A typhoon multi-task estimation method based on physical constraints and correlation modeling

Through methods based on physical constraints and correlation modeling, features are extracted from multimodal meteorological data and the physical connection between typhoon attributes are solved, and the problem of difficulty in estimating the typhoon intensity and wind circle size in the existing technology is solved, achieving efficient Wind speed and wind circle estimates are superior to traditional and other deep learning methods.

CN118332278BActive Publication Date: 2024-11-19ZHEJIANG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410426784.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-10
Publication Date
2024-11-19
Estimated Expiration
2044-04-10

AI Technical Summary

Technical Problem

Existing deep learning methods are difficult to simultaneously effectively estimate the intensity and wind circle size of typhoons, and ignore the dynamics and physical connections of typhoons.

Method used

Using methods based on physical constraints and association modeling, features are extracted from multimodal meteorological data through central extension pooling algorithms, multi-head cross-attention mechanisms, multi-domain recursive convolution modules and general networks, and different attributes are simulated. The physical connection between the wind speed and wind circle are achieved simultaneously.

Benefits of technology

Efficient simultaneous estimation of typhoon intensity and wind ring size is achieved, improving estimation accuracy and better performance indicators than traditional and other deep learning methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118332278B_ABST
    Figure CN118332278B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of typhoon real-time estimation, and discloses a typhoon multi-task estimation method based on physical constraints and association modeling. The center feature is extracted from the cloud top data by using the center expansion pooling algorithm, and the stride feature obtained from the stride sequence is combined to obtain multiple characteristic features through the joint action of the multi-head cross attention mechanism and the category encoder; the multiple characteristic features are respectively input into the multi-domain recursive convolution module to obtain the corresponding conversion features; the time series multi-channel satellite data, the longitude and latitude of the typhoon center and the existence duration are sent to the general network, the position encoder and the time encoder to obtain the shared features; the characteristic features, the conversion features and the shared features of the task are fused, and the wind speed estimation results and the wind circle estimation results are obtained through the estimation module. The present invention solves the problem that the existing technical solutions cannot effectively estimate the typhoon intensity and the size of the wind circle at the same time, and the problem of lack of physical guidance through association modeling and physical constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of typhoon real-time estimation, and in particular relates to a typhoon multi-task estimation method based on physical constraints and association modeling. Background Art

[0002] In recent years, the frequency and intensity of typhoons have been increasing. As a complex and powerful weather system, typhoons, also known as tropical cyclones, are mainly generated on tropical and subtropical seas and are affected by a variety of factors. Typhoon estimation uses meteorological data including satellite data to estimate the current properties of typhoons in real time. Since strong typhoons can cause other natural disasters such as floods and mudslides, in order to prevent such weather disasters, it is very necessary to estimate the intensity and size of the typhoon in real time.

[0003] In general, traditional methods focus on estimating a single attribute, such as intensity or wind circle radius. Therefore, with the continuous development of artificial intelligence technology, many studies have proposed various deep learning methods to simultaneously estimate multiple attributes of typhoons, which will help mitigate the disasters caused by strong typhoons. Although most of the current deep learning-based methods have achieved simultaneous estimation of typhoon wind speed and wind circle, they have not fully considered the negative migration caused by task conflicts. In addition, existing deep learning methods often rely only on data-driven, ignoring the importance of typhoon-related dynamics and physical connections in deep learning. Therefore, the present invention proposes a typhoon multi-task estimation method based on physical constraints and association modeling. Summary of the invention

[0004] The purpose of the present invention is to provide a typhoon multi-task estimation method based on physical constraints and association modeling, which solves the problem that the existing technical solutions cannot effectively estimate the typhoon intensity (wind speed) and wind circle size at the same time, as well as the problem of lack of physical guidance through association modeling and physical constraints.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A typhoon multi-task estimation method based on physical constraints and associated modeling, the typhoon multi-task estimation method based on physical constraints and associated modeling comprising:

[0007] The center feature is extracted from the cloud top data using the center expansion pooling algorithm, and the stride feature obtained from the stride sequence is combined with the multi-head cross attention mechanism and the category encoder to obtain the characteristic features of multiple tasks, including wind speed features and wind circle features;

