A deep learning-based automatic identification method for kunming quasi-stationary front

Abstract

The present application relates to the technical field of front surface path tracking, in particular to a method for automatically identifying Kunming quasi-stationary front based on deep learning, comprising: obtaining historical position data and meteorological data of Kunming quasi-stationary front; obtaining a Kunming quasi-stationary front label set by using the historical position data and meteorological data of Kunming quasi-stationary front; obtaining an RGB image data set by using the meteorological data; inputting the Kunming quasi-stationary front label set and the corresponding RGB image data set of the same time into a DETR model for training to obtain a trained DETR model; inputting an RGB image of any time period into the trained DETR model to output an identification result of Kunming quasi-stationary front of the any time period. The present application simplifies the process of automatically identifying Kunming quasi-stationary front, saves business time, realizes automation of Kunming quasi-stationary front analysis in business forecast, and improves the efficiency and accuracy of automatic identification of Kunming quasi-stationary front in meteorological research work.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of front surface path tracking, in particular to a Kunming quasi-stationary front automatic identification method based on deep learning. BACKGROUND

[0002] The Kunming quasi-stationary front is also called the southwest quasi-stationary front, is an important weather system, and often causes large-scale freezing rain and snow (in winter and spring) in the east region of the front surface and heavy rain, strong convective weather (in summer and autumn) near the front surface. At present, the identification of the Kunming quasi-stationary front is mainly artificial analysis. Artificial analysis can refer to various meteorological elements for comprehensive analysis, but requires higher requirements for forecasters, and a large amount of time is needed for long-term data analysis. Although the artificial analysis method is very accurate for a single front surface system, if the weather system is complex, the subjectively identified front surface may have some deviations, and the experience and understanding of the weather system of different forecasters will also cause some differences in the front surface position identification.

[0003] In recent years, the application of deep learning (DL) method in meteorology has made great progress, and a large amount of meteorological data is conducive to improving the training effect and model performance of DL. Scholars have realized the identification of cold front and warm front by using a deep learning model and achieved good results, but the identification of the Kunming quasi-stationary front is still lacking, and there is no long-term stationary front data set. Therefore, a Kunming quasi-stationary front automatic identification method based on deep learning is needed, which can simplify the process of automatically identifying the Kunming quasi-stationary front, save valuable business time, provide a reference for forecast work, and make a positive contribution to the automation of the Kunming quasi-stationary front in business forecasting. SUMMARY

[0004] The purpose of the present application is to provide a Kunming quasi-stationary front automatic identification method based on deep learning, which simplifies the process of automatically identifying the Kunming quasi-stationary front, saves business time, and realizes the automation of the Kunming quasi-stationary front analysis in business forecasting.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] A Kunming quasi-stationary front automatic identification method based on deep learning comprises:

[0007] Obtaining Kunming quasi-stationary front historical position data and meteorological data;

[0008] Obtaining a Kunming quasi-stationary front label set by using the Kunming quasi-stationary front historical position data and the meteorological data;

[0009] Obtaining an RGB image data set by using the meteorological data;

[0010] inputting the Kunming quasi-stationary front label set and the RGB image data set corresponding to the time into a DETR model for training, to obtain a trained DETR model;

[0011] inputting an arbitrary period RGB image into the trained DETR model, and outputting an identification result of the Kunming quasi-stationary front in the arbitrary period.

[0012] Optionally, obtaining the Kunming quasi-stationary front label set comprises:

[0013] According to the historical position data of the Kunming quasi-stationary front, a position image of the Kunming quasi-stationary front is obtained through a meteorological information comprehensive analysis processing system (MICAPS), and longitude and latitude data of the Kunming quasi-stationary front are calculated.

[0014] According to the meteorological data, a meteorological element image is obtained, and the longitude and latitude data of the Kunming quasi-stationary front are plotted on a meteorological element image corresponding to the time, to obtain a meteorological element distribution and a front line position map.

