A method for identifying an atmospheric front

Abstract

The application discloses a method for identifying an atmospheric front, and relates to the field of meteorological observation. The method obtains six types of meteorological elements of a region to be identified; six types of meteorological element data are fused through a meteorological element fusion module to obtain a comprehensive feature map of the six types of meteorological elements; the region to be identified is divided according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected front region frames; the plurality of suspected front region frames are screened, and overlapping suspected front region frames are removed to obtain a plurality of front suggestion frames; for each of the plurality of front suggestion frames, a regression operation is performed on the front suggestion frame to correct the category, and a front suggestion frame with the highest front category probability score after category correction is determined as a front boundary frame; front boundary frames with a number of grid points less than a threshold value are removed from the plurality of front boundary frames, and an atmospheric front is extracted from the remaining front boundary frames. The method can improve the prediction accuracy of atmospheric front identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of meteorological observation, and in particular to a method for identifying an atmospheric front. BACKGROUND

[0002] An atmospheric front is an interface or transition zone between two air masses of different densities, which is generated by the encounter of two air masses with different thermodynamic properties, and usually occurs in a low pressure trough. The baroclinicity of the atmosphere near the atmospheric front is large, which leads to frequent occurrence of severe weather changes near the atmospheric front. For example, the area near the atmospheric front is often accompanied by severe weather phenomena such as heavy rain and storms, and the atmospheric front is often the center of thunderstorms. According to statistics, the amount of precipitation on the front accounts for 51% of the global extreme precipitation, and up to 90% of extreme precipitation events are related to mid-latitude fronts. Therefore, accurate identification of the front is beneficial to weather analysis and forecasting.

[0003] Convolutional Neural Networks (CNNs) are applied to the front identification task. First, the entire identification region is divided into several local regions, and then each local region is classified into a front category. However, this method lacks consideration of the correlation between local regions and the global consideration of the front system, resulting in loss of front feature information, and as a result, a large number of objective long-wave fronts are detected as multiple short-wave fronts. When using Fully Convolutional Networks (FCN) for atmospheric front identification, the FCN network is directly applied to front identification, and the input is multiple meteorological factors, which causes a conflict between meteorological factors, resulting in low identification accuracy. Similarly, when a U-shaped network is used for automatic front identification, the input data is not effectively processed, resulting in low accuracy of atmospheric front identification.

[0004] Therefore, there is an urgent need for a method that can improve the accuracy of atmospheric front identification. SUMMARY

[0005] Therefore, it is necessary to provide a method for identifying an atmospheric front to solve the above technical problems. The method can improve the prediction accuracy of the atmospheric front.

[0006] The present application adopts the following technical solutions:

[0007] The present application provides a method for identifying an atmospheric front, comprising:

[0008] Six types of meteorological elements of the region to be identified are obtained; the six types of meteorological elements include temperature, air pressure, humidity, wind direction, wind speed and dew point temperature;

[0009] The six types of meteorological element data are fused by a fusion meteorological element module to obtain a comprehensive feature map of the six types of meteorological elements; the fusion meteorological element module comprises a normalization layer, a weight layer and a double convolution layer connected in sequence;

[0010] The to-be-recognized region is divided according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected frontal region frames;

[0011] The plurality of suspected frontal region frames are screened and overlapping suspected frontal region frames are removed to obtain a plurality of frontal suggestion frames;

[0012] For each of the plurality of frontal suggestion frames, the frontal suggestion frame is corrected in class by a regression operation, and the frontal suggestion frame with the highest frontal class probability score after class correction is determined as a frontal boundary frame;

[0013] The frontal boundary frame with a grid point number less than a threshold is removed from the plurality of frontal boundary frames, and an atmospheric front is extracted from the remaining frontal boundary frames.

[0014] Preferably, the six types of meteorological element data are fused by a fusion meteorological element module to obtain a comprehensive feature map of the six types of meteorological elements, specifically comprising:

[0015] The six types of meteorological element data are linearly normalized by the normalization layer;

[0016] The six types of meteorological elements after normalization are assigned different weights at each grid point by the weight layer;

[0017] The six types of meteorological elements after weight assignment are processed by the double convolution layer to obtain a comprehensive feature map of the six types of meteorological elements.

[0018] Preferably, the calculation formula of the weight is:

[0019] ;

[0020] wherein, m is the i-th type of meteorological element, m is the weight value of the i-th type of meteorological element at the j-th row and k-th column grid point position in the to-be-recognized region, is the data value of the i-th type of meteorological element at the j-th row and k-th column grid point position in the to-be-recognized region. m i j m i j

[0021] Preferably, the to-be-recognized region is divided according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected frontal region frames, specifically comprising: ​​​​​​​

[0022] Segment the to-be-identified region based on the comprehensive feature map of the six types of meteorological elements to obtain a plurality of sub-regions;

[0023] Calculate the similarity of all adjacent sub-regions in the plurality of sub-regions to obtain a sub-region similarity set;

[0024] Merge the two sub-regions with the largest similarity in the sub-region similarity set to obtain a new sub-region;

[0025] After obtaining the new sub-region, update the sub-region similarity set until the sub-region similarity set is empty, and determine the sub-region corresponding to the empty sub-region similarity set as a plurality of suspected frontal region frames.

