Method and system for identifying two-dimensional structure of tropical external cyclone

Abstract

The invention discloses a tropical external cyclone two-dimensional structure identification method and system, and relates to the technical field of weather system automatic identification and deep learning. The invention discloses an automatic cyclone two-dimensional structure identification method based on a DETR model. The method comprises the steps that data are downloaded, and a cyclone training data set is made; uploading the data set to an MXMACA operation platform, and training a DETR model network to obtain a recognition cyclone model weight; and predicting the two-dimensional structure of the cyclone by using the trained model. According to the invention, automatic identification of the cyclones can be realized, subjectivity of manual analysis of the cyclones is reduced to a certain extent, and a certain reference is provided for a predictor in weather forecast business work.

Description

Technical Field

[0001] The present invention relates to the field of weather system automatic identification and deep learning technology, and in particular to a method and system for identifying the two-dimensional structure of an extratropical cyclone. Background Art

[0002] Extratropical cyclones are important weather systems that affect mid- and high-latitude weather and climate change, and are the main carriers of water vapor and energy transport. Cyclone systems are usually accompanied by regional climate anomalies, severe storms, heavy precipitation and other disaster events, and are closely related to social economy. The identification of cyclones, especially the horizontal impact range, is an important part of weather forecasting. Currently, manual analysis is the main method in weather forecasting, which is time-consuming and laborious. Automatic identification of weather systems has become an important trend. Automatic identification of cyclones is conducive to quickly obtaining information on the development and evolution of cyclones, providing strong support for meteorological forecasting and disaster response.

[0003] Early cyclone system identification mainly relied on manual analysis of weather maps, but it was gradually eliminated due to subjective and time-consuming factors. With the continuous development of computing resources and the support of various meteorological data, a variety of cyclone automatic identification and tracking methods have been proposed one after another. At present, most of the mainstream cyclone identification algorithms are based on the Lagrangian viewpoint, and sea level pressure, low-altitude potential height or vorticity are selected as basic indicators to identify cyclones. However, traditional algorithms often rely on the positioning of the center point of the cyclone. When the data resolution is too high and the center point cannot be accurately retrieved, it is not conducive to the identification of the cyclone system. With the development of deep learning methods in the field of meteorology, through learning operations and feature extraction of a large amount of data, it is possible to directly identify weather systems from pictures, and better depict the shape and movement of weather systems. The method based on deep learning can more accurately depict the two-dimensional structural characteristics of cyclones through the refined analysis of two-dimensional image features. Therefore, it is necessary to propose a cyclone deep learning identification method that can use domestic computing chips for calculation and accurately depict the shape, improve the accuracy of cyclone automatic analysis, and realize localization in hardware and platform. Summary of the invention

[0004] As we can see from the background, manual identification of cyclones is often not objective and time-consuming. Traditional algorithms rely more on the location of the cyclone center point. When the center point cannot be accurately retrieved, the cyclone system will be missed. Using deep learning methods to perform image segmentation to identify the location and outline of the cyclone system can improve the accuracy of weather system recognition. By using the domestic deep learning platform Muxi MXMACA computing platform and the domestic deep learning computing chip Muxi GPU, the localization of hardware and platform can be achieved in cyclone recognition.

[0005] To achieve the above objective, the present invention provides a method for identifying a two-dimensional structure of an extratropical cyclone, characterized in that the steps include:

[0006] S1. Acquire 850 hPa geopotential height data, sea level pressure data and cyclone data, wherein the cyclone data includes the location and shape information of extratropical cyclones;

[0007] S2. The obtained cyclone data and 850 hPa potential height data are interpolated into a 147×147 equal-area grid with a grid resolution of 100 km×100 km; the cyclone data are used as segmentation instances to generate training labels one by one; the sea level pressure is plotted as a grayscale image, and all are made into a cyclone data set in a coco-like format;

[0008] S3. Upload the cyclone data set to the Muxi MXMACA computing platform, create a mirror image for DETR model training, use the cyclone data set to train the DETR model network, and obtain the DETR cyclone identification model weight;

[0009] S4. Download the sea level pressure data, make the sea level pressure field at different times into a grayscale map, put it into the DETR model network, and use the DETR cyclone identification model weights to make predictions to obtain automatically identified cyclone shape data.

[0010] Preferably, the obtained 850 hPa geopotential height and sea level pressure data have a resolution of 0.25°×0.25° and a time interval of 6 hours.