[0015] According to the meteorological element distribution and the front line position map, the position image of the Kunming quasi-stationary front is revised, to obtain a revised Kunming quasi-stationary front position image.

[0016] The revised Kunming quasi-stationary front position image is preprocessed, to obtain the Kunming quasi-stationary front label set.

[0017] Optionally, preprocessing the revised Kunming quasi-stationary front position image comprises:

[0018] The Kunming quasi-stationary front line in the revised Kunming quasi-stationary front position image is extracted hour by hour, to obtain a Kunming quasi-stationary front position image corresponding to a plurality of times, and the range, pixel size and color of the Kunming quasi-stationary front position image corresponding to the plurality of times are preset.

[0019] Optionally, calculating the longitude and latitude data of the Kunming quasi-stationary front comprises: according to the drawing range and the relative position of the front line grid point of the Kunming quasi-stationary front position image, the longitude and latitude data of the Kunming quasi-stationary front are calculated.

[0020] Optionally, obtaining the RGB image data set by using the meteorological data comprises:

[0021] The meteorological data is subjected to Gaussian smoothing processing, the smoothed meteorological data is subjected to gray scale processing, and the gray scale values are respectively taken as pixel values in R, G and B channels of an RGB picture, to obtain the RGB image data set.

[0022] Optionally, the meteorological data comprises wind field, sea level pressure and temperature data.

[0023] Optionally, inputting the Kunming quasi-stationary front label set and the RGB image data set corresponding to the time into the DETR model for training comprises:

[0024] Inputting the Kunming quasi-stationary front label set and the RGB image data set corresponding to the time into the DETR model for training, and obtaining the optimal weight according to the loss function value;

[0025] Putting the optimal weight into the DETR model as a frozen weight, iteratively training the mask segmentation part, and obtaining the trained DETR model.

[0026] Optionally, the loss function value comprises loss calculation of the target detection part and the target segmentation part, wherein the loss function of the target detection part is calculated according to the Hungarian algorithm, and the target segmentation part is calculated according to the binary cross-entropy loss function.

[0027] The beneficial effects of the present application are: the present application constructs the label data required by the deep learning model based on the front position data, ERA5 reanalysis data and Kunming quasi-stationary front analysis method, draws the RGB image of the meteorological element data set concerned by the front recognition, and then identifies the Kunming quasi-stationary front from the image through training the DETR model, which simplifies the process of automatically identifying the Kunming quasi-stationary front, saves business time, realizes the automation of the Kunming quasi-stationary front analysis in the business forecast, and also improves the efficiency and accuracy of the automatic identification of the Kunming quasi-stationary front in the meteorological research work. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0029] Figure 1 A flow chart of a Kunming quasi-stationary front automatic identification method based on deep learning according to an embodiment of the present application;

[0030] Figure 2 A meteorological element distribution and Kunming quasi-stationary front position map on April 19, 2012, 06 (world time) according to an embodiment of the present application;

[0031] Figure 3 A processed Kunming quasi-stationary front label indication map on February 3, 2017, 06 according to an embodiment of the present application;

[0032] Figure 4 A comparison chart of manual identification and automatic identification of Kunming quasi-stationary front on November 16, 2018, 06 for evaluating the identification effect according to an embodiment of the present application;

[0033] Figure 5 The contrast chart of manual identification and automatic identification of Kunming quasi-stationary front on March 14, 2019 at 06:00 is selected for evaluating the identification effect of the embodiment of the present application. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be apparently and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0035] In order to make the above-mentioned objects, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0036] As shown in the figure, the embodiment provides a deep learning-based automatic identification method of Kunming quasi-stationary front, comprising: Figure 1

[0037] obtaining historical position data and meteorological data of the Kunming quasi-stationary front;

[0038] obtaining a Kunming quasi-stationary front label set by using the historical position data and meteorological data of the Kunming quasi-stationary front;

[0039] obtaining an RGB image data set by using the meteorological data;

[0040] inputting the Kunming quasi-stationary front label set and the corresponding RGB image data set of the time into a DETR model for training, and obtaining a trained DETR model;

[0041] inputting an arbitrary time period RGB image into the trained DETR model, and outputting an identification result of the Kunming quasi-stationary front in the arbitrary time period.