[0008] The wind speed features and wind circle features are respectively input into the multi-domain recursive convolution module. The multi-domain recursive convolution module simulates the physical connection between different task attributes by converting the characteristic features of one task into the characteristic features of another task, and obtains two conversion features corresponding to wind speed and wind circle estimation;

[0009] The time series multi-channel satellite data, the longitude and latitude of the typhoon center and its existence duration are respectively sent to the corresponding general network, position encoder and time encoder for feature extraction and learning to obtain shared features;

[0010] The characteristic features of the task and their corresponding conversion features are input into the fully connected layer. After concatenating the output of the fully connected layer and the shared features, they are input into the estimation module of the corresponding task to obtain the wind speed estimation results and wind circle estimation results.

[0011] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution, but are merely further supplements or preferences. Under the premise that there are no technical or logical contradictions, each optional method can be combined with the above-mentioned overall solution separately, and multiple optional methods can also be combined.

[0012] Preferably, the extracting center features from cloud top data using a center expansion pooling algorithm includes:

[0013] The center expansion pooling algorithm consists of a loop algorithm, a maximum pooling layer, an average pooling layer and a convolutional layer. The loop algorithm first defines two empty lists as a flat pooling feature list and a maximum pooling feature list, takes the dimension of the cloud top data as N, and the loop variable i is 1. When the loop starts, it is first determined whether the loop variable i is less than or equal to N / 2. If not, the average pooling features in the flat pooling feature list and the maximum pooling features in the maximum pooling feature list are input into the convolutional layer, and the convolutional layer outputs the center feature; if so, the starting index is calculated to be N / 2-i, the ending index is N / 2+i, and the slicing range is [starting index: ending index+1]. The cloud top data is sliced ​​according to the slicing range to obtain the pooling feature, the pooling feature is input into the average pooling layer to obtain the average pooling feature under the current cycle and put it into the average pooling feature list, the pooling feature is input into the maximum pooling layer to obtain the maximum pooling feature under the current cycle and put it into the maximum pooling feature list, the loop variable i is increased by 1, and the loop is restarted.

[0014] Preferably, if the dimension of the cloud top data is N, the stride sequence is defined as [1~N / 2].

[0015] Preferably, the category encoder is composed of two identical multi-layer perceptrons, the intensity level of the typhoon at the previous moment is input into the category encoder, the intensity level information of the typhoon at the previous moment is extracted and two weight features are output, the two weight features are respectively called wind speed weight feature and wind circle weight feature.

[0016] Preferably, the multi-head cross attention mechanism and the category encoder are used together to obtain characteristic features of multiple tasks, including:

[0017] The center feature and the stride feature are respectively passed through a fully connected layer to obtain corresponding query features, key features, and value features, and the key features corresponding to the center feature and the query features and value features corresponding to the stride feature are input into a first multi-head cross attention mechanism, and the output of the first multi-head cross attention mechanism is multiplied by the wind speed weight feature tensor after convolution to extract the wind speed feature;

[0018] The key feature corresponding to the stride feature and the query feature and value feature corresponding to the center feature are input into the second multi-head cross attention mechanism. The output of the second multi-head cross attention mechanism is convolved and multiplied with the wind circle weight feature tensor to extract the wind circle feature.

[0019] Preferably, the multi-domain recursive convolution module consists of 6 2D convolutions, 2 1D convolutions and forward and inverse Fourier transforms;

[0020] The characteristic features of the task are input into the multi-domain recursive convolution module and then processed by the forward Fourier transform to obtain the real and imaginary features. The real features are input into three 2D convolutions at the same time. The outputs of the three 2D convolutions are tensor multiplied, and the multiplication results are transposed and input into one 1D convolution. At the same time, the imaginary features are input into another three 2D convolutions at the same time. The outputs of these three 2D convolutions are tensor multiplied, and the multiplication results are transposed and input into another 1D convolution. The outputs of the two 1D convolutions are stacked and soft-shrunk, and then the converted features are output by the inverse Fourier transform.

