Method for identifying atmospheric frontal surface

Abstract

The invention discloses an atmospheric frontal surface identification method, and relates to the field of meteorological observation. The method comprises the steps of obtaining six types of meteorological elements of a to-be-identified area; fusing 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; dividing a to-be-identified area according to the comprehensive feature map of the six types of meteorological elements to generate a plurality of suspected frontal area frames; screening the plurality of suspected frontal area frames, and removing overlapped suspected frontal area frames to obtain a plurality of frontal suggestion frames; for each of the plurality of frontal face suggestion boxes, performing category correction on the frontal face suggestion boxes through regression operation, and determining the frontal face suggestion box with the highest frontal face category probability score after category correction as a frontal face boundary box; and removing the frontal surface bounding boxes of which the number of lattice points is smaller than a threshold value from the plurality of frontal surface bounding boxes, and extracting the atmospheric frontal surface from the remaining frontal surface bounding boxes. The method can improve the prediction precision of atmospheric frontal surface recognition.

Description

Technical Field

[0001] The present invention relates to the field of meteorological observation, and particularly to a method for identifying atmospheric fronts. Background Art

[0002] An atmospheric front is an interface or transition zone between two air masses with different densities, formed by the encounter of two air masses with different thermodynamic properties, and usually appears in a trough of low pressure. The baroclinicity of the atmosphere near the atmospheric front is relatively large, resulting in frequent occurrence of severe weather changes in its vicinity. For example, severe weather phenomena such as heavy precipitation and storms often occur in the areas near the atmospheric front, and the atmospheric front is often the center of thunderstorms. According to statistics, frontal precipitation 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 fronts is beneficial to weather analysis and forecasting.

[0003] Convolutional Neural Networks (CNNs) have been applied to the task of front identification. First, the entire identification area is divided into several local areas, and then the classification of front types is carried out for each local area. However, this type of method lacks consideration of the correlation between local areas and the global consideration of the front system, resulting in the loss of front feature information. 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, since the FCN network is directly applied to front identification and the inputs are all multiple meteorological factors, there is a problem of multiple feature conflicts between meteorological factors, resulting in the disadvantage of low identification accuracy. Similarly, when using the U-shaped network for automatic front identification, due to the ineffective processing of the network input data, the identification accuracy of atmospheric fronts is low.

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

[0005] Based on this, it is necessary to provide a method for identifying atmospheric fronts to solve the above technical problems. This method can improve the prediction accuracy of atmospheric fronts.

[0006] The present invention adopts the following technical solutions: The present invention provides a method for identifying atmospheric fronts, including: Obtaining six types of meteorological elements in the area to be identified; the six types of meteorological elements include temperature, pressure, humidity, wind direction, wind speed, and dew point potential temperature; Fusing the data of the six types of meteorological elements 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 in sequence; Divide the area to be recognized according to the comprehensive feature map of six types of meteorological elements to generate multiple suspected frontal area frames; Screen multiple suspected frontal area frames, and remove the overlapping suspected frontal area frames to obtain multiple frontal area proposal frames; For each of the multiple frontal area proposal frames, perform category correction on the frontal area proposal frame through regression operation, and determine the frontal area proposal frame with the highest frontal category probability score after category correction as the frontal boundary frame; Remove the frontal boundary frames with the number of grid points less than the threshold from the multiple frontal boundary frames, and extract the atmospheric frontal surface from the remaining frontal boundary frames.

[0007] Preferably, 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, specifically including: Perform linear normalization processing on the six types of meteorological element data through a normalization layer; Assign different weights to the six types of meteorological elements after normalization processing at each grid point through a weight layer; Process the six types of meteorological elements after weight assignment through a double convolutional layer to obtain a comprehensive feature map of the six types of meteorological elements.

[0008] Preferably, the calculation formula for the weight is: ; where, m is the m th type of meteorological element, is the weight value of the m th type of meteorological element at the grid point position of the i rd row and j th column in the area to be recognized, is the data value of the m th type of meteorological element at the grid point position of the i th row and j th column in the area to be recognized.

[0009] Preferably, divide the area to be recognized according to the comprehensive feature map of six types of meteorological elements to generate multiple suspected frontal area frames, specifically including: Based on the comprehensive feature map of six types of meteorological elements, segment the area to be recognized to obtain multiple sub-areas; Calculate the similarity of all adjacent sub-areas in the multiple sub-areas to obtain a sub-area similarity set; Perform a merging operation on the two sub-areas with the largest similarity in the sub-area similarity set to obtain a new sub-area; After obtaining the new sub-regions, update the set of sub-region similarities until the set of sub-region similarities is empty, and determine the sub-regions corresponding to the empty set of sub-region similarities as multiple suspected frontal region boxes.

