Island shoreline erosion risk prediction method based on computer vision

Abstract

The invention provides an island shoreline erosion risk prediction method based on computer vision, and belongs to the technical field of island shoreline erosion risk analys.The island shoreline erosion risk prediction method comprises the steps that firstly, a high-resolution remote sensing image is obtained, island contour information is extracted, and a shoreline initial model is constructed through a small grid approximation method; generating multi-level grid representation through grid merging operation and storing the multi-level grid representation as a shoreline feature matrix; then collecting a multi-time sequence remote sensing image, calculating a change matrix set of adjacent time points, and analyzing a shoreline evolution mode; establishing a coast erosion kinetic equation model by combining the wave action intensity, the tide amplitude, the sea level rising rate, the geological structure coefficient and the shoreline slope angle; a shoreline change rule is analyzed by applying a map attention grid evolution model, and finally, an uncertainty analysis module is introduced, a multi-scale erosion risk assessment map is generated, a high-risk area is identified and output as a prediction result, so that the technical problem that the complicated island shoreline erosion risk is difficult to accurately predict is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of island shoreline erosion risk analysis. Specifically, it relates to a method for predicting island shoreline erosion risk based on computer vision. Background Art

[0002] Island shoreline erosion is an important environmental issue under the background of global climate change. Traditional coastal erosion risk assessments mainly rely on field monitoring and mathematical model simulations. Common methods include erosion rate calculation based on empirical formulas, shoreline change analysis based on remote sensing images, and hydrodynamic model simulations, etc. These methods have played an important role in large-scale coastal zone management and protection, especially in disaster prevention and mitigation work in island countries and densely populated coastal areas.

[0003] However, there are obvious defects in the application of traditional technologies. Existing remote sensing image analysis methods usually use simple linear representations of shorelines and are difficult to accurately capture the details of complex and tortuous island shorelines; although physical models consider factors such as waves and tides, they usually ignore the influence of geological structure differences and local topographic features; statistical models based on historical data lack effective expression of multi-scale changes, resulting in insufficient accuracy of prediction results in local areas.

[0004] Facing islands with different shoreline characteristics and change laws, existing technologies are difficult to simultaneously consider the overall changes of large-scale shorelines and the local erosion details of small scales. Especially under complex topographic conditions, they cannot effectively balance the importance of global and local features, thereby affecting the prediction accuracy. That is to say, there is a technical problem in the existing technology that it is difficult to accurately predict the erosion risk of complex island shorelines. This core technical problem seriously restricts the accurate assessment of island shoreline erosion risk and the implementation of effective protection measures. Summary of the Invention

[0005] In view of this, the present invention provides a method for predicting island shoreline erosion risk based on computer vision, which can solve the technical problem in the existing technology that it is difficult to accurately predict the erosion risk of complex island shorelines.

[0006] The present invention is implemented as follows: The present invention provides a method for predicting the erosion risk of island shorelines based on computer vision, which includes obtaining high-resolution remote sensing images of islands and extracting island contour information; constructing an initial model of the island shoreline based on the small grid approximation method; performing a merging operation on the small grids to generate a multi-level grid representation; collecting multiple time-series remote sensing images and calculating the change in the shoreline feature matrix; establishing a coastal erosion dynamics equation model, combining the wave action intensity, tidal amplitude, sea level rise rate, island geological structure coefficient, and shoreline slope angle, to quantitatively calculate the erosion rate at different grid positions; applying a graph attention grid evolution model to analyze the shoreline change law, where the graph attention grid evolution model adjusts the attention distribution mechanism through the grid area weight coefficient to balance global changes and local details; introducing an uncertainty analysis module, calculating the confidence interval of the prediction results, generating a multi-scale erosion risk assessment map, identifying high-risk areas, and outputting the prediction results.

[0007] Among them, obtaining high-resolution remote sensing images of islands and extracting island contour information includes denoising, enhancing, registering, and segmenting the images, and using machine vision algorithms to extract island contour information.

[0008] Among them, constructing an initial model of the island shoreline based on the small grid approximation method includes dividing the island boundary into multiple square grids with a side length of the first dimension value to form an island shoreline grid set.

[0009] Among them, the small grid approximation method represents the island boundary as a set of square grids with equal side lengths, and realizes the accurate expression of complex shoreline forms by adjusting the grid density. Each grid records the existence state and attribute information of the shoreline in the area.

[0010] Among them, the first dimension value is the initial grid side length determined according to the complexity of the island shoreline and the resolution of the remote sensing image, and the numerical range is from 5 meters to 50 meters.

[0011] Among them, performing a merging operation on the small grids to generate a multi-level grid representation includes finding the globally largest square grid and storing all grid information as a shoreline feature matrix.

[0012] Among them, the shoreline feature matrix is a two-dimensional data structure, and its element values represent the attributes of the corresponding position grids, including whether it is a shoreline point, shoreline type, elevation value, and stability index. The matrix dimension is determined by the grid division accuracy.

[0013] Among them, the coastal erosion dynamics equation is used to simulate the physical mechanism in the process of coastal erosion. Based on the principle of mass conservation and fluid dynamics theory, considering the balance relationship between the impact force of waves on the shoreline and the resistance of shoreline materials, the erosion rate distribution of grids at different positions is obtained through integral calculation.

[0014] Among them, the structure of the spectral attention grid evolution model is a fusion architecture of a multi-layer graph convolutional neural network and a multi-head attention mechanism. The shoreline grid is represented as graph-structured data, and the connection relationship between grids corresponds to edges. The spectral attention grid evolution model includes an encoding layer, a multi-layer graph convolutional layer, a multi-head attention layer, and a decoding layer.

[0015] Among them, the core of the spectral attention grid evolution model is to introduce a grid area adaptive attention allocation mechanism. The grid area adaptive attention allocation mechanism dynamically adjusts the attention distribution through three key parameters: the grid density parameter, the grid merging level, and the historical evolution rate of the change matrix in the shoreline feature matrix, so as to effectively capture the shoreline change characteristics at different scales.

[0016] The present invention constructs an initial shoreline model through a small grid approximation method, and combines the spectral attention grid evolution model to achieve accurate prediction of the erosion risk of island shorelines. This method can not only accurately express complex shoreline shapes, but also adaptively allocate attention resources according to the grid size, effectively solving the problem of insufficient accuracy caused by simple linear representation in traditional technologies.

[0017] The present invention comprehensively considers multi-dimensional factors such as wave action intensity, tidal amplitude, sea level rise rate, geological structure coefficient, and shoreline slope angle, and establishes a more comprehensive physical model through the coastal erosion dynamics equation. In particular, the introduced spectral attention grid evolution model can automatically balance the global changes of large-area grids and the local details of small-area grids, and dynamically adjust the attention distribution through the grid area adaptive attention allocation mechanism, overcoming the deficiencies of traditional methods in expressing multi-scale changes.

[0018] Through the multi-level grid representation and area weight adjustment mechanism, the present invention successfully solves the technical problem of difficult to accurately predict the erosion risk of complex island shorelines in the prior art, providing a more reliable scientific basis for the protection of island shorelines and disaster risk management. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0020] Figure 2 It is a schematic diagram of the multi-level grid representation in Embodiment 2.

[0021] Figure 3 It is a graph showing the relationship between the shoreline erosion rate and wave action intensity in different regions in Embodiment 2. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] As shown Figure 1 in the figure, it is a flowchart of a method for predicting the erosion risk of island shorelines based on computer vision provided by the present invention. This method includes the following steps: S01. Obtain high-resolution island remote sensing images and perform image preprocessing, including denoising, enhancement, registration, and segmentation, and extract island contour information using machine vision algorithms; S02. Construct an initial model of the island shoreline based on the small grid approximation method, divide the island boundary into multiple square grids with a side length of the first dimension value to form an island shoreline grid set; S03. Perform a merging operation on the small grids, find the global maximum square grid, generate a multi-level grid representation, and store all grid information as a shoreline feature matrix; S04. Collect island remote sensing images of multiple time series, repeat steps S01 to S03 for each time point to obtain the shoreline feature matrix corresponding to the time point; S05. Calculate the change matrix of the shoreline feature matrix at adjacent time points, construct a time series change matrix set, and analyze the shoreline evolution pattern; S06. Establish a coastal erosion dynamics equation model, combine the wave action intensity, tidal amplitude, sea level rise rate, island geological structure coefficient, and shoreline slope angle to quantitatively calculate the erosion rate at different grid positions; S07. Apply the graph attention grid evolution model to analyze the shoreline change law. The graph attention grid evolution model adjusts the attention distribution mechanism through the grid area weight coefficient to optimize the feature extraction efficiency; S08. Introduce an uncertainty analysis module, calculate the confidence interval of the prediction result, generate a multi-scale erosion risk assessment map, identify high-risk areas and output them as the prediction result of the island shoreline erosion risk.

