A deep learning-driven method for dynamic monitoring of coastline changes

Abstract

This paper provides a deep learning-driven method for monitoring coastline dynamics. This method belongs to the field of image understanding and recognition technology. It constructs a multi-source satellite data fusion network to form a temporal attention matrix. A fourth-order temporal convolutional neural network is used to extract coastline morphological characteristics and construct a first fluctuation matrix. A spatiotemporal attention mechanism is used to perform dimensionality reduction to generate a second fluctuation matrix, which is then converted into quantitative indicators using a fluctuation conversion matrix. The method inputs data into a pretrained CoastDynNet model to analyze physical mechanisms. A coastline evolution optimization function is applied for fine-tuning. A WaveletTempNet network is used for multi-scale decomposition to separate long-term and short-term variations. Finally, a Bayesian model averaging method is used to construct a multi-model integration framework to generate a comprehensive coastline dynamics assessment report.

Description

Technical Field

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

[0002] As the boundary between land and sea, monitoring the dynamic changes of the coastline is crucial for coastal management, disaster prevention and mitigation, and ecological protection. Traditional coastline monitoring relies primarily on single-source satellite remote sensing data. Coastline data is extracted through image segmentation and feature extraction, followed by multi-temporal comparisons to analyze its dynamics. While these methods can achieve basic monitoring under ideal conditions, their accuracy is limited in complex environments. Some studies have attempted to improve predictive capabilities by combining numerical simulation methods with physically driven coastline evolution models, such as GENESIS and Delft3D.

[0003] However, traditional technologies have significant drawbacks: First, the limited temporal resolution of single-source satellite data makes it difficult to capture short-term fluctuations; second, purely image processing methods struggle to accurately distinguish natural fluctuations from long-term trends; and third, purely physical models are computationally complex and parameter-sensitive, making them difficult to rapidly apply to large-scale monitoring. Furthermore, existing methods often divorce data-driven approaches from physical mechanisms, lacking effective fusion mechanisms and resulting in insufficient model accuracy.

[0004] Currently, dynamic monitoring of coastline change urgently faces the technical challenges of efficiently integrating multi-source data, accurately decomposing long- and short-term variations, and co-optimizing physical constraints and data-driven approaches. Especially in the context of climate change and increasing human activity, developing a comprehensive coastline change monitoring method that combines physical rationality with predictive accuracy has become a common challenge faced by both academics and engineers. New technical solutions that can improve the accuracy of dynamic monitoring of coastline change are urgently needed. In other words, existing technologies for dynamic monitoring of coastline change suffer from insufficient accuracy. Summary of the Invention

[0005] In view of this, the present invention provides a deep learning-driven dynamic monitoring method for coastline changes, 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 deep learning-driven dynamic monitoring method for coastline changes, including: constructing a multi-source satellite data fusion network, combining satellites with different orbital periods to obtain time series images, and forming a time series attention matrix; applying a fourth-order time series convolutional neural network to extract coastline morphological change characteristics, and recursively establishing a first fluctuation matrix; using a spatiotemporal attention mechanism to reduce the dimension of the first fluctuation matrix, generate a second fluctuation matrix and convert it into a quantitative coastline change indicator; inputting a pre-trained coastal dynamics model CoastDynNet, and using the sediment transport equation to analyze the physical mechanism of coastline change; applying a coastline evolution optimization function to fine-tune the time interpolation results; using a pre-trained time series decomposition network WaveletTempNet to perform multi-scale decomposition of the coastline change sequence; designing a spectrum analysis module, and calculating the coastline stability index and change frequency matrix based on the separation results; and using the Bayesian model averaging method to construct a multi-model integration framework.

[0007] 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, which is used to capture key change points in the time series.

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

[0009] Among them, 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.

[0010] Among them, the sediment transport equation is a physical mechanism equation that describes the law of sediment movement 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 sediment migration amount and the rate of change of coastline position per unit time.

[0011] Among them, the coastline evolution optimization function is used to constrain the optimization of deep learning prediction results based on physical constraints. The input includes the initial coastline morphology, predicted coastline morphology, terrain slope, sediment characteristics, and extreme event intensity. The output is the revised coastline morphology that meets the physical constraints.

[0012] The CoastDynNet model has 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.

[0013] 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 variation 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 coastline's ability to resist external interference, and is calculated by 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 position in a certain frequency interval.

[0015] Among them, the Bayesian model averaging method is a statistical method based on the principle of Bayesian reasoning to perform weighted averaging of 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 integration framework is determined, thereby achieving model advantage complementarity and prediction uncertainty quantification.

[0016] This method constructs a temporal attention matrix using a multi-source satellite data fusion network, combines it with a fourth-order temporal convolutional neural network to extract coastline morphological characteristics, and applies a spatiotemporal attention mechanism for feature dimensionality reduction and optimization, achieving high-precision detection of coastline changes. This method cleverly integrates the pre-trained coastal dynamics model CoastDynNet with the temporal decomposition network WaveletTempNet, while also incorporating a Bayesian model averaging mechanism to form a monitoring framework that deeply integrates data-driven and physical mechanisms.

[0017] This method effectively solves many defects of 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 guaranteed, solving the problem of insufficient accuracy of pure data-driven methods.

[0018] This invention organically combines deep learning with coastal dynamics to construct a complete method for dynamic monitoring of coastline changes. It not only achieves high-precision, multi-scale change monitoring, but also reveals the physical mechanism behind the changes, providing a scientific basis for coastal management and protection, solving the technical problem of insufficient accuracy in dynamic monitoring of coastline changes, and significantly improving the accuracy and reliability of dynamic monitoring of coastline changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention.

[0020] Figure 2 This is a diagram showing the temporal changes in coastline and typhoon event responses monitored by multi-source satellites in Example 2.

[0021] Figure 3 This is the multi-scale decomposition result diagram of coastline change in Example 2.

[0022] Figure 4 This is the result of the spectrum analysis of coastline changes in Example 2. DETAILED DESCRIPTION

[0023] In order to make the purpose, 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] like Figure 1 FIG. 1 is a flowchart of a deep learning-driven coastline change dynamic monitoring method provided by the present invention. The method includes the following steps:

[0025] S01. Build 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;

[0026] S02. Apply a fourth-order temporal convolutional neural network to extract the coastline morphological change characteristics and establish the first wave matrix through a recursive method;

[0027] S03. Use the spatiotemporal attention mechanism to perform dimensionality reduction processing on the first fluctuation matrix to generate a second fluctuation matrix;

[0028] S04. Establish a fluctuation conversion matrix to convert the second fluctuation matrix into a quantitative coastline change index;

[0029] S05. Input the pre-trained coastal dynamics model CoastDynNet and use the sediment transport equation to analyze the physical mechanism of coastline change;

[0030] S06. Apply the coastline evolution optimization function to fine-tune the time interpolation results to improve the short-term prediction accuracy;

[0031] S07. Use the pre-trained time series decomposition network WaveletTempNet to perform multi-scale decomposition of the coastline change series to separate long-term trends from short-term fluctuations;

[0032] S08. Design a spectrum analysis module to calculate the coastline stability index and change frequency matrix based on the separation results;

[0033] S09. Use the Bayesian model averaging method to build a multi-model integration framework, integrate the outputs of the physical model and deep learning model according to the weighted prediction error, and generate a comprehensive coastline dynamic assessment report;

[0034] Among them, the temporal attention matrix specifically refers to the two-dimensional feature representation matrix formed by weighted fusion of satellite image data at different time points through a multi-head attention mechanism, which is used to capture key change points in the time series.

[0035] Among them, the first fluctuation matrix specifically refers to the primary feature representation of coastline morphological changes extracted from the original satellite image, which includes a feature matrix composed of multi-dimensional indicators such as position offset, area change rate, and morphological complexity.

[0036] 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.

[0037] Among them, the fluctuation conversion matrix specifically refers to the transformation matrix that maps the abstract feature space to the coastline change indicator space with clear physical meaning, realizing the mapping relationship from features to actual changes.

[0038] The sediment transport equation specifically refers to the physical mechanism equation that describes the movement of sediments around the coastline. The inputs include wave energy density, incident angle, sediment particle size distribution, tidal amplitude, and longshore current velocity. The output is the net amount of sediment transported per unit time and the rate of change of coastline position.

[0039] The coastline evolution optimization function is used to constrain optimization of deep learning prediction results based on physical constraints. The input includes the initial coastline morphology, predicted coastline morphology, terrain slope, sediment characteristics, and extreme event intensity. The output is the revised coastline morphology that meets the physical constraints.

[0040] The specific structure of the pre-trained 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 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. The steps of establishing the training data set in the pre-training process of the CoastDynNet model specifically include collecting multi-year satellite observation data of different types of coastlines around the world, extracting water resources in the corresponding area, and analyzing the influence of tides on the coastline. The CoastDynNet model pre-training steps specifically include pre-training on physical simulation data to learn basic dynamic laws, fine-tuning on global satellite observation data to learn actual coastal response patterns, and then performing transfer learning to optimize local adaptability for coastal types similar or close to the test area. Finally, the model parameters are optimized by comparing the morphological differences between the actual observation sequence and the model prediction sequence through error back propagation.