[0010] Preferably, screen the multiple suspected frontal region boxes and remove the overlapping suspected frontal region boxes to obtain multiple frontal proposal boxes, specifically including: Use 4 support vector machines to predict the probabilities of four types of fronts for the multiple suspected frontal region boxes; Sort each column of the multiple suspected frontal region boxes after probability prediction in descending order of the intersection-over-union score; the columns represent the categories of the fronts; For each column in descending order, starting from the column with the maximum score, calculate the intersection-over-union of the column with the maximum score and the columns behind it to obtain the intersection-over-union calculation result; If the intersection-over-union calculation result is greater than the preset threshold, remove the column with the smaller intersection-over-union score in the column with the maximum score and the columns behind it; After processing all columns of the multiple suspected frontal region boxes in sequence, obtain multiple frontal proposal boxes.

[0011] Preferably, perform category correction on the frontal proposal boxes through regression operations, specifically including: Perform regression operations on the frontal proposal boxes through a regressor; the regression operations include translation operations and scale scaling operations; the translation amount corresponding to the translation operation is: , ; Among them, is the translation amount of the abscissa of the center point of the frontal proposal box, is the translation amount of the ordinate of the center point of the frontal proposal box, is the abscissa of the center point of the true-value frontal proposal box, is the abscissa of the center point of the predicted-value frontal proposal box, is the width of the predicted-value frontal proposal box, is the ordinate of the center point of the true-value frontal proposal box, is the ordinate of the center point of the predicted-value frontal proposal box, is the height of the predicted-value frontal proposal box; The formula corresponding to the scale scaling operation is: , ; Among them, is the width translation amount of the frontal proposal box, is the height translation amount of the frontal proposal box, is the width of the true-value frontal proposal box, The width of the frontal suggestion box for the predicted value The height of the frontal suggestion box for the true value The height of the frontal suggestion box for the predicted value; Construct an objective function, and use the category corresponding to the maximum value of the objective function as the corrected category; the objective function is: ; Among them, is the feature vector composed of the predicted values of the height and width of the frontal suggestion box obtained by the regression operation, is the learning parameter, is the objective function.

[0012] Preferably, remove the frontal bounding boxes with the number of grid points less than the threshold from multiple frontal bounding boxes, and extract the atmospheric front from the remaining frontal bounding boxes, specifically including: For each of the multiple frontal bounding boxes, calculate the number of grid points within the frontal bounding box. If the number of grid points is less than the preset threshold, then remove the frontal bounding box; For each of the remaining frontal bounding boxes after the removal operation, calculate the dew point potential temperature gradient of each grid point within the remaining frontal bounding box to obtain the dew point potential temperature gradient value; Select the grid point with the minimum dew point potential temperature gradient value in each row within the frontal bounding box; Connect the grid points with the minimum dew point potential temperature gradient value in each row as the atmospheric front.

[0013] The present invention provides an identification device for an atmospheric front, including: An acquisition module for acquiring six types of meteorological elements in the area to be identified; the six types of meteorological elements include temperature, pressure, humidity, wind direction, wind speed, and dew point potential temperature.

[0014] A fusion module for fusing 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 includes a normalization layer, a weight layer, and a double convolutional layer connected in series in sequence; A division module for dividing the area to be identified according to the comprehensive feature map of the six types of meteorological elements to generate multiple suspected frontal area boxes; A screening module for screening multiple suspected frontal area boxes and removing overlapping suspected frontal area boxes to obtain multiple frontal suggestion boxes; A correction module for, for each of the multiple frontal suggestion boxes, performing category correction on the frontal suggestion box through a regression operation, and determining the frontal suggestion box with the highest frontal category probability score after category correction as the frontal bounding box; An extraction module is used to remove frontal bounding boxes with the number of grid points less than a threshold from multiple frontal bounding boxes, and extract atmospheric fronts from the remaining frontal bounding boxes.

[0015] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned method for identifying an atmospheric front.

[0016] The present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above-mentioned method for identifying an atmospheric front is implemented.

