Typhoon Location Method Based on Social Media

Abstract

The present invention discloses a typhoon positioning method based on social media data, mainly solving the problem of low typhoon positioning accuracy in the prior art when dealing with massive and complex social media data. The implementation scheme is as follows: obtaining the spatio-temporal information in the social media data and encoding it into a spatio-temporal encoding data set; using the BERT model to extract the text semantic features of the spatio-temporal encoding data set, and performing semantic and spatial reconstruction on the semantic features according to the spatio-temporal encoding to obtain a reconstructed semantic feature data set; constructing a convolutional neural network based on a three-dimensional convolutional layer and a convolutional-long short-term memory artificial neural layer, and using the reconstructed semantic feature data set to predict the distance from the center of each grid to the typhoon center; through a spatio-temporal grid positioning algorithm based on location-based services, using the prediction results of the convolutional neural network to determine the latitude and longitude positions of the typhoon center. The present invention improves the accuracy of typhoon positioning by reconstructing the social media data features and can be used for the defense of natural disasters.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and particularly relates to a typhoon positioning method, which can be used for the defense of natural disasters. Background Art

[0002] The frequent occurrence of natural disasters poses a threat to people's lives and property safety, resulting in a large number of casualties and property losses. Typhoon positioning has gradually become a research hotspot in feature engineering and deep learning technologies. With the development of the Internet, social media has become an important platform for information sharing. On this platform, users continuously post content and share geographical information. This voluntary geographical information (VGI) provides a new information source for solving complex problems such as disaster management. Therefore, how to accurately and efficiently extract and integrate valuable information, especially spatio-temporal information related to typhoons, from massive social media data and achieve typhoon positioning using social media data is an urgent problem to be solved.

[0003] Existing methods for typhoon positioning using social media data first explore social media text data in multiple dimensions, then extract features including sentiment analysis, crisis detection, real-time messages, entity extraction, and the number of transmissions, etc., and use deep learning algorithms for trajectory positioning prediction after quantifying the features. However, when extracting features, these methods ignore many abstract semantic features, fail to capture the global spatio-temporal correlation of the data, and cannot accurately predict the typhoon center position.

[0004] Resch B, F, Havas C. Combining machine-learning topic models and spatiotemporal analysis of social media data for disaster footprint and damage assessment. Cartography and Geographic Information Science, 2018, 45(4): 362 - 376 proposed a topic machine learning and spatiotemporal analysis method for disaster trajectory prediction based on social media data. By creating a regular grid composed of multiple grid cells, all social media data related to disasters, such as the total number of transmissions in each region, can be summarized in the corresponding grid cells. Each grid cell has a size of 1 km × 1 km. According to the data in each grid, first extract the number of transmissions of social media data and convert it into the population number, and then use a spatial data analysis algorithm method for hotspot analysis to study the local spatial autocorrelation of the disaster trajectory and the damage it causes, so as to determine the disaster trajectory. However, this method has the following two deficiencies:

[0005] First, since it usually uses intermediate features extracted from text such as the number of transmissions and sentiment index, ignoring the spatio-temporal information and rich semantic features of the text itself, the extracted features are not accurate and comprehensive enough to achieve high-granularity event location prediction.

[0006] Second, since it fails to fully utilize the global spatio-temporal correlation of social media data and lacks an effective mechanism to capture the spatio-temporal relationship between data, it limits the spatio-temporal correlation capture and prediction ability of the positioning model. Summary of the Invention

[0007] The object of the present invention is to propose a typhoon positioning method based on social media to improve the accuracy of feature extraction of social media data through spatio-temporal grid coding and spatio-temporal reconstruction of semantic features, and to improve the capture and prediction ability of global spatio-temporal correlation by constructing a convolutional neural network model, aiming at the deficiencies of existing methods.