[0024] Among them, the small grid approximation method specifically represents the island boundary as a set of square grids with equal side lengths, and realizes the accurate expression of complex shoreline forms by adjusting the grid density. Each grid records the existence state and attribute information of the shoreline in the area.

[0025] Among them, the first dimension value is specifically the initial grid side length determined according to the complexity of the island shoreline and the resolution of the remote sensing image, and the numerical range is 5 meters to 50 meters, which is used to ensure that the grid division can capture shoreline details and maintain calculation efficiency.

[0026] Among them, the large area grid specifically refers to a grid unit with an area greater than 625 square meters formed by merging multiple adjacent small grids, which is usually used to represent relatively straight shoreline areas and obtains a higher global change weight in the graph attention grid evolution model.

[0027] Among them, the small - area grid specifically refers to the basic grid unit with an area less than 100 square meters, which is usually distributed in areas with a large shoreline curvature or complex morphology, and is mainly responsible for the accurate expression of local detail changes in the atlas attention grid evolution model.

[0028] Among them, the shoreline feature matrix is specifically a two - dimensional data structure, and its element values represent the attributes of the corresponding position grid, including whether it is a shoreline point, shoreline type, elevation value, and stability index. The matrix dimension is determined by the grid division accuracy.

[0029] Among them, the wave action intensity is specifically a physical quantity representing the impact energy of waves on the shoreline, with the unit of kilowatt per meter, which is calculated from the wave height, period, and incident angle, and is an important input parameter of the coastal erosion dynamics equation model.

[0030] Among them, the tidal amplitude specifically refers to the maximum value of the water - level change within the tidal cycle, with the unit of meter, which is obtained from measured data or tidal prediction models and affects the vertical distribution range of the shoreline eroded area.

[0031] Among them, the sea - level rise rate specifically refers to the height of the regional sea - level rise per year, with the unit of millimeter per year, which is derived from long - term sea - level monitoring data and is a key parameter for predicting long - term shoreline changes.

[0032] Among them, the island geological structure coefficient is specifically a dimensionless parameter quantifying the erosion resistance of the island shoreline material, which is comprehensively determined according to the rock hardness, soil structure, and vegetation coverage, and the numerical range is from 0 to 1. The larger the value, the stronger the erosion resistance.

[0033] Among them, the shoreline slope angle specifically refers to the angle formed by the shoreline and the horizontal plane, with the unit of degree, which is obtained by calculating elevation data and affects the distribution pattern of wave energy on the shoreline and the erosion efficiency.

[0034] Among them, the coastal erosion dynamics equation is used to simulate the physical mechanism in the coastal erosion process. The inputs include wave action intensity, tidal amplitude, sea - level rise rate, island geological structure coefficient, and shoreline slope angle, and the outputs are the shoreline retreat distance and the volume of eroded material per unit time. The coastal erosion dynamics equation is based on the principle of mass conservation and fluid dynamics theory, considering the balance relationship between the impact force of waves on the shoreline and the resistance of shoreline materials, and calculates the erosion rate distribution of grids at different positions through integration.

[0035] Among them, the grid area weight coefficient specifically refers to the weight parameter for allocating attention resources according to the grid area size in the atlas attention grid evolution model, which is calculated by the ratio of the grid area to the total area and is used to balance the influence degrees of global changes and local details in prediction.

[0036] Among them, the specific structure of the atlas attention grid evolution model is a fusion architecture of a multi-layer graph convolutional neural network and a multi-head attention mechanism. The shoreline grid is represented as graph-structured data, and the connection relationship between grids corresponds to edges. The atlas attention grid evolution model includes an encoding layer, a multi-layer graph convolutional layer, a multi-head attention layer, and a decoding layer. Among them, the multi-head attention mechanism assigns different weights according to the grid area. Larger-area grids obtain higher attention weights in global change prediction, and smaller-area grids maintain accurate expression in local detail change prediction. The core of the atlas attention grid evolution model is to introduce a grid area adaptive attention allocation mechanism. The grid area adaptive attention allocation mechanism dynamically adjusts the attention distribution through three key parameters: the grid density parameter, the grid merging level, and the historical evolution rate of the change matrix in the shoreline feature matrix, so as to effectively capture shoreline change features at different scales.

[0037] Among them, the grid density parameter specifically refers to the number of grids per unit area, with the unit of number per square kilometer, which is calculated from the first size value and is the basic parameter for adjusting the grid area adaptive attention allocation mechanism.

[0038] Among them, the grid merging level specifically refers to the number of levels at which small grids are gradually merged into large grids, with a numerical range of 1 to 10, which is used to characterize the complexity of the grid structure and affects the attention allocation strategy.

[0039] Among them, the historical evolution rate of the change matrix specifically refers to the speed of change of the shoreline feature matrix between adjacent time points, with the unit of square meters per year, which is obtained through historical data analysis and is an important basis for predicting future change trends.

[0040] Among them, the steps for establishing the training data set of the atlas attention grid evolution model specifically include collecting historical remote sensing image sequences of typical island shorelines around the world and annotating the change situations; preprocessing the original images and constructing grid representations according to unified standards; calculating the change matrix between different time points and annotating the influencing factors; constructing supervised learning labels according to the measured erosion results; dividing the data into training sets, validation sets, and test sets to ensure balanced data distribution and coverage of various island types and climate conditions.

[0041] Among them, the steps for training the atlas attention grid evolution model specifically include initializing the model parameters; optimizing the model parameters according to the mini-batch stochastic gradient descent method; using the cross-entropy loss function to evaluate the difference between the prediction result and the true label; introducing regularization techniques to prevent overfitting; implementing an early stopping strategy to avoid overtraining; setting weight coefficients for different island types respectively; improving the generalization ability of the model in new island scenarios through transfer learning methods; freezing the parameters of the feature extraction layer after training and retaining the key layer parameters for fine-tuning.

[0042] Among them, the multi-scale erosion risk assessment map is specifically a risk distribution map generated based on different spatial scales and time spans, representing the erosion risk degree through color gradients, and realizing the intuitive display and spatial analysis of risk areas by combining geographic information system technology.

[0043] The specific implementation manners of the above steps are described in detail below.

[0044] The specific implementation manner of step S01 is to extract the island contour information by obtaining high-resolution island remote sensing images and performing a series of image preprocessing operations. First, the wavelet transform denoising algorithm is used to denoise the remote sensing image to eliminate sensor noise and atmospheric interference; then, the histogram equalization method is applied to enhance the image contrast to make the boundary between the island and the sea clearer; next, the feature point matching algorithm is used for multi-temporal image registration to ensure the accurate correspondence of the spatial positions of images at different time points; finally, the OTSU segmentation algorithm based on threshold is used to separate the island and sea areas, and the threshold selection range is between 0.4 and 0.6, and the morphological opening and closing operations are combined to optimize the segmentation result; on this basis, the Canny edge detection algorithm is used to extract the island contour, and the high and low threshold ratio of edge detection is 2:1, and the high threshold is recommended to be set to 1.5 times the average value of the image gradient. The purpose of this step is to provide accurate island boundary information for subsequent shoreline grid division to ensure the accuracy and reliability of shoreline extraction.