[0041] The specific structure of the pre-trained 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 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. The steps of establishing a training dataset during the pre-training of the WaveletTempNet model specifically include collecting time-scale change sequences of typical coastlines around the world, standardizing and pre-processing the sequences to eliminate dimensionality effects, manually labeling different scale change components as training targets, constructing a mixed sequence containing long-term trends and short-term fluctuations as input, and performing stratified sampling according to different change patterns and frequency characteristics to form training and validation sets. The steps of pre-training the WaveletTempNet model specifically include initial training on synthetic data to learn basic decomposition capabilities, then performing supervised learning on real coastline sequences to improve generalization performance, then optimizing parameters by comparing the consistency of outputs of different decomposition levels with theoretical decomposition results, and finally verifying the reconstruction accuracy to ensure that information in the decomposition and reconstruction processes is losslessly preserved.

[0042] Among them, the multi-layer wavelet transform network specifically refers to a neural network structure designed based on the wavelet transform principle. By learning wavelet basis functions of different scales, it realizes multi-resolution analysis of signals in the time and frequency domains.

[0043] Among them, the stability index specifically refers to an indicator that quantifies the coastline's ability to resist external interference, which is calculated by the ratio of long-term change trend to short-term fluctuation amplitude.

[0044] 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 of the corresponding positions in a certain frequency interval.

[0045] Among them, the Bayesian model averaging method specifically refers to a statistical method based on the principle of Bayesian reasoning to perform weighted averaging of 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, thereby achieving complementary model advantages and quantification of prediction uncertainty.

[0046] The specific implementation of the above steps is described in detail below.

[0047] The specific implementation of step S01 is to achieve effective integration of temporal information by establishing a multi-source satellite data fusion network. First, multi-temporal images from satellites with different orbital periods (including polar-orbiting satellites and geosynchronous satellites) are collected, and radiometric and geometric corrections are performed on each image to ensure data consistency. Then, a multi-head self-attention mechanism is used to construct a temporal attention matrix, with each attention head responsible for learning the changing characteristics at different time scales. Specifically, for the first Image features at each time point With the Image features at each time point , calculate the attention weight ,form dimensional temporal attention matrix A, where is the length of the time series. This matrix can adaptively highlight the image features at the key change moments and reduce the impact of noise. When the value is greater than 0.7, it indicates that there is a significant correlation between the two time points and requires special attention. The purpose of this step is to establish a high-quality time series feature representation by fusing multi-source remote sensing data, laying the data foundation for subsequent coastline change analysis.

[0048] Step S02 specifically applies a fourth-order temporal convolutional neural network to extract coastline morphological change characteristics. First, four temporal convolutional layers are designed, each using a different-sized temporal convolution kernel (3, 5, 7, and 9, respectively) to capture multi-scale temporal features. The convolution results are then processed through a recurrent neural network, specifically a long short-term memory (LSTM) architecture with a hidden layer dimension of 256 and a time step equal to the image sequence length. The output of the recurrent network is mapped through a fully connected layer to generate a first fluctuation matrix F containing multidimensional indicators such as position offset, area change rate, and morphological complexity. Position offset is calculated by calculating the Euclidean distance between corresponding coastline points at adjacent time points; area change rate represents the percentage change in the land-sea boundary area per unit time; and morphological complexity is calculated using the fractal dimension, which ranges from 1 to 2, with larger values indicating more complex coastlines. Areas with position offset exceeding 10 meters per year or area change rate exceeding 5% per year are identified as areas of significant change. The goal of this step is to extract primary features of coastline morphological change from the original imagery and capture key information about these changes.

[0049] The specific implementation of step S03 is to use the spatiotemporal attention mechanism to reduce the dimensionality of the first fluctuation matrix. First, a spatiotemporal attention module is constructed, which includes two attention sub-modules: the time dimension and the space dimension. The time dimension attention sub-module identifies the key time points by calculating the correlation matrix of the features at different time points; the space dimension attention sub-module identifies the key spatial areas by calculating the correlation matrix of the features at different spatial positions. The outputs of the two attention sub-modules are then weighted and fused, and the weight coefficients are automatically optimized through back propagation. Finally, the principal component analysis algorithm is applied to reduce the dimensionality of the fusion results, retaining the principal components with an explained variance ratio of 95%, to form a second fluctuation matrix G with low dimension and high information density. When the maximum principal component contribution rate exceeds 60%, it indicates that there is a dominant change pattern. The purpose of this step is to remove redundant information and noise in the first fluctuation matrix, extract the core change pattern, and provide a more streamlined feature representation for subsequent quantitative analysis.

[0050] The specific implementation method of step S04 is to establish a fluctuation conversion matrix to map the abstract feature space to a coastline change index space with clear physical meaning. First, a three-layer fully connected neural network is designed, with the number of input layer nodes equal to the dimension of the second fluctuation matrix, the number of hidden layer nodes is 128, and the number of output layer nodes equal to the number of target change indicators. Then, using known ground-based measured data as supervision information, 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 final fluctuation conversion matrix T can convert the second fluctuation matrix G into a series of quantitative coastline change indicators, including the average annual retreat rate, seasonal fluctuation amplitude, 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 conversion matrix has achieved acceptable accuracy. The purpose of this step is to convert the abstract features extracted by the deep learning model into change indicators with actual physical meaning, which is convenient for professionals to understand and support decision-making.

[0051] The specific implementation of step S05 is to input the quantitative indicators into the pre-trained coastal dynamics model CoastDynNet, and analyze the physical mechanism of coastline change through the sediment transport equation. First, the change indicators obtained in step S04 are input into the CoastDynNet model together with auxiliary data (such as wave parameters, tidal data, and meteorological data). Then, the sediment dynamics physical constraint layer within the model calculates the coastal sediment transport rate based on the sediment transport equation. and vertical sediment transport rate ,when A ratio greater than 3 indicates that coastal sediment transport is the dominant process. The wave dynamics calculation unit then assesses wave erosion capacity based on the ratio of average wave height to critical erosion wave height. When this ratio exceeds 1.2, the coastline is in an erosive state. Finally, the morphological evolution prediction layer integrates the outputs of each module to generate a physical interpretation report of coastline changes, including the dominant dynamic mechanism, sediment balance status, and predicted trend of change. This step combines data-driven statistical analysis with physical mechanism models to reveal the underlying causes of coastline changes and improve the interpretability of monitoring results.

[0052] The specific implementation method of step S06 is to apply the coastline evolution optimization function to make fine adjustments to the time interpolation results. First, an optimization objective function is constructed based on physical constraints, including terrain slope restrictions, sediment characteristics constraints, and extreme event influencing factors. The constraints are then converted into penalty terms using the Lagrange multiplier method, which together with the prediction error term form a complete optimization function. The gradient descent algorithm is then used for iterative optimization, and the iteration is stopped when the maximum difference between the results of two adjacent iterations is less than 0.5 meters. Finally, a corrected coastline morphology that meets the physical constraints is obtained. For areas significantly affected by tides, additional corrections are required based on tide level changes, and the tide level correction coefficient ranges from 0.8 to 1.2. The purpose of this step is to correct the purely data-driven prediction results by introducing physical constraints, so that the prediction results are more in line with the actual coastal dynamics process and improve the accuracy of short-term predictions.

[0053] The specific implementation of step S07 involves using a pre-trained time series decomposition network, WaveletTempNet, to perform multi-scale decomposition of the coastline change sequence. First, the corrected coastline change sequence is input into the wavelet transform encoding layer of the WaveletTempNet model. Based on the principles of discrete wavelet transform, this layer uses wavelet basis functions of different scales (such as Haar wavelets and Daubechies wavelets) to perform multi-resolution analysis of the input signal. Next, the time series feature extraction layer extracts temporal features at each decomposition level using a multi-head self-attention mechanism. When the input sequence length is n, the number of attention heads is set to min(8, n / 4). The extracted features are then further processed by the long-term short-term memory network layer to capture long-term dependencies. Finally, the reconstruction decoding layer reconstructs the processed results, separating the long-term trend component and the short-term fluctuation component. The decomposition quality is considered acceptable when the reconstruction error is less than 5% of the standard deviation of the original sequence. This step decomposes the complex coastline change sequence into components at different time scales, facilitating the separate analysis of long-term evolutionary trends and short-term dynamic changes.

[0054] The specific implementation of step S08 is to design a spectrum analysis module to further process the separated time series. First, a fast Fourier transform is applied to the short-term fluctuation component to obtain a frequency domain representation and identify the main periodic components. When the energy proportion exceeds 15%, the frequency is identified as a significant period. Then, the ratio of the long-term trend amplitude to the short-term fluctuation standard deviation is calculated as the stability index SI. When SI is greater than 2, it indicates that the coastline of this section is relatively stable, and when SI is less than 0.5, it indicates instability. Then, a change frequency matrix is constructed, where the rows represent 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 of the corresponding position in a certain frequency interval. Finally, based on the spectrum analysis results, potential periodic change patterns and abnormal change events are identified. The purpose of this step is to quantitatively evaluate the dynamic characteristics of the coastline and provide a scientific basis for management decisions.

