A method and system for estimating the intensity of a tropical cyclone

Abstract

The application discloses a method and system for estimating the intensity of a tropical cyclone, and the method comprises the following steps: obtaining an original satellite image of the tropical cyclone; extracting physical information and a convection core from the original satellite image; generating a convection core image according to the convection core; inputting the original satellite image, the convection core image and the physical information into a pre-constructed intensity estimation model to obtain an intensity estimation result of the tropical cyclone. The application performs twice feature extraction and weighting on the image, extracts more accurate features, inputs the features and the core physical information of the tropical cyclone into a full connection layer, directly provides an indication of the intensity of the cyclone through the physical information, realizes a physically guided method, can accurately estimate the intensity of the tropical cyclone, improves the focusing ability on the most relevant channel and the focusing ability on a specific spatial region in the image, can perform more accurate feature extraction, and effectively identifies the importance of each feature.

Description

Technical Field

[0001] This invention relates to the field of tropical cyclone technology, and in particular to a method and system for estimating the intensity of tropical cyclones. Background Technology

[0002] Tropical cyclones (TCs) are among the most destructive weather systems on Earth. The intensity of a TC is defined as the maximum sustained winds near its center. Accurate TC intensity estimation is crucial for understanding TC development and evolution, improving TC forecasting, and disaster prevention and mitigation. In traditional meteorological applications, the main methods for TC intensity estimation include the Dvorak technique, the Advanced Dvorak technique (ADT), the Deviation Angle Variance technique (DAV), and the Satellite Agreement Concordance technique (SATCON).

[0003] With the rapid development of deep learning technology in image processing, especially the development of convolutional neural networks (CNNs), TC intensity estimation has ushered in a new breakthrough. CNNs have the ability to automatically learn features, extracting spatial features from satellite images without human intervention, providing a completely new solution for TC intensity estimation. An existing technology designs a CNN-based architecture for classifying TC based on its intensity. Compared with state-of-the-art techniques using satellite images for intensity estimation, this model achieves better accuracy and lower RMSE, and allows visualization of the features and their deconstruction at different layers to understand the learning process. A new dataset from remote sensing satellites (four channels: VIS, PMW, IR, WV) has also been released for the TC image-to-intensity estimation task, and a new model based on AlexNet has been proposed. For the first time, domain knowledge (rotation invariance) has been incorporated into the convolutional model to achieve better performance. Incorporating environmental factors such as latitude and longitude into the model results in even better accuracy. Notably, rotation invariance is utilized in the testing phase, with samples rotated multiple times before being fed into the model for training and then averaged. In recent research, researchers have begun to consciously extract features more closely related to intensity.

[0004] While the aforementioned studies yielded more reliable estimates, they also have some problems. They failed to fully consider relevant knowledge already available or in use in traditional meteorological fields. Their work often incorporates data containing a large amount of information, such as satellite imagery, into the model for learning. Even with attention mechanisms, the model is easily influenced by irrelevant information, leading to biases. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for analyzing the intensity of tropical cyclones. By performing feature extraction and weighting twice on the image, more accurate features are extracted. These features, along with the core physical information of the tropical cyclone, are input into a fully connected layer, realizing a physical-guided method that can accurately estimate the intensity of tropical cyclones.

[0006] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0007] On one hand, the present invention provides a method for estimating the intensity of a tropical cyclone, comprising:

[0008] Acquire raw satellite images of tropical cyclones;

[0009] Physical information and convection cores are extracted from the raw satellite images;

[0010] Generate a convection core image based on the described convection core;

[0011] The original satellite images, convection core images, and physical information are input into a pre-built intensity estimation model to obtain the intensity estimation results of the tropical cyclone.

[0012] The processing steps of the intensity estimation model include:

[0013] Features are extracted from the original satellite image and the convection core image respectively to obtain feature maps of the original satellite image and the convection core image.

[0014] The feature maps of the original satellite image and the convective core image are subjected to secondary feature extraction and weighting respectively to obtain the weighted feature map of the original satellite image and the weighted feature map of the convective core image;

[0015] The weighted feature map of the original satellite image and the weighted feature map of the convection core image are superimposed to obtain the sum of the image features;

[0016] The sum of the features of the image and the physical information are input into a fully connected layer to output the intensity estimate of the tropical cyclone.