[0045] The specific implementation manner of step S02 is to construct an initial model of the island shoreline by using the small grid approximation method. First, a suitable first size value is determined as the grid side length, and according to the complexity of the island shoreline and the resolution of the remote sensing image, a suitable value is selected within the range of 5 meters to 50 meters. For complex shorelines, 5 to 15 meters is appropriate, and for straight shorelines, 20 to 50 meters can be selected; then, based on the island contour extracted in step S01, sampling points are set at fixed intervals along the shoreline, and the sampling interval is equal to 1 / 2 of the first size value; next, with each sampling point as the center, a square grid with a side length of the first size value is generated, and it is judged whether the grid contains shoreline points. If it contains, the grid is included in the shoreline grid set; finally, the grids are sorted and numbered, and the topological relationship between the grids is established, and the spatial position and adjacent grid information of each grid are recorded. The purpose of this step is to discretize the continuous island shoreline into a grid representation that is convenient for computer processing, laying a foundation for subsequent multi-level grid construction.

[0046] The specific implementation of step S03 is to perform a merging operation on small grids to generate a multi-level grid representation. First, apply the maximum square detection algorithm to find adjacent grids that can be merged from the initial grid set. The judgment condition is that the similarity of adjacent grid attributes is higher than 0.85 and a larger square is formed after merging. Then, perform grid merging according to the bottom-up strategy. Small grids are preferentially merged into medium grids, and medium grids are further merged into large grids, ensuring that the side length of the merged grid is an integer multiple of the side length of the initial grid. Next, construct a hierarchical grid tree structure to record the inclusion relationship between grids, with large grids as parent nodes and the included small grids as child nodes. Finally, convert all grid information into a shoreline feature matrix. The matrix dimension corresponds to the size of the original remote sensing image, and the matrix elements record the attribute information of the grids at the corresponding positions, including grid type (shoreline / non-shoreline), grid size (small / medium / large), elevation value, and stability index (range 0 to 1). The purpose of this step is to reduce the redundancy of the shoreline representation while retaining the detailed information of key areas and improving the subsequent calculation efficiency.

[0047] The specific implementation of step S04 is to collect remote sensing images of islands in multiple time series and repeat the operation process of steps S01 to S03 for each time point. First, collect a sequence of remote sensing images of islands with a time span of at least 5 years, and the time interval is recommended to be 3 months to 1 year to ensure the capture of seasonal and interannual changes. Then, apply the preprocessing method of step S01 to each time-series image to ensure that all image processing standards are consistent. Next, perform the small grid approximation method of step S02 on each processed image to construct the initial shoreline model for the corresponding time point. Finally, apply the grid merging strategy of step S03 to generate the multi-level grid representation and shoreline feature matrix for each time point, with the matrix dimension remaining consistent for subsequent comparative analysis. The purpose of this step is to construct a time-series shoreline feature data set to provide basic data for analyzing the evolution law of the shoreline.

[0048] The specific implementation of step S05 is to calculate the change matrix of the shoreline feature matrix at adjacent time points and construct a set of time series change matrices. First, the difference between the shoreline feature matrices at adjacent time points is calculated using matrix difference operation to obtain the initial change matrix. Then, a spatial filtering algorithm is applied to remove the noise in the change matrix, and the filter window size is set from 3×3 to 5×5, and the change signals with a confidence level exceeding 0.75 are retained. Next, the change rates at each grid position are calculated, including the area change rate and the morphological change rate, and the significantly changed areas are marked, and the change rate threshold is set to ±10% of the value at the previous moment. Finally, the change matrices corresponding to all time points are organized into a time series data structure to analyze the shoreline evolution pattern, identify the erosion acceleration area, the erosion slowdown area and the stable area, and the determination criteria are that the average annual recession distance is greater than 1 meter, the average annual recession distance is less than 0.2 meter, and the average annual change distance is less than 0.05 meter respectively. The purpose of this step is to quantify the spatio-temporal change characteristics of the shoreline and provide a basis for dynamic changes for subsequent erosion risk prediction.

[0049] The specific implementation of step S06 is to establish a coastal erosion dynamics equation model to quantitatively calculate the erosion rate at different grid positions. First, the model input parameters are collected, including the wave action intensity (unit: kilowatts per meter, typical value range 1 to 15), the tidal amplitude (unit: meter, typical value range 0.5 to 5), the sea level rise rate (unit: millimeters per year, typical value range 2 to 10), the island geological structure coefficient (dimensionless, value range 0 to 1), and the shoreline slope angle (unit: degree, typical value range 5 to 45). Then, based on the principle of mass conservation and fluid dynamics theory, a dynamics equation describing the shoreline erosion process is established, and this equation considers the balance relationship between wave energy input and geological structure resistance. Next, the erosion equation is applied to calculate the erosion rate for different types of grids. A simplified calculation mode is adopted for large-area grids, and a refined calculation mode is adopted for small-area grids, and the calculation accuracy threshold is set to 0.1 meter per year. Finally, a shoreline grid erosion rate distribution map is obtained, which intuitively represents the erosion risk levels at different positions. The purpose of this step is to quantify the erosion process through a physical model and provide theoretical support for risk prediction.

[0050] The specific implementation of step S07 is to analyze the law of shoreline change by applying the atlas attention grid evolution model. First, convert the shoreline grid data into a graph structure representation, with the grid center points as graph nodes and the connection relationships between adjacent grids as graph edges. Then, construct an atlas attention network model, including an encoding layer, multiple graph convolutional layers (the number of layers is 3 to 5), a multi-head attention layer (the number of heads is 4 to 8), and a decoding layer. Next, introduce an attention allocation mechanism based on grid area. Larger area grids (greater than 625 square meters) are assigned higher global attention weights (weight coefficients are 0.6 to 0.8), and smaller area grids (less than 100 square meters) focus on local detail expression (weight coefficients are 0.2 to 0.4). Finally, use the trained model to predict the future evolution trend of the shoreline, and output three core indicators including grid position change, morphological change, and rate change. The purpose of this step is to integrate data-driven and physical model methods to improve the accuracy and adaptability of prediction.

[0051] The specific implementation of step S08 is to introduce an uncertainty analysis module, calculate the confidence interval of the prediction results, and generate a multi-scale erosion risk assessment map. First, use the Monte Carlo simulation method to randomly perturb the key input parameters, with the perturbation range being ±2 times the standard deviation of the parameters, and generate 1000 to 5000 groups of simulation parameters. Then, run the atlas attention grid evolution model to obtain multiple groups of prediction results, and calculate the probability distribution of the predicted values at each grid position. Next, determine the 90% confidence interval based on the probability distribution, and visualize the results as a risk level map, using a 5-level risk level classification (extremely high risk, high risk, medium risk, low risk, extremely low risk), and the risk thresholds are an annual erosion rate greater than 2 meters, 1 to 2 meters, 0.5 to 1 meter, 0.1 to 0.5 meter, and less than 0.1 meter respectively. Finally, generate a multi-scale erosion risk assessment map, including short-term (1 to 3 years), medium-term (3 to 10 years), and long-term (10 to 30 years) risk assessments, and identify high-risk areas that need to be protected first. The purpose of this step is to quantify the uncertainty of the prediction and provide risk decision support.

