Tropical cyclone concentric eye wall identification method, device, equipment and medium

Abstract

The application relates to a concentric eye wall recognition method, device, equipment and medium for a tropical cyclone, and particularly relates to the meteorological field. The method comprises the following steps: searching a target microwave brightness temperature image containing a tropical cyclone from a historical microwave brightness temperature image set, wherein the historical microwave brightness temperature image set comprises microwave brightness temperature images scanned by a polar orbit satellite; performing sample labeling on the target microwave brightness temperature image to obtain sample labeling data, wherein the sample labeling data comprises positive samples of a concentric eye wall type and negative samples of a non-concentric eye wall type; constructing a deep learning model based on transfer learning; performing model training on the deep learning model by using the sample labeling data to obtain a trained deep learning model; and inputting a to-be-recognized microwave brightness temperature image containing a tropical cyclone into the trained deep learning model to output a concentric eye wall recognition result of the tropical cyclone. Based on the technical scheme provided in the application, the concentric eye wall can be efficiently and accurately recognized.

Description

Technical Field

[0001] This invention relates to the field of meteorology, specifically to a method, apparatus, equipment, and medium for identifying concentric eyewalls of tropical cyclones. Background Technology

[0002] Concentric eyewalls (CEs) are a type of double-wall structure in tropical cyclones. Observation is the foundation for research and forecasting of concentric eyewall mechanisms and is of great significance for understanding the causes and evolution mechanisms of concentric eyewalls.

[0003] Generally, concentric eyewalls can be observed using microwave satellite remote sensing: polar-orbiting satellites orbit the Earth's North and South Poles and receive microwave signals, which can be used to obtain microwave brightness temperature (TB) images through inversion.

[0004] A solution is urgently needed to process microwave brightness temperature images to accurately identify concentric eyewalls. Summary of the Invention

[0005] This application provides a method, apparatus, equipment, and medium for identifying concentric eyewalls of tropical cyclones, and the technical solution is as follows.

[0006] On the one hand, a method for identifying concentric eyewalls of tropical cyclones is provided, the method comprising:

[0007] Find target microwave brightness temperature images containing tropical cyclones from a historical microwave brightness temperature image set, which includes microwave brightness temperature images obtained by polar-orbiting satellite scans;

[0008] The target microwave brightness temperature image is labeled with samples to obtain sample labeling data, which includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls.

[0009] Construct deep learning models based on transfer learning;

[0010] The deep learning model is trained using the labeled sample data to obtain a trained deep learning model.

[0011] Input the microwave brightness temperature image containing the tropical cyclone to be identified into the trained deep learning model, and output the concentric eyewall recognition result of the tropical cyclone.

[0012] In another aspect, a concentric eyewall identification device for tropical cyclones is provided, the device comprising:

[0013] The target image search module is used to search for target microwave brightness temperature images containing tropical cyclones from a historical microwave brightness temperature image set, which includes microwave brightness temperature images obtained by polar-orbiting satellite scanning.

[0014] The sample annotation module is used to annotate the target microwave brightness temperature image to obtain sample annotation data, which includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls.

[0015] The model building module is used to build deep learning models based on transfer learning.

[0016] The model training module is used to train the deep learning model using the sample labeled data to obtain a trained deep learning model.

[0017] The image recognition module is used to input microwave brightness temperature images containing tropical cyclones into a trained deep learning model and output the concentric eyewall recognition results of the tropical cyclones.

[0018] In another aspect, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, at least one program, code set or instruction set being loaded and executed by the processor to implement the above-described method for identifying concentric eyewalls of tropical cyclones.

[0019] In another aspect, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored therein, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the above-described method for identifying concentric eyewalls of tropical cyclones.

[0020] In another aspect, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned method for identifying concentric eyewalls of tropical cyclones.

[0021] The technical solution provided in this application may include the following beneficial effects:

[0022] Target microwave brightness temperature (MBT) images containing tropical cyclones are identified from a historical collection of microwave brightness temperature (MBT) images. These images are then labeled to obtain labeled data, which includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls. This labeled data is used to train a deep learning model based on transfer learning, resulting in a trained deep learning model. This model can be used to identify concentric eyewalls in MBT images containing tropical cyclones, simplifying the identification process into a binary image classification task, thus achieving efficient and reliable concentric eyewall identification. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of a microwave brightness temperature image of a tropical cyclone according to an exemplary embodiment.

[0025] Figure 2 This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment.

[0026] Figure 3 This is a schematic diagram of a microwave brightness temperature image of a tropical cyclone according to an exemplary embodiment.