[0017] Optionally, the physical information includes the number of convective cores, the temperature of the convective cores, and geographical factors.

[0018] Optionally, the extraction of the convection core is as follows:

[0019] ;

[0020] In the formula, For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The coordinates of the surrounding pixels; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core on the surface.

[0021] Optionally, features of the original satellite image can be extracted using a 3D convolutional neural network, and features of the convection core image can be extracted using a 2D convolutional neural network.

[0022] Optionally, a pre-constructed improved attention module is used to perform secondary feature extraction and weighting on the feature maps of the original satellite image and the convection core image, respectively. The improved attention module includes an improved channel attention module and an improved spatial attention module, specifically:

[0023] In the improved channel attention module, feature information of the feature map of the original satellite image is extracted and flattened to obtain the feature map of the flattened original satellite image. The feature information extracted from the feature map of the original satellite image is used for channel learning to obtain channel attention weights. The channel attention weights Feature maps weighted to the flattened original satellite image In the process, the feature map of the weighted original satellite image is obtained. ;

[0024] In the improved spatial attention module, feature information of the feature map of the convection core image is extracted and flattened to obtain the feature map of the flattened convection core image. The feature information extracted from the feature map of the convection core image is used for channel learning to obtain spatial attention weights. The spatial attention weights Feature map of the convection core image after weighting and flattening In the process, the feature map of the weighted convection core image is obtained. .

[0025] Optionally, the construction of the improved channel attention module includes:

[0026] Obtain the original channel attention module; delete the max pooling layer of the original channel attention module, add a 3D convolutional layer before the global average pooling layer of the original channel attention module, and add multiple fully connected layers after the global average pooling layer of the original channel attention module to obtain the improved channel attention module.

[0027] The construction of the improved spatial attention module includes:

[0028] Obtain the original spatial attention module; add multiple feature extraction layers before the input gate of the original spatial attention module, and add a Sigmoid activation function after the multiple feature extraction layers to obtain the improved spatial attention module.

[0029] Optionally, the feature maps of the weighted images can be superimposed, as shown in the formula:

[0030] ;

[0031] In the formula, The sum of the features of the image; It is a superposition function; The feature map is a weighted representation of the original satellite image. This is a weighted feature map of the convection core image.

[0032] Optionally, the training steps of the intensity estimation model include: optimizing the parameters of the intensity estimation model using the root mean square error loss function and the mean absolute error loss function.

[0033] Optionally, the root mean square error loss function and the mean absolute error loss function are:

[0034] ;

[0035] ;

[0036] ;

[0037] In the formula, The root mean square error loss function; The mean absolute error loss function; The number of image samples; For image samples; This represents the true intensity of the tropical cyclone. This represents the estimated intensity of a tropical cyclone. This is the intensity estimation function; The sum of the features of the image; It is physical information.

[0038] On the other hand, the present invention provides a tropical cyclone intensity estimation system, comprising:

[0039] Image acquisition module, used to acquire raw satellite images of tropical cyclones;

[0040] Intensity estimation models are used to estimate the intensity of tropical cyclones.

[0041] The image feature extraction module is used to extract features from the raw satellite images and the convection core images;

[0042] An improved attention module is used to perform secondary feature extraction and weighting on the feature maps of the original satellite image and the feature maps of the convection core image.

[0043] Beneficial effects

[0044] Compared with existing technologies, the beneficial effects achieved in this aspect are as follows:

[0045] (1) The present invention performs feature extraction and weighting on the image twice to extract more accurate features, and inputs them and the core physical information of the tropical cyclone into the fully connected layer. The physical information directly provides an indication of the cyclone intensity, realizing a physical guidance method that can accurately estimate the intensity of the tropical cyclone.

[0046] (2) In response to the specific needs of image feature extraction, the present invention performs the first feature extraction through a multi-branch convolutional neural network and performs the second feature extraction and weighting through two different attention modules, which improves the ability to focus on the most relevant channels and the ability to focus on specific spatial regions in the image, enabling more accurate feature extraction and effectively identifying the importance of each feature. Attached Figure Description

[0047] Figure 1 The diagram shown is a flowchart of one embodiment of the tropical cyclone intensity estimation method of the present invention.

[0048] Figure 2 The diagram shown is a schematic representation of the tropical cyclone intensity estimation system of the present invention in one embodiment.