[0052] The detailed structure of the atlas attention grid evolution model and the specific implementation of establishing its training dataset are as follows: The core structure of the atlas attention grid evolution model adopts an architecture that combines a graph neural network and an attention mechanism, including four main components. First, the encoding layer is responsible for converting the input shoreline feature matrix into a graph structure representation. Each grid corresponds to a graph node, and the node features include grid type, area, position coordinates, and historical change rate. Adjacent grids are connected by edges, and the edge weights are calculated based on the similarity between grids. The similarity threshold is set to 0.7. Then, the multi-layer graph convolutional layer consists of 3 layers of graph convolutional networks, with 64 convolutional kernels in each layer. The size of the convolutional kernel is 3×3, and the activation function uses ReLU. Batch normalization and residual connections are added after each layer to capture the spatial dependencies between grids. Next, the multi-head attention layer contains 8 attention heads, and each attention head independently learns the importance distribution of different feature subspaces. The attention weights are adaptively adjusted by the grid area, and the adjustment coefficient is the logarithmic ratio of the area to the reference area (100 square meters) to balance global changes and local details. Finally, the decoding layer converts the graph representation back to the grid matrix format and outputs the predicted changes at each grid position, including position offset and morphological changes. The specific implementation of establishing the training dataset includes five core steps: First, collect historical remote sensing image sequences of 50 to 100 typical island shorelines globally, with a time span of no less than 10 years, a spatial resolution better than 5 meters, and covering different climate conditions and geological types. Second, perform standardized preprocessing on all images, including atmospheric correction, geometric correction, and radiometric correction, to ensure the consistency of data quality. Third, construct a unified grid representation, with the grid size being comparable among different islands, and record the grid attribute information at each time point. Fourth, calculate the change matrix of adjacent time points and collect environmental factor data for the corresponding time period, including wave, tide, and meteorological data. Fifth, label the supervised learning labels according to the measured shoreline change results and divide the training set, validation set, and test set according to the ratio of 7:2:1 to ensure balanced and representative data distribution. The model is trained using the mini-batch stochastic gradient descent method, with the batch size set to 32 to 64, the initial learning rate set to 0.001, and the cosine annealing strategy is applied for dynamic adjustment. The number of training epochs is 100 to 200 epochs, and the early stopping patience parameter is set to 20 epochs.

[0053] The following details the mathematical models or calculation processes involved in the present invention.

[0054] In step S05, calculate the change matrix of the shoreline feature matrix at adjacent time points, and its specific representation is as follows: ; In the formula, is the change matrix corresponding to time ; is the time The shoreline feature matrix; At time The shoreline feature matrix.

[0055] The calculation formula of the change rate matrix is as follows: ; In the formula, Is the change rate matrix; Is the time interval, with the unit of year; The zero elements in Are replaced by a constant Less than the minimum non-zero value (it is recommended to take the value of

[0056] In step S06, the core formula of the coastal erosion dynamics equation model is expressed as follows: ; In the formula, Is the erosion rate at the grid position , with the unit of m / year; Is the wave action intensity, with the unit of kW / m; Is the tidal amplitude, with the unit of m; Is the sea level rise rate, with the unit of mm / year; Is the island geological structure coefficient, dimensionless, with the value range from 0 to 1; Is the shoreline slope angle, with the unit of degree; Is the erosion coefficient, with the default value of 0.25, calibrated through historical data; Is the geological resistance coefficient, with the default value of 0.75, obtained through experiments; Is the stability factor, with the default value of 0.05, used to avoid the denominator being zero.

[0057] The acquisition methods of the parameters in the erosion equation are as follows: Obtained through wave observation data or wave models, and the calculation formula is: ; In the formula, Is the seawater density, with the default value of 1025 kg / m³; Is the gravitational acceleration, with the default value of 9.8 m / s²; Is the significant wave height, with the unit of m; Is the wave group velocity, with the unit of m / s; Is the wave incident angle, with the unit of degree.

[0058] Obtained from tidal observation station data or tidal prediction models, and can be obtained by spatial interpolation of data from the nearest tidal station. The calculation formula is: ; In the formula, is the tidal amplitude of the th tidal station; is the spatial weight, and the calculation formula is: ; In the formula, is the grid position to the th tidal station, and the unit is kilometers.

[0059] Obtained from sea level monitoring data, and the calculation formula is: ; In the formula, is the global average sea level rise rate, and the default value is 3.3 mm / year; is the regional sea level change, obtained from satellite altimetry data; is the local sea level change, considering the influence of crustal vertical movement and human activities, and obtained from land survey and GPS data.

[0060] Obtained from geological surveys and rock hardness tests, and the calculation formula is: ; In the formula, is the rock hardness factor, with a range of 0 to 1; is the soil structure factor, with a range of 0 to 1; is the vegetation cover factor, with a range of 0 to 1; is the weight coefficient, satisfying , and the default values are 0.5, 0.3, and 0.2 respectively.

[0061] Calculated from elevation data, and the formula is: ; In the formula, is the elevation difference in the grid along the coastline direction, and the unit is meters; is the corresponding horizontal distance, and the unit is meters; is the coefficient for converting radians to degrees.

[0062] In step S07, the calculation formula for the grid area weight coefficient of the atlas attention grid evolution model is: ; In the formula, is the area weight coefficient of the grid ; is the grid area, with the unit of square meters; is the minimum grid area, with the default value of 25 square meters; is the maximum grid area, with the default value of 625 square meters; is the minimum weight, with the default value of 0.2; is the maximum weight, with the default value of 0.8.

[0063] The attention distribution calculation formula of the grid area adaptive attention allocation mechanism is: ; In the formula, is the attention score of the grid ; is the grid density parameter weight, and the calculation formula is: ; In the formula, is the grid density of the area where the grid is located, with the unit of number per square kilometer; is the optimal grid density, with the default value of 400 number per square kilometer; is the adjustment parameter, with the default value of 200.

[0064] is the historical change rate weight, and the calculation formula is: ; In the formula, is the historical change rate of the grid , with the unit of square meters per year; is the weight enhancement coefficient, with the default value of 0.5; is the hyperbolic tangent function, used to limit the weight change range; is the scaling factor, with the default value of 0.1.

[0065] In step S08, the calculation formula of the risk score is: ; In the formula, is the comprehensive risk score of the grid ; respectively represent the short-term ( ), medium-term ( ), and long-term ( ) risk scores; is the time scale weight, satisfying , and the default values are 0.5, 0.3, and 0.2 respectively.

[0066] The calculation formula for the risk score of each time scale is as follows: ; In the formula, is the probability of occurrence, with a range of 0 to 1; is the impact intensity, with a range of 0 to 10; is the vulnerability, with a range of 0 to 1.

[0067] The calculation formula for the confidence interval is: ; In the formula, is the confidence interval of the grid erosion rate; is the average predicted value of the erosion rate, with the unit of m / year; is the standard deviation; is the critical value of the standard normal distribution with a significance level of , and the corresponding to the 95% confidence interval is 1.96.

[0068] The reason for using the power relationship in the coastal erosion dynamics equation is that the wave energy is proportional to the square of the wave height, which conforms to the basic principles of hydrodynamics; the 0.5th power of the tidal amplitude indicates that the tidal influence increases with the increase of the amplitude, but the growth rate slows down, which is consistent with the actual observation results; the 0.3rd power of the sea - level rise rate indicates that its promoting effect on erosion exists, but it is not as significant as the wave action, which is in line with the research results of coastal dynamics. The cosine function is used in the denominator to describe the influence of the slope angle, reflecting the phenomenon that the steep coastline receives less wave energy and the erosion rate is relatively low. Adding the stability factor avoids the numerical instability problem caused by the denominator approaching zero. This equation comprehensively considers the marine dynamic factors (waves, tides) and geological factors (geological structure, slope), and can more comprehensively describe the physical mechanism of the coastal erosion process.

[0069] The logarithmic relationship is adopted for the grid area weight coefficient because the range of grid area changes is large. A direct linear relationship will lead to too high weights for large - area grids. After logarithmic transformation, it can more reasonably allocate attention resources, enabling the model to pay attention to both large - scale change trends and not ignore local details. The Gaussian function form is adopted for the grid density parameter weight, reflecting the characteristic that there is an optimal value for the grid density. Too high or too low grid density will reduce the model performance. The hyperbolic tangent function is adopted for the historical change rate weight, which can give more attention to the areas with significant historical changes while keeping the function values bounded. This helps the model capture key change areas.

[0070] The risk scoring model adopts the form of the product of three factors: probability, impact intensity, and vulnerability, which conforms to the basic framework of risk assessment and can comprehensively consider the possibility, severity, and tolerance of erosion events. The confidence interval calculation adopts the standard normal distribution hypothesis, and the distribution characteristics of the prediction results are obtained through Monte Carlo simulation, providing quantitative support for decision-making regarding uncertainty.