[0017] The above at least one technical solution adopted by the present invention can achieve the following beneficial effects: Obtain six types of meteorological elements in the area to be identified; fuse the data of the six types of meteorological elements through a meteorological element fusion module to obtain a comprehensive feature map of the six types of meteorological elements, and capture the multi-dimensional physical characteristics of the front through a feature fusion strategy to overcome the limitations of single-element analysis; divide the area to be identified according to the comprehensive feature map of the six types of meteorological elements to generate multiple suspected front area boxes; screen the multiple suspected front area boxes and remove the overlapping suspected front area boxes to obtain multiple front proposal boxes. By removing the overlapping suspected front area boxes, redundant detection boxes are effectively removed and the false positive rate is reduced; for each of the multiple front proposal boxes, perform category correction on the front proposal box through a regression operation, and determine the front proposal box with the highest frontal category probability score after category correction as the frontal bounding box. Adjust the position offset of the proposal box through a regression operation to align the frontal boundary with the real meteorological field; remove frontal bounding boxes with the number of grid points less than a threshold from multiple frontal bounding boxes, and extract atmospheric fronts from the remaining frontal bounding boxes. Retain large-scale frontal structures through area threshold filtering, which conforms to the definition of a frontal system in meteorology. This method can improve the prediction accuracy of atmospheric front identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present invention, and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings: Figure 1 It is a schematic flow chart of a method for identifying an atmospheric front provided by the present invention; Figure 2 It is a schematic diagram of a front identification model provided by the present invention; Figure 3 It is a schematic diagram of a method for training a front identification model provided by the present invention; Figure 4 It is a schematic diagram of the ALexNet network structure provided by the present invention; Figure 5 Schematic diagram of an apparatus for identifying an atmospheric front provided by the present invention; Figure 6 Schematic diagram of a computer device for a method of identifying an atmospheric front provided by the present invention. Detailed implementation manners

[0019] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0020] Devices such as desktop computers, servers, and laptop computers that can execute the solutions of the present invention. For the sake of convenience of description, only a server will be used as the execution subject for description below.

[0021] At present, although the intelligent level of automatic front identification is much higher than that of traditional methods, the reliability is still poor and the robustness is weak. It is necessary to continue to improve and continuously enhance its reliability, robustness and identification accuracy.

[0022] Traditional front identification methods mainly include the Numerical Frontal Analysis (NFA) method, which relies on selected parameters, artificially set functions and thresholds to diagnose fronts, mainly including the gradient method ( ), the Thermal Front Parameter (TFP) method, the Wind Shear (WS) method, and the diagnosis method (F). Among them, Directly obtaining the gradient of the parameter τ and determining the frontal result through a threshold, this method is the simplest but has the lowest accuracy; the TFP method takes a second derivative on the basis of the gradient method and defines the front as the directional derivative of the thermodynamic quantity τ along the horizontal gradient of its gradient. However, the disadvantage is that calculating the second derivative on a finite grid is prone to generating a noise field, which easily leads to large systematic deviations; the WS method uses the changes in wind direction and wind speed, especially the 6-hour change in the meridional wind, to identify the front. The advantage is that it can better detect fronts with weak baroclinicity causing wind changes and the fusion between two anticyclones. The disadvantage is that it almost never detects warm fronts; the F diagnostic method combines the advantages of the thermal method and the wind change method, and diagnoses the area where the normalized product of the isobaric relative vorticity and the horizontal temperature gradient exceeds the threshold as the front. However, its disadvantage is the same as that of the TFP. It requires additional derivatives to determine the warm edge of the frontal formation area as the front, and the identification effect will deteriorate when the potential temperature parameter is replaced with other parameters. Recently, a method for diagnosing fronts based on the Dynamic State Index (DSI) of non-adiabatic processes has been proposed. However, DSI still has the defects of the NFA method, relying on a single physical model and threshold to determine the front, and it is very easy to miss fronts with weak meteorological factor diagnostic values caused by special weather conditions and special terrains, or misjudge the characteristics of non-frontal meteorological factors with high diagnostic values as frontal characteristics.

[0023] Although the current machine learning methods have made great progress in the automatic identification of atmospheric fronts, they are basically based on semantic segmentation networks. There is also a defect in semantic segmentation networks, that is, they can only classify the fronts for each grid point, but cannot provide which specific front the frontal grid points belong to.

[0024] The following will, in conjunction with the accompanying drawings, elaborate on the technical solutions provided by each embodiment of the present invention.

[0025] Figure 1 It is a schematic flowchart of a method for identifying an atmospheric front in the present invention, specifically including the following steps: S101: Obtain six types of meteorological elements in the area to be identified; the six types of meteorological elements include temperature, pressure, humidity, wind direction, wind speed, and dew point potential temperature.

[0026] The six types of meteorological elements in the area to be identified specifically include the temperature, pressure, humidity, wind direction, wind speed, and dew point potential temperature in the area to be identified.

[0027] S102: Fuse the data of the six types of meteorological elements through the 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 in sequence.

[0028] In an exemplary embodiment, six types of meteorological element data are fused through a fused meteorological element module to obtain a comprehensive feature map of the six types of meteorological elements, specifically including: linearly normalizing the six types of meteorological element data through a normalization layer; assigning different weights to the six types of meteorological elements after normalization at each grid point through a weight layer; and processing the six types of meteorological elements after weight assignment through a double convolutional layer to obtain a comprehensive feature map of the six types of meteorological elements.