[0008] To achieve the above object, the technical solution of the present invention includes the following steps:

[0009] (1) Obtain a social media text data set S with spatio-temporal information in the typhoon-affected area during the typhoon from a social media platform;

[0010] (2) Obtain a real-time trajectory data set Track during the typhoon generation from the public platform of the meteorological bureau;

[0011] (3) According to the spatio-temporal information of the social media data, convert the original social media text data set S into a spatio-temporal coding data set S c :

[0012] S c ={(T1, (x1, y1), tc1), …, (T i , (x i , y i ), tc i ), …, (T n , (x n , y n ), tc n )}

[0013] where T i is text data, (x i , y i ) represents the spatial grid coding of each piece of data, where 0 ≤ x i ≤ 3, 0 ≤ y i ≤ 3, and tc i represents the result of time coding;

[0014] (4) Perform time encoding on the typhoon track dataset Track to obtain the time-encoded typhoon track dataset Track c :

[0015] Track c = {(La1, Ln1, tc1),..., (La i , Ln i , tc i ),..., (La m , Ln m , tc m )}

[0016] where tc i is the time encoding value, and tc m is the last time encoding value;

[0017] (5) Use the BERT language model to extract the 16-dimensional semantic features of the text T of each piece of data in the spatio-temporal encoding data S c to obtain the semantic feature dataset W: i where,

[0018]

[0019] represents the j-th dimension of the semantic feature extracted from the text T i ; the j-th dimension;

[0020] (6) According to the semantic feature dataset W, sum the text semantic features with the same spatio-temporal encoding and perform feature space reconstruction to obtain the reconstructed semantic feature dataset C:

[0021] C = {W0,..W tc ,..., W l}

[0022] where represents the spatially reconstructed semantic feature under the tc time encoding, l is the maximum value of the time encoding tc, represents the reconstructed semantic feature at the time encoding tc and the spatial encoding (x, y), and all the semantic features with the same spatio-temporal encoding in the semantic feature dataset W are summed and then reconstructed into a 4X4 matrix size;

[0023] (7) According to the spatio-temporal encoding method, calculate the distance dataset from the center of each spatial grid to the typhoon center under the same time encoding:

[0024] D = {D0,..., D tc ,..., D l}

[0025] Among them represents the distance from the center of all spatio-temporal grids to the typhoon center under the time code tc, represents the distance from the center of the area represented by the spatial code (x, y) to the typhoon center under the time code tc;

[0026] (8) Construct a convolutional neural network model G composed of five layers of three-dimensional convolutional layers and two layers of convolutional-long short-term memory artificial neural layers cross-connected, and use the root mean square error RMSE function as the loss function of this network;

[0027] (9) Use the dataset C as the feature data, the dataset D as the true value, combine the reconstructed semantic feature dataset C and the typhoon center distance dataset D into the typhoon distance semantic dataset B, and then divide this dataset B into a training set Train and a test set Test according to 3:7;

[0028] (10) Input the feature data of the training set Train into the convolutional neural network G, and perform iterative training on it by the backpropagation method until the loss function converges to obtain a trained convolutional neural network model;

[0029] (11) Input the feature data in the test set Test into the trained convolutional neural network to obtain the predicted value of the distance from each spatio-temporal grid center to the typhoon center

[0030]

[0031] Among them represents the predicted value of the distance from the spatial center represented by the spatio-temporal code [tc, (x, y)] to the typhoon center;

[0032] (12) According to the predicted value of the distance from each spatio-temporal grid center to the typhoon center Use the spatio-temporal grid positioning algorithm based on location service to locate the typhoon, and determine the latitude and longitude positions of the predicted typhoon center.

[0033] Compared with the existing inventions, the advantages of the present invention are as follows:

[0034] First, the present invention encodes the dataset S through the spatio-temporal grid encoding method and performs spatio-temporal reconstruction of semantic features to obtain the reconstructed semantic feature dataset C, which can make full use of the abstract semantic features of the dataset, extract data features related to typhoon trajectories to a greater extent, and solve the problem of data loss caused by ignoring abstract semantic features during feature extraction in the prior art.