[0011] Preferably, in step S2, the cyclone data is randomly divided into a training set and a validation set, and the ratio of the training set to the validation set is 8:2; the 850 hPa potential height element value is proportionally scaled to the range of 0-255, and a grayscale image is output to produce corresponding training images and validation images; the 147×147 grid is supplemented and expanded to 192×192 pixels, and 0 is filled in the surrounding area.

[0012] Preferably, in step S3, the model backbone uses resnet50, the number of categories is 1+1, the picture length is 192, the width is 192, the number of GPUs is 4, the number of pictures processed at a single time is 4, and training is performed for 200 rounds; wherein, the training weights are saved every 10 rounds, and verification judgments are added during the training process.

[0013] The present invention also provides a system for identifying the two-dimensional structure of an extratropical cyclone, the system being used to implement the above method, comprising: a collection module, a processing module, a training module and an identification module;

[0014] The acquisition module is used to obtain 850 hPa potential height data, sea level pressure data and cyclone data, and the cyclone data includes the location and shape information of extratropical cyclones;

[0015] The processing module is used to interpolate the acquired cyclone data and 850 hPa potential height data into a 147×147 equal-area grid with a grid resolution of 100 km×100 km; the cyclone data are used as segmentation instances to generate training labels one by one; the sea level pressure is drawn as a grayscale image, and all are made into a cyclone data set in a coco-like format;

[0016] The training module is used to upload the cyclone data set to the Muxi MXMACA computing platform, and to create a mirror image for DETR model training, and to train the DETR model network using the cyclone data set to obtain the DETR cyclone identification model weight;

[0017] The identification module is used to download sea level pressure data, make the sea level pressure field at different times into grayscale images, put them into the DETR model network, and use the DETR cyclone identification model weights to make predictions to obtain automatically identified cyclone shape data.

[0018] Preferably, the obtained 850 hPa geopotential height and sea level pressure data have a resolution of 0.25°×0.25° and a time interval of 6 hours.

[0019] Preferably, the workflow of the processing module includes: randomly dividing the cyclone data into a training set and a validation set, with the ratio of the training set to the validation set being 8:2; scaling the 850 hPa potential height element value proportionally to the range of 0-255, and outputting a grayscale image to produce corresponding training images and validation images; supplementing and expanding the 147×147 grid to 192×192 pixels, and filling the surrounding area with 0.

[0020] Preferably, the model backbone adopted by the training module is resnet50, the number of categories is 1+1, the picture length is 192, the width is 192, the number of GPUs is 4, the number of pictures processed at a single time is 4, and the training is performed for 200 rounds; wherein the training weights are saved once every 10 rounds, and verification judgments are added during the training process.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The present invention discloses a method for automatically identifying the two-dimensional structure of a cyclone based on a DETR model, and the method comprises a plurality of steps: downloading data and making a cyclone training data set; uploading the data set to an MXMACA computing platform and training a DETR model network to obtain the weight of the cyclone identification model; and using the trained model to predict the two-dimensional structure of the cyclone. Through the present invention, the automatic identification of cyclones can be realized, and the subjectivity of manual analysis of cyclones can be reduced to a certain extent, providing a certain reference for forecasters in their weather forecasting work.

[0023] The model training calculation of the present invention uses domestic chips, and the training and prediction can all be completed on the domestic platform MXMACA. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0025] Figure 1 The overall flow diagram of a cyclone identification method based on the DETR model using the Muxi MXMACA computing platform provided by the present invention;

[0026] Figure 2 To select the 850 hPa geopotential height field and cyclone training set labels at 00:00 on January 3, 2000;

[0027] Figure 3 The grayscale image generated for the 850 hPa geopotential height field at 00:00 on January 3, 2000 is the training image;

[0028] Figure 4 This is a schematic diagram of extracting cyclone labels one by one at 00:00 on January 3, 2000 and converting them into JSON format annotations;

[0029] Figure 5 This is the final result image of the cyclone predicted by the DETR model on the sea level pressure field at 00:00 on January 3, 2000.