[0029] Specifically, first, the most important part of the six types of meteorological element data in EAR5 is linearly Min-Max normalized through a normalization layer, and the normalization calculation formula is as shown in formula (1): (1); where, 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.

[0030] Next, the weight layer is used to assign different weights to the six types of meteorological elements at each grid point, and the weights are continuously adjusted using labels during training to solve the problem that when the six types of meteorological elements are in special weather conditions and special geographical locations, strong temperature features that are not frontal are recognized as frontal features, while the meteorological element features of weak fronts are not recognized. The calculation formula for the weights of the six types of meteorological elements at each grid point is as shown in formula (2): (2); where, m is the m th type of meteorological element, is the m th type of meteorological element at the grid point position ij in the area to be recognized, is the m th type of meteorological element at the grid point position ij in the area to be recognized,

[0031] Finally, the comprehensive features of the 6 types of meteorological elements are obtained through 2 double convolutional operations and activation operations: 2 groups of fully convolutional layers and the ReLU activation function are connected in series, and each group of fully convolutional layers is continuously convolved 2 times. The convolutional kernel of the convolutional operation is 3, the stride is 2, and the padding is 0. Another function of the fused meteorological element module is to convert the input 6-channel meteorological element data into 3-channel output.

[0032] Adopting a normalization processing method, the output result of the fused meteorological element module is normalized to the color space of 0-255.

[0033] S103: Divide the area to be recognized according to the comprehensive feature map of six types of meteorological elements, and generate multiple suspected frontal area frames.

[0034] In an exemplary embodiment, dividing the area to be recognized according to the comprehensive feature map of six types of meteorological elements and generating multiple suspected frontal area frames specifically includes: based on the comprehensive feature map of six types of meteorological elements, segmenting the area to be recognized to obtain multiple sub-areas; calculating the similarity of all adjacent sub-areas among the multiple sub-areas to obtain a sub-area similarity set; merging the two sub-areas with the largest 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 frontal area frames.

[0035] Specifically, the process of dividing the area to be recognized mainly includes: Step 1, area segmentation: According to the comprehensive characteristics of 6 types of meteorological elements, use the algorithm of Felzenswalb and Huttenlocher for the entire area to be recognized to generate multiple small areas, achieving the purpose of preliminary area segmentation.

[0036] Step 2, similarity calculation: For the segmented sub-areas, calculate the similarity of adjacent sub-areas, and perform set operations according to the similarity, thereby obtaining a similarity set of sub-areas.

[0037] First, calculate the color similarity between adjacent sub-areas: Use L1-norm normalization to evenly divide the color space of 0-255 into 25 small color spaces, that is, histograms, so that each area obtains a 75 (25x3)-dimensional vector. The calculation formula for the color similarity between adjacent sub-areas is as shown in formula (3): , (3); Among them, are 2 adjacent sub-areas, is the histogram for calculating the new area during the sub-area merging process, k is the k th grid point in 2 color adjacent sub-areas, n is the number of grid points in the color adjacent sub-area, is the k th grid point in the color adjacent sub-area, is the k th grid point in the color sub-area, is the color sub-area, is the color adjacent sub-area, is the color similarity between adjacent sub-regions.

[0038] Secondly, calculate the texture similarity between adjacent sub-regions: Calculate the Gaussian differential with variance σ = 1 in 8 different directions for each color channel, and use L1-norm normalization to obtain 10 histograms for each direction of each color channel of the image, so as to obtain a 240 (10x8x3)-dimensional vector , the calculation formula of the texture similarity between sub-regions is as shown in formula (4): (4); Among them, is the texture similarity between sub-regions, k is the k th grid point in the texture adjacent sub-region, n is the number of grid points in the texture adjacent sub-region, is the k th grid point in the texture sub-region, is the k th grid point in the texture adjacent sub-region, is the i th sub-region, is the j th sub-region.

[0039] The calculation method of the texture feature of the new region after merging is the same as that of the color feature.

[0040] Then, calculate the similarity of the small region size to preferentially merge small regions: In order to make the multi-scale application global, give more weights to small regions to ensure that it is multi-scale at each position of the image before merging. The calculation formula of the region size similarity is as shown in formula (5): (5); Among them, is the region size similarity, is the number of grid points in the region, is the i th sub-region, is the j th sub-region.