[0035] Second, the present invention constructs a convolutional neural network G with cross-connections between three-dimensional convolution and convolution-long short-term memory artificial neural layers, and uses this convolutional neural network G to predict the distance from the center of the spatio-temporal grid to the typhoon center, which can better capture the global spatio-temporal correlation of data features, improve the positioning accuracy, and solve the problem of inaccurate positioning in the prior art due to poor capture ability of the spatio-temporal relationship of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is the implementation flowchart of the present invention;

[0037] Figure 2 is a schematic diagram of typhoon positioning in the present invention;

[0038] Figure 3 is a schematic diagram of the simulation experiment results of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] The present invention will be described in detail below with reference to the accompanying drawings:

[0040] The typhoon positioning method based on social media in this embodiment can use social media data to predict the position of the typhoon center during the typhoon approaching. Here, we take the 2012 Sandy typhoon as an example to introduce how to use the typhoon positioning method based on social media.

[0041] Refer to Figure 1 , the specific implementation steps of this embodiment are as follows:

[0042] Step 1: Obtain social media data with spatio-temporal information.

[0043] Obtain the text social media data set with time stamps and GPS information during the Sandy typhoon from the social platform Twitter:

[0044] S = {(T1, La1, Ln1, tc1), (T2, La2, Ln2, tc2),..., (T n , La n , Ln n , tc n )}},

[0045] where, T i represents the text data, which is the text data related to the typhoon published by users in the typhoon-affected area on social media;

[0046] La i represents longitude and Ln iDenote the latitude, which is the GPS data of the social media data published by users in the typhoon - affected area. Its latitude range is [40.5°N, 41°N], and the longitude range is [-74°W, -73.5°W]; where N is the unit of latitude for south latitude, and W is the unit of longitude for west longitude.

[0047] tc i Denote the timestamp, which is the time when users in the typhoon - affected area publish social media data. Its time interval is between the generation and the end of the hurricane, that is, from 15:00 on October 22, 2012 to 21:00 on October 29, 2012.

[0048] Let n denote the number of the dataset S, which is the number of all typhoon - related social media published by users in the affected area during the typhoon generation to the end time.

[0049] Step 2: Obtain the real - time typhoon track dataset.

[0050] Obtain the real - time track dataset Track during the typhoon generation from the public platform of the meteorological bureau, which is expressed as follows:

[0051] Track = {(La1, Ln1, t1),..., (La i , Ln i , t i ),..., (La m , Ln m , t m )}

[0052] Where La i , Ln i represent the longitude and latitude information of the typhoon center at time t i . t i represents the timestamp, whose value ranges from the generation time of the typhoon at 15:00 on October 22, 2012 to the end at 21:00 on October 29, 2012. The interval between each value t i and t i+1 is 30s. Therefore, the length m of the typhoon track dataset Track is 20880.

[0053] Step 3: Perform spatio - temporal encoding on the social media dataset.

[0054] This step is to perform spatio - temporal encoding on the data according to the timestamp t and the GPS information La i and Ln i representing the latitude respectively, and the implementation is as follows:

[0055] (3.1) Taking the starting time of the typhoon as the starting time of time encoding, after converting the timestamp t of the dataset S into seconds, divide it at 30 - second intervals to obtain the time encoding tc of each data:

[0056]

[0057] Among them, 1350889200 is the number of seconds in the computer when the starting time of the typhoon in this example is 15:00 on October 22, 2012;

[0058] (3.2) According to the latitude and longitude data La i and Ln i of the data set S, in the space represented by the data set S, the latitude range is [40.5°N, 41°N], and the longitude range is [-74°W, -73.5°W]. It is divided into 4x4 spatial grids, where the height width is

[0059] (3.3) Divide the latitude and longitude information Ln i and La i of each data in the data set S into the corresponding grid to obtain the spatial encoding (x i , y i ):

[0060]

[0061] (3.4) Perform spatial encoding and time encoding on each data in the data set S to obtain the spatial encoding (x i , y i ) and time encoding tc i of each data in the data set S, and combine it with the text data T i of each data to form the spatio-temporal encoding data set S c :

[0062] S c = {(T1, (x1, y1), tc1), …, (T i , (x i , y i ), tc i ), …, (T n , (x n , y n ), tc n )}.