[0055] The specific implementation of step S09 is to construct a multi-model integration framework using the Bayesian model averaging method. First, prediction results from physical models (such as CoastDynNet) and deep learning models (such as WaveletTempNet) on a historical validation set are collected. Then, based on the principles of Bayesian inference, the prediction error distribution of each model is calculated. The mean squared error is used when the error follows a normal distribution, and the mean absolute percentage error is used when the error follows a long-tail distribution. Next, the credibility weight of each model is determined based on the error distribution. The weight calculation formula is an exponential function of the model's accuracy, ensuring that models with higher accuracy receive greater weight. Finally, the outputs of each model are weighted averaged according to the weights to generate a comprehensive coastline dynamics assessment report, including trend forecasts, uncertainty estimates, and risk level assessments. When the prediction difference between models exceeds 20%, expert review is required. This step aims to fully leverage the complementary advantages of different model types, improve the accuracy and reliability of predictions, and quantify the uncertainty of the prediction results.

[0056] Furthermore, the CoastDynNet model is structured as a multi-layer encoder-decoder architecture, consisting of four core modules. The first module is the sedimentation dynamics physical constraint layer, which calculates the longshore sediment transport rate using the improved CERC formula and establishes a constraint relationship based on 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, strong longshore transport parameters are used, while when the ratio is less than 0.5, strong vertical transport parameters are used. The second module is the tidal influence attention module, which processes tidal time series data using a gated recurrent unit network. Its parameters are determined by the ratio of the tidal period to the shoreline response time. When the ratio is less than 0.1, the tidal influence is ignored. The third module is the wave dynamics calculation unit, which calculates wave erosion based on the 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, 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 an improved one-dimensional ecological cellular automaton algorithm to predict coastline morphological evolution. The steps for establishing the CoastDynNet model pre-training dataset include: first, collecting multi-year satellite observation data of different types of coastlines around the world (including sandy coasts, reef coasts, mud 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 organizing 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 sedimentation 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 performed on similar or similar coastal types in the region 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.

[0057] Furthermore, the WaveletTempNet model's specific structure consists of an end-to-end deep learning network consisting 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 of the input time series. The wavelet basis function can be selected from the Haar, Daubechies, or Symlet series, and its parameters are determined by the coastline variation period: Haar wavelets are used for periods less than one month, Daubechies wavelets for periods between one and 12 months, and Symlet wavelets for periods greater than 12 months. The second layer is the time series feature extraction layer, which uses a multi-head self-attention mechanism to capture temporal dependencies. The number of attention heads is determined by the length of the input sequence, 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. The hidden state dimension is determined by the feature complexity, set to 64 for simple variation patterns and 256 for complex variation patterns. The fourth layer is the reconstruction decoding layer, which reassembles the processed features into a time series through an inverse wavelet transform and separates long-term trends from short-term fluctuations. The steps for establishing the WaveletTempNet model pre-training dataset include: first, collecting 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, using the Z-score method to eliminate dimensionality effects; 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 and validation sets, with a training set to validation set ratio of 8:2. The pre-training steps of the WaveletTempNet model include: initial training on synthetic data to learn basic decomposition capabilities, with no fewer than 200 initial training rounds; supervised learning on real coastline sequences to improve generalization performance, with no fewer 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 fewer than 50 parameter optimization rounds; finally, reconstruction accuracy verification is performed to ensure that information in the decomposition and reconstruction processes is losslessly preserved. When the reconstruction error is less than 3% of the standard deviation of the original signal, the training is considered to have met the requirements.

[0058] The primary challenge facing traditional coastline monitoring is the long satellite revisit period. For example, the Landsat series of satellites has a revisit period of 16 days, a time interval far exceeding the characteristic timescales of many coastal dynamics. Due to this inherent temporal resolution limitation, extreme meteorological events such as storm surges and wind surges often significantly impact coastlines within hours to days. Traditional single-satellite observation systems are unable to capture these short-term changes in a timely manner. This makes it difficult to fully record the changes in coastline morphology before and after extreme events, and key change points are missed, resulting in a "temporal blind spot" in coastline dynamic monitoring. Furthermore, even under ideal circumstances, data acquisition within the prescribed revisit period is often further reduced by factors such as cloud cover and sensor failures. The effective observation interval can extend to months, making the monitoring of short-term changes even more difficult. This limited temporal resolution severely hinders the accurate monitoring and mechanistic understanding of coastline dynamics, particularly the assessment of the impacts of catastrophic extreme events.

[0059] This method effectively addresses the problem of insufficient temporal resolution caused by the long revisit periods of traditional satellites by constructing a multi-source satellite data fusion network (step S01). This method no longer relies on a single satellite data source, but instead integrates time series imagery acquired by multiple satellites with different orbital periods (such as the optical satellites Landsat and Sentinel-2, and the radar satellite Sentinel-1) to form a high-density time series with complementary coverage. The temporal attention matrix constructed using a multi-head self-attention mechanism can adaptively fuse imagery data from different time phases and sensors, highlighting features at key moments of change and effectively improving temporal resolution. Furthermore, the fourth-order temporal convolutional neural network (step S02) employs a multi-scale temporal convolution kernel design, which can capture variations across different time scales and exhibits high sensitivity to short-term abrupt events. The coastline evolution optimization function (step S06) incorporates physical constraints to enable reasonable interpolation within time intervals where observation data is missing, thus filling in "temporal blind spots." Furthermore, the multi-model integration framework (step S09) constructed using the Bayesian model averaging method can leverage the strengths of both physical models and deep learning models. In data-sparse areas, the prior knowledge provided by the physical models can improve forecast reliability and effectively address sudden changes caused by extreme events. The combined application of these technical approaches enables this invention to significantly improve the temporal resolution of coastline monitoring, effectively capturing short-term changes in coastline caused by extreme events such as storm surges.

[0060] The second difficulty in monitoring coastline changes is the difficulty in effectively distinguishing long-term trends caused by climate change from seasonal fluctuations or short-term changes caused by extreme events. Dynamic changes in coastlines are the result of multiple time-scale processes, including long-term interdecadal trends caused by factors such as sea level rise and changes in sediment supply, periodic seasonal changes caused by factors such as monsoon alternation and tidal cycles, and short-term, dramatic fluctuations caused by extreme events such as typhoons and storm surges. These processes at different time scales overlap and influence each other in the observational data. If they cannot be effectively separated, they will lead to biased trend analysis, inaccurate assessments 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 makes it difficult to accurately separate complex multi-scale change components. This is especially true when data is missing or sampled unevenly, making the separation effect even less ideal.

[0061] To address the challenge of separating multi-scale signals of coastline change, this invention utilizes a pre-trained time series decomposition network, WaveletTempNet (step S07), to effectively separate long-term trends from short-term fluctuations. Designed based on wavelet transform principles, this network uses a multi-scale wavelet transform encoding layer to perform multi-resolution analysis of coastline change sequences. Wavelet basis functions at different scales accurately extract components of variation at different periods. The network's time series feature extraction layer uses a multi-head self-attention mechanism to capture temporal features at each decomposition level. A long-short-term memory (LSTM) layer further processes the sequence information to capture long-term dependencies. Finally, a reconstruction decoding layer reconstructs the processed results, accurately separating long-term trends from short-term fluctuations. Furthermore, the spectrum analysis module (step S08) further processes the separated time series, identifying the main periodic components through fast Fourier transforms, calculating a stability index to quantitatively assess coastline stability, and constructing a frequency matrix to describe the frequency characteristics of variations in different segments. Furthermore, the CoastDynNet model (step S05) analyzes the physical mechanisms of coastline change using sediment transport equations, identifying the drivers of change at different time scales. The coordinated application of these technologies enables the present invention to effectively distinguish between long-term trends caused by climate change and short-term fluctuations caused by seasonality or extreme events, providing reliable technical support for dynamic assessment and prediction of coastlines.

[0062] The mathematical model or calculation process involved in the present invention is described in detail below.

[0063] In step S01, the calculation process of the multi-head self-attention mechanism to construct the temporal attention matrix is specifically expressed as follows:

[0064] ;

[0065] ;

[0066] ;

[0067] Where, For the The image at each time point The attention weight of the image at each time point; is the unnormalized attention score; and Respectively and The image feature vector at each time point; and They are the query transformation matrix and the key transformation matrix, respectively, obtained through model training; is the dimension of the feature vector; is the length of the time series; for dimensional temporal attention matrix.

[0068] The parameter acquisition method is: and It is obtained by extracting features from the original satellite image, specifically including step 1: performing radiation 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 a feature vector. and It is obtained through back propagation algorithm optimization during the model training process, and the training goal is to minimize the error between the predicted coastline and the actual coastline. The value range is usually 64 to 256.

[0069] The calculation principle of this attention mechanism is based on dot product similarity, which measures the correlation between image features at different time points by calculating the inner product of the features and dividing them by Scaling is performed to avoid the vanishing gradient problem. An exponential function and normalization transform the attention scores into probability distributions, ensuring that all weights sum to 1. This design adaptively highlights image features at key moments of change, effectively improving temporal modeling capabilities. Compared to traditional linear weighting methods, this mechanism can capture nonlinear temporal dependencies and is more robust to outliers.