[0021] Preferably, the time-series multi-channel satellite data, the longitude and latitude of the typhoon center and its existence duration are respectively sent to the corresponding general network, position encoder and time encoder for feature extraction and learning to obtain shared features, including:

[0022] The position encoder is composed of four fully connected layers, and the longitude and latitude of the typhoon center are input into the position encoder to extract the position features of the typhoon;

[0023] The position features of the typhoon are spliced ​​with the time-series multi-channel satellite data tensor as the input of the universal network, and the universal network is composed of four feature extraction layers and two channel residual connections, each feature extraction layer includes a 2D convolution and a maximum pooling layer, and the output of the second feature extraction layer is connected to the output of the third feature extraction layer through a channel residual connection, and then input to the fourth feature extraction layer; the input of the fourth feature extraction layer is connected to the output of the fourth feature extraction layer through a channel residual connection, and serves as the output feature of the universal network;

[0024] The time encoder is composed of two fully connected layers. The existence duration of the typhoon is input into the time encoder to extract the development duration characteristics of the typhoon.

[0025] The output features of the general network are concatenated with the typhoon development duration features as tensors and used as shared features.

[0026] Preferably, the estimation module consists of four fully connected layers followed by a Relu activation function, and is used to output two regression values ​​of the typhoon's wind speed and wind circle.

[0027] The present invention provides a typhoon multi-task estimation method based on physical constraints and association modeling, and the data required for estimation include one-dimensional data and two-dimensional satellite data related to the typhoon. The one-dimensional data are the central longitude and latitude of the typhoon, the existence duration of the typhoon, and the intensity level at the previous moment; the two-dimensional satellite data include the bright temperature data and cloud top data of multiple infrared channels. Therefore, the present invention can extract features from multi-modal meteorological data and model and physically constrain the associations between different attributes to be estimated, and then fuse the features, and finally give the estimation results of the typhoon wind speed and wind circle radius at the same time. Compared with the prior art, the method of the present invention can extract the characteristics of the typhoon itself and the association characteristics of different attributes of the typhoon from multi-modal meteorological data, and can use these characteristics to establish physical constraints at the association and feature levels for the intensity and wind circle, and finally achieve simultaneous and efficient estimation of the intensity and wind circle. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flowchart of a typhoon multi-task estimation method based on physical constraints and association modeling of the present invention;

[0029] Figure 2 It is a structural schematic diagram of the estimation network model of the present invention;

[0030] Figure 3 It is a flow chart of the central expansion pooling algorithm of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0033] like Figure 1 As shown, this embodiment provides a typhoon multi-task estimation method based on physical constraints and association modeling, comprising the following steps:

[0034] Step S1, association modeling: The center feature is extracted from the cloud top data using the center expansion pooling algorithm, and the stride feature obtained from the stride sequence is combined to obtain the characteristic features of multiple tasks through the joint action of the multi-head cross attention mechanism and the category encoder. The characteristic features of the multiple tasks include wind speed features and wind circle features.

[0035] The typhoon multi-task estimation method based on physical constraints and association modeling in this embodiment is to construct an estimation network model through training, and then use the trained estimation network model to process real-time typhoon data to estimate the intensity and wind circle at the current moment.

[0036] The estimation network model constructed in this embodiment is as follows Figure 2 As shown in the figure, it includes an association modeling structure, a physical constraint structure, a task commonality learning structure and two estimation modules. The association modeling structure consists of three parts: center expansion pooling algorithm, category encoder and multi-head cross attention mechanism; the physical constraint structure consists of two parts: multi-domain recursive convolution module and two fully connected layers; the task commonality learning structure consists of three parts: position encoder, time encoder and general network.

[0037] Specifically, Figure 3 The center expansion pooling algorithm shown is composed of 1 loop algorithm, 1 maximum pooling layer, 1 average pooling layer and 1 convolution layer. The cloud top data is input into the center expansion pooling algorithm to extract the central features of the typhoon. Before the algorithm starts, the embodiment predefines two empty lists for storing feature results of different pooling ranges. Within the number of cycles, the cloud top data starts from the center position and expands the pooling area outward with a step increase of 1, and the two pooling results are stored in the list in turn, and finally sent to a convolution layer to obtain the central features. The center expansion pooling of this embodiment is implemented based on a window that expands outward from the center, which is conducive to extracting more representative central features.