[0063] Step 4: Perform time encoding on the typhoon track data set.

[0064] This step performs spatio-temporal encoding on the data according to the time stamp t, and the implementation is as follows:

[0065] (4.1) According to the timestamp \(t\) of the typhoon track dataset Track, with 15:00 on October 22, 2012 as the starting time of time encoding for the data, after converting the timestamp \(t\) of the dataset Track into seconds, it is divided at 30 - second intervals to obtain the time encoding \(tc\) of each piece of data in the typhoon track dataset Track:

[0066]

[0067] Among them, 1350889200 is the number of seconds in the computer with 15:00 on October 22, 2012 as the starting time of the typhoon in this example;

[0068] (4.2) Perform spatial encoding and time encoding on each piece of data in the dataset Track to obtain the time encoding \(tc\) of each piece of data i , and combine it with the longitude and latitude information of the typhoon center of each piece of data to form the time - encoded typhoon track dataset Track c :

[0069] Track c = \(\{(La1, Ln1, tc1),..., (La i , Ln i , tc i ),..., (La m , Ln m , tc m )\}\).

[0070] Step 5: Use the BERT model to extract the semantic features of the text data.

[0071] This step will extract the semantic features of the text data in the spatio - temporal encoding dataset S c , and the implementation is as follows:

[0072] (5.1) Use the existing pre - trained BERT model to tokenize, encode and extract the 768 - dimensional semantic features from the text in the spatio - temporal encoding data

[0073] (5.2) Use the principal component analysis method to reduce the semantic features to 16 - dimensional semantic features

[0074]

[0075] For example, for the spatio - temporal encoding dataset S cOne of the text data in it, "People that hate rainy weather obviously don't have rear - drive cars", after being extracted and dimension - reduced by the BERT model, obtains a 16 - dimensional output [0.27990687, - 0.2208395, 0.3265599, 0.1997981, 0.17458375, - 1.1405083, 0.39390883, 0.01827345, - 0.26899743, - 0.60412407, 0.12131958, 0.5419242, - 0.2532862, - 0.2393016, - 0.11166136, 0.3595541];

[0076] (5.3) For the spatio - temporal encoded dataset S c For each text T in the data i Execute step (5.1) and step (5.2) to obtain the semantic feature dataset W:

[0077]

[0078] Among them, Represents the semantic feature extracted from the text T i Of the j - th dimension.

[0079] Step 6: Reconstruct the semantic features according to the spatio - temporal encoding.

[0080] This step will perform spatial reconstruction on the text semantic features with the same spatio - temporal encoding according to the semantic feature dataset W, as follows:

[0081] (6.1) Add up each dimension of all the semantic features with the same spatio - temporal encoding [tc, (x, y)] in the dataset W To obtain the overall semantic feature with the spatio - temporal encoding [tc, (x, y)]

[0082]

[0083] Among them, Represents The j - th dimension data of, and k is the number of text semantic features with the spatio - temporal encoding [tc, (x, y)];

[0084] (6.2) Reconstruct the overall semantic feature with the same spatio - temporal encoding [tc, (x, y)] obtained in (6.1) Into a 4x4 spatial - reconstructed semantic feature ​

[0085]

[0086] (6.3) Reconstruct the semantic features of space with the same time code tc According to the order of dividing the 4*4 grid by space encoding, for Reconstruct again to obtain the reconstructed semantic feature W tc :

[0087]

[0088] (6.4) Sum and reconstruct twice all the text semantic features with the same time code in the semantic feature dataset W according to steps (6.1)-(6.3) to obtain the reconstructed semantic feature dataset C:

[0089] C = {W0,..W tc ,…,W l}.

[0090] Step 7: Calculate the distance from the center of each grid to the typhoon center according to the spatio-temporal encoding.

