A typhoon track prediction method based on deep learning and particle filtering

Abstract

The application discloses a typhoon track prediction method based on deep learning and particle filtering, and comprises the following steps: constructing a typhoon track prediction model based on deep learning and particle filtering, which is based on a Bayesian framework and is divided into typhoon observation and typhoon updating; in the typhoon observation, a typhoon probability map is obtained by means of deep learning and considering typhoon features, the probability map is used to correct and constrain the predicted typhoon range in the typhoon track prediction model; in the typhoon updating, a multi-modal neural network is used to predict the typhoon position in combination with historical environmental factors, the prediction process simultaneously comprises particle prediction, the newly obtained particle weight is updated by means of the probability map obtained by the typhoon observation, the particle weighted average is used to obtain the posterior estimation of the typhoon position at the next moment, and the posterior estimation is used as the input of the next prediction; the typhoon observation and the typhoon updating are continuously cycled to obtain the posterior estimation of the typhoon eye position and obtain the typhoon track prediction result. The typhoon path prediction problem under the condition that the real path is not real-time available is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to intelligent typhoon track prediction, in particular to a typhoon track prediction method based on deep learning and particle filtering. BACKGROUND

[0002] Typhoon is a tropical cyclone occurring in the western Pacific and its adjacent sea area, which is a common extreme weather phenomenon, threatening human life safety and production and life. Typhoon track prediction is a typical problem of typhoon research. The formation of typhoon is affected by many factors such as meteorological environment, cyclone thermodynamics and dynamics. In the traditional method, the typhoon track prediction method mainly includes numerical prediction model and statistical prediction model. Numerical model has high calculation demand to simulate the complex internal structure of typhoon. With the progress of computer technology, numerical model has been further developed, but there are still problems of high calculation complexity and low prediction accuracy. Statistical model mainly predicts based on historical record characteristics. With the improvement of ocean observation station, ground station and meteorological satellite system, the amount of data available for statistical model has doubled, resulting in big data problem in selecting available historical features from massive historical data. In recent years, artificial intelligence technology has flourished, and deep learning algorithm has been gradually applied to image processing, target detection and other fields, especially in dealing with complex data and nonlinear relationship. On the one hand, the efficient feature extraction capability of deep learning can extract hidden features in massive data set and improve the generalization ability, such as using CNN to extract spatial information. On the other hand, deep learning has an advantage in dealing with continuous long-time data set, such as GRU which can extract time information in time series.

[0003] In recent years, deep learning algorithm has been applied to image processing, natural language processing and target detection, especially in the field of meteorology with complex influencing factors and rich data such as typhoon. Deep learning algorithm can extract hidden features from large variable data set, so as to establish an effective mapping between marine environmental elements and typhoon track. The existing typhoon track prediction method based on deep learning mainly uses the strong feature extraction ability of network to extract time information and spatial information. The time information mainly includes the change of relevant marine elements such as pressure, potential and wind speed between the historical track of typhoon and the previous historical node, which reflects the historical law during the movement of typhoon. The spatial information mainly considers the correlation between all sea area nodes, which reflects the overall characteristics of a certain sea area at a fixed time.

[0004] The typhoon historical position information plays an important role in typhoon prediction research, can further converge the typhoon prediction range, greatly improve the typhoon trajectory prediction accuracy, and has important theoretical value. However, in the actual sea state, the typhoon is strong in time-varying, and the typhoon position at a certain time is often obtained through complex observation and calculation, and cannot be obtained in real time, and then the typhoon path under the condition that the real path is not real-time available cannot be predicted. SUMMARY

[0005] Therefore, the application provides a typhoon trajectory prediction method based on deep learning and particle filtering to solve the problem of typhoon path prediction under the condition that the real path is not real-time available.

[0006] To this end, the application provides the following technical solutions.