[0071] Shoreline feature matrix The elements are represented as follows: ; In the formula, is the eigenvector at time at position ; is the grid type, with values of 0 (non-shoreline) or 1 (shoreline); is the grid size, with values of 1 (small), 2 (medium), or 3 (large); is the elevation value, in meters; is the stability index, ranging from 0 to 1.

[0072] It should be noted that the coastal erosion dynamics equation is constructed based on physical mechanisms, considering the balance relationship between wave impact force and coastal resistance. The wave action intensity term uses a 1.5th power, emphasizing the significant impact of extreme wave events on erosion; the tidal amplitude term uses a 0.5th power, representing the non-linear contribution of tides to erosion; the sea level rise rate term uses a 0.3rd power, reflecting the cumulative effect of long-term sea level changes on erosion. These three dynamic factors together constitute the driving force for erosion, while the geological structure coefficient and slope angle represent the resistance. The selection of these power relationships in the equation is based on fitting a large amount of measured data and theoretical analysis, and can better reflect the relative contributions of various factors to the erosion process. Compared with existing technologies, this equation simultaneously considers short-term (waves, tides) and long-term (sea level rise) dynamic factors and introduces regional characteristics (geological structure, slope), having stronger adaptability and prediction ability.

[0073] It should be noted that the grid area adaptive attention allocation mechanism determines the attention allocation of each grid through the product of three key weight factors, achieving effective capture of different scale change characteristics. The grid density parameter weight adopts the form of a Gaussian function, which can ensure expression accuracy while maintaining computational efficiency; the historical change rate weight enhances the attention to regions with significant historical changes through the hyperbolic tangent function, improving the sensitivity of the model to key regions. Compared with traditional fixed weight models, this mechanism can dynamically adjust the attention distribution according to the characteristics of different regions, improving the adaptability and prediction accuracy of the model.

[0074] Specifically, the principle of the present invention is as follows: The core technical principle of the present invention lies in combining complex shoreline characterization with a multi-scale attention mechanism, and achieving accurate prediction through a small grid approximation method and a graph attention grid evolution model. First, the small grid approximation method breaks through the traditional linear characterization mode and represents the island shoreline as a set of square grids with equal side lengths. This characterization method can flexibly adjust the grid density according to the complexity of the shoreline, thereby accurately capturing the detailed features of the shoreline. Through the merging operation of small grids, a multi-level grid representation is formed, which not only retains local detailed information but also extracts global structural features.

[0075] Secondly, the coastal erosion dynamics equation model established in the present invention is based on the principle of mass conservation and fluid dynamics theory, and quantifies the equilibrium relationship between wave impact force and shoreline material resistance into a mathematical model. This model not only considers conventional factors such as wave action intensity, tidal amplitude, and sea level rise rate, but also introduces the island geological structure coefficient and shoreline slope angle, making the model more in line with physical reality and providing a theoretical explanation for the differences in shoreline erosion rates in different regions.

[0076] Most innovatively, the graph attention grid evolution model represents the shoreline grid as graph-structured data and realizes deep learning of shoreline change laws through the fusion architecture of a multi-layer graph convolutional neural network and a multi-head attention mechanism. Its core is a grid area adaptive attention allocation mechanism, which dynamically adjusts the attention distribution through three key parameters: the grid density parameter, the grid merging level, and the historical evolution rate of the change matrix, so that large-area grids obtain higher global change prediction weights, while small-area grids maintain accurate expression of local details. This adaptive mechanism conforms to the law of human visual attention allocation, can effectively balance global and local information, and thus improves the prediction accuracy.

[0077] In addition, the present invention also introduces an uncertainty analysis module to calculate the confidence interval of the prediction result, generate a multi-scale erosion risk assessment map, and further improve the reliability and practicality of the prediction result. This comprehensive technical solution solves the core technical problems in the prediction of complex island shoreline erosion risks in principle.

[0078] The following provides a specific Embodiment 1 of the present invention, and the specific implementation manners of each step in this Embodiment 1 are described in detail as follows.

[0079] The specific implementation of step S01 is to extract the island contour information by obtaining high - resolution island remote sensing images and performing a series of image pre - processing operations. First, the wavelet transform denoising algorithm is used to denoise the remote sensing image to eliminate sensor noise and atmospheric interference. Then, the histogram equalization method is applied to enhance the image contrast, making the boundary between the island and the sea clearer. Next, the feature point matching algorithm is used for multi - temporal image registration to ensure the accurate spatial correspondence of images at different time points. Finally, the OTSU segmentation algorithm based on threshold is used to separate the island and sea areas, and the threshold selection range is between 0.4 and 0.6, and the morphological opening - closing operation is combined to optimize the segmentation result. On this basis, the Canny edge detection algorithm is used to extract the island contour, and the high - to - low threshold ratio of edge detection is 2:1, and the high threshold is recommended to be set to 1.5 times the average value of the image gradient. The purpose of this step is to provide accurate island boundary information for subsequent shoreline grid division, ensuring the accuracy and reliability of shoreline extraction.

[0080] The specific implementation of step S02 is to construct an initial model of the island shoreline using the small - grid approximation method. First, a suitable first - size value is determined as the grid side length. According to the complexity of the island shoreline and the resolution of the remote sensing image, a suitable value is selected in the range of 5 meters to 50 meters. For complex shorelines, 5 to 15 meters is appropriate, and for straight shorelines, 20 to 50 meters can be selected. Then, based on the island contour extracted in step S01, sampling points are set at fixed intervals along the shoreline, and the sampling interval is equal to 1 / 2 of the first - size value. Next, with each sampling point as the center, a square grid with a side length of the first - size value is generated, and it is judged whether the grid contains shoreline points. If it does, the grid is included in the shoreline grid set. Finally, the grids are sorted and numbered, and the topological relationship between the grids is established, recording the spatial position and adjacent grid information of each grid. The purpose of this step is to discretize the continuous island shoreline into a grid representation that is convenient for computer processing, laying the foundation for subsequent multi - level grid construction.

[0081] The specific implementation of step S03 is to perform a merging operation on the small grids to generate a multi - level grid representation. First, the maximum square detection algorithm is applied to find adjacent grids that can be merged from the initial grid set. The judgment condition is that the similarity of adjacent grid attributes is higher than 0.85 and a larger square is formed after merging. Then, the grids are merged according to the bottom - up strategy. Small grids are preferentially merged into medium - sized grids, and medium - sized grids are then merged into large grids, keeping the side length of the merged grid an integer multiple of the initial grid side length. Next, a hierarchical grid tree structure is constructed to record the inclusion relationship between the grids. The large grid is used as the parent node, and the included small grids are used as child nodes. Finally, all grid information is converted into a shoreline feature matrix, and the matrix elements record the attribute information of the grids at the corresponding positions. The elements of the shoreline feature matrix are represented as follows: ; where, At time the eigenvector at position ; is the grid type, taking values of 0 (non-shoreline) or 1 (shoreline); is the grid size, taking values of 1 (small), 2 (medium) or 3 (large); is the elevation value in meters; is the stability index, ranging from 0 to 1. The purpose of this step is to reduce the redundancy of the shoreline representation while retaining the detailed information of key areas and improving the subsequent calculation efficiency.

[0082] The specific implementation of step S04 is to collect remote sensing images of islands in multiple time series and repeat the operation process of steps S01 to S03 for each time point. First, collect a sequence of remote sensing images of islands spanning at least 5 years, with a time interval recommended to be from 3 months to 1 year to ensure capturing seasonal and interannual changes; then, apply the preprocessing method of step S01 to each time-series image to ensure consistent image processing standards; next, execute the small-grid approximation method of step S02 on each processed image to construct the initial shoreline model for the corresponding time point; finally, apply the grid merging strategy of step S03 to generate the multi-level grid representation and shoreline feature matrix for each time point, with the matrix dimensions kept consistent for subsequent comparative analysis. The purpose of this step is to construct a time-series shoreline feature dataset to provide basic data for analyzing the laws of shoreline evolution.

