A multi-modal based method for predicting storm surge disasters in coastal areas

Abstract

The present invention discloses a method for predicting storm surge disasters in coastal areas based on multi-modal data, which combines multi-source meteorological environment data, three-dimensional geospatial data, and historical storm surge occurrence data. On the one hand, a multi-modal network is used to conduct long-term prediction and trend fitting of future storm surge conditions, and the linear component and non-linear component related to storm surge prediction are obtained by fitting. The linear component and non-linear component are combined and then converted into the storm surge prediction probability P1 through an activation function. On the other hand, for the three-dimensional geospatial data, the predicted values of supervised learning for its labeled scenes and the predicted values of hierarchical supervised learning for its unlabeled scenes are obtained respectively; and the average of the predicted values of the two is converted into the storm surge prediction probability P2 through an activation function; finally, the storm surge prediction probability P1 and the storm surge prediction probability P2 are weighted to obtain the final storm surge prediction value.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological disaster risk prediction for coastal areas, and particularly relates to a multi-modal based storm surge disaster prediction method for coastal areas. Background Art

[0002] With the acceleration of the urbanization process, especially the rapid development of infrastructure in coastal areas, how to cope with the impact of natural disasters (such as storm surges, tidal fluctuations, etc.) on coastal infrastructure has become an urgent problem to be solved. Traditional disaster prevention and mitigation technologies have certain limitations in real-time data monitoring, environmental change response, infrastructure risk assessment, etc. Especially in a dynamic environment (such as scenarios with frequent changes in tides, storm surges, etc.), how to efficiently and accurately conduct disaster early warning, risk assessment and infrastructure monitoring has become the key to improving the disaster response ability.

[0003] However, the existing technologies mainly focus on static model construction and infrastructure management, lack sufficient consideration for the dynamic risk assessment brought about by environmental changes, and it is difficult to achieve in-depth integration and predictive analysis of real-time data. Therefore, how to construct a method that can cope with dynamic environmental changes and accurately predict and assess the storm surge disasters in coastal areas has become an urgent technical problem to be solved. Summary of the Invention

[0004] The main purpose of the present invention is to overcome the deficiencies of the existing technologies, and provides a multi-modal based storm surge disaster prediction method for coastal areas.

[0005] The present invention is realized through the following technical solutions:

[0006] A multi-modal based storm surge disaster prediction method for coastal areas, comprising the following steps:

[0007] Step 1, collect meteorological environment data, three-dimensional geographical space data and historical storm surge occurrence data of the target coastal area to be evaluated;

[0008] Step 2, according to the data collected in Step 1, use a multi-modal network to conduct long-term prediction and trend fitting on the future storm surge situation;

[0009] Step 2.1: Fit the linear relationship of the data collected in Step 1 to obtain the linear component related to storm surge prediction ;

[0010] Step 2.2: Fit the non-linear relationship of the data collected in Step 1 to obtain the non-linear component related to storm surge prediction ;

[0011] Step 2.3: The linear component related to storm surge prediction obtained in Step 2.1 and the non - linear component related to the storm surge prediction obtained in step 2.2 Add them together to obtain new storm surge prediction data ;

[0012] Step 2.4: Convert the storm surge prediction data obtained in step 2.3 , through an activation function, into a storm surge prediction probability ;

[0013] Step 3: Combine the three - dimensional geospatial data obtained in step 1, and use a supervised learning method to predict the storm surge;

[0014] Step 3.1: Perform random data shuffling for data augmentation on the three - dimensional geospatial data;

[0015] Step 3.2: Use the Jenks Natural Breaks algorithm (automatic segmentation algorithm) to generate dynamic thresholds for the three - dimensional geospatial data processed in step 3.1, and obtain three thresholds for each category ; ;

[0016] Step 3.3: Based on the three thresholds in step 3.2 , generate hierarchical supervision for the pseudo - labels of the mining objects in the unlabeled scenes in the three - dimensional geospatial data, and divide the hierarchical pseudo - labels into three categories: high confidence, fuzzy, and low confidence;

