as the "noise", and the ratio of the two reflects the stability of the coastline. The change frequency matrix constructs a two-dimensional matrix to describe the spatio-temporal change characteristics of the coastline by calculating the energy distribution at different spatial positions in different frequency intervals.

[0085] The specific implementation of step S09 is to construct a multi-model integration framework using the Bayesian model averaging method. The mathematical expression of the integration process is: ; ; ; where, is the weight of the th model; is the mean squared error of the th model on the validation set; is the temperature parameter that controls the concentration of the weight distribution, and its value range is usually 0.1 - 10; is the total number of models; is the th prediction result of the th model; is the integrated prediction result; is the prediction variance of the th model; is the variance of the integrated prediction, representing the uncertainty of the prediction. First, collect the prediction results of physical models (such as CoastDynNet) and deep learning models (such as WaveletTempNet) on the historical validation set. Then, based on the principle of Bayesian inference, calculate the prediction error distribution of each model. When the error follows a normal distribution, use the mean square error; when it follows a long-tailed distribution, use the mean absolute percentage error. Next, determine the credibility weights of each model according to the error distribution. The weight calculation formula is an exponential function of the model accuracy, ensuring that models with higher accuracy obtain greater weights. Finally, perform a weighted average of the outputs of each model to generate a comprehensive coastline dynamic assessment report, including change trend prediction, uncertainty estimation, and risk level assessment. When the prediction difference between models exceeds 20%, expert review is required. The role of this step is to make full use of the complementary advantages of different types of models to improve the accuracy and reliability of the prediction, while quantifying the uncertainty of the prediction results. The Bayesian model averaging method regards the model weights as a function of the model conditional probability, and uses a negative exponential form of the weight calculation formula to make models with better performance obtain higher weights, making full use of the complementary advantages of different types of models.

[0086] The specific structure of the CoastDynNet model is a multi-layer encoding and decoding architecture, which consists of four core modules. The first module is the sediment dynamics physical constraint layer, which uses an improved CERC formula to calculate the longshore sediment transport rate and establishes a constraint relationship by combining the vertical sediment transport rate. Its parameters are determined by the ratio of the longshore sediment transport rate to the vertical sediment transport rate. When the ratio is greater than 3, a strong longshore sediment transport parameter configuration is adopted. When the ratio is less than 0.5, a strong vertical sediment transport parameter configuration is adopted. The second module is the tidal influence attention module, which uses a gated recurrent unit network to process tidal time series data. Its parameters are determined by the ratio of the tidal period to the coastline response time. When the ratio is less than 0.1, the tidal influence can be ignored. The third module is the wave dynamics calculation unit, which calculates the wave erosion effect based on an improved Battjes-Janssen wave energy dissipation model. Its parameters are determined by the ratio of the mean wave height to the critical erosion wave height. When the ratio is greater than 1.5, a strong erosion mode is activated. The fourth module is the morphological evolution prediction layer, which integrates the outputs of the first three modules and applies an improved one-dimensional ecological cell automata algorithm to predict the morphological evolution of the coastline. The steps for establishing the pre-training dataset of the CoastDynNet model include: First, collect multi-year satellite observation data of different types of coastlines (including sandy coasts, rocky coasts, muddy coasts, etc.) around the world, with a time span of no less than 10 years; then extract the hydrometeorological and ocean dynamics parameters of the corresponding regions, including wave spectra, tidal curves, wind field data, sediment particle size distributions, etc.; then construct physical simulation data of the combined action of tides and waves, with a simulation time step of no more than 1 hour; then mark the occurrence periods of extreme events such as typhoons and storm surges, and extract the morphological change characteristics of the coast before and after; finally, classify and organize according to different coast types and dynamic conditions to form a training sample library with no less than 1000 samples. The pre-training steps of the CoastDynNet model include: First, pre-train on the physical simulation data to learn the basic sediment dynamics laws, with no less than 100 pre-training rounds; then fine-tune on the global satellite observation data to learn the actual coast response patterns, with no less than 50 fine-tuning rounds; then perform transfer learning for coast types with similar or close regions to optimize local adaptability, with no less than 30 transfer learning rounds; finally, by comparing the morphological differences between the real observation sequence and the model prediction sequence, use the root mean square error as the loss function to perform error backpropagation to optimize the model parameters, with no less than 20 optimization rounds.