[0027] Figure 4 This is a schematic diagram illustrating the change of the number of labeled samples and the CE ratio over time according to an exemplary embodiment.

[0028] Figure 5 This is a schematic diagram illustrating the extraction of microwave brightness temperature images of tropical cyclones from microwave data of polar-orbiting satellites, according to an exemplary embodiment.

[0029] Figure 6 This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment.

[0030] Figure 7 This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment.

[0031] Figure 8 This is a schematic diagram illustrating an image processing flow according to an exemplary embodiment.

[0032] Figure 9This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment.

[0033] Figure 10 This is a structural block diagram of a concentric eyewall identification device for tropical cyclones, according to an exemplary embodiment.

[0034] Figure 11 This is a schematic diagram of a computer device provided according to an exemplary embodiment. Detailed Implementation

[0035] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] It should be understood that the term "instruction" mentioned in the embodiments of this application can be a direct instruction, an indirect instruction, or an indication of a relationship. For example, A instructing B can mean that A directly instructs B, such as B being able to obtain information through A; it can also mean that A indirectly instructs B, such as A instructing C, so B can obtain information through C; or it can mean that there is a relationship between A and B.

[0037] In the description of the embodiments of this application, the term "correspondence" may indicate that there is a direct or indirect correspondence between two things, or that there is an association between two things, or that there is a relationship of instruction and being instructed, configuration and being configured, etc.

[0038] In the embodiments of this application, "predefined" can be achieved by pre-storing corresponding codes, tables or other means that can be used to indicate relevant information in the device (e.g., including terminal devices and network devices). This application does not limit the specific implementation method.

[0039] Observation is fundamental to the study and forecasting of concentric eyewalls, and is of great significance for understanding their formation and evolution mechanisms. The vast majority of concentric eyewalls form over tropical ocean surfaces far from land, making in-situ observation difficult. Currently, observation relies mainly on satellite remote sensing, aerial surveys, and ground-based radar detection. Among these methods, microwave satellite signals, which can penetrate high cloud cover, are currently the most effective and commonly used observation method.

[0040] Microwave brightness temperature images can be obtained by inverting microwave signals. These images effectively reflect convective activity around tropical cyclones. Low-value areas correspond to strong convective structures such as the inner and outer eyewalls and spiral rainbands, while high-value areas correspond to the eye and downdraft regions. For example,... Figure 1As shown, microwave brightness temperature images can clearly show the strong convection region (inner and outer eyewalls) and downdraft region (eye and moat) of Typhoon Amber (1997), thus allowing observers to accurately determine Amber's concentric eyewall structure.

[0041] Concentric eyewall identification refers to the extraction of concentric eyewall structures from microwave brightness-temperature images. Due to the lack of a unified and objective definition of concentric eyewalls, current identification primarily relies on subjective analysis methods, requiring the manual setting of a completeness threshold standard. This approach suffers from the following drawbacks: subjective methods require human intervention, resulting in low identification efficiency and difficulty in automation. Due to these shortcomings, existing concentric eyewall identification methods cannot meet the needs of scientific research and operational observation; therefore, there is a need to develop automated and efficient identification methods.

[0042] To address the above problems, the technical solution provided in this application simplifies concentric eyewall recognition into an image binary classification task, and develops an efficient and accurate concentric eyewall recognition method based on a large visual model and transfer learning. The technical solution provided in this application will be further explained below.

[0043] Figure 2 This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment. The method is applied in a computer device. Figure 2 As shown, the method for identifying the concentric eyewalls of this tropical cyclone may include the following steps:

[0044] Step 210: Find the target microwave brightness temperature image containing the tropical cyclone from the historical microwave brightness temperature image set.

[0045] The historical microwave brightness temperature image collection includes microwave brightness temperature images obtained from polar-orbiting satellite scans.

[0046] In this embodiment of the application, the polar-orbiting satellite obtains multiple microwave brightness temperature images through microwave satellite remote sensing. Tropical cyclones may or may not be observed in the microwave brightness temperature images. These microwave brightness temperature images constitute a historical microwave brightness temperature image set. The microwave brightness temperature image containing a tropical cyclone is selected from the historical microwave brightness temperature image set and recorded as the target microwave brightness temperature image.

[0047] Step 220: Perform sample annotation on the target microwave brightness temperature image to obtain sample annotation data. The sample annotation data includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls.

[0048] In this embodiment of the application, the target microwave brightness temperature image is labeled as a positive sample of the concentric eyewall type or a negative sample of the non-concentric eyewall type, and then standardized to obtain sample labeling data.