[0017] Step 3.4: Train the objective loss function, and the total training loss = the supervised learning loss of the labeled scenes + the hierarchical supervised learning loss of the unlabeled scenes ; Through the objective loss function, obtain the prediction value of the supervised learning of the labeled scenes and the prediction value of the hierarchical supervised learning of the unlabeled scenes ;

[0018] Step 3.5: Convert the prediction value , through an activation function, into a storm surge prediction probability ;

[0019] Step 4: Weight the storm surge prediction probability obtained in step 2 and the storm surge prediction probability obtained in step 3 , to obtain the final storm surge prediction value .

[0020] In the above technical solution, the meteorological environment data includes: data such as wind speed, seawater temperature, sea level height, sea level pressure, and air humidity with time attributes, forming a meteorological environment data set represented by a time series; three-dimensional geospatial data includes: three-dimensional geospatial data of the coastline change, marine environment, and infrastructure of the target coastal area with time attributes; storm surge historical occurrence data includes: storm surge water level rise height data with time attributes.

[0021] In the above technical solution, in step 3.3:

[0022] Pseudo-labels of high-confidence categories , ;

[0023] Pseudo-labels of fuzzy categories , ;

[0024] Pseudo-labels of low-confidence categories , ;

[0025] Represents the confidence score of the pseudo-label x.

[0026] In the above technical solution, in step 3.4:

[0027] 1) Supervised learning loss of the annotation scenario:

[0028] ;

[0029] : Total number of samples;

[0030] : Number of classes; the number of classes is 2, the first class is the occurrence of a storm surge, represents the first class; the second class is the non-occurrence of a storm surge, represents the second class;

[0031] : Sample The true label of the sample on class c, which is one-hot encoded; Indicates that sample i belongs to class c, Indicates that sample i does not belong to class c;

[0032] : Sample The predicted probability of the sample on class c;

[0033] 2) Hierarchical supervised learning loss of the unlabeled scenario:

[0034] ;

[0035] : is the weight coefficient;

[0036] High-confidence loss: ;

[0037] ;

[0038] : is the loss function;

[0039] : the pseudo-label of sample i;

[0040] : the predicted value of sample i:

[0041] Fuzzy loss:

[0042] ;

[0043] : the balance factor, used to adjust the importance of positive and negative samples;

[0044] : the adjustment factor, used to reduce the influence of easily classified samples and increase the attention to difficult-to-classify samples;

[0045] : increase the loss contribution to samples with a smaller predicted probability;

[0046] Low-confidence loss: ;

[0047] ;

[0048] : the confidence threshold, values less than this indicate that the predicted value is close to the pseudo-label and is considered valid;

[0049] : indicates discarding the gradient update of this sample or setting its loss to zero.

[0050] In the above technical solution, according to the storm surge predicted value , conduct disaster risk classification and decision-making:

[0051] If , the risk level = high risk;

[0052] If , the risk level = medium risk;

[0053] If , the risk level = low risk;

[0054] = Threshold for high risk;

[0055] = Threshold for low risk.

[0056] The advantages and beneficial effects of the present invention are as follows:

[0057] The present invention combines multi-source meteorological environment data, three-dimensional geospatial data, and historical storm surge occurrence data. On the one hand, a multi-modal network is used to perform long-term prediction and trend fitting on future storm surge conditions, and the linear and non-linear components related to storm surge prediction are obtained by fitting. The linear and non-linear components are combined and then converted into a storm surge prediction probability P1 through an activation function. On the other hand, for the three-dimensional geospatial data, the predicted values of supervised learning for its labeled scenes and the predicted values of hierarchical supervised learning for its unlabeled scenes are obtained respectively; and the average of the predicted values of the two is converted into a storm surge prediction probability P2 through an activation function; finally, the storm surge prediction probability P1 and the storm surge prediction probability P2 are weighted to obtain the final storm surge prediction value. Description of the Drawings

[0058] Figure 1 is a flowchart of the method for predicting storm surge disasters in coastal areas based on multi-modal of the present invention.