[0087] The specific structure of the WaveletTempNet model is an end-to-end deep learning network composed of four key layers. The first layer is the multi-scale wavelet transform encoding layer, which uses the discrete wavelet transform principle to perform multi-scale decomposition on the input time series. The wavelet basis function can choose the Haar, Daubechies or Symlet series, and its parameters are determined by the coastline change cycle. When the cycle is less than 1 month, the Haar wavelet is selected, when the cycle is 1 to 12 months, the Daubechies wavelet is selected, and when the cycle is greater than 12 months, the Symlet wavelet is selected. The second layer is the time series feature extraction layer, which uses a multi-head self-attention mechanism to capture the time dependency. The number of attention heads is determined by the length of the input sequence, which is generally set to 1 / 4 to 1 / 8 of the sequence length. The third layer is the long short-term memory network layer, which uses a bidirectional LSTM structure to process sequence information. Its hidden state dimension is determined by the feature complexity, with a simple change mode set to 64 and a complex change mode set to 256. The fourth layer is the reconstruction decoding layer, which recombines the processed features into a time series through an inverse wavelet transform and separates the long-term trend and short-term fluctuation components. The steps for establishing the pre-training dataset of the WaveletTempNet model include: first, collecting the change sequences of typical coastlines around the world at different time scales, including daily, monthly, annual and decadal scale data; then, standardizing and preprocessing the sequences, and using the Z-score method to eliminate the dimension effect; then, experts manually annotate the change components at different scales as training targets; then, construct a mixed sequence containing long-term trends and short-term fluctuations as input, with a mixing ratio ranging from 0.1 to 10; finally, stratified sampling is performed according to different change patterns and frequency characteristics to form training sets and validation sets, with a ratio of 8:2 between the training set and the validation set. The pre-training steps of the WaveletTempNet model include: first, initial training on synthetic data to learn basic decomposition capabilities, with no less than 200 initial training rounds; then, supervised learning on real coastline sequences to improve generalization performance, with no less than 100 supervised learning rounds; then, by comparing the consistency of the outputs of different decomposition levels with the theoretical decomposition results, parameter optimization is performed using the root mean square error and structural similarity index as loss functions, with no less than 50 rounds of parameter optimization; finally, reconstruction accuracy verification is performed to ensure that the information of the decomposition and reconstruction processes is losslessly retained, and the training is considered to have met the standard when the reconstruction error is less than 3% of the standard deviation of the original signal.

[0088] It should be noted that in step S01, the multi-head self-attention mechanism measures the correlation between image features at different time points by calculating the inner product of them and divides them by Scale it to avoid the problem of gradient vanishing. The exponential function and normalization operation transform the attention scores into a probability distribution, ensuring that the sum of all weights is 1. This mechanism can adaptively highlight the image features at key change moments, effectively improving the temporal modeling ability. Compared with the traditional linear weighting method, this mechanism can capture non-linear time dependencies and is more robust to outliers.

[0089] It should be noted that in step S02, the position offset calculates the displacement of the corresponding sampling points at adjacent time points based on the Euclidean distance; the area change rate reflects the relative change of the sea-land interface area, and using the percentage form facilitates comparison between different regions; the morphological complexity is based on the fractal dimension theory and is calculated by the logarithmic ratio of the coastline length to the area, which can quantitatively describe the geometric complexity of the coastline. These three indicators together constitute a multi-dimensional feature representation of the coastline morphological changes, which can depict the dynamic change characteristics of the coastline more comprehensively than the traditional single indicator.

[0090] It should be noted that in step S03, the spatio-temporal attention mechanism decomposes the attention calculation into two independent calculation processes in the time dimension and the space dimension, and then integrates the information of the two dimensions through weighted fusion. The attention in the time dimension focuses on the dependencies between different time points, and the attention in the space dimension focuses on the correlations between different spatial positions. This decomposition design enables the model to capture the key information in both the time and space dimensions simultaneously, significantly improving the feature extraction ability. The weighting coefficient is determined by automatic optimization and can flexibly adjust the importance of the two dimensions according to the specific data characteristics. Compared with the traditional single-dimensional attention mechanism, this method can more effectively handle the complex dependencies of spatio-temporal data.