[0070] In step S02, the calculation process of the position offset, area change rate and morphological complexity is specifically expressed as follows:

[0071] ;

[0072] ;

[0073] ;

[0074] Where, for The position offset at the moment, in meters; and They are Moment The x- and y-coordinates of the sampling points; is the number of sampling points along the coastline; for The area change rate at the moment, in percentage; for The area of land and sea at the moment, in square meters; for The morphological complexity of the moment, dimensionless; for The length of the coastline at the moment, in meters.

[0075] The parameter acquisition method is: and It is obtained by extracting the coastline from satellite images, which specifically includes step 1: using threshold segmentation or deep learning methods to extract the water-land boundary; step 2: vectorizing the boundary to obtain the coordinates of equally spaced sampling points. It is obtained by calculating the area of the water body region, specifically including step 1: counting pixels of the segmentation result; step 2: multiplying the number of pixels by the square of the pixel resolution to obtain the actual area. It is obtained by calculating the geometric length of the coastline, which specifically includes step 1: calculating the length of the vectorized coastline; step 2: multiplying the pixel length by the resolution to obtain the actual length.

[0076] The calculation principles of these indicators are based on Euclidean distance, percentage change rate, and fractal geometry, respectively. Position offset reflects the degree of change in the coastline's position, calculating the displacement of corresponding sampling points at adjacent time points using Euclidean distance. Area change rate reflects the relative change in the land-sea interface area, expressed as a percentage to facilitate comparison across different regions. Morphological complexity, based on fractal dimension theory, is calculated as the logarithmic ratio of coastline length to area, quantitatively describing the geometric complexity of the coastline. Together, these three indicators constitute a multidimensional representation of coastline morphological change, providing a more comprehensive picture of coastline dynamics than traditional single indicators.

[0077] In step S03, the calculation process of the spatiotemporal attention mechanism is specifically expressed as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] Where, Output of attention in time dimension; 、 、 They are the query matrix, key matrix and value matrix of the time dimension respectively; is the spatial dimension attention output; 、 、 are the query matrix, key matrix, and value matrix of the spatial dimension respectively; is the feature dimension; is the result of spatiotemporal attention fusion; is the attention weight coefficient of the time dimension, ranging from 0 to 1.

[0082] The parameter acquisition method is: 、 、 、 、 、 The first wave matrix Perform linear transformation to obtain, specifically including step 1: design six different linear transformation matrices 、 、 、 、 、 ; Step 2: The first wave matrix Multiplying these transformation matrices yields the corresponding query, key, and value matrices. It is obtained through automatic optimization by the back propagation algorithm, and the initial value is set to 0.5.

[0083] The design principle of the spatiotemporal attention mechanism is to decompose the attention calculation into two independent calculation processes, the time dimension and the space dimension, and then integrate the information of the two dimensions through weighted fusion. The time dimension attention focuses on the dependencies between different time points, while the space dimension attention focuses on the correlation between different spatial locations. This decomposition design enables the model to capture key information in both time and space dimensions at the same time, significantly improving feature extraction capabilities. Weighting coefficient Through automatic optimization, the importance of the two dimensions can be flexibly adjusted according to the specific characteristics of the data. Compared with the traditional single-dimensional attention mechanism, this method can more effectively handle the complex dependencies of spatiotemporal data.

[0084] In step S04, the calculation process of the fluctuation conversion matrix is specifically expressed as follows:

[0085] ;

[0086] ;

[0087] Where, is the fluctuation transformation matrix; 、 、 are the weight matrices of the three layers of the neural network; 、 、 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; It is a quantitative indicator vector of coastline change after transformation, including multiple indicators such as annual average retreat rate, seasonal fluctuation amplitude, and intensity of extreme event response.

[0088] The parameter acquisition method is: 、 、 、 、 、 It is obtained through supervised learning, specifically including step 1: collecting known ground-truth data as training samples, including actual measured annual average retreat rate, seasonal fluctuation amplitude and other indicators; step 2: converting the second fluctuation matrix of the corresponding time period into As input, the measured indicators are used as output targets; Step 3: Use the least squares method to optimize the neural network parameters, the objective function is ,in is the measured indicator vector, is the regularization coefficient, ranging from 0.001 to 0.1, is the Frobenius norm of the weight matrix.

[0089] The design principle of the wave transformation matrix is to establish a mapping relationship from an abstract feature space to an index space with clear physical meaning through a deep neural network. The three-layer neural network structure can effectively fit complex nonlinear mapping relationships. The activation function introduces nonlinear transformation capabilities and addresses the vanishing gradient problem. The least-squares loss function measures the difference between the predicted and measured metrics, and the regularization term prevents overfitting. This design transforms the abstract features extracted by deep learning into meaningful change metrics, improving the model's interpretability and practicality. Compared to traditional linear transformation methods, this method can capture more complex relationships between features and metrics.

[0090] In step S05, the calculation process of the sediment transport equation is specifically expressed as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] Where, is the coastal sediment transport rate, in units of ; is the vertical sediment transport rate, in units of ; 、 、 is the empirical coefficient; is the wave height in the wave breaking zone, in units of ; is the wave incident angle in the wave breaking zone, in radians; is the beach slope, dimensionless; is the coastal gradient of wave height; is the rate of change of coastline position, in units of ; is the critical water depth for wave action, in units of ; is the coastal distance coordinate in units of ; is the distance coordinate perpendicular to the baseline, in units of ; is the time, the unit is .

[0095] The parameter acquisition method is: It is obtained through wave monitoring data or wave model calculation, specifically including step 1: collecting nearshore wave observation data or using the third-generation wave model (such as the SWAN model) to calculate wave propagation; step 2: calculating the evolution of waves near the shore according to the shallow water equation and determining the wave height in the wave breaking area. It is obtained through wave direction observation data, specifically including step 1: collecting deep-water wave direction data; step 2: calculating the incident angle of the wave in the breaking zone after refraction according to Snell's law. It is obtained through seabed topography measurement, specifically including step 1: conducting seabed topography measurement; step 2: calculating the average slope of the nearshore area. 、 、 It is an empirical coefficient and needs to be calibrated by measured data. The value range is 0.1~0.4, The value range is 0.01~0.05, The value range is 0.1~0.3. Available through Estimated, of which is the effective wave height.

[0096] Among them, considering the coastal sediment transport rate It is proportional to the 2.5th power of the wave height and the sine of the incident angle, reflecting the influence of wave energy and incident angle on coastal sediment transport; vertical sediment transport rate It consists of two components: 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 shoreline position change rate equation is based on the principle of sediment mass conservation and calculates the rate of change of shoreline position using the spatial gradient of the alongshore sediment transport rate and the vertical sediment transport rate. This physical model can quantitatively describe the dynamic mechanisms of shoreline evolution and provide a physical explanation for shoreline changes. Compared with purely data-driven methods, this physical model has better interpretability and extrapolation capabilities.

[0097] In step S06, the calculation process of the coastline evolution optimization function is specifically expressed as follows:

[0098] ;

[0099] ;

[0100] ;

[0101] Where, Optimize functions for coastline evolution; is the optimized coastline position vector; is the coastline position vector initially predicted by the model; is the spatial gradient of coastline position; 、 、 、 is the weight coefficient; is the physical constraint condition; and are the maximum and minimum allowable slopes of the coastline, respectively; is the final coastline position vector after optimization; is the number of discrete sampling points along the coastline; For the The coastline locations of the sampling points.

[0102] The parameter acquisition method is: Obtained through preliminary predictions of deep learning models; 、 、 、 is the weight coefficient, determined by cross-validation method, The value range is 0.5~2.0, The value range is 0.1~0.5, The value range is 0.05~0.2, The value range of is 1.0~5.0; and Determined according to the actual coastal terrain characteristics, The value range is usually 30°~45°. The value range of is usually 0°~5°. The optimization process uses the gradient descent algorithm, which includes the following steps: 1. Initialization ; Step 2: Calculate the objective function right Gradient; Step 3: Update in the opposite direction of the gradient , the step size is (The value range is 0.01 to 0.1); Step 4: Repeat steps 2 and 3 until convergence or the maximum number of iterations is reached.

[0103] The design principle of the coastline evolution optimization function is to introduce physical constraints on the basis of data-driven prediction. Using the Lagrange multiplier method, the constrained problem is transformed into an unconstrained optimization problem. The objective function consists of four components: the first is the data term, which ensures that the optimization result is close to the initial prediction; the second is the smoothing term, which prevents unnatural jagged fluctuations in the coastline; the third is the curvature term, which controls the curvature of the coastline; and the fourth is the physical constraint term, which ensures that the coastline slope remains within a reasonable range. This design combines the flexibility of deep learning with the constraints of physical models, generating coastline predictions that are consistent with both data patterns and physical constraints. Compared with purely data-driven methods, this method can effectively prevent overfitting and improve the physical plausibility of the predictions.