[0059] Figure 2 is a graph showing the change trend of the error performance of the method for predicting storm surge disasters in coastal areas based on multi-modal of the present invention with the number of training rounds in terms of mean error (ME), mean absolute error (MAE), root mean square error (RMSE), mean absolute scaled error (MASE), and U-Theil coefficient, etc.

[0060] For those of ordinary skill in the art, without creative efforts, other related drawings can be obtained based on the above drawings. Detailed Embodiments

[0061] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution of the present invention will be further described below in conjunction with specific embodiments.

[0062] The present invention designs a method for predicting storm surge disasters in coastal areas based on multi-modal, and the specific steps are as follows:

[0063] Step 1. Data collection.

[0064] Step 1.1: Through sensors, collect meteorological environment data of the target coastal area to be evaluated, including: data such as wind speed, seawater temperature, sea level height, sea level pressure, and air humidity with time attributes, and form a meteorological environment data set represented by a time series :

[0065] ;

[0066] : Wind speed (unit: );

[0067] : Seawater temperature (unit: );

[0068] : Sea level height (unit: m);

[0069] : Sea level air pressure (unit: hPa);

[0070] : Air humidity (percentage).

[0071] Step 1.2: Obtain 3D geospatial data of the coastline changes, marine environment, and infrastructure in the target coastal area to be evaluated through remote sensing technology (satellites, drones, etc.), and form a 3D geospatial dataset represented by a time series :

[0072] ;

[0073] : Satellite remote sensing data;

[0074] : UAV remote sensing data;

[0075] : Remote sensing data fusion and modeling function.

[0076] Step 1.3: Obtain historical storm surge occurrence data and form a historical storm surge occurrence dataset represented by a time series :

[0077] ; i represents the i-th historical storm surge occurrence data, and N is the total number of historical storm surge occurrence data;

[0078] : Timestamp (time when the storm surge occurred);

[0079] : Storm surge water level rise height (unit: m, representing the height of the water level rise).

[0080] Step 2: According to the data collected in Step 1, use a multimodal network to perform long-term prediction and trend fitting on future storm surge conditions.

[0081] Step 2.1: Fit the linear relationship of the data collected in Step 1 (meteorological environment data, three-dimensional geospatial data, and historical storm surge occurrence data) to obtain the linear components related to storm surge prediction. ;

[0082] ;

[0083] : Represents the linear component related to storm surge prediction at time (for example: the long-term trend or average change law of storm surge, the linear part in periodic fluctuations, etc.);

[0084] : The difference operator, indicating the -order difference operation on the original time series. is the order of difference, and its purpose is to make the non-stationary time series stationary for subsequent modeling and analysis;

[0085] : The constant term, similar to the intercept, represents the average level or reference value of the sequence. In this embodiment, it represents the basic value of the long-term average storm surge occurrence in the region;

[0086] : The autoregressive order, are the autoregressive coefficients, represents the values of the time series at the past moments, represents the initial value, representing the original collected data. and and reflect the influence degree of the past storm surge occurrence on the current storm surge occurrence. For example, when , represents 's influence coefficient on , reflecting the linear correlation degree of the previous moment's storm surge occurrence on the current moment's storm surge occurrence;

[0087] : The moving average order, are the moving average coefficients, are the error terms at the past moments. The moving average part is mainly used to handle the noise, short-term fluctuations and other irregular components in the time series, so that the model can better fit the actually observed historical storm surge data. The coefficient determines the weight distribution of each error term;

[0088] : Is the time White noise, which represents some uncontrollable and random factors that affect the occurrence of storm surges and cannot be explained by the model.

[0089] Step 2.2: Fit the non-linear relationship of the data collected in Step 1 (meteorological environment data, three-dimensional geospatial data, and historical storm surge occurrence data) to obtain the non-linear components related to storm surge prediction. .