[0038] The specific loop algorithm first defines two empty lists as the average pooling feature list and the maximum pooling feature list, takes the dimension of the cloud top data as N, and the loop variable i as 1. When starting the loop, first determine whether the loop variable i is less than or equal to N / 2. If not, the average pooling features in the average pooling feature list and the maximum pooling features in the maximum pooling feature list are input into the convolution layer, and the convolution layer outputs the center features; if so, the starting index is calculated to be N / 2-i, the ending index is N / 2+i, and the slicing range is [starting index: ending index+1]. According to the slicing range, the cloud top data is sliced ​​to obtain the pooling features, and the pooling features are input into the average pooling layer to obtain the average pooling features under the current cycle and put them into the average pooling feature list. The pooling features are input into the maximum pooling layer to obtain the maximum pooling features under the current cycle and put them into the maximum pooling feature list. The loop variable i is increased by 1, and the loop is restarted.

[0039] In a specific embodiment, the number of loops of the loop algorithm is 78, the convolution kernel size, convolution step size and padding width of the two pooling layers are 2, 2, and 0, and the convolution kernel size of the convolution layer is 1.

[0040] The category encoder is composed of two identical multilayer perceptrons. The intensity level (wind speed) data of the typhoon at the previous moment is input into the category encoder to extract the intensity level characteristics of the typhoon. In a specific embodiment, the category extraction module is composed of two identical multilayer perceptrons, each of which includes two fully connected layers, which are used to extract the intensity level information of the typhoon at the previous moment and output two weight features C_1 and C_2, respectively called wind speed weight features and wind circle weight features, which are embedded in the development network as auxiliary features, thereby obtaining development features suitable for two estimation tasks. The two multilayer perceptrons have the same structure, and the number of neurons in the two internal fully connected layers is set to 2 and 4 respectively, and the weights and biases of the two fully connected layers are defined as floating-point variables. It should be noted that all the above-mentioned fully connected layers are followed by the Tanh activation function.

[0041] The multi-head cross attention mechanism consists of 2 fully connected layers, tensor multiplication and 1 activation function. After the center feature and stride feature are input into the cross attention mechanism and a 2D convolution to obtain the corresponding output features, they are multiplied with the wind speed weight feature and wind circle weight feature output by the category encoder respectively to extract the wind speed feature and wind circle feature.

[0042] In a specific embodiment, the weight of the entire cross-attention mechanism is W_CA, which is defined as a floating-point variable without bias. Among them, in order to correspond to the stride feature and the center feature, according to the size of the side length (dimension) N of the cloud top data, the stride sequence is predefined as [1~N / 2]. The stride feature (N / 2, 4) is obtained by expanding the stride sequence and converting the second dimension to 4. It is passed through two fully connected layers with the center feature obtained by the center expansion pooling algorithm to obtain the corresponding query feature, key feature and value feature. The corresponding number of neurons in the fully connected layer is set to 4 and 8 respectively, and its weights and bias are defined as floating-point variables. The input 2D convolution (number of input and output channels, convolution kernel size, convolution step and padding width) is set to (156, 78, 1, 1, 0) before multiplying with the weight feature. It is worth noting that in Figure 2 In the figure, key features 1 and 2 obtained from the stride feature and the center feature are used interchangeably in the two attention mechanisms. Key feature-oriented is a key idea in association modeling. For example, by performing attention calculation on key feature 1 obtained from the stride feature and the query feature and value feature obtained from the center feature, some features representing the wind circle can be fused into the center feature representing the wind speed.

[0043] It should be noted that the wind circle estimated by the present invention is the size of the wind circle corresponding to the maximum wind speed. There is a corresponding relationship between the wind circle and the intensity. The input intensity level can make the network model know what stage of development the typhoon is currently in, etc. The output of the sensor is obtained by mapping and transforming the intensity level, and the result of the sensor output is related to the intensity and the wind circle. At the same time, the stride and center features represent the potential characteristics and numerical range of the wind speed and the wind circle, respectively. Therefore, the mapping of their features is respectively transformed by the calculation of the cross-attention mechanism, and the relationship between them can be modeled, which helps to better estimate the wind speed and the size of the wind circle.