[0026] Preferably, the plurality of suspected frontal region frames are screened and overlapping suspected frontal region frames are removed to obtain a plurality of frontal suggestion frames, specifically including:

[0027] Using four support vector machines to perform probability prediction on the four types of fronts of the plurality of suspected frontal region frames;

[0028] After probability prediction, each column of the plurality of suspected frontal region frames is sorted in descending order of intersection over union score; the list represents the category of the front;

[0029] In descending order, for each column, starting from the column with the maximum score, the intersection over union of the column with the maximum score and the column after the column with the maximum score is calculated to obtain an intersection over union calculation result;

[0030] If the intersection over union calculation result is greater than a preset threshold, the column with the smaller intersection over union score is removed from the column with the maximum score and the column after the column with the maximum score;

[0031] After processing all columns of the plurality of suspected frontal region frames in turn, a plurality of frontal suggestion frames are obtained.

[0032] Preferably, the frontal suggestion frames are corrected in category by a regression operation, specifically including:

[0033] The frontal suggestion frames are subjected to a regression operation by a regressor; the regression operation includes a translation operation and a scale operation; the translation amount corresponding to the translation operation is:

[0034] , ;

[0035] wherein, is the translation amount of the horizontal coordinate of the center point of the frontal suggestion frame, is the translation amount of the vertical coordinate of the center point of the frontal suggestion frame, is the horizontal coordinate of the center point of the frontal suggestion frame with the true value, a horizontal coordinate of a center point of the frontal proposal box for a predicted value, a width of the frontal proposal box for a predicted value, a vertical coordinate of a center point of the frontal proposal box for a true value, a vertical coordinate of a center point of the frontal proposal box for a predicted value, a height of the frontal proposal box for a predicted value;

[0036] The formula corresponding to the scale operation is:

[0037] , ;

[0038] wherein, a width translation amount of the frontal proposal box, a height translation amount of the frontal proposal box, a width of the frontal proposal box for a true value, a width of the frontal proposal box for a predicted value, a height of the frontal proposal box for a true value, a height of the frontal proposal box for a predicted value;

[0039] The target function is constructed, and the category corresponding to the maximum target function value is taken as the corrected category; the target function is:

[0040] ;

[0041] wherein, a feature vector composed of the height and width prediction values of the frontal proposal box obtained through the regression operation, a learning parameter, a target function.

[0042] Preferably, the frontal boundary boxes with a number of grid points less than a threshold value are removed from the plurality of frontal boundary boxes, and the atmospheric front is extracted from the remaining frontal boundary boxes, specifically including:

[0043] For each of the plurality of frontal boundary boxes, the number of grid points in the frontal boundary box is calculated, and if the number of grid points is less than a preset threshold value, the frontal boundary box is removed;

[0044] For each of the remaining frontal boundary boxes after the removal operation, the dew point temperature gradient of each grid point in the remaining frontal boundary box is calculated to obtain a dew point temperature gradient value;

[0045] The grid point with the minimum dew point temperature gradient value is selected in each row in the frontal boundary box;

[0046] The grid points with the minimum dew point temperature gradient value in each row are connected to form the atmospheric front.

[0047] The application provides a front of an atmospheric front recognition device, comprising:

[0048] An acquisition module is configured to acquire six types of meteorological elements of a region to be recognized, including temperature, air pressure, humidity, wind direction, wind speed and dew point potential temperature.

[0049] A fusion module is configured to fuse the six types of meteorological element data by a fusion meteorological element module to obtain a comprehensive feature map of the six types of meteorological elements, wherein the fusion meteorological element module comprises a normalization layer, a weight layer and a double convolution layer connected in sequence.

[0050] A division module is configured to divide the region to be recognized according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected front region frames.

[0051] A screening module is configured to screen the plurality of suspected front region frames and remove overlapping suspected front region frames to obtain a plurality of front suggestion frames.

[0052] A correction module is configured to correct the class of each of the plurality of front suggestion frames by a regression operation, and determine the front suggestion frame with the highest front class probability score after class correction as a front boundary frame.

[0053] An extraction module is configured to remove the front boundary frame with a grid point number less than a threshold from the plurality of front boundary frames, and extract an atmospheric front from the remaining front boundary frames.

[0054] The application provides a computer readable storage medium, wherein the storage medium stores a computer program, and the computer program is executed by a processor to implement the above-mentioned atmospheric front recognition method.

[0055] The application provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned atmospheric front recognition method when executing the program.