[0091] It should be noted that in step S04, the design of the wave transformation matrix establishes a mapping relationship from the abstract feature space to the index space with clear physical meaning through a deep neural network. The three-layer neural network structure can effectively fit complex non-linear mapping relationships, the activation function introduces non-linear transformation ability and solves the problem of gradient vanishing. The least squares loss function measures the difference between the predicted index and the measured index, and the regularization term prevents overfitting. This design transforms the abstract features extracted by deep learning into change indexes with actual physical meaning, improving the interpretability and practicality of the model. Compared with the traditional linear transformation method, this method can capture the more complex relationships between features and indexes.

[0092] It should be noted that in step S05, the longshore sediment transport rate is proportional to the 2.5th power of the wave height and the sine value of the incident angle, reflecting the influence of wave energy and incident angle on longshore sediment transport; the vertical sediment transport rate It consists of two parts. The first term represents the onshore sediment transport caused by waves, which is proportional to the square of the wave height and the slope. The second term represents the offshore sediment transport caused by the wave height gradient. The equation for the rate of change of the coastline position is based on the principle of sediment mass conservation and calculates the rate of change of the coastline position through the spatial gradient of the longshore sediment transport rate and the vertical sediment transport rate. This physical model can quantitatively describe the dynamic mechanism of coastline evolution and provide a physical explanation for coastline changes. Compared with pure data-driven methods, this physical model has better interpretability and extrapolation ability.

[0093] It should be noted that in step S06, the design principle of the coastline evolution optimization function is to introduce physical constraint conditions on the basis of data-driven prediction, and transform the constrained problem into an unconstrained optimization problem through the Lagrange multiplier method. The objective function consists of four parts: the first term is the data term, which ensures that the optimization result is close to the preliminary prediction; the second term is the smoothing term, which prevents unnatural sawtooth fluctuations of the coastline; the third term is the curvature term, which controls the curvature change of the coastline; the fourth term is the physical constraint term, which ensures that the coastline slope is within a reasonable range. This design combines the flexibility of deep learning with the constraint of the physical model, and can generate coastline prediction results that conform to both data laws and physical constraints. Compared with pure data-driven methods, this method can effectively suppress overfitting and improve the physical rationality of prediction.

[0094] It should be noted that in step S07, the calculation principle of the stability index is based on the concept of signal-to-noise ratio. The long-term trend amplitude is regarded as the "signal", and the standard deviation of short-term fluctuations is regarded as the "noise". The ratio of the two reflects the stability of the coastline. The higher the stability index, the more dominant the long-term trend is and the more stable the coastline change is; the lower the index, the more dominant the short-term fluctuation is and the more unstable the coastline change is. The calculation principle of the change frequency matrix is based on spectral analysis and spatial discretization. By calculating the energy distribution in different frequency intervals at different spatial positions, a two-dimensional matrix is constructed to describe the spatio-temporal change characteristics of the coastline. These indicators provide a quantitative assessment of the dynamic characteristics of the coastline and can help identify high-risk areas and key change patterns. Compared with traditional single-index assessment methods, this method provides a more comprehensive and detailed description of change characteristics.

[0095] It should be noted that in step S08, the design principle of the Bayesian model averaging method is based on the Bayesian statistical framework, and the model weights are regarded as a function of the model conditional probability. The weight calculation formula adopts a negative exponential form, so that models with better performance (smaller errors) obtain higher weights, and the temperature parameter Control the "hardness" of weight distribution. The integrated prediction uses the weighted average method, taking into account the prediction results and corresponding weights of each model. The prediction variance calculation formula consists of two parts: the first part reflects the uncertainty within each model, and the second part reflects the uncertainty between models. This design makes full use of the complementary advantages of different types of models, considering both the flexibility of data-driven models and the theoretical basis of physical models, thus improving the accuracy and reliability of the prediction. Compared with the simple average or single model method, this method can provide more accurate prediction results and more reasonable uncertainty estimates.

[0096] Optionally, for the specific implementation of the CoastDynNet model, a sediment dynamics physical constraint layer can also be considered, and its expression is as follows: ; In the formula, is the longshore sediment transport rate, with the unit of ; is the sediment transport coefficient, and its value range is 0.1 - 0.4; is the sediment grain size correction coefficient, and its value range is 0.01 - 0.05; is the median sediment grain size, with the unit of ; is the reference grain size, with a value of 0.2 ; is the wave height in the wave breaking zone, with the unit of ; is the wave incidence angle in the wave breaking zone, with the unit of radians; is the tidal level influence coefficient, and its value range is 0.5 - 2.0; is the relative tidal level, dimensionless.