[0042] Specifically, the label data required by the deep learning model is constructed based on the front position data in the historical weather chart, ERA5 reanalysis data and the Kunming quasi-stationary front analysis method, the RGB image required for training is drawn, and the Kunming quasi-stationary front is identified from the image by training the deep learning model DETR, thereby simplifying the process of automatically identifying the Kunming quasi-stationary front, saving business time, and realizing the automation of the Kunming quasi-stationary front analysis in the business forecast, wherein the DETR (Detection Transformer) model is an end-to-end target detection model based on the Transformer architecture.

[0043] Further, obtaining the Kunming quasi-stationary front label set comprises: ​

[0044] According to the historical position data of the Kunming quasi-stationary front, the position image of the Kunming quasi-stationary front is obtained through the meteorological information comprehensive analysis processing system MICAPS, and the longitude and latitude data of the Kunming quasi-stationary front is calculated;

[0045] According to the meteorological data, the meteorological element image is obtained, and the longitude and latitude data of the Kunming quasi-stationary front is drawn on the corresponding meteorological element image at the time, and the meteorological element distribution and the position image of the front line are obtained;

[0046] According to the meteorological element distribution and the position image of the front line, the position image of the Kunming quasi-stationary front is revised, and the revised position image of the Kunming quasi-stationary front is obtained;

[0047] The revised position image of the Kunming quasi-stationary front is preprocessed, and the Kunming quasi-stationary front label set is obtained.

[0048] Further, the preprocessing of the revised position image of the Kunming quasi-stationary front includes:

[0049] The Kunming quasi-stationary front line in the revised position image of the Kunming quasi-stationary front is extracted time by time, and the Kunming quasi-stationary front position image corresponding to several times is obtained, and the range, pixel size and color of the Kunming quasi-stationary front position image corresponding to several times are pre-set.

[0050] Further, the calculation of the longitude and latitude data of the Kunming quasi-stationary front includes: according to the drawing range and the relative position of the front line grid of the position image of the Kunming quasi-stationary front, the longitude and latitude data of the Kunming quasi-stationary front is calculated.

[0051] Specifically, the Kunming quasi-stationary front label data includes:

[0052] Step S1, obtaining the historical position data of the Kunming quasi-stationary front: collecting the position data of the Kunming quasi-stationary front in the "Kunming Quasi-stationary Front Yearbook" compiled by Yunnan Institute of Meteorological Science;

[0053] Step S2, downloading meteorological data: downloading 10-meter wind field, sea level pressure and ground 2-meter temperature data with resolution of 0.25°x0.25° from European Center for Medium-Range Weather Forecasts (ECMWF), and drawing the distribution image of the above meteorological elements;

[0054] Step S3, using the meteorological information comprehensive analysis and processing system MICAPS (Meteorological Information Comprehensive Analysis and Processing System) independently developed by China Meteorological Administration and widely used in meteorological forecast business to draw the position image of each time in the Kunming quasi-stationary front example in any long period in step S1;

[0055] Step S4, the position of the front is extracted and drawn in the weather element image in step S2, and the position of the Kunming quasi-stationary front is corrected and improved by using the Kunming quasi-stationary front analysis method (since the Kunming quasi-stationary front in the yearbook is mainly based on station data, the data resolution is low, and the resolution of the weather data is higher, the position of the Kunming quasi-stationary front is corrected), and finally the revised Kunming quasi-stationary front position image is obtained;

[0056] Step S5, the Kunming quasi-stationary front position image obtained in step S4 is preprocessed, the image is processed to be within the range of 8-40°N, 90-122°E, and the pixel is 128x128, the position of the front is extracted to generate a black and white image (the grid points of the front position are filled with white color, and the grid points of other regions are black), and the thickening treatment is performed to form a Kunming quasi-stationary front label set in png format.