[0044] Step S2, physical constraints: The wind speed features and wind circle features are respectively input into the Multi-Domain Recursive Convolutional Module, which simulates the physical connection between different task attributes by converting the characteristic features of one task into the characteristic features of another task, and obtains two conversion features corresponding to wind speed and wind circle estimation. Finally, the wind speed feature and conversion feature 1, and the wind circle feature and conversion feature 2 are fused respectively through two fully connected layers.

[0045] Specifically, Figure 2As shown in the lower left corner, the multi-domain recursive convolution module consists of 6 2D convolutions, 2 1D convolutions, and forward and inverse Fourier transforms. The characteristic features of the task are input into the multi-domain recursive convolution module and processed by forward Fourier transform to obtain real and imaginary features. The real features are simultaneously input into 3 2D convolutions. The outputs of the 3 2D convolutions are tensor multiplied, and the multiplication results are transposed and input into 1 1D convolution. At the same time, the imaginary features are simultaneously input into another 3 2D convolutions. The outputs of these 3 2D convolutions are tensor multiplied, and the multiplication results are transposed and input into another 1D convolution. The outputs of the two 1D convolutions are stacked and soft-contracted, and then the conversion features are output by inverse Fourier transform.

[0046] The weight of the whole multi-domain recursive convolution module is W_DC, which is defined as a floating-point variable without bias. The (input and output channel numbers, convolution kernel size, convolution step size and padding width) of the 6 2D convolution layers are all set to (64, 64, 1, 1, 0); the (input and output channel numbers, convolution kernel size) of the 2 1D convolution layers are all set to (4, 4, 1); the dimension of the forward and inverse Fourier transform calculations is (1, 2).

[0047] After the wind speed features and wind circle features are extracted by the association modeling structure, they are fused with the conversion features extracted by the multi-domain recursive convolution module to form wind speed characteristic features and wind circle characteristic features for estimating the specific task of typhoons. The fusion method is realized through the feature transformation of two fully connected layers. Specifically, the number of neurons in the two fully connected layers used for fusion is set to 256, and their weights and biases are defined as floating-point variables.

[0048] It should be noted that there is a physical connection between the wind speed and wind circle estimated by the present invention through polynomial fitting. A multi-domain recursive convolution module is used to convert the features of different tasks at the corresponding feature level, thereby realizing physical constraints between different attributes and further improving the accuracy of estimating wind speed and wind circle.

[0049] Step S3, task commonality learning: the time series multi-channel satellite data, the longitude and latitude of the typhoon center and its existence duration are sent to the corresponding general network, position encoder and time encoder for feature extraction and learning to obtain shared features.

[0050] Specifically, the position encoder is composed of four fully connected layers, and the longitude and latitude data of the typhoon center are input into the position encoder to extract the position features of the typhoon. In a specific embodiment, the number of neurons in the four fully connected layers is set to 5, 25, 58 and 156 respectively, the four weights are W_P_1 to W_P_4, defined as floating-point variables, and the biases are b_p_1 to b_p_4, defined as floating-point variables.

[0051] The time encoder is composed of two fully connected layers. The typhoon's existence duration data is input into the time encoder to extract the typhoon's development duration characteristics. In a specific embodiment, the number of neurons in the two fully connected layers is set to 2 and 4 respectively, the two weights are W_T_1 and W_T_2, defined as floating-point variables, and the biases are b_t_1 and b_t_2, defined as floating-point variables.

[0052] It should be noted that all fully connected layers in the above encoder are followed by a Tanh activation function.

[0053] In this embodiment, the location characteristics of the typhoon are spliced ​​with the time-series multi-channel satellite data tensor as the input of the general network. In a specific embodiment, Figure 2 As shown in the lower right corner, the general network module is mainly composed of four feature extraction layers and two channel residual connections. Each feature extraction layer includes a 2D convolution and a maximum pooling layer. All convolution layers in the general network module are followed by a Relu activation function. The convolution layer is used to extract features from two-dimensional satellite data, and the pooling layer is used for feature downsampling. The output of the second feature extraction layer is connected to the output of the third feature extraction layer through a channel residual connection, and then input to the fourth feature extraction layer; the input of the fourth feature extraction layer is connected to the output of the fourth feature extraction layer through a channel residual connection as the output feature of the general network module.