[0090] Optimize the parameters of the time series neural network by simulating the hunting behavior of humpback whales. Assume is the hyperparameter combination vector of the time series neural network, and the parameters are updated through iteration:

[0091] ;

[0092] ;

[0093] Among them, , Gradually decreases from 2 to 0, and are random vectors within;

[0094] : When searching for the best hyperparameters of the time series neural network, it reflects the distance metric between the hyperparameter combination and the optimal hyperparameter set. For example, the overall difference size between the hyperparameter combination and the optimal combination in dimensions such as the number of neurons, learning rate, and batch length;

[0095] : In the context of hyperparameter optimization, it is the hyperparameter combination vector that currently finds the best performance (such as the minimum) of the time series neural network, including parameter values such as the number of neurons in the hidden layer, learning rate, and batch length, and continuously updates towards the global optimum with the iteration of the algorithm;

[0096] : The hyperparameter combination vector of the time series neural network currently being evaluated, which is continuously adjusted and updated during the algorithm search process. Its initial value is randomly set within the hyperparameter value range and is optimized through interaction and iteration with ;

[0097] and : Coefficient vector, controls the step size scaling for the individual to approach the optimal solution, and its value is affected by and and decreases with the iteration to make the step size change from large to small, guiding the search to gradually focus; is mainly used to adjust the search randomness, and is related to They jointly affect the randomness of individual position updates, balance exploration and exploitation in hyperparameter search, and avoid falling into local optima. For example, when searching for the number of neurons, the search range and step size are reasonably adjusted.

[0098] The non-linear output of the final time series data is:

[0099] ;

[0100] where and are activation functions (Sigmoid and Tanh), is the hyperparameter combination vector of the time series neural network, at time the non-linear components related to storm surge prediction (such as: random changes affected by meteorological environment, non-linear parts in periodic fluctuations, etc.); represents the non-linear components related to storm surge prediction at time .

[0101] Step 2.3: Add the linear components related to storm surge prediction obtained in Step 2.1 and the non-linear components related to storm surge prediction obtained in Step 2.2 to get the new storm surge prediction data :

[0102] .

[0103] Step 2.4: Convert the storm surge prediction data obtained in Step 2.3 into storm surge prediction probability through an activation function (such as the Sigmoid function): ;

[0104] : Storm surge prediction probability (i.e., the predicted probability of storm surge occurrence);

[0105] : Sigmoid activation function to ensure that the value of the output is in the range of .

[0106] Step 3: Combine the three-dimensional geospatial data obtained in Step 1 and use the supervised learning method to predict storm surges.

[0107] Step 3.1: Perform random shuffling data augmentation processing on the three-dimensional geospatial data.

[0108] The main operation process is to crop and compress the point cloud scene into a bird's-eye view grid, further divide the bird's-eye view grid into small units, and achieve data augmentation by shuffling these small units. The shuffled patches are then used to extract features through the backbone network, enabling the network to learn data features under different combinations and orders, enhancing the network's understanding of data and feature extraction ability. For example, when detecting small-sized objects such as small reefs along the coast, this way of shuffling patches allows the network to better learn their feature representations and avoid learning limitations caused by conventional orders, etc.

[0109] Subsequently, features are extracted through the backbone network and the original positions are restored. Among them, the point cloud scene is a data form collected by devices such as lidar, which contains numerous discrete points and digitally represents the real scene (such as the three-dimensional spatial situation of coastal terrain and coastal infrastructure, etc.). Each point in it has relevant attribute information such as its spatial coordinates, which is the data basis for subsequent operations. The bird's-eye view grid is a two-dimensional grid form obtained by converting the three-dimensional point cloud scene through specific projections, divisions, etc., similar to the perspective from above, facilitating operations such as segmentation and shuffling. Each grid can correspond to the summary of point cloud information in a certain spatial area. For example, after converting the point cloud data of a coastal area into a bird's-eye view grid, it is possible to more clearly see the distribution of terrain, buildings, etc. at different positions (presented in the form of a grid).

[0110] Step 3.2: Use the Jenks Natural Breaks algorithm (automatic segmentation algorithm) to generate dynamic thresholds for the three-dimensional geospatial data processed in Step 3.1, and obtain three thresholds for each category c of the three thresholds .