[0091] This step calculates the distance from the center of each grid to the typhoon center according to the spatio-temporal encoding, and the implementation is as follows:

[0092] (7.1) Calculate the distance from the center of the area represented by the time code tc and the space code (i, j) to the typhoon center at this time code

[0093]

[0094] where La tc , Ln tc represent the longitude and latitude of the typhoon center at the spatio-temporal code tc moment, represents the longitude and latitude of the center of the area represented by the space code (x, y);

[0095] (7.2) Calculate the distance from the center of the grid with the time code tc to the typhoon center for all grids according to step (7.1) and combine into a matrix D tc :

[0096]

[0097] (7.3) Calculate each time code to obtain the distance dataset D from the center of the grid to the typhoon center:

[0098] D = {D0,…,D tc ,…,D l}

[0099] Where l represents the last time code, and in this example, the value of l is 20879.

[0100] Step 8: Construct the convolutional neural network model G.

[0101] (8.1) Set the parameters of the five-layer three-dimensional convolutional layer:

[0102] The first-layer three-dimensional convolutional layer a contains 32 convolutional kernels, and the size of the convolutional kernels is 3x16x16;

[0103] The second-layer three-dimensional convolutional layer b contains 16 convolutional kernels, and the size of the convolutional kernels is 3x8x8;

[0104] The third-layer three-dimensional convolutional layer c contains 8 convolutional kernels, and the size of the convolutional kernels is 3x4x4;

[0105] The fourth-layer three-dimensional convolutional layer d contains 8 convolutional kernels, and the size of the convolutional kernels is 3x4x4;

[0106] The fifth-layer three-dimensional convolutional layer e contains 1 convolutional kernel, and the size of the convolutional kernels is 3x4x4;

[0107] (8.2) Set the parameters of the two-layer convolutional-long short-term memory artificial neural layer:

[0108] The first-layer convolutional-long short-term memory artificial neural layer f contains 16 convolutional kernels, and the size of the convolutional kernels is 16x16;

[0109] The second-layer convolutional-long short-term memory artificial neural layer g contains 16 convolutional kernels, and the size of the convolutional kernels is 4x4;

[0110] (8.3) Cross-connect the five-layer three-dimensional convolutional layer and the two-layer convolutional-long short-term memory artificial neural layer in the order a, b, c, f, d, g, e to form the convolutional neural network model G;

[0111] (8.4) Select the rectified linear unit ReLU as the activation function of the convolutional neural network G, and use the root mean square error RMSE as the loss function of the network

[0112] ReLU(x) = max(0, x)

[0113]

[0114] Step 9: Combine the typhoon distance semantic dataset B and divide it into a training set and a test set.

[0115] Use dataset C as the feature data, dataset D as the true value, and combine the reconstructed semantic feature dataset C and the typhoon center distance dataset D into the typhoon distance semantic dataset B;

[0116] Then divide dataset B into a training set Train and a test set Test at a ratio of 3:7.

[0117] Step 10: Train the convolutional neural network G.

[0118] (10.1) Randomly initialize the parameters of the convolutional neural network. In this example, the initial learning rate is 0.0001, and the loss function convergence threshold is 0.001;

[0119] (10.2) Input the feature data of the training set Train into the input layer of the convolutional neural network G, perform forward propagation through the convolutional layer and activation function of the network to obtain the prediction result, and then compare the prediction result with the true value of the training set Train to calculate the loss function of the convolutional neural network;

[0120] (10.3) Through the backpropagation algorithm, calculate the gradient of the loss function with respect to the network parameters, multiply the calculated gradient by the learning rate to update the network parameters, and use the Adam optimizer to adaptively adjust the learning rate according to the first-order moment estimate and second-order moment estimate of the gradient. At the same time, use the momentum technique to accelerate the optimization process;

[0121] (10.4) Repeat steps (10.2) to (10.3) continuously iterate until the loss function converges to obtain the trained convolutional neural network G.

[0122] Step 11: Predict the distance from each spatio-temporal grid to the typhoon center.

[0123] Input the feature data in the test set Test into the trained convolutional neural network to obtain the predicted value of the distance from the center of each spatio-temporal grid to the typhoon center

[0124]

[0125] Where represents the predicted value of the distance from the spatial center represented by the spatio-temporal encoding [tc, (x, y)] to the typhoon center.