[0049] In one possible implementation, step 220 includes: labeling target microwave brightness temperature images that meet the concentric eyewall judgment criteria as positive samples of the concentric eyewall class; labeling target microwave brightness temperature images that do not meet the concentric eyewall judgment criteria as negative samples of the non-concentric eyewall class; wherein the concentric eyewall judgment criteria include: the outer convection ring of the tropical cyclone in the image occupies at least 2 / 3 of the circle, and there is a descending airflow region between the outer and inner convection rings of the tropical cyclone in the image.

[0050] In this implementation, the criteria for determining concentric eyewalls are: the outer convection ring occupies at least 2 / 3 of the circle, and there is a clear downdraft region (i.e., the Moat region) between the inner and outer convection rings. This criterion can be directly observed, and the process is simple.

[0051] For example, such as Figure 3 As shown, Figure 3 Examples (a) and (b) in the above criteria meet the above judgment criteria and are labeled as positive samples of the concentric eye wall category. Figure 3 In (c) and (d), the external strong convection rings did not reach 2 / 3 of the circumference, which did not meet the above judgment criteria, and were labeled as negative samples of non-concentric eyewall type.

[0052] In one possible implementation, after step 220, the following steps are also performed: the sample labeled data is divided into a training set, a validation set, and a test set using a random hold-out method.

[0053] In this implementation, a random hold-out method is used to divide the labeled sample data into a training set (e.g., 70%), a validation set (e.g., 15%), and a test set (e.g., 15%). This is because the quantity and CE (Complete Validation) ratio of labeled sample data fluctuate significantly across different years, making the random hold-out method more reasonable than chronological division. For example, as shown... Figure 4 As shown, the number of labeled data and the ratio of positive to negative samples fluctuate significantly in different years. If the dataset is divided according to time order, it is easy to produce differences in data distribution. For example, the proportion of CE in the last three years is significantly higher than the average level. If it is divided into validation and test sets, it will produce distribution differences with the training set.

[0054] Step 230: Construct a deep learning model based on transfer learning.

[0055] In this embodiment of the application, a deep learning model is pre-trained to construct a deep learning model based on transfer learning, and the transfer learning method is used to improve the model performance.

[0056] Step 240: Train the deep learning model using the sample labeled data to obtain the trained deep learning model.

[0057] Step 250: Input the microwave brightness temperature image containing the tropical cyclone to be identified into the trained deep learning model, and output the concentric eyewall recognition result of the tropical cyclone.

[0058] From an image processing perspective, concentric eyewall recognition can be understood as a target detection or image classification task. Extracting concentric eyewall samples using target detection methods involves two steps: target localization and classification. First, the model reads microwave brightness temperature images scanned around the Earth by polar-orbiting satellites (…). Figure 5 The method (a) determines the location and extent of a tropical cyclone from the target, but this target occupies less than 5% of the entire area, and the vast majority of it is redundant information unrelated to tropical cyclones. After completing the target localization, the model also needs to complete the classification task based on the target's characteristics. This method requires processing a large amount of redundant information and providing labeled samples containing the location and extent of tropical cyclones, resulting in problems such as low efficiency, high labeling costs, and susceptibility to interference from similar targets such as temperature cyclones.

[0059] Another approach is to use image classification methods for identification. This involves obtaining microwave brightness temperature images of tropical cyclones beforehand. Figure 5 (b)). Based on this, the deep learning model only needs to determine whether the microwave brightness temperature image of a tropical cyclone belongs to the concentric eyewall or non-concentric eyewall category. Compared with the previous approach, this method has the following advantages: First, the deep learning model only processes local data blocks centered on the tropical cyclone, which can effectively improve computational efficiency; second, the model only needs the image category label, reducing the workload of sample labeling; finally, this method is free from interference from similar targets.

[0060] In summary, the concentric eyewall recognition method for tropical cyclones provided in this embodiment searches for target microwave brightness temperature images containing tropical cyclones from a historical microwave brightness temperature image set, and performs sample annotation on the target microwave brightness temperature images to obtain sample annotation data. The sample annotation data includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls. The sample annotation data is used to train a deep learning model based on transfer learning to obtain a trained deep learning model. This deep learning model can be used to perform concentric eyewall recognition on microwave brightness temperature images containing tropical cyclones, simplifying concentric eyewall recognition into an image binary classification task, thereby achieving efficient and reliable concentric eyewall recognition.

[0061] In an illustrative embodiment, the target microwave brightness temperature image can be pre-selected from a historical microwave brightness temperature image set based on the scanning information of polar-orbiting satellites and the trajectory information of tropical cyclones.