[0054] Finally, the output features of the general network and the typhoon development duration features are concatenated as tensors and used as shared features.

[0055] The weight of the entire general network module is W_SN, which is defined as a floating-point variable without bias. The (number of input and output channels, convolution kernel size, convolution step and padding width) of the four 2D convolutional layers are set to (9, 16, 7, 3, 0), (16, 32, 5, 1, 0), (32, 64, 3, 1, 1) and (96, 192, 3, 1, 1) respectively. The convolution kernel size, convolution step and padding width of the first three maximum pooling layers are all 3, 2, 0; the convolution kernel size, convolution step and padding width of the last pooling layer are set to 2, 2, 0.

[0056] In this embodiment, before the time series multi-channel satellite data is sent to the general network for feature extraction, the position features generated by the position extraction module will be early fused with it, so that the general network pays more attention to the key position during the feature extraction process. After the general network finishes extracting features, the development duration features extracted by the time extraction model will be late fused with its output features to form shared features for multi-task estimation. Both fusion methods for different periods are achieved through splicing operations.

[0057] Step S4, typhoon estimation: The task characteristic features (wind speed characteristic features and wind circle characteristic features) are spliced ​​with the shared features respectively, and finally they are input into the estimation module of the corresponding task respectively to obtain two estimation results of wind speed and wind circle.

[0058] Specifically, the output features of the task commonality learning structure and the physical constraint structure are used to fuse and feed into the corresponding estimation modules to simultaneously estimate the typhoon's wind speed and wind circle radius. This method obtains simultaneous estimation of typhoon intensity and size under the guidance of association modeling and physical constraints, which is extremely important for disaster prevention applications of typhoons and the natural disasters they cause.

[0059] The estimation module consists of four fully connected layers followed by a Relu activation function, which is used to output two regression values ​​of the typhoon's wind speed and wind circle. The number of neurons in the four fully connected layers is set to 1024, 512, 128 and 1 respectively. The four weights of the two decoders are W_DI_1 to W_DI_4 and W_DS_1 to W_DS_4, defined as floating-point variables, and the biases are b_di_1 to b_di_4 and b_ds_1 to b_ds_4, defined as floating-point variables.

[0060] The present invention estimates the network model, and its training process is as follows:

[0061] First, preprocess the data set. Divide the data set into three parts: training set, validation set, and test set. The training set contains 160 typhoon data that occurred in the longitude and latitude range of 60S-60N, 80E-160W from 2015 to 2020. The validation set and test set contain 46 typhoon data that occurred in 2021-2022, which are randomly divided in a 1:1 ratio. Normalize the one-dimensional data and two-dimensional data in the data set to make the data more conducive to model training. Rotate the data in the training set by 90 degrees, 180 degrees, and 270 degrees clockwise and counterclockwise, respectively, for data enhancement.

[0062] While training:

[0063] The parameters of the convolutional and fully connected layers in the estimated network model were initialized with random weights, and a total of 600 complete trainings were performed; the learning rate of the model optimizer SGD was set to 0.0005 and the momentum was set to 0.4.

[0064] Part of the typhoon's one-dimensional data (typhoon latitude and longitude, typhoon's existence duration) and two-dimensional time series multi-channel satellite data are input into the task commonality learning structure; the stride sequence, two-dimensional cloud top data and the typhoon's intensity level at the previous moment are input into the association modeling structure.

[0065] The associated modeling structure, physical constraint structure and task commonality learning structure respectively extract features from the input multimodal data, and the features of the three modules are mixed and fused.

[0066] Based on the design idea of ​​soft sharing in multi-task learning, the shared representation and the characteristic representation suitable for intensity and wind circle estimation are fused and sent to two estimation modules respectively, and the estimated values ​​of the typhoon's wind speed and wind circle are output at the same time.

[0067] The error between the estimated typhoon and the true value of the typhoon is calculated to obtain two L1 losses, which are added 1:1, and the total error is fed back to the estimation network model. The parameters of the estimation network model are adjusted by minimizing the loss function through the SGD algorithm to optimize the estimation performance.

[0068] The estimation network model is trained until convergence, and the best performance estimation network model is obtained for typhoon estimation.