[0126] In this example, for the spatial grid with a time stamp of 18:00 on October 25, 2012 and a time encoding of 9000, the distance R from the center of each grid to the typhoon center obtained by predicting through the neural convolutional network 9000 :

[0127]

[0128] Step 12: Determine the longitude and latitude of the typhoon center.

[0129] In this step, based on the predicted distance from the center of each spatio-temporal grid to the typhoon center The typhoon is located using the spatio-temporal grid positioning algorithm based on location-based services, and the implementation is as follows:

[0130] (12.2) Taking the center of each spatio-temporal grid as the center of a circle, draw a circle with the radius being the distance prediction value output by the convolutional neural network to obtain the intersection dataset E of all circles. As shown in Figure 2 taking the grid centers B 1,1 , B 1,2 , B 1,4 and B 2,4 as the centers of circles, and the intersections of each circle are the black dots in the figure;

[0131] (12.2) Use the k-means clustering algorithm to find the cluster with the largest number of intersections in the dataset E:

[0132] Divide the intersections in the dataset E into k intersection clusters, and find the cluster with the largest number of intersections by counting the number of intersections in each cluster;

[0133] Take the centroid of the cluster with the largest number of intersections as the typhoon center at the time encoding tc.

[0134] In this example, for the black dots in Figure 2 , analyze them using the clustering method to obtain the largest cluster, that is, the intersections contained in the black dashed box. Then the centroid of this cluster is the predicted typhoon center position, that is, Figure 2 the position of the black triangle in

[0135] The effects of the present invention can be further illustrated by the following simulation experiments:

[0136] First, simulation experiment conditions:

[0137] The operating system used in this experiment is Ubuntu16.04, the deep learning framework is Pytorch1.0.1, the hardware conditions are that the CPU is an eight-core Intel Xeon E5-2630 v4, the memory is 32G, the GPU is an Nvidia Tesla P100, and the video memory is 16G;

[0138] The experimental data are the typhoon-related social media data released by users in the typhoon-affected area during the 2012 US Sandy typhoon and the 2012 US Sandy typhoon track data.

[0139] Second, simulation experiment content and results:

[0140] Under the above simulation experiment conditions, use the present invention to predict the typhoon center trajectory, and the results are as followsFigure 3 , where the solid broken line is the actual moving track of the typhoon center, and the dashed broken line is the predicted typhoon center track in the simulation experiment of the present invention.

[0141] From Figure 3 It can be seen that the predicted typhoon track of the present invention is consistent with the trend of the actual typhoon track, and the tracks are roughly coincident. At the same time, when the moving direction of the actual typhoon center position changes, the moving direction of the predicted typhoon track also changes accordingly. When the typhoon center is far from the affected area, the prediction effect of the typhoon track is poor. When the hurricane center is closer to the affected area, the prediction effect is better, because users can more strongly feel the impact of the disaster, so more intuitive reactions of people to typhoons can be captured in the semantic features.

[0142] At the same time, from Figure 3 it can also be seen that the predicted track is closest to the actual track at the inflection point of the track, but the distance between the predicted track and the actual track increases afterwards. This is because the typhoon track changes significantly in direction at the inflection point, and the model needs to adjust the direction of the typhoon track prediction at this moment, which will cause the distance to increase, but the error distance will gradually decrease after the adjustment.