[0062] Figure 6 This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment. The method is applied in a computer device. Figure 6As shown, step 210 above can be replaced by the following steps:

[0063] Step 610: Obtain scanning information from polar-orbiting satellites and trajectory information from tropical cyclones.

[0064] In this embodiment of the application, scanning information of polar-orbiting satellites and trajectory information of tropical cyclones corresponding to historical microwave brightness temperature image sets are obtained.

[0065] Understandably, since tropical cyclone warning agencies conduct observation operations and release observation results in real time—for example, the National Centers for Environmental Prediction (NCEP) provides global tropical cyclone vital statistics records (TC-Vitals) every 6 hours, and the Cooperative Institute for Meteorological Satellite Studies (CIMSS) at the University of Wisconsin releases tropical cyclone tracks every 30 minutes—and historical tracks can be referenced in the International BestTrack Archive for Climate Stewardship version 4 (IBTrACS v4), real-time and historical track information of tropical cyclones is freely available through public channels and will not become a limiting factor in the identification of the concentric eyewall.

[0066] Step 620: Based on the overlapping areas in the scanning information and trajectory information, determine the target microwave brightness temperature image from the historical microwave brightness temperature image set.

[0067] In this embodiment of the application, if the satellite scan strip in the scanning information and the movement trajectory of the tropical cyclone in the trajectory information overlap at the same time point, then the location corresponding to that time point is the overlapping area, and the microwave brightness temperature image with the overlapping area is recorded as the target microwave brightness temperature image.

[0068] For example, the data processing flow for microwave brightness temperature images is as follows:

[0069] (1) Use open-source microwave brightness temperature data to obtain scanning information of polar-orbiting satellites, and use IBTACS v4 to obtain trajectory information of tropical cyclones at 3-hour intervals.

[0070] (2) An algorithm is designed to automatically search for overlapping areas between the scanning information of polar-orbiting satellites and the trajectory information of tropical cyclones. If there is overlap, a two-dimensional data block centered on the eye of the cyclone and with a longitudinal and latitudinal coverage of 8° is extracted from the microwave brightness temperature data of the satellite's vertical polarization. Missing values ​​in the data block are uniformly filled with 270K. If the missing data exceeds 50%, the data block is discarded.

[0071] (3) The two-dimensional microwave brightness temperature data blocks are uniformly interpolated to a 0.02°×0.02° grid using the natural neighborhood method, thereby eliminating the granular texture caused by the irregularity of the original grid. After interpolation, a two-dimensional matrix of size 401×401 is obtained.

[0072] (4) Use two-dimensional microwave brightness temperature data to draw RGB images, and set the output size to 384×384 pixels.

[0073] In summary, the concentric eyewall identification method for tropical cyclones provided in this embodiment determines the overlapping region based on the scanning information of polar-orbiting satellites and the trajectory information of tropical cyclones. Based on the overlapping region, the target microwave brightness temperature image containing the tropical cyclone can be accurately found from the historical microwave brightness temperature image set.

[0074] In an illustrative embodiment, a large version of the visual large model, the Swin Transformer (Swin-L), is selected as the backbone network to construct a deep learning model based on transfer learning.

[0075] Figure 7 This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment. The method is applied in a computer device. Figure 7 As shown, step 230 above can be replaced by the following steps:

[0076] Step 710: Use Swin-L as the backbone network to build a deep learning model.

[0077] Swin-L is a larger version of the Swin Transformer. The Swin Transformer is a deep learning model for computer vision tasks, which has gained widespread attention for its excellent performance in tasks such as image classification, object recognition, and image segmentation. It is a general-purpose large vision model.

[0078] In this embodiment of the application, in order to improve the classification effect, the Swin-L version containing 197 million trainable parameters was selected as the backbone network, the input image size was 384×384 pixels, and the window size for calculating self-attention was 12×12.

[0079] For example, the process of using Swin-L for image processing to identify concentric eye walls is as follows: Figure 8 As shown:

[0080] The microwave brightness temperature image of the tropical cyclone is an RGB color image with an original size of 384×384×3. After standardization, the image is input into Swin-L. First, it undergoes patch partitioning and linear embedding, implemented in code as a 4×4 convolution operation with a stride of 4 and linear embedding. After this processing, the image becomes intermediate feature 1 with a size of 96×96×192. Then, it is processed by the Swin Transformer Block, which calculates self-attention for both fixed and sliding windows. The fixed window reduces computational complexity, while the sliding window increases information transfer between adjacent windows, thus expanding the receptive field. The feature size remains unchanged in the Swin Transformer Block, so the output is intermediate feature 2, which also has a size of 96×96×192. Intermediate feature 2 undergoes a patch-merging operation. The core idea of ​​this operation is to downsample intermediate feature 2, halving its length and width while doubling its size in the channel direction. After patch-merging, intermediate feature 3 with a size of 48×48×384 is obtained. Intermediate feature 3 is then input into the Swing Transformer Block to calculate the self-attention of both the fixed and sliding windows, resulting in intermediate feature 4 with a size of 48×48×384. The patch-merging and Swing Transformer Block calculation process is then repeated twice, reducing the length and width of intermediate feature 4 to 1 / 4 and quadrupling the channel direction, resulting in an output size of 12×12×1536, thus providing the model with more abstract image features. Finally, global average pooling and fully connected operations are used to output the classification result, i.e., the probability that the microwave brightness temperature image belongs to concentric eyewalls or non-concentric eyewalls.

[0081] Step 720: Using a dataset containing targets to be identified with preset shapes, pre-train a deep learning model to obtain a deep learning model based on transfer learning. The preset shapes include: ring and sphere.

[0082] In this embodiment of the application, Swin-L is pre-trained using a dataset containing targets to be identified with preset shapes (such as the ImageNet 21K dataset). The preset shapes include: ring-shaped and spherical.

[0083] Understandably, as a general-purpose large-scale vision model, Swin-L has a trainable parameter scale of 190 million. After pre-training, it can have general feature extraction capabilities and accurately represent the detailed information of the image's underlying layers, including edges, corners, colors, pixels, gradients, etc. This general data representation capability can not only significantly improve the prediction effect of similar or related target tasks, but also has good applicability to different domains, different types, and different distributions of data or tasks. Therefore, the Swin-L pre-trained model can be transferred to various tasks such as image classification, object detection, semantic segmentation, video detection, and image super-resolution.

[0084] Furthermore, the core task of the deep learning model in this embodiment is to extract semantic features from the concentric eyewall structure—the eye (high TB ​​value area), inner eyewall (low TB value area), moat area (high TB ​​value area), and outer eyewall (low TB value area)—from the microwave brightness temperature image. These features include TB gradient, edge, and contour information. The TB RGB image data distribution is similar to that of ImageNet. Additionally, the ImageNet21K dataset contains various ring-shaped or spherical targets similar to the concentric eyewall structure, such as car tires, lifebuoys, donuts, rings, circular tracks, basketballs, soccer balls, tennis balls, ping-pong balls, rugby balls, dome roofs, and spherical egg cells. These similar target categories help improve the transferability of the pre-trained model in the concentric eyewall recognition task. Therefore, the large-scale visual model Swin-L, pre-trained on the ImageNet 21K dataset, not only improves training speed but also effectively enhances the recognition performance of concentric eyewalls.

[0085] In one possible implementation, a deep learning model is trained using labeled sample data to obtain a trained deep learning model, including:

[0086] The hyperparameters of the deep learning model after transfer learning are fine-tuned to obtain multiple fine-tuned models; the multiple fine-tuned models are trained separately using sample labeled data to obtain multiple trained fine-tuned models; the model weights of the multiple trained fine-tuned models are averaged to obtain the final trained deep learning model.

[0087] In this implementation, the optimal combined model is found based on the ensemble learning method: by adjusting hyperparameters such as the Swin-L learning rate (5.0e-5, 1.0e-4, 1.5e-4), learning rate decay factor (5.0e-6, 1.0e-5, 1.5e-5), and training epochs (45, 60, 75), multiple sets of fine-tuning experiments are designed. After each set of fine-tuning experiments, the weights of the model with the highest accuracy on the validation set are saved, and the average of these model weights is calculated to obtain the optimal combined model. The optimal combined model is used as the final deep learning model for application, thereby improving the model's stability and generalization ability.

[0088] Understandably, in related technologies, the best-performing single fine-tuned model is generally selected as the final solution, while the results of other models are discarded. This approach has two main drawbacks: first, single-model predictions suffer from high uncertainty and weak anti-interference capabilities; second, discarding 26 fine-tuned models wastes computational resources. However, conventional multi-model ensembles increase training costs and inference time. Compared to a single best fine-tuned model, the optimal combination model found using ensemble learning methods is obtained by averaging the weights of multiple fine-tuned models. This method does not increase computational complexity or inference time, and can effectively improve the accuracy of concentric eyewall recognition without increasing inference time.