[0041] Finally, calculate the fitness distance of the small region: If the region is included, merge it first. If it is difficult to connect, do not merge them together. Defining the fitness distance of the region is mainly to measure whether two regions are more "fitting". The index is the minimum rectangle that can enclose the region after merging Bounding Box, The smaller it is, the higher the degree of fitting, that is, the closer the similarity is to 1. The calculation method of the fitness distance of the small region is as shown in formula (6): (6); Among them, is the fitness distance, is the smallest rectangle that can enclose the area, is the number of grid points in the area, is the i sub-region, is the j sub-region.

[0042] Calculate the sum of the above four types of similarities: Calculate and obtain the sum of the above four types of similarities. The calculation method of the sum of the above four types of similarities is shown in formula (7): (7); Among them, is the sum of the four types of similarities, , is the weight parameter of the color similarity, is the weight parameter of the texture similarity, is the weight parameter of the area size similarity, is the weight parameter of the fitness of the area.

[0043] Step 3, region merging. Merge the two sub-regions with the largest similarity in the similarity set to form a new sub-region, and then update the similarity set of the sub-regions again.

[0044] Step 4, iterative merging, sorting and output. Repeat the above steps 2 and 3 until the similarity set of the sub-regions is empty. At this time, the obtained region set is the final segmentation region of the frontal recognition region. Through the operation of step 4, the images of the comprehensive characteristics of the six meteorological elements generate 1000 suspected frontal region boxes.

[0045] Specifically, reduce the size of the suspected frontal region box. After reducing the region box of the suspected frontal surface to a size of 227 * 227, input it into the ALexNet network.

[0046] S104: Screen multiple suspected frontal region boxes and remove the overlapping suspected frontal region boxes to obtain multiple frontal proposal boxes.

[0047] Obtain the fifth-generation reanalysis dataset of the frontal recognition region, and divide the fifth-generation reanalysis dataset into a training set, a validation set, and a test set; the fifth-generation reanalysis dataset includes six meteorological elements.

[0048] Taking the identification of the frontal surface in the research area as an example, first, download the frontal surface images. Second, use software to create labels for training the network. Use the image annotation tool labelme to annotate the downloaded frontal surface images, and then convert the annotated frontal surface images into the COCO (Common Objects in Context) data format for image recognition and annotation.

[0049] The input data for the network to extract the frontal surface bounding box is as follows: Download the fifth-generation reanalysis dataset ERA5 of the research area, with a longitude range from 73°E to 135°E and a latitude range from 3°N to 53°N. For correspondence with the labels, both longitude and latitude are rounded. The network input is the dew point temperature at 2 meters, the temperature at 1000 hPa height, U-wind, V-wind, specific humidity, and sea-level mean pressure data, with a horizontal resolution of 0.25°×0.25° and a time interval of 3 hours.

[0050] Divide the obtained ERA5 dataset corresponding to the labels into a training set, a validation set, and a test set.

[0051] Construct a frontal surface recognition model, as Figure 2 shown is the frontal surface recognition model provided by the present invention. The frontal surface recognition model includes a meteorological element fusion module and a network for extracting the frontal surface bounding box. The meteorological element fusion module includes a normalization module, a weight layer module, a double convolutional layer module, and a normalization module. The network for extracting the frontal surface bounding box includes a selective search module, an AlexNet module, an SVM module, an NMS module, and a regressor module.

[0052] Specifically, the process of training the frontal surface recognition model is as follows: Use the training set to train the frontal surface recognition model, and use the trained frontal surface recognition model as the network model for frontal surface recognition applications. The number of training epochs (Epoch) for the network for extracting the frontal surface bounding box is set to 100. The optimizer uses the Stochastic Gradient Descent (SGD) method, the learning rate of the network parameters is set to 0.0001, and the momentum value is set to 0.7.

[0053] The method for training the frontal surface recognition model is as Figure 3 shown. The main steps are as follows: After every 10 training epochs, select the model corresponding to the network parameters with the minimum loss, calculate the average loss on the validation set, and save the parameters of the network model at this time. After 100 epochs, among the 10 network models, select the network model parameters with the minimum average loss on the test set as the parameters of the trained frontal surface recognition model, and use the trained frontal surface recognition model in the recognition stage of engineering practice.

[0054] The loss function of the network is obtained by comparing the predicted value and the true value. The loss function is as shown in formula (8): (8); Wherein, Loss is the loss function value, represents x, y, w, h (the coordinates of the frontal boundary box), x, y is the center point coordinate of the frontal boundary box, w is the width of the frontal boundary box, h is the height of the frontal boundary box, i is the i th box, N is the total number of frontal boundary boxes, is the translation amount of the frontal boundary box, is the feature vector, is the parameter to be learned, is regularization.