[0049] Figure 3 The diagram shown is a flowchart illustrating the feature extraction process of the original satellite image in one embodiment of the present invention.

[0050] Figure 4 The diagram shown is a flowchart illustrating the feature extraction process of the convection core image in one embodiment of the present invention.

[0051] Figure 5 The diagram shown is a structural schematic of the intensity estimation model of the present invention in one embodiment;

[0052] Figure 6 The diagram shown is a structural schematic of the improved channel attention module of the present invention in one embodiment;

[0053] Figure 7 The diagram shown is a structural schematic of the improved spatial attention module of the present invention in one embodiment;

[0054] Figure 8 The image shown is a schematic diagram of the original satellite image of the present invention in one embodiment;

[0055] Figure 9 The diagram shown is a structural schematic of the convection core image of the present invention in one embodiment;

[0056] Figure 10 The image shows a scatter plot of the intensity estimation results of the tropical cyclone according to one embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0058] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0059] Example 1

[0060] like Figure 1 As shown in the figure, this embodiment introduces a method for estimating the intensity of a tropical cyclone, including:

[0061] Obtain raw satellite images of tropical cyclones. For example... Figure 8 The image shown is the original satellite image of a tropical cyclone.

[0062] Physical information and convection cores are extracted from the raw satellite images.

[0063] A convection core image is generated based on the described convection core. For example... Figure 9 The image shown is a convection core image.

[0064] The extraction of the convection core is as follows:

[0065] ;

[0066] In the formula, For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The coordinates of the surrounding pixels; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core on the surface.

[0067] Satellite images are composed of many different pixels. Essentially, they are obtained by finding pixels that satisfy a formula from satellite infrared images and defining these pixels as convection cores. The physical information is the number of these pixels and their minimum and maximum temperatures.

[0068] The physical information includes the number of convective cores, their temperature, and geographical factors. Some characteristics of the convective cores constitute physical information.

[0069] The raw satellite images, convective core images, and physical information are input into a pre-built intensity estimation model to obtain the intensity estimation results of the tropical cyclone. For example... Figure 5 The diagram shows the structure of the intensity estimation model.

[0070] The processing steps of the intensity estimation model include:

[0071] Features of the original satellite image are extracted using a 3D convolutional neural network to obtain a feature map of the original satellite image; features of the convection core image are extracted using a 2D convolutional neural network to obtain a feature map of the convection core image. The use of the 3D convolutional neural network allows the model to process multiple image frames simultaneously, preserving the spatiotemporal context crucial for accurate intensity estimation. Here, 2D convolution is used to focus on the spatial features of the convection core without the additional complexity of handling the temporal dimension.

[0072] We used 3D convolutional neural networks (3D-CNN) and 2D convolutional neural networks (2D-CNN) to extract and learn features from the original satellite images and the convection core images, respectively. This is because the three channels of the satellite cloud image contain a lot of information, and 2D convolutional neural networks (2D-CNN) cannot effectively obtain the individual information of each channel. Therefore, we used 3D convolutional neural networks (3D-CNN) to extract information from the three-channel images during feature extraction.

[0073] To identify and learn more valuable information, we utilize an improved attention module to further extract features from the features extracted by the two neural networks. The improved attention module performs secondary feature extraction and weighting on the feature maps of the original satellite image and the convection core image, respectively, to obtain the weighted feature map of the original satellite image. Feature map weighted by convection core image .

[0074] The improved attention module includes an improved channel focus module (ICFM) and an improved spatial focus module (ISFM), such as... Figure 6 and Figure 7 The diagrams shown are schematic representations of the Improved Channel Focus Module (ICFM) and the Improved Spatial Focus Module (ISFM). Specifically:

[0075] like Figure 3 As shown, in the Improved Channel Focus Module (ICFM), the feature information of the feature map of the original satellite image is extracted and flattened to obtain the feature map of the flattened original satellite image. The feature information extracted from the feature map of the original satellite image is used for channel learning to obtain channel attention weights. The channel attention weights Feature maps weighted to the flattened original satellite image In the process, the weighted feature map of the original satellite image is obtained. .

[0076] The weighted feature map of the original satellite image The calculation formula is:

[0077] .