[0104] In step S08, the calculation process of the stability index and the change frequency matrix is specifically expressed as follows:

[0105] ; ;

[0106] Where, is the stability index, dimensionless; is the long-term trend amplitude, and the unit depends on the specific variable; is the standard deviation of short-term fluctuations, and its unit is same; is the element of the frequency matrix, indicating the spatial position In the frequency range the intensity of the change; For spectrum analysis The energy of the frequency components; For the The frequency of the frequency component; is an indicator function, which takes 1 when the condition is met, otherwise it takes 0; is the Dirac function, used for spatial positioning; is the total number of frequency components; For the The location of the spatial sampling points.

[0107] The parameter acquisition method is: It is obtained by calculating the difference between the start and end points of the long-term trend component and dividing it by the time interval, specifically including step 1: obtaining the long-term trend component separated by the WaveletTempNet model; step 2: calculating the rate of change of the trend component over the entire time span. It is obtained by calculating the standard deviation of the short-term fluctuation component, specifically including step 1: obtaining the short-term fluctuation component separated by the WaveletTempNet model; step 2: calculating the standard deviation of the component. and It is obtained by performing fast Fourier transform on the short-term fluctuation component, specifically including step 1: applying the FFT algorithm to the short-term fluctuation component; step 2: calculating the energy spectrum and corresponding frequency of each frequency component. and It is the predefined frequency interval boundary. Common interval settings include interannual variation (1 to 5 years), seasonal variation (0.25 to 1 year), tidal cycle variation (0.5 to 1 day), etc.

[0108] The stability index is calculated based on the concept of signal-to-noise ratio, using the amplitude of the long-term trend as the "signal" and the standard deviation of short-term fluctuations as the "noise." The ratio of the two reflects the stability of the coastline. A higher stability index indicates that long-term trends dominate and the coastline is more stable; a lower index indicates that short-term fluctuations dominate and the coastline is more unstable. The change frequency matrix is calculated based on spectral analysis and spatial discretization. By calculating the energy distribution at different spatial locations in different frequency intervals, a two-dimensional matrix is constructed to describe the spatiotemporal variation 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-indicator assessment methods, this method provides a more comprehensive and detailed description of the variation characteristics.

[0109] In step S09, the calculation process of the Bayesian model averaging method is specifically expressed as follows:

[0110] ; ;

[0111] ;

[0112] Where, For the The weight of each model; For the The mean squared error of the model on the validation set; is the temperature parameter, which controls the concentration of weight distribution; is the total number of models; For the The prediction results of the model; To integrate the prediction results; For the The prediction variance of the model; is the variance of the ensemble prediction, which represents the uncertainty of the prediction.

[0113] The parameter acquisition method is: 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 value and the actual value of each model on the validation set. Determined by cross-validation method, the value range is usually 0.1 to 10, with larger The value makes the weight distribution more concentrated towards the best performing models. For deep learning models, the Monte Carlo dropout method can be used for estimation, which specifically includes step 1: keeping dropout activated during the prediction stage; step 2: performing multiple forward propagations and calculating the variance of the prediction results; for physical models, the parameter perturbation method can be used for estimation, which specifically includes step 1: performing small random perturbations on the parameters of the physical model; step 2: counting the variance of the prediction results under different parameter combinations.

[0114] The design principle of the Bayesian model averaging method is based on the Bayesian statistical framework, which regards the model weight as a function of the model conditional probability. The weight calculation formula uses a negative exponential form, so that the model with better performance (smaller error) gets a higher weight. The temperature parameter Control the "hardness" of weight distribution. Ensemble prediction uses a weighted average approach, 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 fully utilizes 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 predictions. Compared with simple averaging or single model methods, this method can provide more accurate prediction results and more reasonable uncertainty estimates.

[0115] 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 technological breakthrough by constructing a coastline change monitoring framework with multi-level linkage 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, and effectively integrates the observation advantages of satellites with different orbital periods through the time series attention matrix mechanism, thereby enhancing the information richness of the original data. The fourth-order time series convolutional neural network can efficiently capture the multi-scale features in time series data, and its recursive structure design enables the model to have long-distance dependent learning capabilities, effectively extracting the inherent laws of coastline morphological changes.

[0116] The introduction of a spatiotemporal attention mechanism and a wave transformation matrix solves the problem of mapping high-dimensional feature space into a space of change indicators with clear physical meaning, ensuring efficient and interpretable feature extraction. The CoastDynNet model is designed based on the principles of sedimentary dynamics. Its multi-layered encoding and decoding architecture closely integrates coastline change prediction with physical process simulation. Through a tidal influence attention module and a wave dynamics calculation unit, it enables an understanding of coastline evolution from a physical perspective. The WaveletTempNet model, based on the principles of wavelet transforms, enables multi-scale decomposition of coastline change sequences, providing a mathematical foundation for distinguishing long-term trends from short-term fluctuations.

[0117] The Bayesian model averaging method solves the problem of multi-model integration from a statistical perspective, determines the credibility weights of different models through the prediction error distribution, and 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 is that it breaks the barriers between data analysis and physical simulation in traditional methods, constructs a complete knowledge chain from data to physics, and makes the monitoring results conform to both data laws and physical constraints, realizing the organic unity of multi-source data fusion, multi-scale feature extraction and physical mechanism analysis, thereby significantly improving the accuracy of dynamic monitoring of coastline changes.

[0118] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0119] In the specific implementation of step S01, the attention weight is calculated by the following formula:

[0120] ; ; ;

[0121] Where, For the The image at each time point The attention weight of the image at each time point; is the unnormalized attention score; and Respectively and The image feature vector at each time point; and They are the query transformation matrix and the key transformation matrix, respectively, obtained through model training; is the dimension of the feature vector, and its value range is usually 64 to 256; is the length of the time series; for dimensional temporal attention matrix. When the attention weight A value greater than 0.7 indicates a significant correlation between the two time points, warranting further attention. This step fuses multi-source remote sensing data to establish a high-quality time series representation, laying the foundation for subsequent coastline change analysis. This mechanism adaptively highlights image features at key moments of change, reduces the impact of noise, and effectively improves time series modeling capabilities.

[0122] The specific implementation of step S02 is to apply a fourth-order temporal convolutional neural network to extract the morphological characteristics of the coastline. First, four temporal convolution layers are designed, each using a temporal convolution kernel of different sizes (3, 5, 7, and 9, respectively) to capture multi-scale temporal features. The convolution results are then processed through 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 length equal to the image sequence length. The output of the recurrent network is mapped through a fully connected layer to generate a 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:

[0123] ;

[0124] ; ;

[0125] Where, for The position offset at the moment, in meters; and They are Moment The x- and y-coordinates of the sampling points; is the number of sampling points along the coastline; for The area change rate at the moment, in percentage; for The area of land and sea at the moment, in square meters; for The morphological complexity of the moment, dimensionless; for The length of the coastline at that moment, in meters. Areas identified as significantly changing are those with a positional offset exceeding 10 meters per year or an area change rate exceeding 5% per year. This step aims to extract primary features of coastline morphological change from the original imagery, capturing key information about these changes. These three metrics together form a multidimensional representation of coastline morphological change, providing a more comprehensive picture of coastline dynamics than traditional single metrics.

[0126] The specific implementation of step S03 is the same as above and will not be described in detail here.

[0127] The specific implementation of step S04 is to establish a fluctuation conversion matrix, mapping the abstract feature space into a space of coastline change indicators with clear physical meaning. First, a three-layer fully connected neural network is designed. The number of input layer nodes is equal to the dimension of the second fluctuation matrix, the number of hidden layer nodes is 128, and the number of output layer nodes is equal to the number of target change indicators. The conversion process is implemented using the following formula:

[0128] ; ;

[0129] Where, is the fluctuation transformation matrix; 、 、 are the weight matrices of the three layers of the neural network; 、 、 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 converted quantitative indicator vector of coastline change, which includes multiple indicators such as the average annual retreat rate, seasonal fluctuation amplitude, and intensity of extreme event response. Then, using known ground-based measured data as supervision information, the network parameters are optimized by the least squares method to learn the mapping relationship from the feature space to the change indicator space. The final fluctuation conversion matrix T can convert the second fluctuation matrix G into a series of quantitative coastline change indicators. When the root mean square error between the predicted value and the measured value is less than 0.15, the conversion matrix is considered to have achieved acceptable accuracy. The purpose of this step is to convert the abstract features extracted by the deep learning model into change indicators with actual physical meaning, which is convenient for professionals to understand and support decision-making. The design of the fluctuation conversion matrix establishes a mapping relationship from the abstract feature space to the indicator space with clear physical meaning through a deep neural network, which improves the interpretability and practicality of the model.