[0056] The above-mentioned at least one technical scheme adopted by the application can achieve the following beneficial effects:

[0057] Six types of meteorological elements of the to-be-identified region are acquired; the six types of meteorological element data are fused through a fusion meteorological element module to obtain a comprehensive feature map of the six types of meteorological elements, the multi-dimensional physical characteristics of the front are captured through a feature fusion strategy, and the limitations of single-element analysis are overcome; the to-be-identified region is divided according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected front region frames; the plurality of suspected front region frames are screened, and overlapping suspected front region frames are removed to obtain a plurality of front suggestion frames, the redundant detection frames are effectively removed by removing the overlapping suspected front region frames, and the false positive rate is reduced; for each of the plurality of front suggestion frames, the front suggestion frame is corrected in category through a regression operation, and the front suggestion frame with the highest front category probability score after category correction is determined as a front boundary frame, the position offset of the suggestion frame is adjusted through the regression operation, the front boundary is aligned with the real meteorological field, and the front boundary frame with the grid point number less than a threshold value is removed from the plurality of front boundary frames, and an atmospheric front is extracted from the remaining front boundary frames, the large-scale front structure is reserved through an area threshold filtering, and the definition of the front system in meteorology is met. The method can improve the prediction accuracy of atmospheric front identification. BRIEF DESCRIPTION OF DRAWINGS

[0058] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application, and do not limit the application. In the drawings:

[0059] Figure 1 A flowchart of an atmospheric front identification method provided by the application is shown in the figure;

[0060] Figure 2 A front identification model provided by the application is shown in the figure;

[0061] Figure 3 A front identification model training method provided by the application is shown in the figure;

[0062] Figure 4 An ALexNet network structure provided by the application is shown in the figure;

[0063] Figure 5 An atmospheric front identification device provided by the application is shown in the figure;

[0064] Figure 6 A computer device for an atmospheric front identification method provided by the application is shown in the figure. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0066] Devices such as desktop computers, servers, and laptop computers that can execute the solution of the present invention are described below with the server as the execution subject for the sake of convenience.

[0067] Although the current intelligent level of automatic front identification is very high compared with traditional methods, its reliability and robustness are still poor. It is necessary to continue to improve and continuously enhance its reliability, robustness and identification accuracy.

[0068] Traditional front identification methods are mainly numerical frontal analysis (NFA) methods, which rely on selected parameters and artificially set functions and thresholds to diagnose fronts. There are mainly gradient methods ( ), Thermal Front Parameter (TFP), Wind Shear (WS) and Diagnostic Method (F). Directly calculating the gradient of the parameter τ and determining the front result through a threshold is the simplest method but has the lowest accuracy. The TFP method further derives the derivative based on the gradient method and defines the front as the directional derivative of the thermodynamic quantity τ along the horizontal gradient of its gradient. However, its disadvantage is that calculating the second derivative on a finite grid is prone to generate noise fields, which can easily lead to large systematic deviations. The WS method uses changes in wind direction and wind speed, especially the 6-hour changes in the meridional wind, to identify fronts. Its advantage is that it can better detect fronts with weak baroclinicity that cause wind changes and fusion between two anticyclones. Its disadvantage is that warm fronts are almost never detected. The F diagnostic method, taking into account the advantages of the thermal method and the wind change method, diagnoses the area where the normalized product of the isobaric relative vorticity and the horizontal temperature gradient exceeds the threshold as a front. However, its disadvantage is the same as TFP, that additional derivatives are required to determine the warm edge of the front formation area as a front, and the identification effect will deteriorate after replacing the parameter potential temperature with other parameters. Recently, a method for diagnosing fronts based on the Dynamic State Index (DSI) of non-adiabatic processes has been proposed. However, the DSI still has the defects of the NFA method. It relies on a single physical model and threshold to determine the front. It is easy to miss fronts with weak meteorological factor diagnostic values ​​caused by special weather conditions and special terrain, or misjudge non-frontal meteorological factor characteristics with high diagnostic values ​​as frontal characteristics.

[0069] Although the current machine learning method applied to automatic identification of atmospheric front has made great progress, it is basically based on semantic segmentation network, and the semantic segmentation network also has a defect that it can only classify each grid point of the front, and cannot provide which specific front the front grid point belongs to.

[0070] The technical solutions of the embodiments of the present application are described in detail below with reference to the drawings.

[0071] Figure 1 The flowchart of the method for identifying an atmospheric front in the present application specifically comprises the following steps:

[0072] S101: Obtain six types of meteorological elements of a to-be-identified region; the six types of meteorological elements include temperature, air pressure, humidity, wind direction, wind speed and dew point temperature.

[0073] The six types of meteorological elements of the to-be-identified region specifically include temperature, air pressure, humidity, wind direction, wind speed and dew point temperature of the to-be-identified region.

[0074] S102: Fuse the six types of meteorological element data through a meteorological element fusion module to obtain a comprehensive feature map of the six types of meteorological elements; the meteorological element fusion module comprises a normalization layer, a weight layer and a double convolution layer connected in sequence.

[0075] In an exemplary embodiment, the six types of meteorological element data are fused through the meteorological element fusion module to obtain the comprehensive feature map of the six types of meteorological elements, specifically comprising: linearly normalizing the six types of meteorological element data through the normalization layer; assigning different weights to the six types of meteorological elements after normalization at each grid point through the weight layer; processing the six types of meteorological elements after weight assignment through the double convolution layer to obtain the comprehensive feature map of the six types of meteorological elements.