[0007] The application discloses a typhoon trajectory prediction method based on deep learning and particle filtering, comprising:

[0008] A typhoon trajectory prediction model based on deep learning and particle filtering is constructed, the typhoon trajectory prediction model is based on a Bayesian framework, and is divided into a typhoon observation and a typhoon update;

[0009] The typhoon observation part obtains a typhoon probability map by means of deep learning and considering typhoon characteristics, and the probability map corrects and constrains the predicted typhoon range in the typhoon trajectory prediction model;

[0010] The typhoon update part predicts the typhoon position by using a multi-modal neural network combined with historical environmental elements, the prediction process simultaneously includes particle prediction, the newly obtained particle weight is updated by the probability map obtained by the typhoon observation, the particle weighted average is obtained, the posterior estimation of the typhoon position at the next moment is obtained, and this is taken as the input of the next prediction;

[0011] The typhoon observation and the typhoon update are continuously cycled, the posterior estimation about the typhoon eye position is obtained, and the typhoon trajectory prediction result is obtained.

[0012] Further, the typhoon probability map is obtained by means of deep learning and considering typhoon characteristics, comprising:

[0013] An observation model based on a convolutional neural network is constructed; the observation model comprises four convolutional neural network layers Conv2d for extracting features from an input typhoon image; wherein the input tensor accepted by the first layer contains 9 channels, respectively representing the potential and wind speed features at three pressure levels; after each convolutional layer, a ReLU activation function is used for nonlinear conversion, and a maximum pooling MaxPool2d is used for downsampling; then a softmax activation function is defined to convert the output of the neural network into a probability estimate for each category; in addition to the first four convolutional layers, the fifth convolutional layer uses a 1*1 convolution kernel size to reduce the number of feature map channels, and the sixth convolutional layer uses the same size of the convolution kernel to convert the feature map into the output probability of each category; in the last two convolutional layers, 128 channels of input are converted into 32 channels and 2 channels of output, respectively; in the forward propagation function, the forward propagation calculation of the convolutional layer is realized; wherein the input tensor is converted into a feature map in each convolutional layer, and is flattened into a one-dimensional vector after the last layer; finally, the vector is input into the sigmoid activation function to compress the output value into the range of 0 to 1 to represent the probability estimate for each category.

[0014] The observation model is used to obtain the probability of typhoon occurrence at the current time.

[0015] Further, the observation model is used to obtain the probability of typhoon occurrence at the current time, comprising:

[0016] Meteorological samples are collected based on the three-dimensional characteristics of typhoons and meteorological principles to form a typhoon trajectory dataset;

[0017] The region sample is cropped from the typhoon trajectory dataset;

[0018] It is judged whether the region is covered by the typhoon range, if not, the sample is determined as a negative sample of typhoon; if yes, the sample is determined as a positive sample of typhoon;

[0019] The sample is formed into a typhoon positive and negative sample dataset;

[0020] The typhoon positive and negative sample dataset is input into a convolutional neural network, and a typhoon feature map is obtained based on the convolutional neural network;

[0021] The classifier scans the current time global sea area to obtain a binary image;

[0022] A Gaussian distribution is constructed;

[0023] A global sea area typhoon occurrence probability map is obtained based on threshold filtering and Gaussian distribution.

[0024] Further, the typhoon position prediction is performed by using a multi-modal neural network combined with historical environmental factors, comprising:

[0025] A motion model based on a deep neural fusion network is constructed. The deep neural fusion network includes a wind field subnetwork, a potential subnetwork, and a historical trajectory subnetwork, which extract the wind field, potential, and historical trajectory separately. The deep neural fusion network also includes a fusion subnetwork. The fusion subnetwork is used to merge the outputs of the wind field subnetwork, the potential subnetwork, and the historical trajectory subnetwork into a single vector, and outputs the final result through two fully connected layers.

[0026] The motion model is used to predict the displacement of the typhoon at the next moment.

[0027] Furthermore, the wind field sub-network extracts features from the input typhoon data. The wind field sub-network includes three convolutional layers and three fully connected layers. Each convolutional layer is followed by a normalization and ReLU activation function, and outputs a vector of size 2, which is the displacement of the typhoon at the next moment.

[0028] The potential subnetwork extracts features from the input typhoon data. The potential subnetwork includes 3 convolutional layers and 3 fully connected layers. Each convolutional layer is followed by a normalization and ReLU activation function, and outputs a vector of size 2, which is the displacement of the typhoon at the next moment.