[0083] The specific implementation of step S05 is to calculate the change matrix of the shoreline feature matrix at adjacent time points and construct a set of time-series change matrices. First, use matrix difference operation to calculate the difference between the shoreline feature matrices at adjacent time points to obtain the initial change matrix. The specific calculation formula is: ; where is the change matrix corresponding to time ; is the shoreline feature matrix at time ; is the shoreline feature matrix at time . Then, apply a spatial filtering algorithm to remove the noise in the change matrix, with the filter window size set to 3×3 to 5×5, and retain the change signals with a confidence level exceeding 0.75; next, calculate the change rates at each grid position, including the area change rate and the morphological change rate. The calculation formula for the change rate matrix is: ; where is the change rate matrix; is the time interval in years; The zero elements in are replaced with a constant Instead, it is used to avoid division-by-zero errors. Finally, the significantly changed areas are marked, and the change rate threshold is set to ±10% of the value at the previous moment. The change matrices corresponding to all time points are organized into a time series data structure to analyze the shoreline evolution pattern, identify the erosion acceleration area, the erosion deceleration area, and the stable area. The determination criteria are that the annual average recession distance is greater than 1 meter, the annual average recession distance is less than 0.2 meter, and the annual average change distance is less than 0.05 meter, respectively. The purpose of this step is to quantify the spatio-temporal change characteristics of the shoreline and provide a basis for dynamic changes for subsequent erosion risk prediction.

[0084] The specific implementation of step S06 is to establish a coastal erosion dynamics equation model to quantitatively calculate the erosion rate at different grid positions. First, collect the model input parameters, including the wave action intensity (unit: kilowatt per meter, typical value range 1 to 15), the tidal amplitude (unit: meter, typical value range 0.5 to 5), the sea level rise rate (unit: millimeter per year, typical value range 2 to 10), the island geological structure coefficient (dimensionless, value range 0 to 1), and the shoreline slope angle (unit: degree, typical value range 5 to 45); then, based on the principle of mass conservation and fluid dynamics theory, establish a dynamics equation describing the shoreline erosion process, which considers the balance relationship between wave energy input and geological structure resistance. The core formula of the coastal erosion dynamics equation model is expressed as follows: ; In the formula, is the erosion rate at grid position , with the unit of meter / year; is the wave action intensity, with the unit of kilowatt / meter; is the tidal amplitude, with the unit of meter; is the sea level rise rate, with the unit of millimeter / year; is the island geological structure coefficient, dimensionless, with a value range of 0 to 1; is the shoreline slope angle, with the unit of degree; is the erosion coefficient, with a default value of 0.25, calibrated through historical data; is the geological resistance coefficient, with a default value of 0.75, obtained through experiments; is the stability factor, with a default value of 0.05, used to avoid the denominator being zero. The acquisition methods of the parameters in the erosion equation are as follows: is obtained through wave observation data or wave models, and the calculation formula is: ; In the formula, is the seawater density, with a default value of 1025 kg / m³; is the gravitational acceleration, with a default value of 9.8 m / s²; is the significant wave height, with the unit of meter; is the wave group velocity, with the unit of meter / second; is the wave incident angle, with the unit of degree. Obtained from tidal observation station data or tidal prediction models, and can be obtained by spatial interpolation of data from the nearest tidal stations. The calculation formula is: ; In the formula, is the tidal amplitude of the th tidal station; is the spatial weight, and the calculation formula is: ; In the formula, is the grid position to the th tidal station, with the unit of kilometer. Obtained from sea level monitoring data, and the calculation formula is: ; In the formula, is the global average sea level rise rate, with a default value of 3.3 mm / year; is the regional sea level change, obtained from satellite altimetry data; is the local sea level change, considering the influence of crustal vertical movement and human activities, and obtained from land survey and GPS data. Obtained from geological surveys and rock hardness tests, and the calculation formula is: ; In the formula, is the rock hardness factor, with a range of 0 to 1; is the soil structure factor, with a range of 0 to 1; is the vegetation cover factor, with a range of 0 to 1; is the weight coefficient, satisfying , with default values of 0.5, 0.3, and 0.2 respectively. Calculated from elevation data, and the formula is: ; In the formula, is the elevation difference in the grid along the coastline direction, with the unit of meter; is the corresponding horizontal distance, with the unit of meter; is the coefficient for converting radians to degrees. Then, the erosion equation is applied to calculate the erosion rate for different types of grids. The simplified calculation mode is adopted for large-area grids, and the refined calculation mode is adopted for small-area grids. The calculation accuracy threshold is set to 0.1 m / year. Finally, the shoreline grid erosion rate distribution map is obtained, visually representing the erosion risk levels at different locations. The purpose of this step is to quantify the erosion process through physical models and provide theoretical support for risk prediction.

[0085] The specific implementation of step S07 is to apply the graph attention grid evolution model to analyze the law of shoreline change. First, convert the shoreline grid data into a graph structure representation, with the grid center points as graph nodes and the connection relationships between adjacent grids as graph edges; then, construct a graph attention network model, including an encoding layer, multiple graph convolutional layers (the number of layers is 3 to 5), a multi-head attention layer (the number of heads is 4 to 8), and a decoding layer; next, introduce an attention allocation mechanism based on grid area. Larger grids (greater than 625 square meters) are assigned higher global attention weights (the weight coefficient is 0.6 to 0.8), and smaller grids (less than 100 square meters) focus on local detail expression (the weight coefficient is 0.2 to 0.4). The calculation formula for the grid area weight coefficient of the graph attention grid evolution model is: ; where is the area weight coefficient of grid ; is the grid area, in square meters; is the minimum grid area, with a default value of 25 square meters; is the maximum grid area, with a default value of 625 square meters; is the minimum weight, with a default value of 0.2; is the maximum weight, with a default value of 0.8. The calculation formula for the attention distribution of the grid area adaptive attention allocation mechanism is: ; where is the attention score of grid ; is the grid density parameter weight, and the calculation formula is: ; where is the grid density of the area where grid is located, in number per square kilometer; is the optimal grid density, with a default value of 400 per square kilometer; is the adjustment parameter, with a default value of 200. is the historical change rate weight, and the calculation formula is: ; where is the historical change rate of grid in square meters per year; is the weight enhancement coefficient, with a default value of 0.5; is the hyperbolic tangent function, used to limit the weight change range; is the scaling factor, with a default value of 0.1. Finally, use the trained model to predict the future evolution trend of the shoreline, and output three core indicators including grid position change, morphological change, and rate change. The purpose of this step is to integrate data-driven and physical model methods to improve the accuracy and adaptability of prediction.

[0086] The specific implementation of step S08 is to introduce an uncertainty analysis module, calculate the confidence interval of the prediction results, and generate a multi-scale erosion risk assessment map. First, the Monte Carlo simulation method is used to randomly perturb the key input parameters, with the perturbation range being ±2 times the standard deviation of the parameters, and 1000 to 5000 groups of simulated parameters are generated; then, the graph attention grid evolution model is run to obtain multiple groups of prediction results, and the probability distribution of the predicted values at each grid position is calculated; then, the 90% confidence interval is determined according to the probability distribution, and the calculation formula of the confidence interval is: ; In the formula, is the confidence interval of the grid erosion rate; is the average predicted value of the erosion rate, with the unit of m / year; is the standard deviation; is the significance level of the standard normal distribution critical value, and the corresponding to the 95% confidence interval is 1.96. Then, the results are visualized as a risk level map, and the calculation formula of the risk score is: ; In the formula, is the comprehensive risk score of the grid ; respectively represent the short-term ( ), medium-term ( ), and long-term ( ) risk scores; is the time scale weight, satisfying , and the default values are 0.5, 0.3, and 0.2 respectively. The calculation formulas for the risk scores of each time scale are: ; In the formula, is the occurrence probability, with the range of 0 to 1; is the impact intensity, with the range of 0 to 10; is the vulnerability, with the range of 0 to 1. A 5-level risk rating is adopted (extremely high risk, high risk, medium risk, low risk, extremely low risk), and the risk thresholds are an annual erosion rate greater than 2 m, 1 to 2 m, 0.5 to 1 m, 0.1 to 0.5 m, and less than 0.1 m respectively; finally, a multi-scale erosion risk assessment map is generated, including short-term (1 to 3 years), medium-term (3 to 10 years), and long-term (10 to 30 years) risk assessments, and high-risk areas that need to be protected first are identified. The purpose of this step is to quantify the uncertainty of the prediction and provide risk decision support.