[0111] The principle of the Jenks Natural Breaks algorithm is to design a threshold generation strategy based on the confidence scene set and its prediction pairs, comprehensively considering the confidence score, objectness score, and consistency threshold (IoU);

[0112] ;

[0113] where JNB represents the automatic segmentation algorithm of Jenks Natural Breaks, used to divide the data into several natural groups (i.e., categories);

[0114] : the set of confidence scores for category c ; here, category c refers to: the two categories of storm surge occurrence and no storm surge occurrence;

[0115] : the high-confidence threshold;

[0116] : Intermediate confidence threshold;

[0117] : Low confidence threshold.

[0118] Consistency threshold (the overlapping ratio between the predicted storm surge impact area and the area affected by the actual historical storm surge, used to assist in judging the accuracy of the prediction and the degree of fit with the actual situation, and further assist in the generation of the threshold):

[0119] ;

[0120] : The th predicted bounding box and the ground truth bounding box 's intersection over union;

[0121] : IoU consistency threshold;

[0122] : Total number of samples.

[0123] Step 3.3: Based on the three thresholds in Step 3.2 , hierarchical supervised generation of pseudo-labels for the mining objects in the unlabeled scenes in the three-dimensional geospatial data is performed, and the hierarchical pseudo-labels are divided into three categories: high confidence, fuzzy, and low confidence.

[0124] The pseudo-labels are not accurate labels of real manual annotations. They are a temporary identification based on the thresholds in the early stage to roughly divide the categories they belong to, corresponding to the categories of high confidence, fuzzy, and low confidence;

[0125] 1) High confidence category: Represents a relatively reliable and more trustworthy data category;

[0126] ; Indicates the pseudo-label of the high confidence category, Indicates the confidence score of the pseudo-label x.

[0127] 2) Fuzzy category: A category divided based on the threshold, between high confidence and low confidence, representing those data situations where it is not very certain about their accurate conditions but is not completely untrustworthy;

[0128] ; Indicates the pseudo-label of the fuzzy category.

[0129] 3) Low confidence category: represents the data category with low credibility. The input data may contain a lot of interference or inaccuracy, which needs to be processed to avoid adverse effects on network training. For example, some sensors collect abnormal data, and remote sensing images are difficult to identify and unreliable due to occlusion and other reasons.

[0130] ; Pseudo labels representing low confidence classes.

[0131] For pseudo labels of low confidence categories, they are removed as noise points.

[0132] Step 3.4: Train the target loss function.

[0133] Total training loss = supervised learning loss for labeled scenes + Hierarchical supervised learning loss for unlabeled scenes ,Right now: By training the target loss function, we can get the predicted value of supervised learning for the labeled scene. And the predicted values of hierarchical supervised learning for unlabeled scenes .

[0134] The total training loss takes into account the labeled and unlabeled scenarios, and measures the overall degree of deviation between the network training effect and the ideal situation. When conducting machine learning training related to disaster risk assessment of coastal infrastructure, the accuracy of the final disaster risk assessment is improved by continuously optimizing and reducing the total training loss.

[0135] 1) Supervised learning loss for labeled scenes;

[0136] ;

[0137] : total sample size;

[0138] : Number of categories; in the present invention, the number of categories is 2, the first category is the occurrence of storm surge, c=1 represents the first category; the second category is no storm surge, c=2 represents the second category;

[0139] :sample The true label of category c is one-hot encoded; Representation sample Belongs to category c, Representation sample Does not belong to category c;

[0140] : Sample Prediction probability on class c.

[0141] 2) Hierarchical supervised learning loss for unlabeled scenarios:

[0142] ;

[0143] : Losses corresponding to high, ambiguous, and low confidence levels respectively;

[0144] : Corresponding weight coefficients.