[0089] In summary, the concentric eyewall recognition method for tropical cyclones provided in this embodiment selects the large visual model Swin-L as the backbone network and pre-trains Swin-L using a dataset containing targets to be recognized with preset shapes, including rings and spheres, thereby ensuring model performance based on transfer learning.

[0090] Furthermore, ensemble learning methods are used to find the optimal combination of models to improve the model's classification performance.

[0091] In the illustrative embodiment, a combination of offline and online augmentation is used to address the problems of insufficient training samples and class imbalance.

[0092] Figure 9 This is a flowchart illustrating a method for identifying concentric eyewalls of tropical cyclones according to an exemplary embodiment. The method is applied in a computer device. Figure 9 As shown, after step 220, the following steps may also be included:

[0093] Step 910: In the offline stage, for positive samples labeled as concentric eyewalls, supplementary positive samples are generated by random rotation.

[0094] Understandably, the Deviation Angle Variance (DAV) parameter can be used to measure the structural symmetry of tropical cyclones: Choosing the center of the tropical cyclone as a reference point, the angle between the TB gradient direction of each pixel within a specific radius (100-300 km) and the line connecting that pixel to the reference point is calculated. This angle is the deviation angle of that pixel relative to the reference point. The absolute value of the deviation angle quantitatively characterizes the difference between the tropical cyclone structure and the ideal vortex structure. Then, the variance of the deviation angles of all pixels is calculated, i.e., the DAV value. The smaller the DAV value, the higher the degree of symmetry of the convective structure relative to the center of the tropical cyclone.

[0095] Since the average DAV of non-concentric eyewall classes is higher than that of concentric eyewall classes, it indicates that the structural symmetry of concentric eyewalls is significantly higher than that of non-concentric eyewalls, and symmetrical structures are rotationally invariant. Therefore, positive samples of concentric eyewall classes can be generated using random rotation.

[0096] Step 920: In the online phase, for positive samples labeled as concentric eyewalls and negative samples labeled as non-concentric eyewalls, supplementary positive samples and supplementary negative samples are generated online using an adjustment method.

[0097] The online adjustment methods include at least one of the following: random scaling, random panning, random brightness adjustment, random contrast adjustment, and random saturation adjustment.

[0098] Step 930: Add the supplementary positive samples and supplementary negative samples to the sample labeling data.

[0099] To ensure that the deep learning model can learn to distinguish between strong concentric and non-concentric eye samples, sufficient training samples are needed. If only based on... Figure 6 The method shown results in an insufficient number of labeled samples and suffers from sample imbalance.

[0100] In this embodiment, to address the issues of insufficient training samples and class imbalance, a combination of offline and online augmentation is employed. Specifically, since there are significant differences in the degree of symmetry between concentric and non-concentric eyewall structures, CE samples are first generated using random rotation during the offline augmentation stage to increase the proportion of concentric eyewall samples. Then, online augmentation is used to enhance the overall diversity of the samples. The online augmentation methods include random scaling, translation, and random adjustment of brightness, contrast, and saturation.

[0101] In one possible implementation, the ratio of positive to negative samples in the supplemented sample annotation data is set as the target sample ratio; where the target sample ratio is the sample ratio at which the performance and efficiency of the deep learning model are optimal in a sensitivity test of the sample ratio.

[0102] In this implementation, in order to find the optimal ratio of non-concentric eyewall and concentric eyewall samples, a set of sensitivity experiments can be conducted. Based on the results on the test set, if online augmentation and offline augmentation are used simultaneously, the model accuracy is improved, and the CE class ratio that achieves a balance between performance and efficiency is taken as the optimal ratio.

[0103] In summary, the concentric eyewall recognition method for tropical cyclones provided in this embodiment generates CE samples using random rotation during the offline augmentation stage, thereby increasing the proportion of concentric eyewall samples. Then, it enhances the overall diversity of the samples through online augmentation. The online augmentation methods include random scaling, translation, and random adjustment of brightness, contrast, and saturation to solve the problems of insufficient training samples and class imbalance.

[0104] It should be noted that the above method embodiments can be implemented individually or in combination, and this application does not limit them in this regard.

[0105] Figure 10 This is a structural block diagram illustrating a concentric eyewall identification device for tropical cyclones according to an exemplary embodiment. The device includes:

[0106] The target image search module 1001 is used to search for target microwave brightness temperature images containing tropical cyclones from a historical microwave brightness temperature image set, wherein the historical microwave brightness temperature image set includes microwave brightness temperature images obtained by polar-orbiting satellite scanning.

[0107] The sample annotation module 1002 is used to annotate the target microwave brightness temperature image to obtain sample annotation data, which includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls.