[0029] The historical trajectory subnetwork includes a fully connected neural network, and its output is the displacement of the typhoon at the next moment.

[0030] Further, training the deep neural fusion network includes:

[0031] Three independent networks were pre-trained;

[0032] The three trained networks are spliced ​​and merged in the last layer and then trained again to fine-tune the parameters.

[0033] Train the fusion subnetwork.

[0034] Furthermore, the particle prediction includes:

[0035] Particles are set using a two-dimensional Gaussian distribution based on the known current location of the typhoon.

[0036] The displacement at the next moment is predicted by a multimodal fusion network, and the particle position is updated according to the displacement.

[0037] The particle weights are updated based on the probability map, and the weighted average of the particles is used to obtain the posterior estimate of the typhoon's position at the next moment. This estimate is then used as the input for the next prediction, and the particles are resampled every preset number of times.

[0038] Furthermore, the particle weights are updated based on the probability graph, including:

[0039] The probability value corresponding to the position of each particle is calculated based on the probability graph. The weight of all particles is calculated, and the weight of all particles is normalized to obtain the new weight of each particle, thereby obtaining the particle set at the current time.

[0040] Furthermore, the particle resampling includes:

[0041] A resampling strategy using roulette wheel is adopted to eliminate the particle depletion problem and obtain the final posterior distribution set of particles at the current time.

[0042] Furthermore, a posterior estimate of the location of the typhoon eye is obtained, including:

[0043] For a single typhoon eye, the weighted centroid of the posterior distribution particle set is the current position of the typhoon eye. For multiple typhoon eyes, the number of typhoon eyes is determined, and then the posterior distribution particle set is clustered. The weighted centroid of each cluster is the position of the eye.

[0044] Advantages and positive effects of this invention: Unlike other methods that use typhoon location information for prediction, this invention proposes a typhoon trajectory prediction framework based on deep learning and particle filtering. This framework is based on a Bayesian architecture and consists of two parts: typhoon observation and typhoon update. When the historical typhoon trajectory is unknown, the typhoon observation part uses deep learning and considers typhoon characteristics to obtain a typhoon probability map. This probability map corrects and constrains the predicted typhoon range within the overall framework, improving prediction accuracy. In the typhoon update part, a multimodal neural network is used in conjunction with historical environmental factors to predict the typhoon location. These environmental factors do not contain any location information. This prediction process also includes particle prediction. The newly obtained particle weights are updated using the probability map obtained from typhoon observation. The weighted average of the particles yields a posterior estimate of the typhoon's location at the next moment, which is used as the input for the next prediction. Observation and updating are continuously cyclical, solving the problem of typhoon path prediction when the actual path is not available in real time. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a diagram of the typhoon trajectory prediction architecture based on deep learning and particle filtering in an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of a motion model based on a deep neural fusion network in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of an observation model based on a convolutional neural network in an embodiment of the present invention;

[0049] Figure 4 This is a flowchart of the probability graph generation process of the observation model in an embodiment of the present invention. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0051] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0052] The typhoon trajectory prediction method based on deep learning and particle filtering in this embodiment is based on a particle filtering-based typhoon trajectory prediction framework, combined with an observation model based on a convolutional neural network and a motion model based on a deep neural fusion network. The observation model mainly obtains the probability of typhoon occurrence at the current global node through deep learning, while the motion model mainly predicts the typhoon's displacement at the next moment through a deep neural fusion network. Finally, the particle filtering framework is used to correct the final prediction, thereby achieving efficient and high-precision prediction when the typhoon's current position cannot be obtained in real time.

[0053] 1. Sample Construction

[0054] This embodiment requires constructing sufficiently rich samples for both the observation model based on convolutional neural networks and the motion model based on deep neural fusion networks. The samples are sourced from the fifth-generation atmospheric reanalysis dataset (REA5) provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) and typhoon track data for the Northwest Pacific region provided by the China Meteorological Administration (CMA). The reanalysis dataset has a resolution of 1 hour, and the track dataset has a resolution of 6 hours, spanning the period from 1999 to 2020.