[0030] Figure 6 This is the final result image of the cyclone predicted by the untrained DETR model on the sea level pressure field at 00:00 on August 27, 2023. DETAILED DESCRIPTION

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

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Embodiment 1

[0034] This embodiment provides a method for identifying the two-dimensional structure of an extratropical cyclone, and its process diagram is as follows: Figure 1 As shown, the specific steps include:

[0035] S1. Obtain 850 hPa geopotential height data and cyclone data, including the location and shape information of extratropical cyclones.

[0036] Download the 850 hPa potential height and sea level pressure data from the ERA-5 hourly reanalysis data set, with a resolution of 0.25°×0.25° and a time interval of 6 hours. Randomly select 2 years of data as the training set, and use the traditional cyclone identification algorithm to generate a cyclone label set. The number of training hours selected in this embodiment is not fixed, but it cannot be too small, and the selected area can also be dynamically adjusted according to the study area. Since the object identified in this embodiment is the extratropical cyclone in the northern hemisphere (north of 20°N), it is chosen to interpolate the above sea level pressure and cyclone data into a 147×147 equal-area grid with a grid resolution of approximately 100 km×100 km. If you need to focus on a specific area, you can adjust the actual geographical scope appropriately.

[0037] S2. The acquired cyclone data and 850 hPa potential height data are interpolated into 147×147 equal-area grids with a grid resolution of 100 km×100 km. The cyclone data are used as segmentation instances to generate training labels one by one, and the sea level pressure is plotted as a grayscale image. All of them are made into cyclone datasets in a coco-like format.

[0038] According to the cyclone data obtained in S1, the random function in Python is used to randomly divide the selected 2-year data into a training set and a validation set, with a ratio of 8:2 (there is no specific requirement for the ratio, but the training set must be large in number). Python is used to scale the 850 hPa geopotential height element values ​​proportionally to the range of 0-255, and output the grayscale image to produce the corresponding training image and validation image. In order to facilitate training and make the image include the entire extratropical area of ​​the Northern Hemisphere, the 147×147 grid is supplemented and expanded to 192×192 pixels, and the surrounding area is filled with 0. This process does not affect the recognition effect (the research scope can be adjusted according to actual needs, so the pixel size will also change according to the size of the research area). The cyclone label data in S1 are taken as independent segmentation instances one by one, and black and white binary images are drawn respectively (the cyclone area is white, and other areas are black). The pycococreatortools function tool is used to write annotation files in json format (the corresponding library of pycococreatortools can be obtained as public resources on the Github website), that is, the position information of the pixels in the cyclone area in the entire image is marked (including the horizontal coordinate x and the vertical coordinate y), and together with the training images, they form a cyclone training and verification data set in a coco-like format.

[0039] S3. Upload the cyclone dataset to the Muxi MXMACA computing platform, and create a mirror image for DETR model training. Use the cyclone dataset to train the DETR model network and obtain the DETR cyclone identification model weights.

[0040] Copy and upload the prepared cyclone dataset to the Muxi MXMACA computing platform; download the DETR model network training script, which can be obtained from the Github website. Prepare the training image, use Anaconda to create a virtual environment, and install various dependent libraries for running the DETR model script based on the python version, including but not limited to cython, pycocotools, submitit, panopticapi, and scipy. The version of the dependent library must be compatible with the pytorch version. After the environment is adapted, the DETR model is trained using the cyclone dataset on the Muxi MXMACA computing platform. Set appropriate parameters for the model network according to the type of dataset. The backbone of the model uses resnet50, and its related pre-trained weights are obtained from the Github website. The number of categories is 1+1 (i.e., cyclones and background), the image length is 192 and the width is 192, the number of GPUs is 4, the number of images processed at a time (batch size) is 4, and the training is 200 rounds. The training weights are saved every 10 rounds, and verification judgments are added during the training process. The weights with the highest verification scores during the training process are saved as the best weights using the coco_evaluator.coco_eval function. The DETR model needs to be trained for instance detection first and then for segmentation. The trained best instance detection weights are put back into the DETR model as frozen weights to train its mask segmentation head. The best segmentation weights are also saved to obtain the DETR cyclone segmentation model.

[0041] S4. Download the sea level pressure data, make the sea level pressure field at different times into a grayscale map, put it into the DETR model network, and use the DETR cyclone identification model weights to make predictions to obtain automatically identified cyclone shape data.