[0130] The specific implementation of step S05 is to input the quantitative indicators into the pre-trained coastal dynamics model CoastDynNet and analyze the physical mechanism of coastline change through the sediment transport equation. The sediment transport calculation formula is as follows:

[0131] ; ; ;

[0132] Where, is the coastal sediment transport rate, in units of ; is the vertical sediment transport rate, in units of ; 、 、 is the empirical coefficient, The value range is 0.1~0.4, The value range is 0.01~0.05, The value range of is 0.1~0.3; is the wave height in the wave breaking zone, in units of ; is the wave incident angle in the wave breaking zone, in radians; is the beach slope, dimensionless; is the coastal gradient of wave height; is the rate of change of coastline position, in units of ; is the critical water depth for wave action, in units of ; is the coastal distance coordinate in units of ; is the distance coordinate perpendicular to the baseline, in units of ; is the time, the unit is First, the change index obtained in step S04 is input into the CoastDynNet model together with auxiliary data (such as wave parameters, tidal data, and meteorological data). Then, the sediment dynamics physical constraint layer within the model calculates the coastal sediment transport rate based on the sediment transport equation. and vertical sediment transport rate ,when When the ratio is greater than 3, it indicates that coastal sediment transport is the dominant process. The wave dynamics calculation unit then evaluates the wave erosion capacity based on the ratio of the average wave height to the critical erosion wave height. When the 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 coastline changes. The role of this step is to combine data-driven statistical analysis with physical mechanism models to reveal the essential reasons behind coastline changes and improve the interpretability of monitoring results. The sediment transport equation is based on classical coastal dynamics theory, combined with the CERC formula and the vertical sediment transport model. It can quantitatively describe the dynamic mechanism of coastline evolution and provide a physical explanation for coastline changes.

[0133] The specific implementation of step S06 is to apply the coastline evolution optimization function to fine-tune the time interpolation result. The mathematical expression of the optimization function is:

[0134] ;

[0135] ;

[0136] ;

[0137] Where, Optimize functions for coastline evolution; is the optimized coastline position vector; is the coastline position vector initially predicted by the model; is the spatial gradient of coastline position; 、 、 、 is the weight coefficient, The value range is 0.5~2.0, The value range is 0.1~0.5, The value range is 0.05~0.2, The value range of is 1.0~5.0; is the physical constraint condition; and are the maximum and minimum allowable slopes of the coastline, The value range is usually 30°~45°. The value range of is usually 0°~5°; is the final coastline position vector after optimization; is the number of discrete sampling points along the coastline; For the The coastline locations of the sampling points are determined. First, an optimization objective function is constructed based on physical constraints, including terrain slope constraints, sediment property constraints, and extreme event influencing factors. The constraints are then converted into penalty terms using the Lagrange multiplier method, which together with the prediction error term form the complete optimization function. An iterative optimization algorithm is then used, stopping when the maximum difference between two consecutive iterations is less than 0.5 meters. Ultimately, a revised coastline morphology is obtained that satisfies the physical constraints. For areas significantly affected by tides, additional corrections are required based on tidal fluctuations, with tidal correction coefficients ranging from 0.8 to 1.2. This step aims to modify purely data-driven predictions by introducing physical constraints, making them more consistent with actual coastal dynamics and improving short-term prediction accuracy. The design of the coastline evolution optimization function combines the flexibility of deep learning with the constraints of physical models, generating coastline predictions that conform to both data patterns and physical constraints.

[0138] The specific implementation of step S07 is the same as above and will not be described in detail here.

[0139] 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:

[0140] ;

[0141] ;

[0142] Where, is the stability index, dimensionless; is the long-term trend amplitude, and the unit depends on the specific variable; is the standard deviation of short-term fluctuations, and its unit is same; is the element of the frequency matrix, indicating the spatial position In the frequency range the intensity of change; For spectrum analysis The energy of the frequency components; For the The frequency of the frequency component; is an indicator function, which takes 1 when the condition is met, otherwise it takes 0; is the Dirac function, used for spatial positioning; is the total number of frequency components; For the The locations of spatial sampling points are analyzed. First, a fast Fourier transform is applied to the short-term fluctuation component to obtain a frequency domain representation. The main periodic components are identified, and frequencies with an energy contribution exceeding 15% are identified as significant cycles. The stability index (SI) is then calculated: the ratio of the long-term trend amplitude to the short-term fluctuation standard deviation. An SI greater than 2 indicates relative stability of the coastline, while an SI less than 0.5 indicates instability. A variation frequency matrix is then constructed, with rows representing spatial sampling points and columns representing predefined frequency intervals (such as interannual, seasonal, or tidal variations). The matrix elements represent the intensity of variation at the corresponding location within a given frequency interval. Finally, based on the spectral analysis results, potential periodic patterns and anomalous variation events are identified. This step aims to quantitatively assess the dynamic characteristics of the coastline and provide a scientific basis for management decisions. The stability index is calculated based on the signal-to-noise ratio concept, using 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 variation frequency matrix calculates the energy distribution at different spatial locations in different frequency intervals, creating a two-dimensional matrix describing the spatiotemporal variation of the coastline.

[0143] The specific implementation of step S09 is to use the Bayesian model averaging method to build a multi-model integration framework. The mathematical expression of the integration process is:

[0144] ;

[0145] ;

[0146] ;

[0147] Where, For the The weight of each model; For the The mean squared error of the model on the validation set; is a temperature parameter that controls the concentration of weight distribution, and its value range is usually 0.1 to 10; is the total number of models; For the The prediction results of the model; To integrate the prediction results; For the The prediction variance of the model; To integrate the variance of the predictions, we represent the uncertainty of the predictions. First, we collect prediction results from physical models (such as CoastDynNet) and deep learning models (such as WaveletTempNet) on a historical validation set. Then, based on Bayesian inference principles, we calculate the prediction error distribution for each model. The mean squared error is used when the error follows a normal distribution, and the mean absolute percentage error is used when the error follows a long-tail distribution. Next, we assign a credibility weight to each model based on the error distribution. The weight calculation formula is an exponential function of the model's accuracy, ensuring that models with higher accuracy receive a higher weight. Finally, we perform a weighted average of the model outputs according to the weights to generate a comprehensive coastline dynamics assessment report, including trend forecasts, uncertainty estimates, and risk ratings. When the difference in predictions between models exceeds 20%, expert review is required. This step leverages the complementary strengths of different model types, improving prediction accuracy and reliability while also quantifying the uncertainty of the prediction results. The Bayesian model averaging method treats model weights as functions of the model's conditional probability. Using a negative exponential weighting formula, we assign higher weights to models with better performance, leveraging the complementary strengths of different model types.

[0148] The CoastDynNet model is structured as a multi-layer encoder-decoder architecture, consisting of four core modules. The first module is the sedimentary dynamics physical constraint layer, which calculates the longshore sediment transport rate using the improved CERC formula and establishes a constraint relationship based on 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, strong longshore transport parameters are used, while when the ratio is less than 0.5, strong vertical transport parameters are used. The second module is the tidal influence attention module, which processes tidal time series data using a gated recurrent unit network. Its parameters are determined by the ratio of the tidal period to the shoreline response time. When the ratio is less than 0.1, the tidal influence is ignored. The third module is the wave dynamics calculation unit, which calculates wave erosion based on the 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, 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 an improved one-dimensional ecological cellular automaton algorithm to predict coastline morphological evolution. The steps for establishing the CoastDynNet model pre-training dataset include: first, collecting multi-year satellite observation data of different types of coastlines around the world (including sandy coasts, reef coasts, mud 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 organizing 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 sedimentation dynamics laws, with no fewer than 100 pre-training rounds; then fine-tuning on global satellite observation data to learn actual coastal response patterns, with no fewer than 50 fine-tuning rounds; then transfer learning is performed on regionally similar or similar coastal types to optimize local adaptability, with no fewer 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 fewer than 20 optimization rounds.

[0149] The WaveletTempNet model's specific architecture consists of an end-to-end deep learning network consisting of four key layers. The first layer is the multiscale wavelet transform encoding layer, which uses the discrete wavelet transform principle to perform multiscale decomposition on the input time series. The wavelet basis functions can be Haar, Daubechies, or Symlet series, with their parameters determined by the coastline variation period: Haar wavelets are used for periods less than one month, Daubechies wavelets for periods between one and 12 months, and Symlet wavelets for periods greater than 12 months. The second layer is the time series feature extraction layer, which uses a multi-head self-attention mechanism to capture temporal dependencies. The number of attention heads is determined by the length of the input sequence, typically 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 architecture to process sequence information. The hidden state dimension is determined by the feature complexity, set to 64 for simple variation patterns and 256 for complex variation patterns. The fourth layer is the reconstruction decoding layer, which reassembles the processed features into a time series using an inverse wavelet transform and separates long-term trends from short-term fluctuations. The steps for establishing the WaveletTempNet model pre-training dataset include: first, collecting 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, using the Z-score method to eliminate dimensionality effects; 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 and validation sets, with a training set to validation set ratio of 8:2. The pre-training steps of the WaveletTempNet model include: initial training on synthetic data to learn basic decomposition capabilities, with no fewer than 200 initial training rounds; supervised learning on real coastline sequences to improve generalization performance, with no fewer 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 fewer than 50 parameter optimization rounds; finally, reconstruction accuracy verification is performed to ensure that information in the decomposition and reconstruction processes is losslessly preserved. When the reconstruction error is less than 3% of the standard deviation of the original signal, the training is considered to have met the requirements.

[0150] 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 Scaling is performed to avoid the vanishing gradient problem. An exponential function and normalization transform the attention scores into probability distributions, ensuring that all weights sum to 1. This mechanism adaptively highlights image features at key moments of change, effectively improving temporal modeling capabilities. Compared to traditional linear weighting methods, this mechanism can capture nonlinear temporal dependencies and is more robust to outliers.