[0055] This embodiment constructs observation model samples based on convolutional neural networks. Using the optimal typhoon path set for the Northwest Pacific region (0°N-60°N, 100°E-180°E) provided by the China Meteorological Administration's Tropical Cyclone Center as a reference, positive samples are constructed around the typhoon eye, with offsets in various directions around the center. Negative samples are constructed in other areas unaffected by typhoons. Both positive and negative samples are 5°*5° (latitude and longitude), representing an actual area of ​​approximately a 555km*555km rectangle, covering the vast majority of typhoons. To ensure clear typhoon characteristics in the positive samples, the maximum offset is 20%, resulting in a final sample size of 160,000, with a positive-to-negative sample ratio of 1:1.

[0056] Motion model sample construction based on deep neural fusion network. This embodiment uses the time information of the optimal typhoon path set in the Northwest Pacific region provided by the China Meteorological Administration Tropical Cyclone Center as a reference to create an atmospheric reanalysis dataset of the entire sea area during typhoon formation. To capture air motion at different levels of the troposphere to better characterize typhoon features, this embodiment uses low clouds, middle clouds, and high clouds as representatives, and extracts u10 (lateral component of wind speed at 10 meters), v10 (vertical component of wind speed at 10 meters), and geopotential at three atmospheric pressure levels (700 hPa, 500 hPa, and 225 hPa) as sample features. The final constructed dataset contains a total of 553 events, including cyclones and typhoons.

[0057] 2. A Typhoon Trajectory Prediction Framework Based on Particle Filtering

[0058] A typhoon trajectory prediction framework based on particle filtering, such as Figure 1 As shown.

[0059] Bayesian theory explains that the state estimation problem involves estimating a state based on a series of previously available data y. 1:n (Posterior knowledge) Recursively calculate the current state x n The credibility of the probability formula p(x) is the probability of the probability. n |y 1:n This framework is based on Bayesian architecture, and its core is observation and updating.

[0060] The observation component, based on abundant samples, enables the neural network to learn typhoon characteristics. These characteristics are then used to determine whether each sub-region (5°*5°, the same size as the positive and negative samples) in the entire sea area possesses typhoon features. Finally, a probability map of the global sea area is obtained by constructing a Gaussian distribution. Areas with higher probabilities indicate strong typhoon characteristics, while those with lower probabilities indicate normal sea conditions. This probability map serves as a crucial basis for subsequent particle weight updates. Based on the probability map, the prediction particle set is... The weights are updated. During the update, the position of each particle can be calculated based on the probability map. Corresponding probability value Calculate the weights of all particles. After normalizing the weights of all particles, we obtain new weights for each particle. Thus, the set of particles at the current moment is obtained. For particle sets Perform a resampling operation, such as using a roulette wheel resampling strategy, to eliminate the particle depletion problem and obtain the final posterior distribution set of particles at the current time step. For cases with a single typhoon eye, the posterior distribution of the particle set... The weighted centroid represents the current position of the typhoon eye. For multiple typhoon eyes, it's necessary to determine the number and then apply the posterior distribution of the particle set. By performing clustering, the weighted centroid of each cluster is the eye of the storm.

[0061] Using a Bayesian architecture, a posterior estimate of the typhoon eye's location is obtained through iterative updates in two steps: prediction and update. The specific steps include:

[0062] S11. Set the particles in a two-dimensional Gaussian distribution based on the known current typhoon location;

[0063] S12. Predict the displacement at the next moment using a multimodal fusion network, and update the particle position based on the displacement.

[0064] S13. Update the particle weights based on the probability map, calculate the weighted average of the particles to obtain the posterior estimate of the typhoon's position at the next moment, and use this as the input for the next prediction. Perform particle resampling every 5 times (adjustable).