[0042] The sea level pressure field downloaded from S1 is made into a grayscale image, and the format is consistent with that in S2 (there are no specific requirements for data selection and image drawing range, which can be adjusted according to actual needs). The image is converted into the image size for DETR model operation using the torchvision.transforms function, and put into the model network. The DETR cyclone identification model weights obtained from S3 are loaded using torch.load, and the detr.eval function is used for prediction to obtain the cyclone results identified by the DETR model in array format. The cyclone results in array format are used to calculate the latitude and longitude information corresponding to each cyclone based on the relative position of the grid points in the image, and finally the cyclone results identified by the DETR model are obtained.

[0043] Embodiment 2

[0044] The feasibility and accuracy of the cyclone identification method of the present invention will be verified by combining actual cases below.

[0045] The cyclone identification method of the present invention can identify cyclones at all times when data exists, and all of them can be completed on a domestic platform. Here, we take 00:00 on January 3, 2000 as an example to illustrate the identification effect of the method. Figure 2 As shown, at this time, a total of 15 cyclones and their horizontal ranges and structures were analyzed in the extratropical region of the Northern Hemisphere using traditional identification algorithms.

[0046] Using the cyclone identification method of the present invention, the cyclone data output by the DETR model is obtained:

[0047] 1. Create a data set. Randomly select the 850 hPa potential height data in the range north of 20°N from 2000 to 2001 ( Figure 3 ) to create feature images and use traditional recognition algorithms to obtain cyclone labels. Figure 4 As shown, 15 cyclones are extracted one by one into 15 black-and-white binary images, where the positions of the cyclones are white and other positions are black. For the complete content of the data set preparation, please refer to the specific implementation steps S1-S2.

[0048] 2. Upload the data set to the Muxi MXMACA computing platform and put it into the DETR model network training (the details of the model training have been given in the specific implementation step S3) to obtain the trained DETR cyclone identification model weights.

[0049] 3. Input the grayscale image obtained by drawing the surface sea level pressure at 00:00 on January 3, 2000 into the trained DETR cyclone recognition model to obtain the cyclone prediction segmentation result corresponding to this time. Figure 5 Compared with the cyclone distribution analyzed manually at the same time ( Figure 2 ), the cyclones identified by the DETR cyclone model are basically consistent with the training set, the number and location distribution of cyclones are the same, the cyclone morphology is also similar, and it is also consistent with the distribution of sea level pressure field.

[0050] Figure 6 The cyclone distribution information obtained by this method at 00:00 on August 27, 2023 is given. This time is randomly selected and has not been trained by the model in the case. It can be seen that the cyclones identified by the model on the sea level pressure field at 00:00 on August 27, 2023, have a good match with the distribution of the sea level pressure field. This case can illustrate that the cyclone automatic identification method proposed in the present invention can identify the individual position of cyclones more accurately, and the method has a good two-dimensional shape feature description capability for cyclones.

[0051] In terms of efficiency, the DETR cyclone identification model trained in this method can obtain cyclone data for all times of the year within 5 minutes. In the case of 00:00 on August 27, 2023, the model can also accurately identify and locate cyclones for times without cyclone labels.

[0052] Embodiment 3

[0053] The present invention also provides a two-dimensional structure recognition system for an extratropical cyclone, comprising: a collection module, a processing module, a training module and a recognition module. The acquisition module is used to obtain 850 hPa potential height data, sea level pressure data and cyclone data. The cyclone data includes the location and shape information of extratropical cyclones. The processing module is used to interpolate the acquired cyclone data and 850 hPa potential height data into 147×147 equal-area grids with a grid resolution of 100 km×100 km. The cyclone data are used as segmentation instances to generate training labels one by one. The sea level pressure is plotted as a grayscale image, and all are made into cyclone data sets in a coco-like format. The training module is used to upload the cyclone data set to the Muxi MXMACA computing platform, and to make an image for DETR model training. The cyclone data set is used to train the DETR model network and obtain the DETR cyclone recognition model weights. The recognition module is used to download sea level pressure data, make the sea level pressure field at different times into grayscale images, put them into the DETR model network, and use the implementation of the DETR cyclone recognition model weights for prediction to obtain automatically identified cyclone shape data.

[0054] The following will describe in detail how the present invention solves the technical problems in practical work in conjunction with this embodiment.

[0055] Firstly, the acquisition module is used to obtain 850 hPa potential height data and cyclone data, which include the location and shape information of extratropical cyclones.