[0151] It should be noted that in step S02, the position offset is calculated based on the Euclidean distance between corresponding sampling points at adjacent time points. The area change rate reflects the relative change in the land-sea interface area, expressed as a percentage to facilitate comparison across regions. Morphological complexity, based on fractal dimension theory, is calculated as the logarithmic ratio of coastline length to area, quantitatively describing the geometric complexity of a coastline. Together, these three metrics constitute a multidimensional representation of coastline morphological change, providing a more comprehensive description of coastline dynamics than traditional single metrics.

[0152] It should be noted that in step S03, the spatiotemporal attention mechanism decomposes the attention calculation into two independent calculation processes, the time dimension and the space dimension, and then integrates the information of the two dimensions through weighted fusion. The time dimension attention focuses on the dependencies between different time points, and the space dimension attention focuses on the correlation between different spatial positions. This decomposition design enables the model to capture key information in both time and space dimensions at the same time, significantly improving the feature extraction capability. Weighting coefficient Through automatic optimization, the importance of the two dimensions can be flexibly adjusted according to the specific characteristics of the data. Compared with the traditional single-dimensional attention mechanism, this method can more effectively handle the complex dependencies of spatiotemporal data.

[0153] It should be noted that in step S04, the design of the fluctuation conversion 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 nonlinear mapping relationships. The activation function introduces nonlinear transformation capabilities and addresses the vanishing gradient problem. The least-squares loss function measures the difference between the predicted and measured metrics, and the regularization term prevents overfitting. This design transforms the abstract features extracted by deep learning into meaningful change metrics, improving the model's interpretability and practicality. Compared to traditional linear transformation methods, this method can capture more complex relationships between features and metrics.

[0154] It should be noted that in step S05, the coastal sediment transport rate It is proportional to the 2.5th power of the wave height and the sine of the incident angle, reflecting the influence of wave energy and incident angle on coastal sediment transport; vertical sediment transport rate It consists of two components: 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 shoreline position change rate equation is based on the principle of sediment mass conservation and calculates the rate of change of shoreline position using the spatial gradient of the alongshore sediment transport rate and the vertical sediment transport rate. This physical model can quantitatively describe the dynamic mechanisms of shoreline evolution and provide a physical explanation for shoreline changes. Compared with purely data-driven methods, this physical model has better interpretability and extrapolation capabilities.

[0155] It should be noted that in step S06, the design principle of the coastline evolution optimization function is to introduce physical constraints 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 is the data term, which ensures that the optimization result is close to the initial prediction; the second is the smoothing term, which prevents unnatural jagged fluctuations in the coastline; the third is the curvature term, which controls the curvature of the coastline; and the fourth 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 constraints of physical models, generating coastline predictions that are consistent with both data patterns and physical constraints. Compared with purely data-driven methods, this method can effectively suppress overfitting and improve the physical plausibility of the prediction.

[0156] It should be noted that in step S07, the stability index is calculated based on the signal-to-noise ratio concept, using 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. A higher stability index indicates that long-term trends dominate and the coastline is more stable; a lower index indicates that short-term fluctuations dominate and the coastline is more unstable. The calculation principle of the change frequency matrix is based on spectral analysis and spatial discretization. By calculating the energy distribution of different spatial locations in different frequency intervals, a two-dimensional matrix is constructed to describe the spatiotemporal variation characteristics of the coastline. These indicators provide a quantitative assessment of the dynamic characteristics of the coastline, which can help identify high-risk areas and key change patterns. Compared with traditional single-indicator assessment methods, this method provides a more comprehensive and detailed description of the variation characteristics.

[0157] It should be noted that in step S08, the design principle of the Bayesian model averaging method is based on the Bayesian statistical framework, which regards the model weight as a function of the model conditional probability. The weight calculation formula uses a negative exponential form, so that the model with better performance (smaller error) obtains a higher weight. The temperature parameter Control the "hardness" of weight distribution. Ensemble prediction uses a weighted average approach, 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 fully utilizes 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 predictions. Compared with simple averaging or single model methods, this method can provide more accurate prediction results and more reasonable uncertainty estimates.

[0158] Optionally, for the specific implementation of the CoastDynNet model, a deposition dynamics physical constraint layer can also be considered, which is expressed as follows:

[0159] ;

[0160] Where, is the coastal sediment transport rate, in units of ; is the sediment transport coefficient, ranging from 0.1 to 0.4; is the sediment particle size correction coefficient, ranging from 0.01 to 0.05; is the median particle size of sediment, in units of ; is the reference particle size, which is 0.2 ; is the wave height in the wave breaking zone, in units of ; is the wave incident angle in the wave breaking zone, in radians; is the tide level influence coefficient, ranging from 0.5 to 2.0; is the relative tide level, dimensionless.

[0161] Optionally, the vertical sediment transport model uses an improved formulation that takes into account wave-tidal interaction:

[0162] ;

[0163] Where, is the vertical sediment transport rate, in units of ; 、 、 are empirical coefficients, with values ranging from 0.01 to 0.05, 0.1 to 0.3, and 0.05 to 0.2; is the wave-current interaction coefficient, ranging from 0.5 to 1.5; is the Froude number, which indicates the tidal strength and is dimensionless; is the beach slope, dimensionless; is the coastal gradient of wave height; is the tidal velocity, in units of ; The angle between the tidal current and the coastline, measured in radians.

[0164] Optionally, the tidal influence attention module of the CoastDynNet model uses a time-varying weight calculation formula:

[0165] ;

[0166] Where, for Tidal influence weight at the moment; 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 ; is the bias vector.

[0167] Optionally, the morphological evolution prediction layer of the CoastDynNet model uses an improved one-dimensional ecological cellular automaton algorithm, and its state update rule is:

[0168] ;

[0169] Where, For spatial location In time the state of the coastline; is the neighborhood radius, usually ranging from 1 to 3; is the state transfer function, which is realized by neural network; External forcing conditions include waves, wind fields and other factors; and Position In time The coastal sediment transport rate and vertical sediment transport rate.

[0170] Optionally, for a specific implementation of the WaveletTempNet model, the mathematical expression of its multi-scale wavelet transform coding layer is:

[0171] ;

[0172] ;

[0173] Where, is the wavelet coefficient; is the input time series; is the wavelet basis function; is the scale parameter; is the translation parameter; express In actual implementation, the fast algorithm of discrete wavelet transform is used to calculate the wavelet coefficients.

[0174] Optionally, the time series feature extraction layer of the WaveletTempNet model uses a multi-head self-attention mechanism enhanced by positional encoding:

[0175] ;

[0176] ;

[0177] Where, 、 、 They are query matrix, key matrix and value matrix respectively; is the feature dimension; is the position encoding matrix; and is the position index in the sequence; is the scaling factor for position encoding, usually set to 0.1 times the sequence length.

[0178] Optionally, the long short-term memory network layer of the WaveletTempNet model uses a bidirectional LSTM structure with residual connections:

[0179] ;

[0180] ;

[0181] ;

[0182] Where, and are the hidden states of the forward and backward LSTM respectively; and They are forward and backward LSTM units respectively; for Input at the moment; and is the weight matrix; is the bias vector; Represents a vector concatenation operation.

[0183] Optionally, the reconstruction decoding layer of the WaveletTempNet model uses a weighted reconstruction formula:

[0184] ;

[0185] Where, is the reconstructed time series; is the long-term trend component; is the wavelet coefficient after processing; is the wavelet basis function; For the The scale reconstruction weight is obtained through model training; is the maximum scale level of decomposition.

[0186] Optionally, the loss function of WaveletTempNet includes a reconstruction error term, a decomposition consistency term, and a smoothing regularization term:

[0187] ;

[0188] Where, is the total loss; is the original input sequence; To reconstruct the sequence; and They are the theoretical wavelet decomposition and the model prediction. scaled wavelet coefficients; for Long-term trend value at the moment; 、 、 is the weight coefficient, which takes values of 1.0, 0.5 and 0.1 respectively; is the sequence length.

[0189] In summary, in this embodiment, starting from multi-source data fusion, through multi-level feature extraction and conversion, combined with physical models and deep learning methods, high-precision dynamic monitoring of coastline changes is finally achieved. 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 and difficulty in distinguishing long-term trends from short-term fluctuations caused by the long revisit period of traditional satellites, and significantly improve the accuracy and timeliness of coastline change monitoring. In order 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 each year. It is an ideal area for studying the impact of extreme meteorological events on the coastline.

[0190] First, a multi-source satellite data fusion network was constructed. Satellite imagery data from 2010 to 2022 was collected, including observations from multiple satellites, including Landsat 8 (16-day revisit period), Sentinel-2 (5-day revisit period), and GF-1 (4-day revisit period), totaling 457 images. After radiometric and geometric corrections were performed on these images, a temporal attention matrix was constructed using a multi-head self-attention mechanism. The specific satellite data sources and their characteristics are shown in Table 1.