[0055] In an exemplary embodiment, a plurality of suspected frontal region boxes are screened, and overlapping suspected frontal region boxes are removed to obtain a plurality of frontal proposal boxes. Specifically, it includes: using 4 support vector machines to perform probability prediction of four types of fronts on a plurality of suspected frontal region boxes; sorting each column of the plurality of suspected frontal region boxes after probability prediction in descending order according to the intersection over union (IoU) score; the column represents the category of the front; for each column in descending order, starting from the column with the maximum score, calculate the IoU between the column with the maximum score and the columns behind the column with the maximum score to obtain the IoU calculation result; if the IoU calculation result is greater than the preset threshold, then remove the columns with smaller IoU scores in the column with the maximum score and the columns behind the column with the maximum score; after processing all columns of the plurality of suspected frontal region boxes in sequence, a plurality of frontal proposal boxes are obtained.

[0056] Specifically, to remove overlapping suspected frontal region boxes: use the non-maximum suppression (NMS) method to remove overlapping suspected frontal region boxes to obtain the best suspected frontal region boxes, and use these region boxes as frontal proposal boxes. The main steps are as follows: (1) Sort each column of the 1000×4-dimensional suspected frontal region boxes in descending order, where each column represents one category and there are four categories.

[0057] (2) Starting from the region box of the suspected frontal surface with the maximum score in each column, calculate the IoU with the region boxes of the suspected frontal surfaces with scores behind this column respectively. If IoU > the preset threshold, and the preset threshold is set according to specific engineering practices. In the present invention, the preset threshold is taken as 0.4, then remove the suspected frontal region boxes with smaller scores, otherwise it is considered that there are multiple targets of the same type of front in the image.

[0058] (3) Starting from the suspected frontal region box with the second largest score in each column, repeat step (2).

[0059] (4) Repeat step (3) until all suspected frontal area boxes in this column are traversed.

[0060] (5) After traversing all columns of the 1000×4-dimensional suspected frontal area boxes, that is, perform non-maximum suppression on all frontal types.

[0061] Specifically, use the AlexNet network without the Softmax layer to extract the 4096-dimensional feature vectors of the suspected frontal area boxes. The AlexNet network is as Figure 4 shown, including 1 input layer, 5 convolutional layers (C1, C2, C3, C4, C5), 2 fully connected layers (FC6, FC7) with 4096 neurons each, and 1 output layer.

[0062] Predict the probabilities of the four types of fronts for each column of the 1000*4-dimensional matrix output by the AlexNet network: Use 4 Support Vector Machine (SVM) classifiers to predict the probabilities of the four types of fronts for each column of the 1000×4-dimensional matrix output by the AlexNet network. The learning algorithm of the non-linear SVM as the Gaussian radial basis function classifier is as follows: Input: The training data set is set as T = {( , ), ( , ),..., , ).

[0063] Output: The separating hyperplane and the classification decision function.

[0064] Select the kernel function K (x,z) and the penalty parameter C > 0, construct and solve the convex quadratic programming problem. The feature mapping space is as shown in formula (9): (9); s.t. ; 0 , i = 1, 2,...., N ; where, is the penalty parameter, i is the abscissa of the 2D space data, j is the ordinate of the 2D space data, is the data value of multiple meteorological elements, is the penalty parameter, is the penalty parameter,N is the number of data sets.

[0065] The optimal solution is .

[0066] Calculate: Select a component of that meets the condition , calculate .

[0067] Classification decision function: The Gaussian kernel function is shown in formula (10): (10); where x is a sample point in the input space, i.e., the feature vector, z is the center point of the Gaussian kernel function, is a parameter, K is the Gaussian kernel function.

[0068] The classification decision function is shown in formula (11): (11); where x is the feature vector, i is the ordinal number, N is the number of variables, are variables, z is a variable, is a parameter, is the classification decision function.

[0069] S105: For each of the multiple frontal suggestion boxes, perform category correction on the frontal suggestion box through regression operation, and determine the frontal suggestion box with the highest frontal category probability score after category correction as the frontal bounding box.

[0070] In an exemplary embodiment, performing category correction on the frontal suggestion box through regression operation specifically includes: performing regression operation on the frontal 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 value of the objective function as the corrected category.

[0071] Specifically, after performing category correction on the category of the frontal suggestion box, determine the frontal suggestion box with the highest category probability score as the frontal bounding box.

[0072] Specifically, perform the regression operation: use the regressor to perform regression operation on the frontal suggestion boxes of 4 categories. When the regressor corrects the frontal suggestion box, the translation amount ( , ) and scale scaling ( , ).