[0076] Specifically, first, the six types of meteorological element data in the most important part of EAR5 are linearly Min-Max normalized through the normalization layer, and the calculation formula of normalization is as shown in formula (1):

[0077] (1)

[0078] wherein, x is the original meteorological element data, is the minimum value of the meteorological element data, is the maximum value of the meteorological element data, is the normalized meteorological element data.

[0079] Then, different weights are assigned to the six types of meteorological elements at each grid point using the weight layer, and the weights are continuously adjusted using the labels during training to address the problem that the strong temperature characteristics of non-frontal surfaces are identified as frontal surface characteristics, and the meteorological element characteristics of weak frontal surfaces are not identified under special weather conditions and special geographical locations. The calculation formula of the weight of the six types of meteorological elements at each grid point is shown in formula (2):

[0080] (2);

[0081] wherein, m is the i-th m meteorological element, is the weight value of the i-th m meteorological element at the grid point position of the region to be identified, ij is the data value of the i-th meteorological element at the grid point position of the region to be identified. m ij Finally, the comprehensive features of the six types of meteorological elements are obtained through two double convolution operations and an activation operation: two groups of full convolution layers and ReLU activation functions are connected in series, each group of full convolution is convolved twice, the convolution kernel of the convolution operation is 3, the step is 2, and the padding is 0. Another function of the fused meteorological element module is to convert the input six-channel meteorological element data into three-channel output.

[0082] The output result of the fused meteorological element module is normalized to the color space of 0-255 using the normalization processing method.

[0083]

[0084] S103: Dividing the region to be identified according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected frontal surface region frames.

[0085] In an example embodiment, the region to be identified is divided according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected frontal surface region frames, specifically including: segmenting the region to be identified based on the comprehensive feature map of the six types of meteorological elements to obtain a plurality of sub-regions; calculating the similarity of all adjacent sub-regions in the plurality of sub-regions to obtain a sub-region similarity set; performing a merging operation on the two sub-regions with the largest similarity in the sub-region similarity set to obtain a new sub-region; after obtaining the new sub-region, updating the sub-region similarity set until the sub-region similarity set is empty, and determining the sub-region corresponding to the empty sub-region similarity set as the plurality of suspected frontal surface region frames.

[0086] Specifically, the process of dividing the region to be identified mainly includes:

[0087] ​​Step 1: Region segmentation. Based on the comprehensive characteristics of the six types of meteorological elements, the Felzenswalb and Huttenlocher algorithm is used to generate multiple small regions for the entire area to be identified, so as to achieve the purpose of preliminary region segmentation.

[0088] Step 2: similarity calculation. For the segmented sub-regions, the similarities of adjacent sub-regions are calculated, and set operations are performed based on the similarities to obtain a similarity set of sub-regions.

[0089] First, calculate the color similarity between adjacent sub-regions: use L1-norm normalization to evenly divide the 0-255 color space into 25 small color spaces, also known as histograms, so that each region obtains a 75 (25x3) dimensional vector. The calculation formula for the color similarity between adjacent sub-regions is shown in formula (3):

[0090] ,

[0091] (3);

[0092] in, are two adjacent sub-regions, Calculate the histogram of the new region for the sub-region merging process, k The first k grid points, n is the number of grid points in the adjacent sub-regions of color, is the first k grid points, is the first k grid points, is the color sub-region, are color adjacent subregions, is the color similarity between adjacent sub-regions.

[0093] Secondly, the texture similarity between adjacent sub-regions is calculated: the Gaussian differential with variance σ = 1 is calculated for 8 different directions of each color channel, and 10 histograms of each direction of each color channel of the image are obtained using L1-norm normalization to obtain a 240 (10x8x3) dimensional vector , the formula for calculating the texture similarity between sub-regions is shown in formula (4):

[0094] (4);

[0095] in, is the texture similarity between sub-regions, k is the first k grid points,n is the number of grid points in the texture adjacent sub-region, is the number of grid points in the texture sub-region, k is the number of grid points in the texture adjacent sub-region, is the number of grid points in the texture adjacent sub-region, k is the number of grid points in the texture adjacent sub-region, is the number of grid points in the texture adjacent sub-region, i is the number of grid points in the texture adjacent sub-region, is the number of grid points in the texture adjacent sub-region. j

[0096] The texture feature of the new region after merging is calculated in the same way as the color feature.