[0069] The present invention also provides experimental data under the data set of the estimated network model training, and compares the technical solution of the present invention with some traditional technical solutions (ADT, MTCSWA) and other deep learning methods (Intensity-CNN, Size-CNN, TC-MTLNet and DeepTCNet). The comparison performance indicators use the mean absolute error MAE and the root mean square error RMSE. The comparison results are shown in Table 1:

[0070] Table 1 Comparison results of performance indicators

[0071]

[0072] In Table 1, Intensity (kts) represents the wind speed estimation task, and Size (nmi) represents the wind circle estimation task. From the comparison results in Table 1, it can be seen that the method of the present invention can not only realize the estimation of wind speed and wind circle multiple tasks at the same time, but also the mean absolute error and root mean square error of the method of the present invention in each task are significantly smaller than other traditional technical solutions and other deep learning methods. It can be seen that the performance of the method of the present invention is better than that of traditional technical solutions and other deep learning methods.

[0073] ADT refers to the technical solution disclosed in the literature Kenneth Knapp and James Kossin. New global tropical cyclone data set from isccp b1 geostationary satellite observations. Journal of Applied 566 Remote Sensing, 1, 02 2007. MTCSWA refers to the technical solution disclosed in the literature Timothy L. Olander and Christopher S. Velden. The advanced dvorak technique: Continued development of an objective scheme to estimate tropical cyclone intensity using geostationary infrared satellite imagery. Weather and Forecasting, 22(2): 287–298, 2007. Intensity-CNN refers to the technical solution disclosed in the literature Chong Wang, Gang Zheng, Xiaofeng Li, Qing Xu, Bin Liu, and Jun Zhang. Tropical cyclone intensity estimation from geostationary satellite imagery using deep convolutional neural networks. IEEE Transactions on Geoscience and Forecasting, 22(2): 287–298, 2007. Remote Sensing, 60: 1–16, 2022. Size–CNN refers to the deep learning method disclosed in the literature Fan Meng, Pengfei Xie, Ying Li, Han-dan Sun, Danya Xu, and Tao Song. Tropical cyclone size estimation using deep convolutional neural network. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, pages 8472–8475, 2021.; TC-MTLNet refers to the deep learning method disclosed in the literature Wei Tian, ​​Xinxin Zhou, Xianhua Niu, Linhong Lai, Yonghong Zhang, and Kenny Lim Kam Sian. A lightweight multi-task learning model with adaptive loss balance for tropical cyclone intensity and size estimation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, PP: 1–15, 01 2022.; DeepTCNet refers to the deep learning method disclosed in the literature Jing-Yi Zhuo and Zhe-Min Tan. Physics-augmented deep learning to improve tropical cyclone intensity and size estimation from satellite imagery. Monthly Weather Review, 149, 04 2021. .