[0073] Translation amount ( , ) is shown in formula (12): , (12); wherein, is the translation amount of the abscissa of the center point of the front surface suggestion box, is the translation amount of the ordinate of the center point of the front surface suggestion box, is the abscissa of the center point of the front surface suggestion box of the true value, is the abscissa of the center point of the front surface suggestion box of the predicted value, is the width of the front surface suggestion box of the predicted value, is the ordinate of the center point of the front surface suggestion box of the true value, is the ordinate of the center point of the front surface suggestion box of the predicted value, is the height of the front surface suggestion box of the predicted value.

[0074] Scale scaling ( , ) is shown in formula (13): , (13); wherein, is the width translation amount of the front surface suggestion box, is the height translation amount of the front surface suggestion box, is the width of the front surface suggestion box of the true value, is the width of the front surface suggestion box of the predicted value, is the height of the front surface suggestion box of the true value, is the height of the front surface suggestion box of the predicted value.

[0075] wherein, G is the true value (Ground Truth) of the front surface, P is the CNN feature corresponding to the suggestion box.

[0076] Modify the front surface suggestion box using the objective function: The objective function is shown in formula (14): (14); wherein, is the feature vector composed of the height and width prediction values of the front surface suggestion box obtained by the regression operation, is the learning parameter, is the objective function.

[0077] Take the frontal suggestion box with the highest corrected score for each category as the frontal bounding box.

[0078] S106: Remove the frontal bounding boxes with the number of grid points less than the threshold from multiple frontal bounding boxes, and extract the atmospheric front from the remaining frontal bounding boxes.

[0079] In an exemplary embodiment, removing the frontal bounding boxes with the number of grid points less than the threshold from multiple frontal bounding boxes and extracting the atmospheric front from the remaining frontal bounding boxes specifically includes: for each of the multiple frontal bounding boxes, calculate the number of grid points within the frontal bounding box, and if the number of grid points is less than the preset threshold, remove the frontal bounding box; for each of the remaining frontal bounding boxes after the removal operation, calculate the dew point potential temperature gradient of each grid point within the remaining frontal bounding box to obtain the dew point potential temperature gradient value; select the grid point with the minimum dew point potential temperature gradient value in each row within the frontal bounding box; connect the grid points with the minimum dew point potential temperature gradient value in each row as the atmospheric front.

[0080] Specifically, the threshold is set according to specific engineering practices.

[0081] Specifically, remove the frontal bounding boxes with a small range: calculate the number of grid points within each frontal bounding box, and if the number of grid points within the bounding box is less than 6, remove this bounding box.

[0082] Extract the front within the frontal bounding box: First, calculate the gradient of the dew point potential temperature of each grid point within each frontal bounding box, then select the grid point with the minimum dew point potential temperature gradient value in each row within the frontal bounding box, and finally connect the grid points with the minimum dew point potential temperature gradient value in each row as the atmospheric front.

[0083] When applying the method for identifying an atmospheric front provided by the present invention, it is not necessary to execute according to Figure 1 the order of the steps shown. The specific execution order of each step can be determined as needed, and the present invention does not limit this.

[0084] The above is a method for identifying an atmospheric front provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for identifying an atmospheric front, as Figure 5 shown.

[0085] Figure 5 is a schematic diagram for identifying an atmospheric front provided by the present invention, including: An acquisition module 501, configured to acquire six types of meteorological elements of the area to be identified; the six types of meteorological elements include temperature, pressure, humidity, wind direction, wind speed, and dew point potential temperature.

[0086] The fusion module 502 is configured to fuse 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 includes a normalization layer, a weight layer, and a double convolution layer connected in series in sequence.

[0087] The division module 503 is configured to divide the area to be recognized according to the comprehensive feature map of the six types of meteorological elements, and generate multiple suspected frontal area frames.

[0088] The screening module 504 is configured to screen multiple suspected frontal area frames, and remove the overlapping suspected frontal area frames to obtain multiple frontal suggestion frames.

[0089] The correction module 505 is configured to, for each of the multiple frontal suggestion frames, perform category correction on the frontal suggestion frame through a regression operation, and determine the frontal suggestion frame with the highest frontal category probability score after category correction as the frontal boundary frame.

[0090] The extraction module 506 is configured to remove the frontal boundary frames with the number of grid points less than the threshold from the multiple frontal boundary frames, and extract the atmospheric front from the remaining frontal boundary frames.

[0091] For the specific limitations of an apparatus for identifying an atmospheric front, reference may be made to the limitations of a method for identifying an atmospheric front in the foregoing text, which will not be elaborated herein. Each module in the foregoing apparatus for identifying an atmospheric front may be implemented in whole or in part by software, hardware, and their combination. The foregoing modules may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the foregoing modules.