[0108] Model building module 1003 is used to build deep learning models based on transfer learning;

[0109] Model training module 1004 is used to train the deep learning model using the sample labeled data to obtain a trained deep learning model.

[0110] The image recognition module 1005 is used to input the microwave brightness temperature image containing the tropical cyclone to be identified into the trained deep learning model and output the concentric eyewall recognition result of the tropical cyclone.

[0111] In one possible implementation, the model building module 1003 is used for:

[0112] The deep learning model was constructed using Swin-L as the backbone network.

[0113] The deep learning model is pre-trained using a dataset containing targets of preset shapes to be identified, to obtain a deep learning model based on transfer learning. The preset shapes include: ring-shaped and spherical.

[0114] In one possible implementation, the model training module 1004 is used for:

[0115] The hyperparameters of the deep learning model after transfer learning are fine-tuned to obtain multiple fine-tuned models;

[0116] The sample labeled data is used to train multiple fine-tuned models to obtain multiple fine-tuned models after training.

[0117] The model weights of multiple fine-tuned models after training are averaged to obtain the final trained deep learning model.

[0118] In one possible implementation, the target image search module 1001 is used for:

[0119] Acquire the scanning information of the polar-orbiting satellite and the trajectory information of the tropical cyclone;

[0120] Based on the overlapping areas in the scanning information and the trajectory information, the target microwave brightness temperature image is determined from the historical microwave brightness temperature image set.

[0121] In one possible implementation, the sample annotation module 1002 is used for:

[0122] The target microwave brightness temperature image that meets the criteria for judging concentric eyewalls is labeled as a positive sample of the concentric eyewall class;

[0123] The target microwave brightness temperature image that does not meet the concentric eyewall judgment criteria is labeled as a negative sample of the non-concentric eyewall class;

[0124] The criteria for determining the concentric eyewall include: the outer convection ring of the tropical cyclone in the image occupies at least 2 / 3 of the circle, and there is a descending airflow zone between the outer and inner convection rings of the tropical cyclone in the image.

[0125] In one possible implementation, the apparatus further includes: a sample supplementation module; the sample supplementation module is configured to:

[0126] During the offline phase, positive samples labeled as concentric eyewalls are used to generate supplementary positive samples through random rotation.

[0127] During the online phase, for positive samples labeled as concentric eye wall type and negative samples labeled as non-concentric eye wall type, supplementary positive samples and supplementary negative samples are generated using online adjustment methods. The online adjustment methods include at least one of the following: random scaling, random translation, random brightness adjustment, random contrast adjustment, and random saturation adjustment.

[0128] The supplementary positive samples and supplementary negative samples are added to the sample labeling data.

[0129] In one possible implementation, the sample supplementation module is used for:

[0130] Set the ratio of positive to negative samples in the supplemented sample annotation data as the target sample ratio;

[0131] The target sample ratio is the sample ratio at which the performance and efficiency of the deep learning model are optimal in a sensitivity test on the sample ratio.

[0132] In one possible implementation, the apparatus further includes: a sample partitioning module; the sample partitioning module is configured to:

[0133] The sample labeled data is divided into training set, validation set and test set using the random hold-out method.

[0134] It should be noted that the concentric eyewall identification device for tropical cyclones provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0135] Please see Figure 11 This is a schematic diagram of a computer device provided according to an exemplary embodiment of the present application. The computer device includes a memory and a processor. The memory is used to store a computer program. When the computer program is executed by the processor, it implements the above-described method for identifying concentric eyewalls of tropical cyclones.

[0136] The processor can be a central processing unit (CPU). It can also be other general-purpose processors, graphics processing units (GPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations thereof.

[0137] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of this invention. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the methods described in the above embodiments.

[0138] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0139] In one exemplary embodiment, a computer-readable storage medium is also provided for storing at least one computer program, which is loaded and executed by a processor to implement all or part of the steps in the above-described method. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage device, etc.