[0097] Optionally, the vertical sediment transport model adopts an improved formula considering the wave-tide interaction: ; In the formula, is the vertical sediment transport rate, with the unit of ; , , are empirical coefficients, with values of 0.01 - 0.05, 0.1 - 0.3, and 0.05 - 0.2 respectively; is the wave-current interaction coefficient, and its value range is 0.5 - 1.5; is the Froude number, representing the strength of the tidal current, dimensionless; is the beach slope, dimensionless; is the wave height longshore gradient; is the tidal current velocity, with the unit of ; is the angle between the tidal current and the coastline, in radians.

[0098] Optionally, the tidal influence attention module of the CoastDynNet model adopts a time-varying weight calculation formula: ; In the formula, is the tidal influence weight at time is the sigmoid activation function; is the weight matrix, obtained through model training; is the hidden state of the gated recurrent unit; is the tidal phase, ranging from 0 to 2 ; is the tidal amplitude, in units of ; is the bias vector.

[0099] Optionally, the morphological evolution prediction layer of the CoastDynNet model adopts an improved one-dimensional ecological cellular automata algorithm, and its state update rule is: ; In the formula, is the coastline state at spatial position at time ; is the neighborhood radius, usually taking values from 1 to 3; is the state transition function, implemented through a neural network; is the external forcing condition, including factors such as waves and wind fields; and are the longshore sediment transport rate and the vertical sediment transport rate at position at time respectively.

[0100] Optionally, for the specific implementation of the WaveletTempNet model, the mathematical expression of its multi-scale wavelet transform coding layer is: ; ; In the formula, is the wavelet coefficient; is the input time series; is the wavelet basis function; is the scale parameter; is the translation parameter; denotes the complex conjugate of. In actual implementation, the fast algorithm of discrete wavelet transform is used to calculate the wavelet coefficient.

[0101] Optionally, the time series feature extraction layer of the WaveletTempNet model uses a multi-head self-attention mechanism enhanced with positional encoding: ; ; where, , , are the query matrix, key matrix, and value matrix, respectively; is the feature dimension; is the positional encoding matrix; and are the position indices in the sequence; is the scaling factor of the positional encoding, usually set to 0.1 times the sequence length.

[0102] Optionally, the long short-term memory network layer of the WaveletTempNet model adopts a bidirectional LSTM structure with residual connections: ; ; ; where, and are the hidden states of the forward and backward LSTMs, respectively; and are the forward and backward LSTM cells, respectively; is the input at time and are the weight matrices; is the bias vector; represents the vector concatenation operation.

[0103] Optionally, the reconstruction decoding layer of the WaveletTempNet model uses a weighted reconstruction formula: ; where, is the reconstructed time series; is the long-term trend component; is the processed wavelet coefficient; is the wavelet basis function; is the reconstruction weight at the scale, obtained through model training;

[0104] Optionally, the loss function of the WaveletTempNet includes a reconstruction error term, a decomposition consistency term, and a smoothing regularization term: ; In the formula, is the total loss; is the original input sequence; is the reconstructed sequence; and are the wavelet coefficients of the scale of the theoretical wavelet decomposition and model prediction respectively; is the long-term trend value at time , , are weight coefficients, taking values of 1.0, 0.5 and 0.1 respectively; is the sequence length.

[0105] To sum up, in this embodiment, starting from multi-source data fusion, through multi-level feature extraction and transformation, combining physical models and deep learning methods, the high-precision dynamic monitoring of coastline changes is finally realized. In particular, the multi-source satellite data fusion network in S01 and the WaveletTempNet model in S07 effectively solve the technical problems of insufficient temporal resolution caused by the long revisit period of traditional satellites and the difficulty in distinguishing long-term trends from short-term fluctuations, and significantly improve the accuracy and timeliness of coastline change monitoring. To better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: Researchers selected a coastal area to conduct dynamic monitoring of coastline changes. This area is significantly affected by the monsoon climate and is frequently hit by typhoons from June to October every year. It is an ideal area to study the impact of extreme meteorological events on the coastline.