[0145] High-confidence loss (the loss function uses mean squared error when the error is small for smoother results; uses mean absolute error when the error is large to reduce the risk of gradient explosion):

[0146] ;

[0147] ;

[0148] : It is an existing loss function for deep learning;

[0149] : Sample Pseudo-label of

[0150] : Sample Predicted value of

[0151] Ambiguous loss (used for classification tasks of ambiguous pseudo-labels):

[0152] ;

[0153] : Balancing factor, used to adjust the importance of positive and negative samples;

[0154] : Adjustment factor, used to reduce the influence of easily classified samples and increase the attention to difficult-to-classify samples;

[0155] : Increase the loss contribution of samples with relatively small prediction probabilities (difficult to classify).

[0156] Low-confidence loss (used for denoising tasks of low-confidence pseudo-labels):

[0157] ;

[0158] ;

[0159] : Confidence threshold. A value less than this indicates that the predicted value is close to the pseudo-label and can be considered valid.

[0160] : Indicates discarding the gradient update of this sample or setting its loss to zero to avoid interference with model training.

[0161] Step 3.5: Output the predicted storm surge probability.

[0162] The predicted value is converted into the probability output of storm surge prediction through an activation function (such as the Sigmoid function):

[0163] ;

[0164] : Storm surge prediction probability;

[0165] is the Sigmoid activation function, ensuring that the output value range is within .

[0166] Step 4: Weight the predicted probabilities obtained in Step 2 and Step 3 to obtain the final storm surge prediction value for the coastal area , realizing the quantitative assessment of disaster risk;

[0167] ;

[0168] Among them, : Weight value.

[0169] Furthermore, according to the storm surge prediction value , disaster risk classification and decision-making can be carried out:

[0170] If , risk level = high risk;

[0171] If , risk level = medium risk;

[0172] If , risk level = low risk;

[0173] = Risk threshold (determined based on historical data and experience);

[0174] = Threshold for high risk;

[0175] = Threshold for low risk.

[0176] Experimental verification:

[0177] See AppendixFigure 2 , the variation trend of the multi-modal based storm surge disaster prediction method proposed by the present invention with the number of training rounds in terms of error performance such as mean error (ME), mean absolute error (MAE), root mean square error (RMSE), mean absolute scaled error (MASE), U-theil coefficient, etc. is simulated.

[0178] From Figure 2 , it can be observed that all error metrics show a downward trend during the increase of the number of training rounds. This indicates that the proposed algorithm is continuously optimized during training, the error gradually decreases, and the prediction ability of the model is improving.

[0179] ME (mean error) drops the fastest, rapidly decreasing in the first 10 training rounds and then leveling off, indicating that the bias of this method converges earlier. MAE (mean absolute error) and RMSE (root mean square error) decrease relatively uniformly, but RMSE decreases more slowly than MAE, indicating that it is still affected by some larger error samples (RMSE is more sensitive to large errors). MASE (mean absolute scaled error) drops at a slower rate than ME but also levels off after 30 rounds.

[0180] In the initial stage of training (the first 10 rounds), all error values drop rapidly, indicating that the model has learned the main patterns and features in the early stage. In the middle stage of training (10 - 30 rounds), the downward trend is still obvious but the rate decreases, meaning that the learning curve of the model begins to level off. In the late stage of training (30 - 50 rounds), all error metrics basically stabilize, indicating that the algorithm converges well, and further training has little effect on the decrease of errors, entering a bottleneck period.

[0181] There are a small number of fluctuations in the error curve, which is due to the addition of random noise to simulate the uncertainty factors in the experiment. These fluctuations are small in amplitude, indicating that the training process of the algorithm is relatively stable.

[0182] The above makes an exemplary description of the present invention. It should be noted that without departing from the core of the present invention, any simple deformation, modification, or equivalent substitution that can be made by those skilled in the art without creative labor falls within the protection scope of the present invention.