[0057] Further, the RGB image data set is obtained by using the weather data, including:

[0058] The weather data is subjected to Gaussian smoothing, the smoothed weather data is subjected to gray scale processing, and the gray scale values are respectively taken as pixel values in R, G and B channels of the RGB picture, and the RGB image data set is obtained.

[0059] Specifically, the downloaded 10-meter wind field u component, sea level pressure and ground 2-meter temperature are respectively subjected to Gaussian smoothing with a standard deviation of 1, the smoothed 10-meter wind field u component, sea level pressure and ground 2-meter temperature are subjected to gray scale processing, and the corresponding values are respectively taken as pixel values in R, G and B channels of the RGB picture, and a feature image is generated. The image range is between 8-40°N, 90-122°E, and the pixel is 128x128.

[0060] The formula for gray scale processing in G and B channels is:

[0061] (1);

[0062] is the variable after gray scale processing, is the minimum value of the variable, is the maximum value of the variable, and x is the variable.

[0063] To highlight the characteristics of the difference between the east and west winds near the front, the 10-meter wind field u component in the R channel is processed as follows: the positive value of the 10m wind field u component is mapped to the intensity value of 128-255, the negative value is mapped to the intensity value of 0-127, and the 0 value is mapped to the intensity value of 127.

[0064] Further, the Kunming quasi-stationary front label set and the corresponding RGB image data set are input into the DETR model for training, including:

[0065] The Kunming quasi-stationary front label set and the corresponding RGB image data set are input into the DETR model for training, and when the loss function value decreases to the minimum value and tends to be stable, the weight of the current round is saved as the best weight.

[0066] The best weight is put into the DETR model as a frozen weight, and the mask segmentation part is iteratively trained to obtain the trained DETR model.

[0067] Further, the loss function value includes loss calculation of the target detection part and the target segmentation part, wherein the loss function of the target detection part is calculated according to the Hungarian algorithm, and the target segmentation part is calculated according to the binary cross-entropy loss function.

[0068] Specifically, the model training specifically includes:

[0069] Step S1, the RGB image data set is randomly put into the training set and the validation set according to the proportion of 3:1.

[0070] Step S2, corresponding to the file name of the divided training set and validation set images, the Kunming quasi-stationary front label set is also divided into training set and validation set, and the divided Kunming quasi-stationary front label set training set and validation set are respectively converted into json files, and the json files and the RGB image training set and validation set form a coco format data set.

[0071] Step S3, input the coco format data set into the DETR model for training. According to the type of the data set, set appropriate parameters for the model network, the picture length is 128, the width is 128, the GPU number is 3, the memory is 12GB, the learning rate is set to 1x10 -4 , the weight decay is set to 1x10 -4 , and the dropout rate is set to 0.1. The loss function of the target detection part of the DETR model is calculated based on the Hungarian algorithm, including the weighted sum of the loss of the two parts of classification and bounding box prediction, and the binary cross-entropy loss function is used for processing target segmentation. When the loss function value decreases to the minimum value and tends to be stable, the weight of this round is saved as the best weight. Then the best weight is put into the DETR model again as a frozen weight, and the mask segmentation part is trained. After 200 rounds, the model performs well on the validation set and the test set, reaching the best model, saving the segmentation best weight, and thus obtaining the DETR Kunming quasi-stationary front segmentation model.