[0078] Similar to Figure 4 As shown, in the Improved Spatial Focus Module (ISFM), feature information of the feature map of the convection core image is extracted and flattened to obtain the feature map of the flattened convection core image. The feature information extracted from the feature map of the convection core image is used for channel learning to obtain spatial attention weights. The spatial attention weights Feature map of the convection core image after weighting and flattening In the process, the weighted feature map of the convection core image is obtained. .

[0079] The weighted feature map of the convection core image The calculation formula is:

[0080] .

[0081] like Figure 6 As shown, the construction of the improved channel focus module (ICFM) includes:

[0082] Obtain the original channel attention module.

[0083] Since the max pooling layer weakens the information around the eye of the typhoon, the max pooling layer of the original channel attention module is deleted, a 3D convolutional layer is added before the global average pooling layer of the original channel attention module, and multiple fully connected layers are added after the global average pooling layer of the original channel attention module to obtain an improved channel attention module.

[0084] This design facilitates deeper feature extraction and channel information fusion from three-channel cloud images. 3D convolutional layers can better capture complex features and interactive information in three-channel cloud images. Adding a fully connected layer after each convolutional layer enhances feature extraction capabilities. These fully connected layers further emphasize the importance of each channel, thereby better distinguishing the distribution of features such as cloud thickness and density during training, making the Improved Channel Focus Module (ICFM) more precise in focusing on important channels.

[0085] Existing Channel Focus Modules (CFM) use global average pooling and max pooling to obtain global channel information, and then generate channel attention maps through 1×1 convolutions and activation functions. The channel attention map is usually multiplied element-wise with the input feature map and then directly passed to the next network layer.

[0086] like Figure 7 As shown, the construction of the improved spatial focus module (ISFM) includes:

[0087] Obtain the original spatial attention module.

[0088] By adding multiple feature extraction layers before the input gate of the original spatial attention module and then adding a sigmoid activation function after the multiple feature extraction layers, an improved spatial attention module is obtained.

[0089] This invention employs a hierarchical processing structure. First, the input convection core image undergoes progressive feature extraction, generating multiple feature layers. Finally, a sigmoid activation function is used to generate a spatial attention map. This structure is more conducive to capturing the spatial information of the convection core image, thereby improving the extraction of global features. During training, convection core feature maps of different resolutions are weighted to better extract multi-scale information from the image. The combination of feature maps of different resolutions helps to improve the spatial attention module (ISFM) to more accurately locate the convection core region and enhance the focus on target features.

[0090] Existing spatial attention modules operate directly on global spatial features, such as generating a spatial attention map from global features through convolution. During training, spatial attention... Figure 1 Generally, it is multiplied element-wise with the input feature map and then directly passed to the next layer.

[0091] Weighted feature map of the original satellite image Feature map weighted by convection core image By superimposing the features, we obtain the sum of the image features, as shown in the formula:

[0092] ;

[0093] In the formula, The sum of the features of the image; It is a superposition function; The feature map is a weighted representation of the original satellite image. This is a weighted feature map of the convection core image. It can simultaneously utilize spatial and channel-specific information. This fusion combines the advantages of both branches, improving the overall accuracy of intensity estimation.

[0094] During convolution, the convection core image is gradually broken down by the convolution kernel, which destroys quantitative information. If the lowest and highest temperatures occur at points where these points are broken down, they will also be lost. Therefore, the sum of the image features and the physical information are input into a fully connected layer to output a tropical cyclone intensity estimate. This physical information directly indicates cyclone intensity but is easily overlooked by the model.

[0095] The parameters of the intensity estimation model are optimized using the root mean square error loss function and the mean absolute error loss function.

[0096] To understand the difference between the estimated values ​​and the true values ​​from the intensity estimation model, we use the RMSE and MAE loss functions to calculate the error between the estimated and true values. The root mean square error loss function and the mean absolute error loss function are as follows:

[0097] ;

[0098] ;

[0099] ;

[0100] In the formula, The root mean square error loss function; The mean absolute error loss function; The number of image samples; For image samples; This represents the true intensity of the tropical cyclone. This represents the estimated intensity of a tropical cyclone. This is the intensity estimation function; The sum of the features of the image; It is physical information.