[0106] First, a multi-source satellite data fusion network was constructed. Satellite image data from 2010 to 2022 was collected, including observation data from multiple satellites such as Landsat 8 (16-day revisit period), Sentinel-2 (5-day revisit period), and GF-1 (4-day revisit period), totaling 457 images. After radiometric correction and geometric correction of these images, a temporal attention matrix was constructed using the multi-head self-attention mechanism. The specific satellite data sources and their characteristics are shown in Table 1: Table 1 Multi-source satellite data characteristics table

[0107] By using a fourth-order temporal convolutional neural network to extract the morphological change features of the coastline, the researchers set up four convolutional layers. The temporal convolutional kernel sizes of the first to fourth layers were 3, 5, 7, and 9 respectively, and the number of feature channels were 32, 64, 128, and 256 respectively. Subsequently, the convolutional results were processed by a long short-term memory network with the hidden layer dimension set to 256, generating the first fluctuation matrix containing indicators such as position offset, area change rate, and morphological complexity. The calculation results showed that the change indicators of the coastline before and after the "Goni" typhoon in the study area in 2020 were as shown in Table 2: Table 2 Change indicators of the coastline before and after the "Goni" typhoon

[0108] Figure 2 It shows the change in the coastline position offset of five coastal segments (E01 - E05) in the eastern part of Hainan Island before and after the "Goni" typhoon event from September to December 2020. The horizontal axis in the figure represents time (date), and the vertical axis represents the coastline position offset (meters). The chart uses curves to show the dynamic change process of each coastal segment, where negative values indicate that the coastline retreats towards the land. The observation dates of three different satellites, Landsat 8, Sentinel-2, and GF-1, are also marked in the figure, intuitively showing the improvement in time resolution brought by the fusion of multi-source satellite data. The "Goni" typhoon event is marked with a vertical line, clearly presenting the sharp retreat of the coastline caused by the typhoon and the subsequent recovery process. It can be seen from the figure that the E04 segment was most significantly affected by the typhoon, with the position offset reaching approximately -30 meters, while the E03 segment was relatively less affected.

[0109] Using the spatio-temporal attention mechanism to perform dimensionality reduction on the first fluctuation matrix, 8 heads of temporal dimension attention and 4 heads of spatial dimension attention were adopted, and the weight coefficient was automatically optimized to 0.65 through backpropagation, indicating that the temporal dimension features had a more significant impact on the coastline changes in this study area. After principal component analysis, the first 5 principal components with an explained variance ratio reaching 96.3% were retained, forming the second fluctuation matrix with low dimension and high information density.

[0110] When constructing the fluctuation transformation matrix, a three-layer fully connected neural network was designed. The number of input layer nodes was 5 (corresponding to the dimension of the second fluctuation matrix), the number of hidden layer nodes was 128, and the number of output layer nodes was 6, corresponding to 6 indicators such as the annual average retreat rate, seasonal fluctuation amplitude, and extreme event response intensity. By comparing with the data of 25 field measurement points, the optimized transformation matrix achieved a root mean square error of 0.13 on the validation set, meeting the accuracy requirements. The quantified indicators of the coastline changes after transformation are shown in Table 3: Table 3 Quantified indicators of coastline changes (2010 - 2022)

[0111] The quantified indicators and wave and tide data are input into the pre-trained coastal dynamics model CoastDynNet, and the physical mechanism of coastline change is analyzed through the sediment transport equation. In the calculation of the longshore sediment transport rate, takes a value of 0.35, and the measured wave height in the wave breaking zone has an average of 1.2 m, and the wave incident angle in the wave breaking zone is mainly concentrated in the range of 15° - 35°. In the calculation of the vertical sediment transport rate, takes a value of 0.03, takes a value of 0.21, and the beach slope has an average of 0.08. The analysis results show that the ratios of the longshore sediment transport rate to the vertical sediment transport rate in sections E02 and E04 are 3.6 and 4.2 respectively, indicating that these two sections of the coast are mainly controlled by the longshore sediment transport process and corresponding to the areas with higher erosion risks.