[0140] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0141] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for identifying concentric eyewalls of tropical cyclones, characterized in that, The method includes: Finding target microwave brightness temperature images containing tropical cyclones from a historical microwave brightness temperature image set, wherein finding target microwave brightness temperature images containing tropical cyclones from a historical microwave brightness temperature image set includes: Acquire scanning information from polar-orbiting satellites and trajectory information from tropical cyclones; Based on the overlapping regions in the scanning information and the trajectory information, the target microwave brightness temperature image is determined from the historical microwave brightness temperature image set, wherein determining the target microwave brightness temperature image from the historical microwave brightness temperature image set based on the overlapping regions in the scanning information and the trajectory information includes: Search for the overlapping area between the scanning information of the polar-orbiting satellite and the trajectory information of the tropical cyclone. If there is an overlap, extract a two-dimensional microwave brightness temperature data block centered on the eye of the cyclone with a longitudinal and latitudinal coverage of 8° from the microwave brightness temperature data of the satellite's vertical polarization. Fill missing values ​​in a two-dimensional data block; The natural neighborhood method is used to perform uniform grid interpolation on the two-dimensional microwave brightness temperature data block after missing values ​​are filled. RGB images are plotted using two-dimensional microwave brightness temperature data blocks interpolated with a uniform grid to determine the target microwave brightness temperature image. The historical microwave brightness temperature image set includes microwave brightness temperature images obtained by polar-orbiting satellite scanning. The target microwave brightness temperature image is labeled with samples to obtain sample labeling data, which includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls. Construct deep learning models based on transfer learning; The deep learning model is trained using the labeled sample data to obtain a trained deep learning model. Input the microwave brightness temperature image containing the tropical cyclone to be identified into the trained deep learning model, and output the concentric eyewall recognition result of the tropical cyclone.

2. The method according to claim 1, characterized in that, The construction of a deep learning model based on transfer learning includes: The deep learning model was constructed using Swin-L as the backbone network. The deep learning model is pre-trained using a dataset containing targets of preset shapes to be identified, to obtain a deep learning model based on transfer learning. The preset shapes include: ring-shaped and spherical.

3. The method according to claim 2, characterized in that, The step of training the deep learning model using the labeled sample data to obtain a trained deep learning model includes: The hyperparameters of the deep learning model after transfer learning are fine-tuned to obtain multiple fine-tuned models; The sample labeled data is used to train multiple fine-tuned models to obtain multiple fine-tuned models after training. The model weights of multiple fine-tuned models after training are averaged to obtain the final trained deep learning model.

4. The method according to claim 1, characterized in that, The target microwave brightness temperature image is sample-annotated to obtain sample annotation data. The sample annotation data includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls, including: The target microwave brightness temperature image that meets the criteria for judging concentric eyewalls is labeled as a positive sample of the concentric eyewall class; The target microwave brightness temperature image that does not meet the concentric eyewall judgment criteria is labeled as a negative sample of the non-concentric eyewall class; The criteria for determining the concentric eyewall include: the outer convection ring of the tropical cyclone in the image occupies at least 2 / 3 of the circle, and there is a descending airflow zone between the outer and inner convection rings of the tropical cyclone in the image.

5. The method according to claim 1, characterized in that, The method further includes: During the offline phase, positive samples labeled as concentric eyewalls are used to generate supplementary positive samples through random rotation. During the online phase, for positive samples labeled as concentric eye wall type and negative samples labeled as non-concentric eye wall type, supplementary positive samples and supplementary negative samples are generated using online adjustment methods. The online adjustment methods include at least one of the following: random scaling, random translation, random brightness adjustment, random contrast adjustment, and random saturation adjustment. The supplementary positive samples and supplementary negative samples are added to the sample labeling data.

6. The method according to claim 5, characterized in that, The method further includes: Set the ratio of positive to negative samples in the supplemented sample annotation data as the target sample ratio; The target sample ratio is the sample ratio at which the performance and efficiency of the deep learning model are optimal in a sensitivity test on the sample ratio.

7. The method according to claim 1, characterized in that, The method further includes: The sample labeled data is divided into training set, validation set and test set using the random hold-out method.

8. A concentric eyewall identification device for tropical cyclones, characterized in that, The apparatus is used to implement the method as described in any one of claims 1-7, the apparatus comprising: The target image search module is used to search for target microwave brightness temperature images containing tropical cyclones from a historical microwave brightness temperature image set, which includes microwave brightness temperature images obtained by polar-orbiting satellite scanning. The sample annotation module is used to annotate the target microwave brightness temperature image to obtain sample annotation data, which includes positive samples of concentric eyewalls and negative samples of non-concentric eyewalls. The model building module is used to build deep learning models based on transfer learning. The model training module is used to train the deep learning model using the sample labeled data to obtain a trained deep learning model. The image recognition module is used to input microwave brightness temperature images containing tropical cyclones into a trained deep learning model and output the concentric eyewall recognition results of the tropical cyclones.

9. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set, or instruction set, the at least one instruction, at least one program, code set, or instruction set being loaded and executed by the processor to implement the concentric eyewall identification method for tropical cyclones as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the concentric eyewall identification method for tropical cyclones as described in any one of claims 1 to 7.

Images