[0101] Example 2

[0102] like Figure 1 and Figure 10 As shown in the figure, this embodiment introduces the experimental steps of a method for estimating the intensity of a tropical cyclone, specifically including:

[0103] The satellite imagery data used in this embodiment is an integrated TCIR dataset obtained from two public data sources, GridSat and CMORPH. This dataset provides satellite cloud images in four channels: infrared (IR), water vapor (WV), visible light (VIS), and passive microwave (PMW).

[0104] The IR, WV, and VIS channels are sourced from the GridSat data source, which collects data every 3 hours with a spatial resolution of 0.07°. The PMW channel is sourced from the CMORPH data source, also with a 3-hour interval, but with a spatial resolution of 0.25°. Therefore, linear interpolation is required to scale the PMW channel to 0.07° to ensure uniform channel scale.

[0105] Because the VIS channel is extremely unstable, we discarded it and only used the IR, WV, and PMW channels. The convection core map (CC) was obtained from the IR image using CST technology. Specific satellite image data are shown in Table 1.

[0106] Table 1 Satellite image data

[0107]

[0108] We used TC data from 2003 to 2017. We divided the dataset into training and validation sets in a 9:1 ratio, and used a separate test set for the 2017 data to evaluate the performance of the best model. We also performed centering, random rotation, and normalization operations on the data.

[0109] To verify the effectiveness of each part of the model, we designed five methods (M1, M2, M3, M4, M5) for experiments, and the results are shown in Table 2.

[0110] The M1 method uses a baseline model. Based on experimental results, this method uses 3D-CNN to extract information from a three-channel cloud map and directly enters the fully connected layer for intensity estimation without going through an attention module.

[0111] In the M2 method, we introduce ICFM to further extract global channel information. Experimental results show that the performance is improved by 2.7% after using channel attention, which indicates that our designed attention has learned the interrelationships between channels.

[0112] In the M3 method, we incorporate the convection core map into the intensity estimation model and use 2D-CNN to extract features to assist in intensity estimation. As shown in the table, compared with the M1 method, adding the convection core improves the performance by 6.9%, which indicates that prior knowledge plays a significant role in assisting intensity estimation.

[0113] To verify the effectiveness of ISFM, we added this module to the M4 method. Experimental results show that while the features of the convective core are extracted in the M3 method, more spatial location information is not taken into account, while ISFM successfully enables the intensity estimation model to learn more refined spatial location features of the convective core.

[0114] Finally, we incorporated physical information into the intensity estimation model to obtain the M5 method. The addition of physical information further reduced the error, proving that physical information, namely the number and temperature of convective cores, plays a role in intensity estimation.

[0115] Table 2 shows the relevant configurations and experimental results for methods M1, M2, M3, M4, and M5, where Δ represents the performance improvement in RMSE compared to method M1.

[0116] Table 2 shows the relevant model configurations and experimental results.

[0117]

[0118] Figure 10 The image shows a scatter plot of the optimal path strength and the model-estimated strength results in this embodiment, with an RMSE of 8.13. The red line in the figure represents the linear fit between the two strengths, with the corresponding equation being y = 0.93x + 1.94. This indicates that this embodiment has high accuracy on the independent test set.

[0119] Example 3

[0120] like Figure 2 As shown in the figure, this embodiment introduces a tropical cyclone intensity estimation system, including:

[0121] The image acquisition module is used to acquire raw satellite images of tropical cyclones.

[0122] An intensity estimation model is used to estimate the intensity of tropical cyclones. The intensity estimation model includes an image feature extraction module, an improved attention module, and a fully connected layer.

[0123] The image feature extraction module is used to extract features from the original satellite images and the convection core images.

[0124] An improved attention module is used to perform secondary feature extraction and weighting on the feature maps of the original satellite image and the feature maps of the convection core image.

[0125] The improved attention module includes an improved channel attention module and an improved spatial attention module.

[0126] The improved channel attention module is used to perform secondary feature extraction and weighting on the feature map of the original satellite image.

[0127] The improved spatial attention module is used to perform secondary feature extraction and weighting on the feature map of the convection core image.

[0128] The fully connected layer is used to estimate the intensity of tropical cyclones.

[0129] Example 4

[0130] This embodiment introduces a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the tropical cyclone intensity estimation method described in Embodiment 1 or 2.