[0097] Then, the similarity of the small region size is calculated to preferentially merge small regions: in order to apply multi-scale to the global, more weight is given to small regions, and the calculation formula of the region size similarity is shown in formula (5):

[0098] (5);

[0099] wherein, is the region size similarity, is the number of grid points in the region, is the number of grid points in the texture adjacent sub-region, i is the number of grid points in the texture adjacent sub-region. j

[0100] Finally, the fitness distance of the small region is calculated: if the region is contained, it is first merged, and if it is difficult to interface, it is not merged together. The fitness distance of the region is mainly defined to measure whether two regions are more "fitting", and the index is the minimum rectangle that can frame the region of the region after merging Bounding Box, is smaller, the fitness is higher, that is, the similarity is closer to 1. The calculation method of the fitness distance of the small region is shown in formula (6):

[0101] (6);

[0102] wherein, is the fitness distance, is the minimum rectangle that can frame the region, is the number of grid points in the region, is the number of grid points in the texture adjacent sub-region, i is the number of grid points in the texture adjacent sub-region. j

[0103] ​​​​​The sum of the above four types of similarities is calculated, and the calculation method of the sum of the above four types of similarities is shown in formula (7):

[0104] (7);

[0105] wherein, is the sum of the four types of similarities, , is a weight parameter of color similarity, is a weight parameter of texture similarity, is a weight parameter of region size similarity, is a weight parameter of region suitability.

[0106] Step three, region merging, the two sub-regions with the largest similarity in the similarity set are merged to form a new sub-region, and then the similarity set of the sub-region is updated again.

[0107] Step four, iterative merging, sorting and outputting, repeating the above step two and step three until the similarity set of the sub-region is empty, at this time the region set obtained is the final segmentation region of the frontal surface recognition region. Through the operation of step four, the image of the comprehensive characteristics of the six types of meteorological elements generates 1000 suspected frontal surface region frames.

[0108] Specifically, the size of the suspected frontal surface region frame is reduced, and the suspected frontal surface region frame is reduced to a size of 227*227 and input into the ALexNet network.

[0109] S104: screening a plurality of suspected frontal surface region frames, and removing overlapping suspected frontal surface region frames to obtain a plurality of frontal surface suggestion frames.

[0110] The fifth generation reanalysis data set of the frontal surface recognition region is obtained, and the fifth generation reanalysis data set is divided into a training set, a verification set and a test set; the fifth generation reanalysis data set includes six types of meteorological elements.

[0111] The present application takes the frontal surface of the identified research region as an example, first, downloads the frontal surface picture; second, uses software to make a label of the training network, uses an image labeling tool labelme to label the downloaded frontal surface picture, then converts the labeled frontal surface picture into a COCO (Common Objects in Context, COCO) data format of image recognition and labeling.

[0112] The input data of the front edge boundary box network is: downloading the fifth generation reanalysis dataset ERA5 of the research area, the longitude range is 73°E to 135°E, the latitude range is 3°N to 53°N, and the longitude and latitude are rounded for correspondence with the label; the network input is 2-meter dew point temperature, 1000-pa height temperature, U wind, V wind, specific humidity and sea level average pressure data, the horizontal resolution is 0.25°*0.25°, and the time interval is 3h.

[0113] The obtained ERA5 dataset corresponding to the label is divided into a training set, a validation set and a test set.

[0114] The front edge recognition model is constructed, as shown in the following formula: Figure 2 The front edge recognition model provided by the application comprises a fusion meteorological element module and an extracted front edge boundary box network. The fusion meteorological element module comprises a normalization module, a weight layer module, a double convolution layer module and a normalization module, and the extracted front edge boundary box network comprises a selective search module, an AlexNet module, an SVM module, an NMS module and a regressor module.

[0115] Specifically, the process of training the front edge recognition model is: training the front edge recognition model by using the training set, and using the trained front edge recognition model as the network model in the application of front edge recognition, the training rounds (Epoch) of the extracted front edge boundary box network are set to 100, the optimizer uses the stochastic gradient descent method (Stochastic Gradient Descent, SGD), the learning rate of the network parameter is set to 0.0001, and the momentum value is set to 0.7.

[0116] The method for training the front edge recognition model is as shown in the following formula: Figure 3 The main steps are: selecting the model corresponding to the network parameter with the minimum loss after training every 10 rounds, calculating the average loss on the validation set, saving the parameters of the network model at this time, and after 100 rounds, selecting the network model parameters with the minimum average loss in the 10 network models on the test set as the parameters of the trained front edge recognition model, and using the trained front edge recognition model for the recognition stage of engineering practice.

[0117] The loss function of the network is obtained by comparing the predicted value and the true value, and the loss function is as shown in the following formula (8):

[0118] (8);

[0119] Wherein, Loss is the loss function value, represents x,y,w,h (the coordinates of the front edge boundary box), x,y is the center point coordinate of the front edge boundary box, w is the width of the front edge boundary box,h is a height of the edge boundary frame, i is a first i frame, N is a total number of edge boundary frames, is a translation amount of the edge boundary frame, is a feature vector, is a parameter to be learned, is a regularization.

[0120] In an exemplary embodiment, the plurality of suspected edge region frames are screened and overlapping suspected edge region frames are removed to obtain a plurality of edge proposal frames, specifically comprising: using four support vector machines to perform probability prediction of four types of edges on the plurality of suspected edge region frames; sorting each column of the plurality of suspected edge region frames after probability prediction in descending order of intersection over union (IoU) score; the list represents the category of the edge; in descending order, starting from the maximum score column, the IoU of the maximum score column and the column after the maximum score column is calculated, and the IoU calculation result is obtained; if the IoU calculation result is greater than a preset threshold, the column with a smaller IoU score is removed from the maximum score column and the column after the maximum score column; after sequentially processing all columns of the plurality of suspected edge region frames, the plurality of edge proposal frames are obtained.