[0065] 3. Motion model based on deep neural fusion network

[0066] Motion models based on deep neural fusion networks, such as Figure 2As shown in the figure, this invention designs a spatial diversity parameterization method. In the horizontal direction, a 5°*5° latitude and longitude grid is selected, enabling the neural network to focus more on the characteristics of the typhoon center and extract more effective information from the air mass motion around the typhoon. In the vertical direction, wind field and geopotential features at three representative pressure levels—225 hPa, 500 hPa, and 700 hPa—are selected to capture air motion at three different levels of the troposphere: low, middle, and high clouds. To address the complexity of wind field changes, this invention designs a multimodal fusion feature extraction method. Based on historical trajectories, the continuous motion characteristics of the wind field are extracted; based on meteorological data, the abrupt motion characteristics of the typhoon are extracted. A multimodal fusion network is used to effectively fuse the two features, improving the effectiveness of wind field feature extraction.

[0067] The deep neural fusion network consists of three sub-networks, each extracting wind field, geopotential, and historical trajectory data independently. The wind field and geopotential models both utilize convolutional neural networks to extract features from the input typhoon data. Their structures are similar, including three convolutional layers and three fully connected layers. Each convolutional layer is followed by a normalization and ReLU activation function. Both models ultimately input a vector of size 2, representing the typhoon's displacement at the current moment. The historical trajectory network uses a fully connected neural network to achieve the same function. Before training the fusion network, the three independent networks are pre-trained. Then, the three trained networks are concatenated and fused in the final layer and further trained for parameter fine-tuning. The fusion network is the core of the entire model; it combines the outputs of the three sub-networks into a single vector, which is then output through two fully connected layers to produce the final result.

[0068] The feature extraction network based on multimodal fusion in this embodiment of the invention uses wind speed, geopotential, and historical trajectory for independent prediction. Each independent branch uses a convolutional neural network, and its parameters are trained separately. Finally, the three branches are integrated into a fusion network, which is then adjusted and optimized. The steps are as follows:

[0069] S21. A wind field feature prediction network is established using the wind speed u component, wind speed v component and convolutional neural network. A geopotential feature prediction network and a historical trajectory feature prediction network are established using geopotential and historical trajectory and convolutional neural network. The prediction content of the three branch networks is the displacement of the typhoon at the next moment. The network parameters are adjusted independently.

[0070] S22. The wind field feature prediction network, geopotential feature prediction network and trajectory feature prediction network are integrated into a typhoon trajectory prediction network, and the network parameters are adjusted according to the final prediction results.

[0071] 4. Observation model based on convolutional neural network

[0072] Observation models based on convolutional neural networks, such as Figure 3As shown. The main purpose of the observation model based on convolutional neural networks is to output a probability map, which facilitates the updating of particle weights. The main idea is to pre-train a convolution-based binary classifier to determine whether an input image patch contains a typhoon, thus enabling the trained convolutional neural network to capture the basic features of the typhoon. The trained feature matrix is ​​then used to scan the global sea area sequentially in a sliding window manner, thereby obtaining a binary map of the global sea area. Positive sample areas are identified as typhoon occurrence areas. Because typhoon movement is continuous, positive samples in several time frames are often close together. This embodiment utilizes the continuity of the typhoon to eliminate misclassified areas in the network and proves the effectiveness of the method. Based on the binary map, a Gaussian probability map is constructed in the positive sample area, making the probability higher for areas near the typhoon eye and extending outwards. This operation can further improve prediction accuracy.

[0073] In the network architecture, four convolutional neural network layers (Conv2d) were first used to extract features from the input typhoon image. The first layer received an input tensor with nine channels, representing geopotential and wind speed features at three different pressure levels. After each convolutional layer, a ReLU activation function was used for non-linear transformation, followed by downsampling using max pooling (MaxPool2d). A softmax activation function was then defined to transform the neural network output into probability estimates for each class. Two convolutional layers transformed a 128-channel input into 32-channel and a 2-channel output, respectively. In addition to the first four convolutional layers, the fifth convolutional layer used a 1x1 kernel size to reduce the number of channels in the feature map, and the sixth convolutional layer used the same kernel size to transform the feature map into output probabilities for each class.