[0056] Download the 850 hPa potential height and sea level pressure data from the ERA-5 hourly reanalysis data set, with a resolution of 0.25°×0.25° and a time interval of 6 hours. Randomly select 2 years of data as the training set, and use the traditional cyclone identification algorithm to generate a cyclone label set. The number of training hours selected in this embodiment is not fixed, but it cannot be too small, and the selected area can also be dynamically adjusted according to the study area. Since the object identified in this embodiment is the extratropical cyclone in the northern hemisphere (north of 20°N), it is chosen to interpolate the above sea level pressure and cyclone data into a 147×147 equal-area grid with a grid resolution of approximately 100 km×100 km. If you need to focus on a specific area, you can adjust the actual geographical scope appropriately.

[0057] The processing module then interpolates the acquired cyclone data and 850 hPa potential height data into a 147×147 equal-area grid with a grid resolution of 100 km×100 km; the cyclone data are used as segmentation instances to generate training labels one by one, and the sea level pressure is plotted as a grayscale image, all of which are made into a cyclone dataset in a coco-like format.

[0058] According to the cyclone data obtained by the acquisition module, the random function in Python is used to randomly divide the selected 2-year data into a training set and a validation set, with a ratio of 8:2 (there is no specific requirement for the ratio, but the training set must be large in number). Python is used to scale the 850 hPa geopotential height element values ​​proportionally to the 0-255 range, and output the grayscale image to produce the corresponding training image and validation image. In order to facilitate training and make the image include the entire extratropical area of ​​the Northern Hemisphere, the 147×147 grid is supplemented and expanded to 192×192 pixels, and the surrounding areas are filled with 0. This process does not affect the recognition effect (the research scope can be adjusted according to actual needs, so the pixel size will also change according to the size of the research area). The cyclone label data in S1 are taken as independent segmentation instances one by one, and black and white binary images are drawn respectively (the cyclone area is white, and other areas are black). The pycococreatortools function tool is used to write annotation files in json format (the corresponding library of pycococreatortools can be obtained as public resources on the Github website), that is, the position information of the pixels in the cyclone area in the entire image is marked (including the horizontal coordinate x and the vertical coordinate y), and together with the training images, they form a cyclone training and verification data set in a coco-like format.

[0059] The cyclone data set is uploaded to the training module. In this embodiment, the training module adopts the Muxi MXMACA computing platform and creates a mirror for DETR model training. The DETR model network is trained using the cyclone data set to obtain the DETR cyclone recognition model weights.

[0060] Copy and upload the prepared cyclone dataset to the training module; download the DETR model network training script, which can be obtained from the Github website. Prepare the training image, use Anaconda to create a virtual environment, and install various dependent libraries for running the DETR model script based on the python version, including but not limited to cython, pycocotools, submitit, panopticapi, and scipy. The version of the dependent library must be compatible with the pytorch version. After the environment is adapted, the DETR model is trained using the cyclone dataset on the Muxi MXMACA computing platform. Set appropriate parameters for the model network according to the type of dataset. The model backbone uses resnet50, and its related pre-trained weights are obtained from the Github website. The number of categories is 1+1 (i.e., cyclones and background), the image length is 192 and the width is 192, the number of GPUs is 4, the number of images processed at a time (batch size) is 4, and the training is 200 rounds. The training weights are saved every 10 rounds, and verification judgments are added during the training process. The weights with the highest verification scores during the training process are saved as the best weights using the coco_evaluator.coco_eval function. The DETR model needs to be trained for instance detection first and then for segmentation. The trained best instance detection weights are put back into the DETR model as frozen weights to train its mask segmentation head. The best segmentation weights are also saved to obtain the DETR cyclone segmentation model.

[0061] Finally, the recognition module downloads the sea level pressure data, makes the sea level pressure field at different times into a grayscale image, puts it into the DETR model network, and uses the DETR cyclone recognition model weights for prediction to obtain automatically recognized cyclone shape data.

[0062] The sea level pressure field downloaded by the acquisition module is made into a grayscale image, and the format is consistent with that in the processing module (there are no specific requirements for data selection and image drawing range, which can be adjusted according to actual needs). The image is converted into the image size for DETR model operation using the torchvision.transforms function, and put into the model network. The DETR cyclone identification model weights obtained by the training module are loaded using torch.load, and the detr.eval function is used for prediction to obtain the cyclone results identified by the DETR model in array format. The cyclone results in array format are used to calculate the latitude and longitude information corresponding to each cyclone based on the relative position of the grid points in the image, and finally the cyclone results identified by the DETR model are obtained.