[0121] Specifically, the overlapping suspected edge region frames are removed: the non-maximum suppression (NMS) method is used to remove the overlapping suspected edge region frames to obtain the best suspected edge region frame, and these region frames are used as edge proposal frames, and the main steps are as follows:

[0122] (1) The suspected edge region frames of 1000x4 dimensions are sorted in descending order, and each column represents a category, and there are four categories.

[0123] (2) Starting from the suspected edge region frame with the maximum score in each column, the intersection over union (IoU) is calculated with the suspected edge region frame with a score behind the column, and if the IoU is greater than a preset threshold, the suspected edge region frame with a smaller score is removed, otherwise it is considered that there are multiple targets of the same type of edge in the image.

[0124] (3) Starting from the suspected edge region frame with the second largest score in each column, step (2) is repeated.

[0125] (4) Step (3) is repeated until all suspected edge region frames in the column are traversed.

[0126] (5) All columns of the suspected edge region frames of 1000x4 dimensions are traversed, that is, the non-maximum suppression is performed on all edge types.

[0127] Specifically, a 4096-dimensional feature vector of the suspected frontal surface region frame is extracted by using an ALexNet network without a Softmax layer, and the ALexNet network is as shown in the following table: Figure 4 As shown in the table, the ALexNet network includes 1 input layer (input layer), 5 convolution layers (C1, C2, C3, C4, and C5), 2 fully connected layers (FC6 and FC7) with 4096 neurons, and 1 output layer (output layer).

[0128] Each column of the 1000*4-dimensional matrix output by the ALexNet network is used to predict the probability of four types of fronts: four support vector machine (SVM) classifiers are used to predict the probability of four types of fronts for each column of the 1000*4-dimensional matrix output by the ALexNet network, and the learning algorithm of the nonlinear SVM is a Gaussian radial basis function classifier as follows:

[0129] Input: The training data set is set as T ={( , ),( , ),..., , }.

[0130] Output: Separation hyperplane and classification decision function.

[0131] Select the kernel function K (x,z) and the penalty parameter C >0, construct and solve the convex quadratic programming problem, and the feature mapping space is as shown in formula (9):

[0132] (9);

[0133] s.t. ;

[0134] 0 , i=1,2,....,N ;

[0135] wherein is the penalty parameter, i is the horizontal coordinate of the 2-dimensional space data, j is the vertical coordinate of the 2-dimensional space data, is the data value of the multi-weather element, is the penalty parameter, is the penalty parameter, N is the number of data sets.

[0136] The optimal solution is .

[0137] Calculation: Select A component of Meet the conditions ,calculate .

[0138] Classification decision function:

[0139] The Gaussian kernel function is shown in formula (10):

[0140] (10);

[0141] in, x is a sample point in the input space, i.e., a feature vector. z is the center point of the Gaussian kernel function, is the parameter, K is the Gaussian kernel function.

[0142] The classification decision function is as shown in formula (11):

[0143] (11);

[0144] in, x is the eigenvector, i is an ordinal number, N is the number of variables, is a variable, z is a variable, is the parameter, is the classification decision function.

[0145] S105: For each of the multiple front suggestion boxes, perform category correction on the front suggestion box through a regression operation, and determine the front suggestion box with the highest front category probability score after category correction as the front bounding box.

[0146] In an exemplary embodiment, the category of the front suggestion box is corrected through a regression operation, specifically including: performing a regression operation on the front suggestion box through a regressor; the regression operation includes a translation operation and a scale scaling operation; constructing an objective function, and taking the category corresponding to the maximum objective function value as the corrected category.

[0147] Specifically, after the categories of the front suggestion boxes are corrected, the front suggestion box with the highest category probability score is determined as the front bounding box.

[0148] Specifically, a regression operation is performed: the regression operation is performed on the front suggestion boxes of the four categories using the regressor. When the regressor corrects the front suggestion box, the translation amount ( , ) and scale scaling ( , )。

[0149] translation amount of the horizontal coordinate of the center point of the frontal plane proposal box, , The translation amount of the horizontal coordinate of the center point of the frontal plane proposal box is shown in equation (12):

[0150] , (12);

[0151] wherein, is a translation amount of the horizontal coordinate of the center point of the frontal plane proposal box, is a translation amount of the vertical coordinate of the center point of the frontal plane proposal box, is the horizontal coordinate of the center point of the frontal plane proposal box of the ground truth, is the horizontal coordinate of the center point of the frontal plane proposal box of the prediction value, is the width of the frontal plane proposal box of the prediction value, is the vertical coordinate of the center point of the frontal plane proposal box of the ground truth, is the vertical coordinate of the center point of the frontal plane proposal box of the prediction value, is the height of the frontal plane proposal box of the prediction value.