[0072] The contents of the present embodiment will be further described in combination with the drawings:

[0073] In this embodiment, historical Kunming quasi-stationary front data set and ERA5 meteorological data are used to automatically identify Kunming quasi-stationary front by using deep learning model DETR, and the specific process is as follows:

[0074] Step S1, obtain the historical position data of Kunming quasi-stationary front. Collect the position data of Kunming quasi-stationary front recorded in the Yearbook of Kunming Quasi-Stationary Front edited by Yunnan Institute of Meteorological Science. The yearbook data contains the position image of Kunming quasi-stationary front at 06 o'clock (world time, the same below) of each day from January to April and November to December in 2011-2020. The front position range is 20-30 °N, 100-110 °E;

[0075] Step S2, download meteorological data. Download reanalysis data from 1979 to 2023 from the European Center for Medium-Range Weather Forecasts (ECMWF). The data has a horizontal resolution of 0.25°x0.25°, and the elements are 10-meter wind field, sea level pressure, and ground 2-meter temperature data;

[0076] Step S3, make Kunming quasi-stationary front label data:

[0077] Step S31, use the MICAPS (Meteorological Information Comprehensive Analysis and Processing System) developed by the China Meteorological Administration and widely used in meteorological forecast business to draw the position image of each time of all Kunming quasi-stationary front cases in step S1. One time one picture, process as black background and white front line image, and extract the latitude and longitude coordinates of the white front line for storage. Write the longitude and latitude of the front line into two text files respectively. The calculation method of longitude and latitude is mainly based on the relative position of the drawing range and the front line grid. The longitude and latitude of Kunming quasi-stationary front are calculated one by one. The specific formula for calculating longitude and latitude information is:

[0078] (2);

[0079] Where, lon and lat represent the longitude and latitude of the Kunming quasi-stationary front, respectively. leftlon, rightlon, upperlat, and lowerlat represent the minimum and maximum values of longitude and latitude in the selected data drawing range, respectively. mx and my represent the length and width of the model identification output array, respectively. ix and iy represent the column and row positions of the points in the array segmented by the model.

[0080] Step S32, use the data downloaded in step S2 to draw the spatial distribution image of 10-meter wind field, sea level pressure, and ground 2-meter temperature of each day at 06 o'clock from 2011 to 2018, and draw the front line latitude and longitude data in step S31 in the corresponding image. That is, the distribution of meteorological elements and the position of the front line at each time are obtained.

[0081] Step S33, the Kunming quasi-stationary front surface analysis method (referring to the subjective front surface analysis method of Xu Meiling et al. (2011) and the objective identification method of Kunming quasi-stationary front proposed by Duan Xu et al. (2017)) is used to analyze the meteorological elements and front position in the image obtained in step S32, and the position of the Kunming quasi-stationary front is corrected and improved (since the Kunming quasi-stationary front in the yearbook is mainly based on station data, the data resolution is low, and the meteorological data with higher resolution is used for correction), and finally the corrected Kunming quasi-stationary front position image of 06:00 on January-April and November-December of 2011-2020 is obtained, as shown in FIG. 2. Figure 2 FIG. 2 shows the meteorological element distribution and the Kunming quasi-stationary front position before and after correction at 06:00 on April 19, 2012, Figure 2 The filled area in the middle represents the 2-meter temperature (℃), the black contour line represents the sea level pressure (hPa), and the arrow represents the 10-meter horizontal wind field. The gray line is the Kunming quasi-stationary front position recorded in the yearbook, and the black line is the Kunming quasi-stationary front position after analysis and correction;

[0082] Step S34, the Kunming quasi-stationary front line obtained in step S33 is extracted hour by hour, one image for each hour, processed into an image with black background and white front line (the image without front line is displayed as a pure black image), and the image is processed into an image with a range of 8-40°N, 90-122°E and a pixel of 128×128. The front line is thickened to form the Kunming quasi-stationary front label set, which is in png format, as shown in FIG. 3. Figure 3 FIG. 3 shows the Kunming quasi-stationary front label at 06:00 on February 3, 2017 obtained according to the above method.

[0083] S4, making a feature image. The 10-meter wind field u component, sea level pressure, and ground 2-meter temperature at 06:00 on January-April and November-December of 2011-2020 in step S2 are respectively Gaussian smoothed with a standard deviation of 1. The smoothed 10-meter wind field u component, sea level pressure, and ground 2-meter temperature are gray processed, and their corresponding numerical values are taken as pixel values in the R, G, and B channels of the RGB picture respectively to generate a feature image, which is in jpg format. The image range is between 8-40°N, 90-122°E, and the pixel is 128×128.