[0063] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for identifying the two-dimensional structure of an extratropical cyclone, characterized in that the steps include: S1. Acquire 850 hPa geopotential height data, sea level pressure data and cyclone data, wherein the cyclone data includes the location and shape information of extratropical cyclones; S2. The obtained cyclone data and 850 hPa potential height data are interpolated into a 147×147 equal-area grid with a grid resolution of 100 km×100 km; the cyclone data are used as segmentation instances to generate training labels one by one; the sea level pressure is plotted as a grayscale image, and all are made into a cyclone data set in a coco-like format; S3. Upload the cyclone data set to the Muxi MXMACA computing platform, create a mirror image for DETR model training, use the cyclone data set to train the DETR model network, and obtain the DETR cyclone identification model weight; S4. Download the sea level pressure data, make the sea level pressure field at different times into a grayscale map, put it into the DETR model network, and use the DETR cyclone identification model weights to make predictions to obtain automatically identified cyclone shape data.

2. The method for identifying the two-dimensional structure of an extratropical cyclone according to claim 1, characterized in that: The obtained 850 hPa geopotential height and sea level pressure data have a resolution of 0.25°×0.25° and a time interval of 6 hours.

3. The method for identifying the two-dimensional structure of an extratropical cyclone according to claim 1, characterized in that: In step S2, the cyclone data are randomly divided into a training set and a validation set, and the ratio of the training set to the validation set is 8:2; the 850 hPa potential height element value is proportionally scaled to the range of 0-255, and a grayscale image is output to produce corresponding training images and validation images; the 147×147 grid is supplemented and expanded to 192×192 pixels, and zeros are filled in the surrounding area.

4. The method for identifying the two-dimensional structure of an extratropical cyclone according to claim 1, characterized in that: In step S3, the model backbone uses resnet50, the number of categories is 1+1, the picture length is 192, the width is 192, the number of GPUs is 4, the number of pictures processed at a time is 4, and the training is performed for 200 rounds; the training weights are saved every 10 rounds, and verification judgments are added during the training process.

5. A system for identifying the two-dimensional structure of an extratropical cyclone, the system being used to implement the method according to any one of claims 1 to 4, characterized in that: include: Acquisition module, processing module, training module and recognition module; The acquisition module is used to obtain 850 hPa potential height data, sea level pressure data and cyclone data, and the cyclone data includes the location and shape information of extratropical cyclones; The processing module is used to interpolate the acquired cyclone data and 850 hPa potential height data into a 147×147 equal-area grid with a grid resolution of 100 km×100 km; the cyclone data are used as segmentation instances to generate training labels one by one; the sea level pressure is drawn as a grayscale image, and all are made into a cyclone data set in a coco-like format; The training module is used to upload the cyclone data set to the Muxi MXMACA computing platform, and to create a mirror image for DETR model training, and to train the DETR model network using the cyclone data set to obtain the DETR cyclone identification model weight; The identification module is used to download sea level pressure data, make the sea level pressure field at different times into grayscale images, put them into the DETR model network, and use the DETR cyclone identification model weights to make predictions to obtain automatically identified cyclone shape data.

6. The system for identifying the two-dimensional structure of an extratropical cyclone according to claim 5, characterized in that: The obtained 850 hPa geopotential height and sea level pressure data have a resolution of 0.25°×0.25° and a time interval of 6 hours.

7. The system for identifying the two-dimensional structure of an extratropical cyclone according to claim 5, characterized in that: The workflow of the processing module includes: randomly dividing the cyclone data into a training set and a validation set, with the ratio of the training set to the validation set being 8:2; scaling the 850 hPa potential height element value to the range of 0-255, and outputting a grayscale image to produce corresponding training images and validation images; supplementing and expanding the 147×147 grid to 192×192 pixels, and filling the surrounding area with 0.

8. The system for identifying the two-dimensional structure of an extratropical cyclone according to claim 5, characterized in that: The model backbone used in the training module is resnet50, the number of categories is 1+1, the picture length is 192, the width is 192, the number of GPUs is 4, the number of pictures processed at a single time is 4, and the training is 200 rounds; among them, the training weight is saved once every 10 rounds, and verification judgment is added during the training process.

Images