[0112] The prediction results are refined by applying the coastline evolution optimization function. The weight coefficients 、 、 、 in the optimization function are set to 1.5, 0.3, 0.1, and 2.0 respectively. Through iterative optimization using the gradient descent algorithm, it converges after 78 iterations, and the corrected coastline morphology that meets the physical constraint conditions is obtained. For the area of section E04 that is significantly affected by tides, a tidal level correction coefficient of 1.15 is additionally set, further improving the prediction accuracy.

[0113] The WaveletTempNet is used to perform multi-scale decomposition on the coastline change sequence. The wavelet transform coding layer of this model uses the Daubechies 4 wavelet basis function, the time series feature extraction layer uses a 12-head self-attention mechanism, and the hidden state dimension of the long short-term memory network layer is set to 192. The decomposed results show that the long-term trend of the coastline in section E04 from 2010 to 2022 shows a retreating feature, with an average retreat rate of 2.37 m / year, while the short-term fluctuations are mainly concentrated in the typhoon season, and the maximum fluctuation amplitude can reach 20.5 m.

[0114] Figure 3 Shows the multi-scale decomposition results of the coastline change in section E04, including four components: total change, long-term trend, seasonal fluctuation, and extreme event response, as well as the change in the stability index. Four major typhoon events and their response intensities are marked in the figure, clearly presenting the retreat and recovery processes of the coastline after each extreme event. Figure 4 The results of the spectral analysis are shown, presenting the spectral energy distribution and main change frequencies of five coastal sections. The horizontal axis represents the change frequency (times / year), and the vertical axis represents the spectral energy ( ). The main frequency and stability index of each coastal segment are marked in the figure, and the monsoon-dominated interval (period > 6 months) and typhoon-dominated interval (period < 4 months) are divided. It can be seen from the figure that the E04 segment has the highest main frequency (4.6 times / year) and the lowest stability index (0.47), indicating that this segment of the coast is significantly affected by typhoons and tides and has the most drastic changes. The stability indices and main change frequencies of the five coastal segments calculated by the spectral analysis module are shown in Table 4: Table 4 Results of Stability Analysis of Coastal Segments

[0115] Finally, a multi-model integration framework is constructed using the Bayesian model averaging method, integrating the outputs of the physical model (CoastDynNet) and the deep learning model (WaveletTempNet). The temperature parameter is set to 2.5. According to the performance of each model on the historical validation set, the weights of the physical model and the deep learning model are 0.42 and 0.58 respectively. The root mean square error of the prediction results of the integrated model is 0.78 m, which is 37.6% and 45.5% lower than the errors of using the deep learning model alone (1.25 m) and the physical model (1.43 m) respectively.

[0116] Traditional coastline change monitoring methods mainly rely on a single satellite data source, such as the Landsat series of satellites. Its 16-day revisit cycle cannot capture short-term changes, especially sudden changes caused by extreme events such as typhoons. At the same time, traditional methods often use simple time series analysis techniques and are difficult to effectively distinguish long-term trends and short-term fluctuations. The multi-source satellite data fusion network of the present invention significantly improves the time resolution, achieving an effective observation interval of an average of 3.5 days in the embodiment and successfully capturing the impact of extreme events such as the "Goni" typhoon on the coastline. At the same time, the WaveletTempNet model realizes the multi-scale decomposition of the coastline change sequence, accurately separating the long-term trend and short-term fluctuations. The stability index and change frequency matrix provide a scientific basis for coastal erosion risk assessment. The multi-model integration framework constructed by the Bayesian model averaging method organically combines the physical model and the data-driven model, and the prediction accuracy is improved by more than 40% compared with the traditional single-model method, providing more reliable technical support for coastal zone planning and management.

[0117] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 5, 6, and 7 below.

[0118] Table 5 Variable Explanation Table (Part 1)

[0119] Table 6 Variable Explanation Table (Part 2)

[0120] Table 7 Variable Explanation Table (Part III)

[0121] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A deep learning-driven method for dynamic monitoring of coastline changes, characterized in that: include: Construct a multi-source satellite data fusion network, combine satellites with different orbital periods to obtain time series images, and form a time series attention matrix; A fourth-order time-series convolutional neural network is used to extract the characteristics of coastline morphological changes, and the first fluctuation matrix is ​​established recursively. The first fluctuation matrix is ​​reduced in dimension using the spatiotemporal attention mechanism to generate the second fluctuation matrix and convert it into a quantitative coastline change index. The matrix is ​​input into the pre-trained coastal dynamics model CoastDynNet, and the sediment transport equation is used to analyze the physical mechanism of coastline change. The coastline evolution optimization function is used to fine-tune the time interpolation results. The pre-trained time series decomposition network WaveletTempNet model is used to perform multi-scale decomposition of the coastline change sequence. A spectrum analysis module is designed to calculate the coastline stability index and change frequency matrix based on the separation results. A multi-model integration framework is constructed using the Bayesian model averaging method.