Claims

1. A multi-modal based storm surge disaster prediction method for coastal areas, characterized in that, It includes the following steps: Step 1: Collect meteorological environment data, three-dimensional geospatial data, and historical storm surge occurrence data of the target coastal area to be evaluated; Step 2: According to the data collected in Step 1, use a multimodal network to conduct long-term prediction and trend fitting of future storm surge conditions; Step 2.1: Fit the linear relationship of the data collected in Step 1 to obtain the linear components related to storm surge prediction ; Step 2.2: Fit the non-linear relationship of the data collected in Step 1 to obtain the non-linear components related to storm surge prediction ; Step 2.3: Add the linear component related to storm surge prediction obtained in Step 2.1 and the non-linear component related to storm surge prediction obtained in Step 2.2 to obtain new storm surge prediction data ; Step 2.4: The storm surge prediction data obtained in Step 2.3 , is converted into the storm surge prediction probability through the activation function ; Step 3: Combine the three-dimensional geospatial data obtained in Step 1 and use a supervised learning method to predict storm surges; Step 3.1: Conduct random data shuffling data augmentation processing on the three-dimensional geospatial data; Step 3.2: Using the Jenks Natural Breaks algorithm, generate dynamic thresholds for the three-dimensional geospatial data processed in Step 3.1 to obtain three thresholds for each category ; ; Step 3.3: Based on the three thresholds in Step 3.2 , generate hierarchical supervision for the pseudo-labels of the mining objects in the unlabeled scenes in the three-dimensional geospatial data. The hierarchical pseudo-labels are divided into three categories: high confidence, fuzzy, and low confidence; Step 3.4 Train the target loss function, total training loss = Supervised learning loss of the labeled scenario + Hierarchical supervised learning loss of the unlabeled scenario ; Through the target loss function, obtain the predicted values of supervised learning for the labeled scenario and the predicted values of hierarchical supervised learning for the unlabeled scenario ; Step 3.5: Convert the predicted value to the storm surge prediction probability through the activation function ; Step 4: Weight the storm surge prediction probability obtained in Step 2 and the storm surge prediction probability obtained in Step 3 to obtain the final storm surge prediction value .

2. The multimodal-based storm surge disaster prediction method for coastal areas according to claim 1, wherein: The meteorological environment data includes: wind speed, seawater temperature, sea level height, sea level pressure, and air humidity data with time attributes, forming a meteorological environment dataset represented by a time series; the three-dimensional geospatial data includes: three-dimensional geospatial data of the coastline changes, marine environment, and infrastructure of the target coastal area with time attributes; the historical storm surge occurrence data includes: storm surge water level rise height data with time attributes.

3. The multimodal-based storm surge disaster prediction method for coastal areas according to claim 1, wherein: In Step 3.3: Pseudo-labels of high-confidence categories , ; Pseudo-labels for fuzzy categories , ; Pseudo-labels for low-confidence categories , ; Represents the confidence score of the pseudo-label x.

4. The multimodal-based storm surge disaster prediction method for coastal areas according to claim 1, characterized in that: In Step 3.4: 1) The supervised learning loss of the labeled scenario: ； : Total number of samples; : The number of categories; the number of categories is 2. The first category is the occurrence of storm surge, represents the first category; The second category is that no storm surge occurred, represents the second category; : Sample The true label on class c is one - hot encoded; Indicates that sample i belongs to class c, Indicates that sample i does not belong to class c; : Sample Predicted probability on class c; 2) The hierarchical supervised learning loss of the unlabeled scenario: ； : is the weight coefficient; High-confidence loss: ; ； : is the loss function; : The pseudo-label of sample i; : Predicted value of sample i: Fuzzy loss: ； : Balance factor, used to adjust the importance of positive and negative samples; : A regulatory factor used to reduce the influence of easily classifiable samples and increase the attention to difficult-to-classify samples; : Increase the loss contribution of samples with relatively low predicted probabilities; Low confidence loss: ; ； : Confidence threshold. A value less than this indicates that the predicted value is close to the pseudo-label and is considered valid; : Indicates discarding the gradient update of the sample or setting its loss to zero.

5. The multimodal-based storm surge disaster prediction method for coastal areas according to claim 1, wherein: According to the predicted value of storm surge , conduct disaster risk classification and decision-making: If , the risk level = high risk; If , the risk level = medium risk; If , the risk level = low risk; = Threshold for high risk; = Threshold for low risk.

Images