[0152] scaling (scale) , The scaling is shown in equation (13):

[0153] , (13);

[0154] wherein, is a width translation amount of the frontal plane proposal box, is a height translation amount of the frontal plane proposal box, is the width of the frontal plane proposal box of the ground truth, is the width of the frontal plane proposal box of the prediction value, is the height of the frontal plane proposal box of the ground truth, is the height of the frontal plane proposal box of the prediction value.

[0155] wherein, G is the ground truth of the frontal plane, P is the CNN feature corresponding to the proposal box.

[0156] The frontal plane proposal box is corrected by using the objective function: the objective function is shown in equation (14):

[0157] (14);

[0158] wherein, is a feature vector composed of the height and width prediction values of the frontal plane proposal box obtained by the regression operation, to learn parameters, to a target function.

[0159] The highest score of each category after the correction is taken as the edge boundary box.

[0160] S106: Remove the edge boundary box with the number of grid points less than the threshold value from the plurality of edge boundary boxes, and extract the atmospheric front from the remaining edge boundary boxes.

[0161] In an exemplary embodiment, removing the edge boundary box with the number of grid points less than the threshold value from the plurality of edge boundary boxes, and extracting the atmospheric front from the remaining edge boundary boxes specifically includes: for each of the plurality of edge boundary boxes, calculating the number of grid points in the edge boundary box, and if the number of grid points is less than the preset threshold value, removing the edge boundary box; for each of the remaining edge boundary boxes after the removal operation, calculating the dew point potential temperature gradient of each grid point in the remaining edge boundary box to obtain the dew point potential temperature gradient value; selecting the grid point with the minimum dew point potential temperature gradient value in each row in the edge boundary box; and connecting the grid points with the minimum dew point potential temperature gradient value in each row to obtain the atmospheric front.

[0162] Specifically, the threshold value is set according to specific engineering practice.

[0163] Specifically, the small-range edge boundary box is removed: the number of grid points in each edge boundary box is calculated, and if the number of grid points in the boundary box is less than 6, the boundary box is removed.

[0164] The edge in the edge boundary box is extracted: first, the gradient of the dew point potential temperature of each grid point in each edge boundary box is calculated, then the grid point with the minimum dew point potential temperature gradient value in each row in the edge boundary box is selected, and finally, the grid points with the minimum dew point potential temperature gradient value in each row are connected as the atmospheric front.

[0165] When the atmospheric front recognition method provided by the present application is applied, the number of grid points in the edge boundary box can not be counted according to the number of grid points in the edge boundary box. Figure 1 The execution order of each step shown in the above method can be determined as needed, and the present application does not limit the execution order of each step.

[0166] The above is a method for recognizing an atmospheric front provided by one or more embodiments of the present application. Based on the same idea, the present application also provides a corresponding atmospheric front recognition device, as shown in Figure 5 .

[0167] Figure 5 The atmospheric front recognition schematic diagram provided by the present application includes:

[0168] The acquisition module 501 is configured to acquire six types of meteorological elements of a region to be identified; the six types of meteorological elements include temperature, air pressure, humidity, wind direction, wind speed and dew point potential temperature.

[0169] The fusion module 502 is configured to fuse the six types of meteorological element data by a fusion meteorological element module to obtain a comprehensive feature map of the six types of meteorological elements; the fusion meteorological element module includes a normalization layer, a weight layer and a double convolution layer connected in sequence.

[0170] The division module 503 is configured to divide the region to be identified according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected front region frames.

[0171] The screening module 504 is configured to screen the plurality of suspected front region frames and remove overlapping suspected front region frames to obtain a plurality of front suggestion frames.

[0172] The correction module 505 is configured to, for each of the plurality of front suggestion frames, correct the front suggestion frame by a regression operation, and determine a front boundary frame with the highest front class probability score after class correction as the front boundary frame.

[0173] The extraction module 506 is configured to remove a front boundary frame with a grid point number less than a threshold from the plurality of front boundary frames, and extract an atmospheric front from the remaining front boundary frames.

[0174] The specific limitation of the atmospheric front recognition device can refer to the limitation of the atmospheric front recognition method in the above, which will not be repeated here. Each module in the atmospheric front recognition device described above can be realized by software, hardware and their combination. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in the form of software, so that the processor calls and executes the operations corresponding to each module.

[0175] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the atmospheric front recognition method provided by the present application. Figure 1 The present application also provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the atmospheric front recognition method provided by the present application.

[0176] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the atmospheric front recognition method provided by the present application. Figure 6 The structure diagram of the computer device is shown in FIG. 1. Figure 6 The computer device includes a processor, an internal bus, a network interface, a memory and a non-volatile memory at the hardware level, and of course can also include other hardware required by the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs to realize the atmospheric front recognition method provided by the present application. Figure 1 The structure diagram of the computer device is shown in FIG. 1.