Claims

1. A typhoon positioning method based on social media data, characterized in that, The steps include: (1) Obtain a social media text dataset S with spatiotemporal information from social media platforms in the typhoon-affected area during the typhoon period; (2) Obtain the real-time track dataset Track during the typhoon formation period from the public platform of the Meteorological Bureau; (3) According to the spatio-temporal information of social media data, the original social media text dataset S is transformed into a spatio-temporal encoded dataset S according to the spatio-temporal grid encoding method c : S c ={(T1,(x1,y1),tc1),...,(T i ,(x i ,y i ),tc i ),...,(T n ,(x n ,y n ),tc n )} Where T i is text data, (x i ,y i ) represents the spatial grid encoding of each data where 0≤x i ≤3, 0≤y i ≤3, tc i Indicates the result of time encoding; (4) Perform time encoding on the typhoon track dataset Track to obtain the time-encoded typhoon track dataset Track c : Track c s{(La1,Ln1,tc1),...,(La i ,Ln i ,tc i ),...,(The m ,Ln m ,tc m )} where tc i is the time code value, tc m is the last time code value; (5) Extract spatiotemporal coding data S using the BERT language model c The text of each data T i The sixteen-dimensional semantic features of , get the semantic feature dataset W: in, Indicates the text T i Extracted semantic features The jth dimension of (6) Based on the semantic feature dataset W, the semantic features of the text encoded in the same spatiotemporal encoding are summed and the feature space is reconstructed to obtain the reconstructed semantic feature dataset C: C = {W0,..W tc ,..., W l} in represents the spatial reconstruction semantic feature under the time coding tc, l is the maximum value of the time coding tc, Represents the reconstructed semantic features of the time code tc and the space code (x, y), and all the semantic features under the same time and space coding in the semantic feature dataset W are After summing, reconstruct it into a 4x4 matrix size; (7) According to the space-time coding method, calculate the distance data set from the center of each spatial grid to the typhoon center under the same time coding: D = {D0, ..., D tc , ..., D l} in represents the distance from the center of all time and space grids to the typhoon center under the time code tc, Indicates the distance from the center of the area represented by the time code tc and the space code (x, y) to the typhoon center; (8) Construct a convolutional neural network model G consisting of five layers of three-dimensional convolutional layers and two layers of convolution-long short-term memory artificial neural network layers, and use the root mean square error (RMSE) function as the loss function of the network; (9) Take dataset C as feature data and dataset D as ground truth, and combine the reconstructed semantic feature dataset C and the typhoon center distance dataset D into a typhoon distance semantic dataset B. Then divide the dataset B into a training set (Train) and a test set (Test) according to a ratio of 3:

7. (10) Input the feature data of the training set Train into the convolutional neural network G, and iteratively train it by back propagation until the loss function converges to obtain a trained convolutional neural network model network; (11) Input the feature data in the test set Test into the trained convolutional neural network to obtain the predicted value of the distance from the center of each spatio-temporal grid to the typhoon center in Represents the predicted value of the distance between the center of the space represented by the spatiotemporal code [t, (x, y)] and the typhoon center; (12)Predict the distance from the center of each spatio-temporal grid to the typhoon center Use the spatio-temporal grid positioning algorithm based on location services to locate the typhoon and determine the latitude and longitude positions of the predicted typhoon center.

2. The method according to claim 1, characterized in that, The social media text dataset S obtained in step (1) is represented as follows: S={(T1,La1,Ln1,t1),...,(T i ,The i ,Ln i ,t i ),...,(T n ,The n ,Ln n ,t n )} where T i represents text data, La i represents longitude, Ln i represents latitude information, t i represents the timestamp when the social media information is sent, and n represents the length of the data set.

3. The method according to claim 1, wherein Step (2) obtains the real-time trajectory dataset Track during the typhoon generation period from the public platform of the Meteorological Bureau, which is expressed as follows: Track = {(La1, Ln1, t1),..., (La i , Ln i , t i ),..., (La m , Ln m , t m )} where t i Indicates timestamp, La i , Ln i Indicates that at t i The longitude and latitude information of the typhoon center at the moment, m represents the length of the typhoon track dataset.

4. The method according to claim 1, characterized in that: In step (3), the original social media text dataset S is transformed into a spatio-temporal coding dataset S according to the spatio-temporal grid coding method c , which is achieved as follows: (3a) Based on the timestamp t of the dataset S, the time code tc is obtained by dividing it into 30-second time intervals: where t min represents the starting timestamp in the dataset S; (3b) Based on the longitude and latitude data of the dataset S, the spatial range represented by the dataset S is divided into 4*4 spatial grids, and the longitude and latitude of the dataset S are divided into the corresponding grids to obtain the spatial code (x, y): in Indicates the width of the grid, Indicates the height of the grid, Tweet lat-max and Tweet lat-min Indicates the maximum and minimum values of latitude in the dataset S, Tweet long-max and Tweet long-min Indicates the maximum and minimum values of longitude in the dataset S; (3c)Spatial and temporal encoding is performed on each piece of data in the dataset S to obtain the spatial encoding (x i , y i ) and temporal encoding tc i of each piece of data, and combine them with the text data T i of each piece of data to form the spatio-temporal encoding dataset S c : S c = {(T1, (x1, y1), tc1),..., (T i , (x i , y i ), tc i ),..., (T n , (x n , y n ), tc n )}.