[0074] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A typhoon multi-task estimation method based on physical constraints and association modeling, characterized in that: The typhoon multi-task estimation method based on physical constraints and association modeling includes: The center feature is extracted from the cloud top data using the center expansion pooling algorithm, and the stride feature obtained from the stride sequence is combined with the multi-head cross attention mechanism and the category encoder to obtain the characteristic features of multiple tasks, including wind speed features and wind circle features; The wind speed features and wind circle features are respectively input into the multi-domain recursive convolution module. The multi-domain recursive convolution module simulates the physical connection between different task attributes by converting the characteristic features of one task into the characteristic features of another task, and obtains two conversion features corresponding to wind speed and wind circle estimation; The time series multi-channel satellite data, the longitude and latitude of the typhoon center and its existence duration are respectively sent to the corresponding general network, position encoder and time encoder for feature extraction and learning to obtain shared features; The characteristic features of the task and their corresponding conversion features are input into the fully connected layer. After the output of the fully connected layer and the shared features are concatenated, they are input into the estimation module of the corresponding task to obtain the wind speed estimation results and wind circle estimation results. The method of extracting the center feature from the cloud top data using the center expansion pooling algorithm includes: The center expansion pooling algorithm consists of a loop algorithm, a maximum pooling layer, an average pooling layer and a convolution layer. The loop algorithm first defines two empty lists as a flat pooling feature list and a maximum pooling feature list, takes the dimension of the cloud top data as N, and the loop variable i is 1. When starting the loop, first determine whether the loop variable i is less than or equal to N / 2. If not, the average pooling features in the flat pooling feature list and the maximum pooling features in the maximum pooling feature list are input into the convolution layer, and the convolution layer outputs the center feature; if so, the starting index is calculated to be N / 2-i, the ending index is N / 2+i, and the slicing range is [starting index: ending index+1]. The cloud top data is sliced ​​according to the slicing range to obtain the pooling feature, the pooling feature is input into the average pooling layer to obtain the average pooling feature under the current cycle and put it into the average pooling feature list, the pooling feature is input into the maximum pooling layer to obtain the maximum pooling feature under the current cycle and put it into the maximum pooling feature list, the loop variable i is increased by 1, and the loop is restarted; The category encoder is composed of two identical multi-layer perceptrons, the intensity level of the typhoon at the previous moment is input into the category encoder, the intensity level information of the typhoon at the previous moment is extracted and two weight features are output, the two weight features are respectively called wind speed weight feature and wind circle weight feature; Among them, the multi-head cross attention mechanism and the category encoder work together to obtain the characteristic features of multiple tasks, including: The center feature and the stride feature are respectively passed through a fully connected layer to obtain corresponding query features, key features, and value features, and the key features corresponding to the center feature and the query features and value features corresponding to the stride feature are input into a first multi-head cross attention mechanism, and the output of the first multi-head cross attention mechanism is multiplied by the wind speed weight feature tensor after convolution to extract the wind speed feature; Inputting the key feature corresponding to the stride feature and the query feature and value feature corresponding to the center feature into the second multi-head cross attention mechanism, and the output of the second multi-head cross attention mechanism is multiplied by the wind circle weight feature tensor after convolution to extract the wind circle feature; The time series multi-channel satellite data, the longitude and latitude of the typhoon center and its existence duration are respectively sent to the corresponding general network, position encoder and time encoder for feature extraction and learning to obtain shared features, including: The position encoder is composed of four fully connected layers, and the longitude and latitude of the typhoon center are input into the position encoder to extract the position features of the typhoon; The position features of the typhoon are spliced ​​with the time-series multi-channel satellite data tensor as the input of the universal network, and the universal network is composed of four feature extraction layers and two channel residual connections, each feature extraction layer includes a 2D convolution and a maximum pooling layer, and the output of the second feature extraction layer is connected to the output of the third feature extraction layer through a channel residual connection, and then input to the fourth feature extraction layer; the input of the fourth feature extraction layer is connected to the output of the fourth feature extraction layer through a channel residual connection, and serves as the output feature of the universal network; The time encoder is composed of two fully connected layers. The existence duration of the typhoon is input into the time encoder to extract the development duration characteristics of the typhoon. The output features of the general network are concatenated with the typhoon development duration features as tensors and used as shared features.

2. The typhoon multi-task estimation method based on physical constraints and association modeling according to claim 1 is characterized in that: If the dimension of the cloud top data is N, the stride sequence is defined as [1~ N / 2].

3. The typhoon multi-task estimation method based on physical constraints and association modeling according to claim 1 is characterized in that: The multi-domain recursive convolution module consists of 6 2D convolutions, 2 1D convolutions and forward and inverse Fourier transforms; The characteristic features of the task are input into the multi-domain recursive convolution module and then processed by the forward Fourier transform to obtain the real and imaginary features. The real features are input into three 2D convolutions at the same time. The outputs of the three 2D convolutions are tensor multiplied, and the multiplication results are transposed and input into one 1D convolution. At the same time, the imaginary features are input into another three 2D convolutions at the same time. The outputs of these three 2D convolutions are tensor multiplied, and the multiplication results are transposed and input into another 1D convolution. The outputs of the two 1D convolutions are stacked and soft-shrunk, and then the converted features are output by the inverse Fourier transform.

4. The typhoon multi-task estimation method based on physical constraints and association modeling according to claim 1 is characterized in that: The estimation module consists of four fully connected layers followed by a Relu activation function, and is used to output two regression values ​​of the typhoon's wind speed and wind circle.