[0084] To highlight the characteristics of the difference between the east and west winds near the front, the 10-meter wind field u component in the R channel is processed as follows: the positive value of the 10m wind field U component is mapped to the intensity value of 128-255, the negative value is mapped to the intensity value of 0-127, and the 0 value is mapped to the intensity value of 127.

[0085] S5, model training:

[0086] Step S51, the RGB image data set in 2011-2018 in step S4 is randomly put into the training set and the validation set according to the ratio of 3:1, and the RGB image data set in 2019-2020 is taken as the test set.

[0087] Step S52, corresponding to the file name of the training set and the validation set image divided in S51, the Kunming quasi-stationary front label set in step S34 is also divided into a training set and a validation set respectively, and the training set and the validation set of the divided Kunming quasi-stationary front label set are respectively converted into json files, and the json files and the training set and the validation set of the RGB image form a coco format data set.

[0088] Step S53, the coco format data set generated in step S52 is input into the DETR model for training. According to the type of the data set, the appropriate parameters are set for the model network, the backbone of the model uses resnet50, the related pre-training weight is the public resource obtained from the Github website, the classification number is 1+1 (i.e. Kunming quasi-stationary front 1 class + background 1 class); the training set is 1088, the validation set is 362. The picture length is 128, the width is 128, the number of GPUs is 3, the memory is 12GB, the single processing picture number (batch size) is 2, the learning rate is set to 1×10 -4 , the weight decay is set to 1×10 -4 , the dropout rate is set to 0.1, and the confidence threshold is set to 80%. The loss function of the target detection part of the DETR model is calculated based on the Hungarian algorithm, the loss function value of the target detection part is the weighted sum of the loss of the classification and the bounding box prediction, and the binary cross entropy loss function is used for processing target segmentation. When the loss function value reaches the minimum level and tends to be stable, the weight of this round is saved as the best weight. After 84 rounds of training, the training loss decreases to about 4.6 and tends to be stable, indicating that the best weight has been reached. Then the best weight is put into the DETR model again as the frozen weight, and the mask segmentation part is trained. After 200 rounds, the model performs well on the validation set and the test set, reaching the best model, and the segmentation best weight is saved, thereby obtaining the DETR Kunming quasi-stationary front segmentation model.

[0089] S6, automatic identification of the front. The obtained DETR Kunming quasi-stationary front segmentation model is loaded into the DETR model network, and the RGB image generated according to the method of step S4 in any period from 1979 to 2023 is input, and the result with a score greater than 0.8 is output. The automatic identification of the Kunming quasi-stationary front in any period is realized.

[0090] Figure 4The recognition result of DETR on the feature image at 06:00 on November 16, 2018 is shown, in which the filled color represents the 2-meter temperature (℃), the black contour line represents the sea level pressure (hPa), and the arrow represents the 10-meter horizontal wind field. The gray line is the position of the Kunming quasi-stationary front identified by manual recognition, and the black line is the position of the Kunming quasi-stationary front identified by DETR. It can be seen that the DETR model identifies the Kunming quasi-stationary front from the target boundary of the cold high pressure front, and there is an obvious wind direction shear and temperature gradient near the front. The position of the Kunming quasi-stationary front identified by DETR is basically consistent with the position of the Kunming quasi-stationary front identified by manual labeling.