[0087] The detailed structure of the atlas attention grid evolution model and the specific implementation of establishing its training dataset are as follows: The core structure of the atlas attention grid evolution model adopts an architecture that combines a graph neural network and an attention mechanism, including four main components. First, the encoding layer is responsible for converting the input shoreline feature matrix into a graph structure representation. Each grid corresponds to a graph node, and the node features include grid type, area, position coordinates, and historical change rate. Adjacent grids are connected by edges, and the edge weights are calculated based on the similarity between grids. The similarity threshold is set to 0.7. Then, the multi-layer graph convolutional layer consists of 3 layers of graph convolutional networks, with 64 convolutional kernels in each layer. The size of the convolutional kernel is 3×3, and the activation function is ReLU. Batch normalization and residual connections are added after each layer to capture the spatial dependencies between grids. Next, the multi-head attention layer contains 8 attention heads, and each attention head independently learns the importance distribution in different feature subspaces. The attention weights are adaptively adjusted by the grid area, and the adjustment coefficient is the logarithmic ratio of the area to the reference area (100 square meters) to balance global changes and local details. Finally, the decoding layer converts the graph representation back to the grid matrix format and outputs the predicted changes at each grid position, including position offset and morphological changes.

[0088] The specific implementation of establishing the training dataset includes five core steps: First, collect historical remote sensing image sequences of 50 to 100 typical island shorelines globally, with a time span of no less than 10 years, a spatial resolution better than 5 meters, covering different climate conditions and geological types. Second, perform standardized preprocessing on all images, including atmospheric correction, geometric correction, and radiometric correction, to ensure data quality consistency. Third, construct a unified grid representation, with the grid size being comparable among different islands, and record the grid attribute information at each time point. Fourth, calculate the change matrix between adjacent time points and collect environmental factor data for the corresponding time period, including wave, tide, and meteorological data. Fifth, label the supervised learning labels according to the measured shoreline change results and divide the training set, validation set, and test set in a ratio of 7:2:1 to ensure balanced and representative data distribution.

[0089] The model training adopts the mini-batch stochastic gradient descent method. The batch size is set to 32 to 64, the initial learning rate is 0.001, and the cosine annealing strategy is applied for dynamic adjustment. The number of training epochs is 100 to 200 rounds, and the early stopping patience parameter is set to 20 rounds. During the training process, the cross-entropy loss function is used to evaluate the difference between the prediction results and the true labels, and the L2 regularization technique is introduced to prevent overfitting. The regularization coefficient is set to 0.0001 to 0.001. Weight coefficients are set for different island types respectively, with a focus on the types vulnerable to erosion. For example, the weight coefficient of sandy islands is 1.5, and the weight coefficient of rocky islands is 0.8. The generalization ability of the model in new island scenarios is improved through the transfer learning method. The pre-trained model is trained on the global dataset and then fine-tuned on the target island data. After the training is completed, the parameters of the feature extraction layer are frozen, and the key layer parameters for fine-tuning are retained to adapt to the specific conditions of different islands.

[0090] In this embodiment, the reason for using the power relationship in the coastal erosion dynamic equation is that the wave energy is proportional to the square of the wave height, which conforms to the basic principles of hydrodynamics; the 0.5th power of the tidal amplitude indicates that the tidal influence increases with the increase of the amplitude, but the growth rate slows down, which is consistent with the actual observation results; the 0.3rd power of the sea level rise rate indicates that its promoting effect on erosion exists, but it is not as significant as the wave action, which is in line with the results of coastal dynamics research. The cosine function is used in the denominator to describe the influence of the slope angle, reflecting the phenomenon that the steep coastline receives less wave energy and the erosion rate is relatively low. Adding a stability factor To avoid the numerical instability problem caused by the denominator approaching zero. This equation comprehensively considers the ocean dynamic factors (waves, tides) and geological factors (geological structure, slope), and can more comprehensively describe the physical mechanism of the coastal erosion process.

[0091] In this embodiment, the grid area weight coefficient adopts a logarithmic relationship because the range of grid area changes is large, and a direct linear relationship will lead to too high weights for large-area grids. After logarithmic transformation, it can more reasonably allocate attention resources, enabling the model to pay attention to both large-scale change trends and not ignore local details. The grid density parameter weight adopts a Gaussian function form, reflecting the characteristic that there is an optimal value for the grid density. Too high or too low grid density will reduce the model performance. The historical change rate weight adopts the hyperbolic tangent function, which can give more attention to the areas with significant historical changes while keeping the function values bounded, which helps the model capture the key change areas.

[0092] In this embodiment, the risk scoring model adopts the form of the product of three factors: probability, impact intensity, and vulnerability, which conforms to the basic framework of risk assessment and can comprehensively consider the possibility, severity, and tolerance of erosion events. The confidence interval calculation adopts the standard normal distribution hypothesis, and the distribution characteristics of the prediction results are obtained through Monte Carlo simulation, providing quantitative support for decision-making regarding uncertainty. The innovation of this evaluation system lies in the introduction of a comprehensive risk assessment mechanism for multiple time scales, which can take into account both short-term disaster prevention needs and long-term planning and management.

[0093] To better understand and implement the present invention, the following provides an embodiment 2 of a specific application scenario of the present invention: The research team selects a coral reef island in a certain sea area as the research object and conducts experimental research using the method for predicting the risk of island shoreline erosion based on computer vision. The coral reef island has an area of approximately 2.7 square kilometers, a shoreline length of approximately 8.3 kilometers, a geological structure mainly composed of coral reef limestone, and a sandy beach in some areas. In recent years, due to the influence of climate change and human activities, the phenomenon of shoreline erosion has become increasingly serious.

[0094] First, the team obtained high-resolution satellite remote sensing images of the island from 2015 to 2023, with a resolution better than 2 meters and a collection interval of 6 months. The wavelet transform denoising algorithm was used to eliminate image noise, increasing the signal-to-noise ratio from the original 16.5 dB to 27.8 dB. Then, the histogram equalization method was applied to enhance the image contrast, and the clarity of the boundary between the island and the sea increased by 42% after enhancement. The SIFT feature-based registration algorithm was used to register the multi-temporal images, with an average registration accuracy reaching 0.76 pixels. Finally, the OTSU segmentation algorithm was used to separate the island from the sea, with a segmentation threshold of 0.47, and the Canny operator was used to extract the island contour, with the high and low thresholds set to 1.67 times and 0.83 times the average image gradient, respectively.

[0095] The small grid approximation method was used to construct the initial shoreline model. Considering the complexity of the shoreline and the image resolution, the initial grid side length (the first dimension value) was selected as 12 meters. Sampling points were set at 6-meter intervals along the shoreline, resulting in a total of 1384 sampling points. Square grids were generated around each sampling point, and 1102 grids were included in the shoreline grid set after screening.

[0096] As Figure 2 shown, the merging operation was performed on the small grids to generate a multi-level grid representation. The similarity threshold of adjacent grid attributes was set to 0.87, and a bottom-up strategy was adopted for grid merging, finally forming a three-level grid structure, as shown in Table 1: Table 1 Statistical table of multi-level grids of the island shoreline

[0097] The team applied the above processing flow to the shoreline images at each time point to obtain the shoreline feature matrix corresponding to that time point. The change matrix of the shoreline feature matrices at adjacent time points was calculated, and a 5×5 spatial filter was applied to remove noise, with the confidence threshold set at 0.78. The research team analyzed the changes in the island shorelines from 2015 to 2023 and identified three types of regions, as shown in Table 2: Table 2 Statistical Table of Classification of Shoreline Change Regions

[0098] A coastal erosion dynamics equation model was established, and environmental parameters of the study area were collected. The wave action intensity was calculated based on the data of inshore observation buoys, with an average value of 4.3 kW / m and a maximum value reaching 12.7 kW / m. The tidal amplitude was obtained from the data of the Lingshui tide gauge station, with an average value of 1.2 m. The sea-level rise rate was determined to be 4.6 mm / year based on long-term monitoring data. The island geological structure coefficient was obtained through field investigations, and the calculation formula is , where the values of each factor are shown in Table 3: Table 3 Statistical Table of Island Geological Structure Parameters

[0099] Based on the elevation data, the shoreline slope angles were calculated. The average slope in the eastern coral reef area is 12°, the average slope in the southern sandy beach area is 6°, the average slope in the western mixed area is 15°, and the average slope in the northern cliff area is 36°.