[0074] The forward propagation function implements the forward propagation computation of the convolutional layers. The input tensor is transformed into a feature map in each convolutional layer and flattened into a one-dimensional vector after the last layer. Finally, this vector is fed into the sigmoid activation function to compress the output value to the range of 0 to 1, representing the probability estimate for each class.

[0075] like Figure 4 As shown, the observation network uses a convolutional neural network to learn typhoon characteristics from a large number of meteorological samples, thereby determining the meteorological characteristics of the global sea area. A global typhoon occurrence probability map is then obtained through threshold filtering and the addition of a Gaussian distribution. The specific steps are as follows:

[0076] S41. Collect meteorological samples based on the three-dimensional characteristics of typhoons and meteorological principles to form a dataset of meteorological elements that affect typhoon trajectories;

[0077] S42. Crops typhoon area samples from meteorological element datasets;

[0078] S43. Determine whether the typhoon coverage rate of the area reaches 80%. If not, then the sample is determined as a negative typhoon sample; if so, then the sample is determined as a positive typhoon sample.

[0079] S43. Generate a typhoon positive and negative sample dataset from the samples;

[0080] S44. Obtain typhoon feature maps based on convolutional neural networks;

[0081] S45. The classifier scans the entire sea area at the current moment to obtain a binary map;

[0082] S46. Construct a Gaussian distribution;

[0083] S47. Obtain the global typhoon occurrence probability map based on threshold filtering and Gaussian distribution.

[0084] In the above embodiments, the typhoon trajectory prediction framework based on deep learning and particle filtering, grounded in a Bayesian architecture, is divided into two parts: typhoon observation and typhoon update. When the historical typhoon trajectory is unknown, the typhoon observation part uses deep learning and considers typhoon characteristics to obtain a typhoon probability map. This probability map corrects and constrains the predicted typhoon range within the overall framework, improving prediction accuracy. In the typhoon update part, a multimodal neural network is used in conjunction with historical environmental factors to predict the typhoon location. These environmental factors do not contain any location information. This prediction process also includes particle prediction. The newly obtained particle weights are updated using the probability map obtained from typhoon observation. A weighted average of the particles is used to obtain a posterior estimate of the typhoon's location at the next moment, which is then used as the input for the next prediction. This observation and update cycle is continuously repeated, solving the problem of typhoon path prediction when the actual path is not available in real time.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A typhoon trajectory prediction method based on deep learning and particle filtering, characterized in that, include: A typhoon trajectory prediction model based on deep learning and particle filtering is constructed. The typhoon trajectory prediction model is based on a Bayesian architecture and is divided into typhoon observation and typhoon update. The typhoon observation section uses deep learning and considers typhoon characteristics to obtain a typhoon probability map, which is then used to correct and constrain the predicted typhoon range in the typhoon trajectory prediction model. The typhoon update section uses a multimodal neural network combined with historical environmental factors to predict the typhoon location. The prediction process also includes particle prediction. The newly obtained particle weights are updated using the probability map obtained from typhoon observations. The weighted average of the particles is used to obtain the posterior estimate of the typhoon location at the next moment, and this is used as the input for the next prediction. The typhoon observation and typhoon update are continuously cycled to obtain a posterior estimate of the typhoon eye location and thus the typhoon trajectory prediction result.

2. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 1, characterized in that, Typhoon probability maps are obtained by leveraging deep learning and considering typhoon characteristics, including: An observation model based on a convolutional neural network is constructed. The model comprises four Conv2d layers to extract features from input typhoon images. The first layer receives an input tensor with nine channels, representing geopotential and wind speed features at three different pressure levels. After each convolutional layer, a ReLU activation function is used for non-linear transformation, followed by downsampling using MaxPool2d. A softmax activation function is then defined to transform the neural network output into a probability estimate for each class. In addition to the first four convolutional layers, the fifth convolutional layer uses a 1x1 multiplication table. The kernel size is used to reduce the number of channels in the feature map. The sixth convolutional layer uses the same kernel size to transform the feature map into output probabilities for each class. In the last two convolutional layers, the 128-channel input is transformed into 32-channel and 2-channel outputs, respectively. The forward propagation function performs the forward propagation computation of the convolutional layers. The input tensor is transformed into a feature map in each convolutional layer and flattened into a one-dimensional vector after the last layer. Finally, this vector is input into the sigmoid activation function to compress the output values ​​to the range of 0 to 1 to represent the probability estimate for each class. The probability of a typhoon occurring at the global node at the current moment is obtained using the observation model.

3. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 2, characterized in that, The probability of a typhoon occurring at the current global node is obtained using the observation model, including: Meteorological samples were collected based on the three-dimensional characteristics of typhoons and meteorological principles to form a typhoon trajectory dataset. Regional samples were cropped from the typhoon trajectory dataset; Determine whether the area is covered by the typhoon's range. If not, the sample is identified as a negative typhoon sample; if so, the sample is identified as a positive typhoon sample. The samples will be used to create a dataset of positive and negative typhoon samples. The positive and negative sample datasets of the typhoon are input into a convolutional neural network, and the typhoon feature map is obtained based on the convolutional neural network. The classifier scans the entire sea area at the current moment to obtain a binary map; Construct a Gaussian distribution; A global typhoon occurrence probability map was obtained based on threshold filtering and Gaussian distribution.

4. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 1, characterized in that, Typhoon location prediction using multimodal neural networks combined with historical environmental factors includes: A motion model based on a deep neural fusion network is constructed. The deep neural fusion network includes a wind field subnetwork, a potential subnetwork, and a historical trajectory subnetwork, which extract the wind field, potential, and historical trajectory separately. The deep neural fusion network also includes a fusion subnetwork. The fusion subnetwork is used to merge the outputs of the wind field subnetwork, the potential subnetwork, and the historical trajectory subnetwork into a single vector, and outputs the final result through two fully connected layers. The motion model is used to predict the displacement of the typhoon at the next moment.

5. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 4, characterized in that, The wind field subnetwork extracts features from the input typhoon data. The wind field subnetwork includes 3 convolutional layers and 3 fully connected layers. Each convolutional layer is followed by a normalization and ReLU activation function, and outputs a vector of size 2, which is the displacement of the typhoon at the next moment. The potential subnetwork extracts features from the input typhoon data. The potential subnetwork includes 3 convolutional layers and 3 fully connected layers. Each convolutional layer is followed by a normalization and ReLU activation function, and outputs a vector of size 2, which is the displacement of the typhoon at the next moment. The historical trajectory subnetwork includes a fully connected neural network, and its output is the displacement of the typhoon at the next moment.

6. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 5, characterized in that, Training the deep neural fusion network includes: Three independent networks were pre-trained; The three trained networks are spliced ​​and merged in the last layer and then trained again to fine-tune the parameters. Train the fusion subnetwork.

7. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 1, characterized in that, The particle prediction includes: Particles are set using a two-dimensional Gaussian distribution based on the known current location of the typhoon. The displacement at the next moment is predicted by a multimodal fusion network, and the particle position is updated according to the displacement. The particle weights are updated based on the probability map, and the weighted average of the particles is used to obtain the posterior estimate of the typhoon's position at the next moment. This estimate is then used as the input for the next prediction, and the particles are resampled every preset number of times.

8. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 7, characterized in that, Update particle weights based on the probability graph, including: The probability value corresponding to the position of each particle is calculated based on the probability graph. The weight of all particles is calculated, and the weight of all particles is normalized to obtain the new weight of each particle, thereby obtaining the particle set at the current time.

9. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 7, characterized in that, The particle resampling includes: A resampling strategy using roulette wheel is adopted to eliminate the particle depletion problem and obtain the final posterior distribution set of particles at the current time.

10. The typhoon trajectory prediction method based on deep learning and particle filtering according to claim 7, characterized in that, Obtain a posterior estimate of the location of the typhoon eye, including: For a single typhoon eye, the weighted centroid of the posterior distribution particle set is the current position of the typhoon eye. For multiple typhoon eyes, the number of typhoon eyes is determined, and then the posterior distribution particle set is clustered. The weighted centroid of each cluster is the position of the eye.

Images