5. The method according to claim 1, wherein In step (6), all semantic features with the same spatio-temporal encoding [tc, (x, y)] in the semantic feature dataset W are summed and reconstructed into a 4x4 matrix to obtain the reconstructed semantic feature W tc , which is implemented as follows: (6a) All semantic features with the same spatiotemporal encoding [tc, (x, y)] in the dataset W are Perform and obtain overall semantic features in Represents the text semantic features of the i-th data whose spatiotemporal encoding is [tc, (x, y)] express The j-th dimension data of , k is the number of text semantic features whose spatiotemporal encoding is [tc, (x, y)]. (6b) Encode all time as the overall semantic features of tc Reconstruct semantic features according to the 4x4 space 6. The method according to claim 1, characterized in that In step (6), calculate the distance from the center of each spatio-temporal grid to the typhoon center The formula is as follows: where La tc and Ln tc represent the longitude and latitude of the typhoon center at time tc, represents the longitude and latitude of the center of the area represented by the spatial code (x, y).

7. The method according to claim 1, characterized in that The structural parameters of the five-layer three-dimensional convolutional layer and the two-layer convolution-long short-term memory artificial neural network layer in the convolutional neural network model G in step (8) are as follows: The first three-dimensional convolution layer a contains 32 convolution kernels with a size of 3x16x16; The second three-dimensional convolution layer b contains 16 convolution kernels with a size of 3x8x8; The third three-dimensional convolution layer c contains 8 convolution kernels with a size of 3x4x4; The fourth three-dimensional convolution layer d contains 8 convolution kernels with a size of 3x4x4; The fifth three-dimensional convolution layer e contains one convolution kernel with a size of 3x4x4; The first convolutional layer - long short-term memory artificial neural layer f, contains 16 convolution kernels with a size of 16x16; The second convolutional layer - long short-term memory artificial neural layer g, contains 16 convolution kernels with a convolution kernel size of 4x4; The above five three-dimensional convolutional layers and two convolutional-long short-term memory artificial neural layers are connected in the order of a, b, c, f, d, g, and e.

8. The method according to claim 1, characterized in that The loss function of the convolutional neural network model G is set in step (8) and is expressed as follows: Among them D tc is the distance of time encoding tc in dataset D, R tc The predicted value output by the model.

9. The method according to claim 1, wherein In step (10), the convolutional neural network model G is iteratively trained as follows: (10a) Input the feature data of the training set Train into the network, perform forward propagation to calculate the prediction result, and calculate the loss function based on the difference between the prediction result and the true value of the training set Train; (10b) Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the network parameters, use the result obtained by multiplying this gradient by the learning rate to update the network parameters, and use the Adam optimization algorithm to adaptively adjust the learning rate; (10c) Repeat steps 10a to 10b, continuously perform the iterative process of forward propagation, backpropagation, and parameter update until the loss function converges, and stop the iteration to obtain the trained convolutional neural network G.

10. The method according to claim 1, characterized in that In step (12), the distance prediction value from the center of each spatiotemporal grid to the typhoon center is Typhoon positioning is achieved using a spatiotemporal grid positioning algorithm based on location services, as follows: (12a) With the center of each spatio-temporal grid as the center of a circle, draw a circle with a radius equal to the distance prediction value output by the convolutional neural network to obtain the dataset E of the intersection points of all circles; (12b) Use the k-means clustering algorithm to find the cluster with the largest number of intersection points in the dataset E, and the longitude and latitude of the centroid of this cluster are the longitude and latitude of the typhoon center at the time encoding tc moment.

Images