[0177] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments of the methods. In the embodiments of the present application, any reference to memory, storage, database or other medium can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0178] The technical features of the above embodiments can be combined in any way. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

Claims

1. A method for identifying an atmospheric front, characterized in that: include: Acquire six types of meteorological elements of the area to be identified; the six types of meteorological elements include temperature, air pressure, humidity, wind direction, wind speed and dew point temperature; The six types of meteorological element data are fused through a meteorological element fusion module to obtain a comprehensive feature map of the six types of meteorological elements; the meteorological element fusion module includes a normalization layer, a weight layer and a double convolution layer connected in series; Dividing the area to be identified according to the comprehensive characteristic map of six types of meteorological elements to generate multiple suspected front area frames; screening the plurality of suspected front region frames and removing overlapping suspected front region frames to obtain a plurality of front suggestion frames; For each of the multiple front suggestion boxes, the category of the front suggestion box is corrected through regression operation, and the front suggestion box with the highest front category probability score after category correction is determined as the front bounding box; removing front bounding boxes with a grid point number less than a threshold from a plurality of front bounding boxes, and extracting an atmospheric front from the remaining front bounding boxes; The method of dividing the area to be identified based on the comprehensive feature map of the six meteorological elements to generate multiple suspected front area frames specifically includes: segmenting the area to be identified based on the comprehensive feature map of the six meteorological elements to obtain multiple sub-areas; calculating the similarities of all adjacent sub-areas in the multiple sub-areas to obtain a sub-area similarity set; merging two sub-areas with the greatest similarity in the sub-area similarity set to obtain a new sub-area; after obtaining the new sub-area, updating the sub-area similarity set until the sub-area similarity set is empty, and determining the sub-areas corresponding to the empty sub-area similarity set as multiple suspected front area frames; The screening of the multiple suspected front area frames and removing overlapping suspected front area frames to obtain multiple front suggestion frames specifically includes: using four support vector machines to perform probability prediction of four types of fronts on the multiple suspected front area frames; sorting each column of the multiple suspected front area frames after the probability prediction from large to small according to the intersection-and-union (IoU) score; the columns represent the categories of the fronts; calculating the IoU of the maximum score column with the columns following the maximum score column for each column in order from large to small, starting from the maximum score column, to obtain an IoU calculation result; if the IoU calculation result is greater than a preset threshold, removing the columns with smaller IoU scores between the maximum score column and the columns following the maximum score column; after processing all columns of the multiple suspected front area frames in sequence, obtaining multiple front suggestion frames; The classification correction of the front suggestion box through the regression operation specifically includes: A regression operation is performed on the front suggestion box through a regressor; the regression operation includes a translation operation and a scale operation; the translation amount corresponding to the translation operation is: Among them, t x is the translation of the horizontal coordinate of the center point of the front suggestion box, t y is the vertical coordinate translation of the center point of the front suggestion box, G x is the horizontal coordinate of the center point of the front suggestion box of the true value, P x is the horizontal coordinate of the center point of the front suggestion box of the predicted value, P w The width of the box suggested for the predicted value front, G y is the vertical coordinate of the center point of the front suggestion box of the true value, P y The vertical coordinate of the center point of the front suggestion box of the predicted value, P h Suggests the height of the box for the front of the predicted value; The formula corresponding to the scale scaling operation is: Among them, t w The width of the front suggestion box is translated, t h is the height translation of the front suggestion box, G w The width of the front suggestion box for the true value, P w The width of the box suggested for the predicted value front, G h is the height of the front suggestion box of the true value, P h Suggests the height of the box for the front of the predicted value; Construct an objective function, and take the category corresponding to the maximum objective function value as the corrected category; the objective function is: Among them, Φ5(P) is the feature vector composed of the height and width prediction values ​​of the front suggestion box obtained by the regression operation, is the learning parameter, d * (P) is the objective function.

2. The method according to claim 1, wherein The six types of meteorological element data are fused by the fusion meteorological element module to obtain a comprehensive feature map of the six types of meteorological elements, specifically including: The six types of meteorological element data are linearly normalized through the normalization layer; The weight layer is used to assign different weights to the six types of meteorological elements after normalization at each grid point; The six categories of meteorological elements after weight allocation are processed through a double convolutional layer to obtain the comprehensive feature maps of the six categories of meteorological elements.

3. The method according to claim 2, wherein The calculation formula of the weight is: Among them, m is the mth type of meteorological element, is the weight value of the grid point at the i-th row and j-th column of the m-th type of meteorological element in the area to be identified, is the data value of the grid point at the i-th row and j-th column of the m-th type of meteorological element in the area to be identified.

4. The method according to claim 1, wherein The removing of frontal bounding boxes having a number of grid points less than a threshold from the plurality of frontal bounding boxes and extracting the atmospheric front from the remaining frontal bounding boxes specifically includes: For each of the plurality of front boundary boxes, calculating the number of grid points in the front boundary box, and if the number of grid points is less than a preset threshold, removing the front boundary box; For each of the front boundary boxes remaining after the removal operation, calculating the dew point potential temperature gradient of each grid point in the remaining front boundary box to obtain a dew point potential temperature gradient value; In each row within the front boundary box, the grid point with the smallest dew point potential temperature gradient value is selected; The grid points with the smallest dew point potential temperature gradient value in each row are connected as the atmospheric front.

Images