[0100] Figure 3 shows the relationship between the shoreline erosion rate and the wave action intensity in four different regions of the island (eastern coral reef area, southern sandy beach, western mixed area, and northern cliff area). The chart includes the fitted curves and actual measurement data points (with error bars) for each region, clearly showing the influence of different geological structures on the erosion rate. The chart shows that the erosion rate in the southern sandy beach area increases the fastest with the increase in wave intensity, while the erosion rate in the northern cliff area is relatively low due to the high rock hardness. The critical erosion threshold line (1.5 m / year) is marked in the figure, and physical mechanism change points such as the wave breaking energy increase point and the sediment transport critical point are also marked. The atlas attention grid evolution model was applied to analyze the shoreline change law. The shoreline grid was converted into a graph structure, and a network model containing 4 graph convolutional layers and 6 heads of attention was constructed. In the calculation of the grid area weight coefficient is taken as 144 square meters, is taken as 2304 square meters, is taken as 0.25, is taken as 0.75. In the calculation of the grid density parameter weight, is taken as 420 per square kilometer, is taken as 215. In the calculation of the historical change rate weight, Take it as 0.45, take it as 0.12.

[0101] Use the collected data to train the model. Adopt the mini-batch stochastic gradient descent method, set the batch size to 48, the initial learning rate to 0.001, adopt the cosine annealing strategy, train for 150 rounds, and reach the best performance at the 132nd round. The average prediction error on the validation set is 0.076 m / year. The model training results are shown in Table 4: Table 4 Statistical table of model training performance

[0102] Introduce the uncertainty analysis module to generate a multi-scale erosion risk assessment map. Use the Monte Carlo simulation method to perform 2500 sets of random perturbations on the key parameters and calculate the 95% confidence interval, the value is 1.96. The risk score uses the weights of 0.5, 0.3, and 0.2 for the short-term, medium-term, and long-term time scales respectively. According to the evaluation results, the island shorelines are divided into five risk levels, as shown in Table 5: Table 5 Distribution table of shoreline erosion risk levels

[0103] Traditional island shoreline erosion risk assessment mainly relies on historical satellite image comparison and simplified physical models, which have problems such as low accuracy of shoreline extraction, rough spatial resolution, and large prediction uncertainty. Traditional methods usually adopt manual interpretation of island boundaries, and the error can reach more than 10 meters; and due to the lack of a representation method suitable for complex shoreline morphologies, it is impossible to accurately depict the detailed changes of the shoreline. In addition, traditional physical models usually only consider single or a few influencing factors and ignore the multi-factor coupling effect, resulting in a large deviation between the prediction results and the actual situation, especially limited prediction ability for shoreline changes under the influence of extreme events such as storm surges.

[0104] In contrast, the computer vision-based method for predicting the erosion risk of island shorelines in the present invention has achieved technological progress in the following aspects: First, by adopting the small grid approximation and multi-level merging strategy, the shoreline representation accuracy has been increased by 3.7 times, and it can adapt to various shoreline forms from straight coasts to complex coral reefs; Second, by constructing a coastal erosion dynamics equation considering multiple factors such as waves, tides, and sea level rise, the model prediction accuracy has been improved by 42%; Third, by introducing the atlas attention grid evolution model, differential processing of large-area grids and small-area grids is realized, while capturing the global change trend, local detailed features are retained, and the spatial accuracy of the prediction results has been increased by 2.8 times; Fourth, through uncertainty analysis and multi-scale risk assessment, a more comprehensive and reliable decision-making basis for risk management is provided, and the prediction credibility has been increased by 36%. The experimental results show that this method is particularly suitable for small islands with complex shoreline forms, providing strong technical support for island protection and sustainable development.

[0105] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 6 and 7 below.

[0106] Table 6 Variable Explanation Table (Part 1)

[0107] Table 7 Variable Explanation Table (Part 2)

[0108] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A computer vision-based island shoreline erosion risk prediction method, characterized in that: It includes obtaining high-resolution island remote sensing images and extracting island contour information; constructing an initial model of the island coastline based on a small grid approximation method; merging small grids to generate a multi-level grid representation; collecting multiple time series remote sensing images and calculating changes in the coastline feature matrix; A coastal erosion dynamics equation model is established, and the erosion rates at different grid positions are quantitatively calculated by combining wave intensity, tidal amplitude, sea level rise rate, island geological structure coefficient and coastline slope angle. The atlas attention grid evolution model is used to analyze the law of coastline changes. The atlas attention grid evolution model adjusts the attention allocation mechanism through the grid area weight coefficient to balance global changes and local details. An uncertainty analysis module is introduced to calculate the confidence interval of the prediction results, generate a multi-scale erosion risk assessment map, identify high-risk areas and output prediction results.

2. The island shoreline erosion risk prediction method based on computer vision according to claim 1, characterized in that: Acquiring high-resolution island remote sensing images and extracting island contour information includes image denoising, enhancement, registration and segmentation, and using machine vision algorithms to extract island contour information.

3. The island shoreline erosion risk prediction method based on computer vision according to claim 2, characterized in that: Constructing an initial model of the island coastline based on a small grid approximation method includes dividing the island boundary into a plurality of square grids with a side length of a first size value to form an island coastline grid set.

4. The method for predicting island shoreline erosion risk based on computer vision according to claim 3, characterized in that: The small grid approximation method represents the island boundary as a set of square grids with equal sides. By adjusting the grid density, the complex coastline morphology can be accurately expressed. Each grid records the existence status and attribute information of the coastline in the area.

5. The method for predicting island shoreline erosion risk based on computer vision according to claim 4, characterized in that: The first size value is the initial grid side length determined according to the complexity of the island coastline and the resolution of the remote sensing image, and the value range is 5 meters to 50 meters.

6. The method for predicting island shoreline erosion risk based on computer vision according to claim 5, characterized in that: The merging operation of small grids to generate a multi-level grid representation includes finding the global largest square grid and storing all grid information as a shoreline feature matrix.

7. The method for predicting island shoreline erosion risk based on computer vision according to claim 6, characterized in that: The shoreline feature matrix is ​​a two-dimensional data structure whose element values ​​represent the attributes of the corresponding position grid, including whether it is a shoreline point, shoreline type, elevation value and stability index. The matrix dimension is determined by the grid division accuracy.

8. The method for predicting island shoreline erosion risk based on computer vision according to claim 7, characterized in that: The coastal erosion dynamics equation is used to simulate the physical mechanism of coastal erosion. Based on the principle of mass conservation and fluid dynamics theory, it considers the balance between the impact force of waves on the coastline and the resistance of coastline materials, and obtains the erosion rate distribution of grids at different locations through integral calculation.

9. The method for predicting island shoreline erosion risk based on computer vision according to claim 8, characterized in that: The structure of the graph attention grid evolution model is a fusion architecture of a multi-layer graph convolutional neural network and a multi-head attention mechanism. The shoreline grid is represented as graph structure data, and the connection relationship between grids corresponds to the edge; the graph attention grid evolution model includes an encoding layer, a multi-layer graph convolution layer, a multi-head attention layer and a decoding layer.

10. The method for predicting island shoreline erosion risk based on computer vision according to claim 9, characterized in that: The core of the atlas attention grid evolution model is to introduce a grid area adaptive attention allocation mechanism. The grid area adaptive attention allocation mechanism dynamically adjusts the attention distribution through three key parameters: grid density parameter, grid merging level and historical evolution rate of the change matrix in the shoreline feature matrix, so as to effectively capture the shoreline change characteristics of different scales.

Images