[0191] Table 1 Multi-source satellite data characteristics

[0192]

[0193] The researchers used a fourth-order temporal convolutional neural network to extract the characteristics of coastline morphological changes. They set up four convolutional layers, with the temporal convolution kernel sizes of the first, second, third, and fourth layers being 3, 5, 7, and 9, respectively, and the number of feature channels being 32, 64, 128, and 256, respectively. The convolution results were then processed through a long-short-term memory network with a hidden layer dimension of 256, generating a first fluctuation matrix containing indicators such as position offset, area change rate, and morphological complexity. The calculation results show that the coastline change indicators before and after Typhoon Swan in the study area in 2020 are shown in Table 2:

[0194] Table 2 Coastline change indicators before and after Typhoon Swan

[0195]

[0196] Figure 2 This figure shows the changes in coastline position offset for five coastal segments (E01-E05) on eastern Hainan Island before and after Typhoon Swan from September to December 2020. The horizontal axis represents time (date), and the vertical axis represents coastline position offset (meters). The graph uses curves to illustrate the dynamic changes in each coastal segment, with negative values indicating landward coastline retreat. The observation dates from three different satellite sources, Landsat 8, Sentinel-2, and GF-1, are also marked in the figure, visually demonstrating the improved temporal resolution achieved by multi-source satellite data fusion. The Typhoon Swan event is marked with a vertical line, clearly illustrating the typhoon-induced rapid coastline retreat and subsequent recovery. As can be seen from the figure, segment E04 was most significantly affected by the typhoon, with a position offset of approximately -30 meters, while segment E03 was relatively less affected.

[0197] The first fluctuation matrix is reduced in dimension using the spatiotemporal attention mechanism, with 8-head time dimension attention and 4-head space dimension attention, and the weight coefficient Backpropagation automatically optimized the coefficient to 0.65, indicating that temporal characteristics have a more significant impact on coastline variation in this study area. Principal component analysis retained the top five principal components, accounting for 96.3% of the explained variance, forming a low-dimensional, high-information-density second fluctuation matrix.

[0198] To construct the fluctuation conversion matrix, a three-layer fully connected neural network was designed. The input layer had 5 nodes (corresponding to the dimensions of the second fluctuation matrix), the hidden layer had 128 nodes, and the output layer had 6 nodes. These corresponded to six indicators: annual average retreat rate, seasonal fluctuation amplitude, and extreme event response intensity. Comparing the optimized conversion matrix with data from 25 field measurement points, it achieved a root mean square error of 0.13 on the validation set, meeting the accuracy requirements. The quantitative indicators of shoreline change after conversion are shown in Table 3:

[0199] Table 3 Quantitative indicators of coastline change (2010-2022)

[0200]

[0201] The quantitative indicators are input into the pre-trained coastal dynamics model CoastDynNet together with wave and tide data, and the physical mechanism of coastline change is analyzed through the sediment transport equation. The value is 0.35, and the wave height in the measured wave breaking area is The average is 1.2 m, and the wave incident angle in the wave breaking zone is It is mainly concentrated in the range of 15°-35°. In the calculation of vertical sediment transport rate, The value is 0.03, The value is 0.21, the beach slope The average is 0.08. The analysis results show that the ratios of the alongshore 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 alongshore sediment transport process and correspond to areas with higher erosion risks.

[0202] Apply the coastline evolution optimization function to fine-tune the prediction results and optimize the weight coefficients in the function 、 、 、 The coefficients were set to 1.5, 0.3, 0.1, and 2.0, respectively. An iterative optimization using a gradient descent algorithm converged after 78 iterations, yielding a corrected coastline morphology that satisfied the physical constraints. For the E04 section, which is significantly affected by tides, an additional tidal correction factor of 1.15 was set, further improving prediction accuracy.

[0203] WaveletTempNet was used to perform a multiscale decomposition of the coastline change series. The model's wavelet transform encoding layer employed the Daubechies 4 wavelet basis function, the time series feature extraction layer employed a 12-head self-attention mechanism, and the long-short-term memory network layer had a hidden state dimension of 192. The decomposition results showed that the coastline of section E04 exhibited a long-term trend of retreat from 2010 to 2022, with an average retreat rate of 2.37 m / year. Short-term fluctuations were primarily concentrated during the typhoon season, with maximum amplitudes reaching 20.5 m.

[0204] Figure 3 The multiscale decomposition of coastline change in section E04 is presented, including four components: overall change, long-term trend, seasonal fluctuation, and response to extreme events, as well as changes in the stability index. Four major typhoon events and their intensity are annotated in the figure, clearly illustrating the coastline's retreat and recovery after each extreme event. Figure 4 The spectrum analysis results show the spectrum energy distribution and main change frequency of the five coastal sections. The horizontal axis represents the change frequency (times / year), and the vertical axis represents the spectrum energy ( The figure shows the main frequency and stability index of each coastal segment, and divides it into monsoon-dominated intervals (period > 6 months) and typhoon-dominated intervals (period < 4 months). As can be seen from the figure, segment E04 has the highest main frequency (4.6 times / year) and the lowest stability index (0.47), indicating that this section of the coast is significantly affected by typhoons and tides and changes most dramatically. The stability index and main change frequency of the five coastal segments calculated by the spectrum analysis module are shown in Table 4:

[0205] Table 4. Coastal section stability analysis results

[0206]

[0207] Finally, a multi-model integration framework is constructed using the Bayesian model averaging method to integrate the outputs of the physical model (CoastDynNet) and the deep learning model (WaveletTempNet). The weights for the physical model and deep learning model were set to 2.5, based on the performance of each model on the historical validation set. The ensemble model achieved a root mean square error of 0.78 m, a 37.6% and 45.5% reduction compared to the errors of the deep learning model (1.25 m) and the physical model (1.43 m) used alone, respectively.

[0208] Traditional coastline change monitoring methods mainly rely on a single satellite data source, such as the Landsat series of satellites, whose 16-day revisit period cannot capture short-term changes, especially mutations caused by extreme events such as typhoons. At the same time, traditional methods often use simple time series analysis techniques, which make it difficult to effectively distinguish long-term trends from short-term fluctuations. The present invention significantly improves the temporal resolution through a multi-source satellite data fusion network, and achieves an average effective observation interval of 3.5 days in the embodiment, successfully capturing the impact of extreme events such as Typhoon Swan on the coastline. At the same time, the WaveletTempNet model realizes the multi-scale decomposition of the coastline change sequence, accurately separating long-term trends 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 physical models and data-driven models, and the prediction accuracy is improved by more than 40% compared with the traditional single model method, providing more reliable technical support for coastal planning and management.

[0209] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5, 6 and 7 below.

[0210] Table 5 Variable Explanation Table (Part I)

[0211]

[0212] Table 6 Variable Explanation Table (Part II)

[0213]

[0214] Table 7 Variable Explanation Table (Part 3)

[0215]

[0216] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A deep learning-driven method for dynamic monitoring of coastline changes, characterized by: include: S01. Build 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; S02. Apply a fourth-order temporal convolutional neural network to extract the coastline morphological change characteristics and establish the first wave matrix through a recursive method; S03. Use the spatiotemporal 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 and use the sediment transport equation to analyze the physical mechanism of coastline change; S06. Apply the coastline evolution optimization function to fine-tune the time interpolation results to improve the short-term prediction accuracy; S07. Use a pre-trained time series decomposition network to perform multi-scale decomposition of the coastline change series to separate long-term trends from short-term fluctuations. S08. Design a spectrum analysis module to calculate the coastline stability index and change frequency matrix based on the separation results; S09. Use the Bayesian model averaging method to build a multi-model integration framework, integrate the outputs of the physical model and deep learning model according to the weighted prediction error, and generate a comprehensive coastline dynamic assessment report; The sediment transport equation is as follows: ； ； ； Where, is the coastal sediment transport rate, is the vertical sediment transport rate, 、 、 is the empirical coefficient, is the wave height in the wave breaking zone, is the wave incident angle in the wave breaking zone, is the beach slope, is the coastal gradient of wave height; is the rate of change of coastline position, is the critical water depth for wave action, is the coastal distance coordinate, is the distance coordinate perpendicular to the baseline, For time; Among them, the coastline evolution optimization function is as follows: ； ； ； Where, Optimize functions for coastline evolution; is the optimized coastline position vector; is the preliminary predicted coastline position vector; is the spatial gradient of coastline position; 、 、 、 is the weight coefficient; is the physical constraint condition; and are the maximum and minimum allowable slopes of the coastline, respectively; is the final coastline position vector after optimization; is the number of discrete sampling points along the coastline; For the The coastline locations of the sampling points.

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, which 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 is 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.

5. The deep learning-driven coastline change dynamic monitoring method according to claim 4 is characterized in that: The coastal dynamics model is the CoastDynNet model, which has a multi-layer encoding and decoding architecture and 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 alongshore 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.

6. The deep learning-driven coastline change dynamic monitoring method according to claim 5 is characterized in that: The time series decomposition network is a WaveletTempNet model, which 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 variation 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.

7. The deep learning-driven coastline change dynamic monitoring method according to claim 6 is characterized in that: The stability index is an indicator that quantifies the coastline's ability to resist external interference, and is calculated by 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 position in a certain frequency interval.

8. The deep learning-driven coastline change dynamic monitoring method according to claim 7 is characterized in that: The Bayesian model averaging method is a statistical method that performs weighted averaging of prediction results of multiple models based on the Bayesian reasoning principle.

Images