[0092] The present invention also provides a computer-readable storage medium, which stores a computer program, and the computer program can be used to execute the Figure 1 provided method for identifying an atmospheric front.

[0093] The present invention also provides Figure 6 a schematic structural diagram of the computer device shown in Figure 6 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, other hardware required for other services may also be included. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the Figure 1 provided method for identifying an atmospheric front.

[0094] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memories. 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. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0095] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope recorded by the present invention.

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 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 series in sequence; 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 multiple suspected front area frames and removing overlapping suspected front area frames to obtain multiple front suggestion frames; For each of the multiple front suggestion boxes, the category of the front suggestion box is corrected through a regression operation, and the front suggestion box with the highest front category probability score after the category correction is determined as the front boundary box; The frontal bounding boxes having the number of grid points less than a threshold are removed from the multiple frontal bounding boxes, and the atmospheric front is extracted from the remaining frontal bounding boxes.

2. The method according to claim 1, characterized in that, 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, which specifically includes: The six types of meteorological element data are linearly normalized through the normalization layer; The weight layer is used to assign different weights to each grid point for the six types of meteorological elements after normalization; The six types of meteorological elements after weight allocation are processed through a double convolutional layer to obtain the comprehensive feature maps of the six types of meteorological elements.

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

4. The method according to claim 1, characterized in that, The step of dividing the area to be identified according to the comprehensive characteristic map of six types of meteorological elements and generating multiple suspected front area frames specifically includes: Based on the comprehensive characteristic map of six types of meteorological elements, the area to be identified is segmented to obtain multiple sub-areas; Calculate the similarities of all adjacent sub-regions in multiple sub-regions to obtain a sub-region similarity set; Merging two sub-regions with the greatest similarity in the sub-region similarity set to obtain a new sub-region; After obtaining the new sub-region, the sub-region similarity set is updated until the sub-region similarity set is empty, and the sub-region corresponding to the sub-region similarity set being empty is determined as a plurality of suspected front region frames.

5. The method according to claim 1, characterized in that, 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; After the probability prediction is performed, each column of the plurality of suspected front region frames is sorted from large to small according to the intersection-over-union ratio score; the column represents the category of the front; For each column in order from large to small, starting from the maximum score column, calculate the intersection and union ratio of the maximum score column and the columns following the maximum score column to obtain the intersection and union ratio calculation result; If the intersection-and-union ratio calculation result is greater than a preset threshold, the maximum score column and the columns after the maximum score column with a smaller intersection-and-union ratio score are removed; After all columns of multiple suspected front area boxes are processed in sequence, multiple front suggestion boxes are obtained.

6. The method according to claim 1, characterized in that, The category correction of the front suggestion box through the regression operation specifically includes: Perform a regression operation on the frontal proposal box through a regressor; the regression operation includes a translation operation and a scale scaling operation; the translation amount corresponding to the translation operation is: ， ； Among them, is the translation amount of the abscissa of the center point of the frontal suggestion box, is the translation amount of the ordinate of the center point of the frontal suggestion box, is the abscissa of the center point of the frontal suggestion box of the true value, is the abscissa of the center point of the frontal suggestion box of the predicted value, is the width of the frontal suggestion box of the predicted value, is the ordinate of the center point of the frontal suggestion box of the true value, is the ordinate of the center point of the frontal suggestion box of the predicted value, is the height of the frontal suggestion box of the predicted value; The formula corresponding to the scale scaling operation is: ， ； Among them, is the width translation amount of the frontal suggestion box, is the height translation amount of the frontal suggestion box, is the width of the frontal suggestion box of the true value, is the width of the frontal suggestion box of the predicted value, is the height of the frontal suggestion box of the true value, is the height of the frontal suggestion box of the predicted value; Construct an objective function and use the category corresponding to the maximum objective function value as the corrected category; the objective function is: ； Among them, is the feature vector composed of the predicted height and width values of the frontal surface suggestion box obtained by the regression operation, is the learning parameter, is the objective function.

7. The method according to claim 1, characterized in that, Removing the frontal bounding boxes with the number of grid points less than the threshold from multiple frontal bounding boxes and extracting the atmospheric front from the remaining frontal bounding boxes specifically includes: For each of the multiple frontal bounding boxes, calculate the number of grid points within the frontal bounding box. If the number of grid points is less than the preset threshold, remove the frontal bounding box; For each of the remaining frontal bounding boxes after the removal operation, calculate the dew point potential temperature gradient of each grid point within the remaining frontal bounding box to obtain the dew point potential temperature gradient value; Select the grid point with the minimum dew point potential temperature gradient value in each row within the frontal bounding box; Connect the grid points with the minimum dew point potential temperature gradient value in each row as the atmospheric front.

Images