[0091] Figure 5 For the recognition result of DETR Kunming quasi-stationary front at 06:00 on March 14, 2019, which is not trained, the Kunming quasi-stationary front is also identified at the target boundary. In the figure, the filled color represents the 2-meter temperature (℃), the black contour line represents the sea level pressure (hPa), and the arrow represents the 10-meter horizontal wind field. The gray line is the position of the Kunming quasi-stationary front identified by manual recognition, and the black line is the position of the Kunming quasi-stationary front identified by DETR. The identified Kunming quasi-stationary front is located near the area with large sea level pressure and temperature gradient, and there is an obvious east-west wind direction change near the Kunming quasi-stationary front. It can be seen that for the trained and untrained time, the DETR model can directly identify the Kunming quasi-stationary front line from the image, and has good matching with various meteorological elements.

[0092] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for automatic identification of Kunming quasi-stationary front based on deep learning, characterized by: include: Obtain historical position data and meteorological data of Kunming's quasi-stationary front; Using the Kunming quasi-stationary front historical position data and the meteorological data, a Kunming quasi-stationary front label set is obtained, including: According to the historical position data of the Kunming quasi-stationary front, a position image of the Kunming quasi-stationary front is obtained through the meteorological information comprehensive analysis and processing system MICAPS, and the latitude and longitude data of the Kunming quasi-stationary front are calculated; Obtaining a meteorological element image based on the meteorological data, and plotting the latitude and longitude data of the Kunming quasi-stationary front on the meteorological element image at the corresponding time to obtain a meteorological element distribution and front position map; revising the position image of the Kunming quasi-stationary front according to the meteorological element distribution and front position map to obtain a revised position image of the Kunming quasi-stationary front; Preprocessing the revised Kunming quasi-stationary front position image to obtain the Kunming quasi-stationary front label set; Using the meteorological data, obtaining an RGB image dataset; Inputting the Kunming quasi-stationary front label set and the RGB image dataset of the corresponding time into the DETR model for training to obtain a trained DETR model; The RGB image of any time period is input into the trained DETR model, and the recognition result of the Kunming quasi-stationary front of any time period is output.

2. The deep learning-based automatic identification method for Kunming quasi-stationary front according to claim 1 is characterized in that: Preprocessing the revised Kunming quasi-stationary front position image includes: The Kunming quasi-stationary front line in the revised Kunming quasi-stationary front position image is extracted hour by hour to obtain Kunming quasi-stationary front position images corresponding to several time periods, and the range, pixel size and color of the Kunming quasi-stationary front position images corresponding to the several time periods are preset.

3. The Kunming quasi-stationary front automatic identification method based on deep learning according to claim 1 is characterized in that: Calculating the longitude and latitude data of the Kunming quasi-stationary front includes calculating the longitude and latitude data of the Kunming quasi-stationary front according to a drawing range of the position image of the Kunming quasi-stationary front and relative positions of front line grid points.

4. The Kunming quasi-stationary front automatic identification method based on deep learning according to claim 1 is characterized in that: Using the meteorological data, obtaining an RGB image dataset includes: The meteorological data is subjected to Gaussian smoothing processing, the smoothed meteorological data is subjected to grayscale processing, and the grayscale values ​​are respectively used as pixel values ​​in the R, G, and B channels of the RGB image to obtain the RGB image data set.

5. The Kunming quasi-stationary front automatic identification method based on deep learning according to claim 1 is characterized in that: The meteorological data includes wind field, sea level pressure and temperature data.

6. The Kunming quasi-stationary front automatic identification method based on deep learning according to claim 1 is characterized in that: Inputting the Kunming quasi-stationary front label set and the RGB image dataset of the corresponding time into the DETR model for training includes: Input the Kunming quasi-stationary front label set and the RGB image dataset of the corresponding time into the DETR model for training, and obtain the optimal weight according to the loss function value; The optimal weight is put into the DETR model as a frozen weight, and the mask segmentation part is iteratively trained to obtain the trained DETR model.

7. The deep learning-based automatic identification method for Kunming quasi-stationary front according to claim 6 is characterized in that: The loss function value includes the loss calculation of the target detection part and the target segmentation part, wherein the loss function of the target detection part is calculated according to the Hungarian algorithm, and the target segmentation part is calculated according to the binary cross entropy loss function.

Images