2. The deep learning-driven coastline change dynamic monitoring method according to claim 1 is characterized in that: The temporal attention matrix is ​​a two-dimensional feature representation matrix formed by weighted fusion of satellite image data at different time points through a multi-head attention mechanism, and is used to capture key change points in the time series.

3. The deep learning-driven coastline change dynamic monitoring method according to claim 2 is characterized in that: The first fluctuation matrix is ​​a primary feature representation of coastline morphological changes extracted from original satellite images, and includes a feature matrix composed of multi-dimensional indicators such as position offset, area change rate, and morphological complexity.

4. The deep learning-driven coastline change dynamic monitoring method according to claim 3 is characterized in that: The second fluctuation matrix is ​​a low-dimensional, high-information-density feature representation formed after dimensionality reduction processing, which removes redundant information and noise and retains the core change pattern.

5. The deep learning-driven coastline change dynamic monitoring method according to claim 4 is characterized in that: The sediment transport equation is a physical mechanism equation that describes the movement of sediments around the coastline. The input includes wave energy density, incident angle, sediment particle size distribution, tidal amplitude, and coastal current velocity. The output is the net migration amount of sediment per unit time and the rate of change of coastline position.

6. The deep learning driven coastline change dynamic monitoring method according to claim 5 is characterized in that: The coastline evolution optimization function is used to constrain optimization of deep learning prediction results based on physical constraints. The input includes initial coastline morphology, predicted coastline morphology, terrain slope, sediment characteristics, and extreme event intensity. The output is a revised coastline morphology that meets the physical constraints.

7. The deep learning driven coastline change dynamic monitoring method according to claim 6 is characterized in that: The structure of the CoastDynNet model is a multi-layer encoding and decoding architecture, which includes a sedimentation dynamics physical constraint layer, a tidal impact attention module, a wave dynamics calculation unit and a morphological evolution prediction layer; the parameters of the sedimentation dynamics physical constraint layer in the CoastDynNet model are determined by the ratio of the coastal sediment transport rate to the vertical sediment transport rate, the parameters of the tidal impact attention module are determined by the ratio of the tidal cycle to the coastline response time, and the parameters of the wave dynamics calculation unit are determined by the ratio of the average wave height to the critical erosion wave height.

8. The deep learning driven coastline change dynamic monitoring method according to claim 7 is characterized in that: The structure of the WaveletTempNet model is an end-to-end deep learning network consisting of a multi-scale wavelet transform encoding layer, a time series feature extraction layer, a long short-term memory network layer and a reconstruction decoding layer; the wavelet basis function parameters of the wavelet transform encoding layer in the WaveletTempNet model are determined by the coastline change period, the number of attention heads of the time series feature extraction layer is determined by the input sequence length, and the hidden state dimension of the long short-term memory network layer is determined by the feature complexity.

9. The deep learning driven coastline change dynamic monitoring method according to claim 8, characterized in that: The stability index is an indicator that quantifies the coastline's ability to resist external interference, and is calculated through the ratio of long-term change trends to short-term fluctuation amplitudes; the change frequency matrix is ​​a two-dimensional matrix that records the change frequency characteristics of different sections of the coastline. The rows represent spatial positions, the columns represent change frequency intervals, and the matrix element values ​​represent the change intensity of the corresponding positions in a certain frequency interval.

10. The deep learning driven coastline change dynamic monitoring method according to claim 9, characterized in that: The Bayesian model averaging method is a statistical method based on the principle of Bayesian reasoning to perform weighted averaging on the prediction results of multiple models. By calculating the prediction error distribution of each model on the historical validation set, its credibility weight in the integrated framework is determined to achieve model advantage complementarity and prediction uncertainty quantification.

Images