[0131] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0133] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0135] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for estimating the intensity of a tropical cyclone, characterized in that, include: Obtain raw satellite images of tropical cyclones; Physical information and convection cores are extracted from the raw satellite images; Generate a convection core image based on the described convection core; The original satellite images, convection core images, and physical information are input into a pre-built intensity estimation model to obtain the intensity estimation results of the tropical cyclone. The processing steps of the intensity estimation model include: Features are extracted from the original satellite image and the convection core image respectively to obtain feature maps of the original satellite image and the convection core image. The feature maps of the original satellite image and the convective core image are subjected to secondary feature extraction and weighting respectively to obtain the weighted feature map of the original satellite image and the weighted feature map of the convective core image; The weighted feature map of the original satellite image and the weighted feature map of the convection core image are superimposed to obtain the sum of the image features; The sum of the features of the image and the physical information are input into a fully connected layer to output the intensity estimation result of the tropical cyclone. Features of the original satellite image are extracted using a 3D convolutional neural network, and features of the convection core image are extracted using a 2D convolutional neural network. The feature maps of the original satellite image and the convection core image are subjected to secondary feature extraction and weighting through a pre-constructed improved attention module. The improved attention module includes an improved channel attention module and an improved spatial attention module, specifically: In the improved channel attention module, feature information of the feature map of the original satellite image is extracted and flattened to obtain the feature map of the flattened original satellite image. The feature information extracted from the feature map of the original satellite image is used for channel learning to obtain channel attention weights. The channel attention weights Feature maps weighted to the flattened original satellite image In the process, the weighted feature map of the original satellite image is obtained. ; In the improved spatial attention module, feature information of the feature map of the convection core image is extracted and flattened to obtain the feature map of the flattened convection core image. The feature information extracted from the feature map of the convection core image is used for channel learning to obtain spatial attention weights. The spatial attention weights Feature map of the convection core image after weighting and flattening In the process, the weighted feature map of the convection core image is obtained. ; The construction of the improved channel attention module includes: Obtain the original channel attention module; delete the max pooling layer of the original channel attention module, add a 3D convolutional layer before the global average pooling layer of the original channel attention module, and add multiple fully connected layers after the global average pooling layer of the original channel attention module to obtain the improved channel attention module. The construction of the improved spatial attention module includes: Obtain the original spatial attention module; add multiple feature extraction layers before the input gate of the original spatial attention module, and add a Sigmoid activation function after the multiple feature extraction layers to obtain the improved spatial attention module.

2. The method for estimating the intensity of a tropical cyclone according to claim 1, characterized in that, The physical information includes the number of convective cores, the temperature of the convective cores, and geographical factors.

3. The method for estimating the intensity of a tropical cyclone according to claim 1, characterized in that, The extraction of the convection core is as follows: ； In the formula, For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The coordinates of the surrounding pixels; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core above; For at pixel The temperature of the convective core on the surface.

4. The method for estimating the intensity of a tropical cyclone according to claim 1, characterized in that, The weighted feature map of the original satellite image and the weighted feature map of the convection core image are superimposed, as shown in the formula: ； In the formula, The sum of the features of the image; It is a superposition function; The feature map is a weighted representation of the original satellite image. This is a weighted feature map of the convection core image.

5. The method for estimating the intensity of a tropical cyclone according to claim 1, characterized in that, The training steps of the intensity estimation model include: optimizing the parameters of the intensity estimation model using the root mean square error loss function and the mean absolute error loss function.

6. The method for estimating the intensity of a tropical cyclone according to claim 5, characterized in that, The root mean square error loss function and the mean absolute error loss function are: ； ； ； In the formula, The root mean square error loss function; The mean absolute error loss function; The number of image samples; For image samples; This represents the true intensity of the tropical cyclone. This represents the estimated intensity of a tropical cyclone. This is the intensity estimation function; The sum of the features of the image; It is physical information.

7. A system for estimating the intensity of a tropical cyclone, characterized in that, The system for performing the method according to any one of claims 1 to 6 comprises: Image acquisition module, used to acquire raw satellite images of tropical cyclones; The image feature extraction module is used to extract features from the raw satellite images and the convection core images; An improved attention module is used to perform secondary feature extraction and weighting on the feature maps of the original satellite image and the feature maps of the convection core image; The fully connected layer is used to estimate the intensity of tropical cyclones.

Images