A method for automatic identification of weather fronts based on deep separable convolutional networks

Through the method of separable convolutional network, expanding labels and calculating meteorological element gradients, extracting and fusion features, combining adaptive weights and cross entropy loss functions, the identification bias and noise problems in the existing front automatic analysis method are solved, and the problem of improvement of the identification of the front is improved. The accuracy and robustness of automatic front recognition.

CN115082791BActive Publication Date: 2024-11-01NAT UNIV OF DEFENSE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210714730.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-22
Publication Date
2024-11-01
Estimated Expiration
2042-06-22

AI Technical Summary

Technical Problem

The existing front automatic analysis method has single judgment rules and noise when calculating gradient derivatives. The network structure is too simple, the misjudgment rate of recognition results is high and the generalization ability is weak, especially the improper handling of label class imbalance problems leads to the non-recognition results. Front type.

Method used

The automatic recognition method of meteorological fronts based on the depth separable convolution network is adopted. Gradient features, downsampling, fusion and upsampling features are extracted by extending labels, calculating meteorological factor gradients, and deep separable convolutions. The front probability is predicted by combining the Softmax function. In network training, adaptive weights and cross entropy loss functions are used to eliminate small area fronts and finally extract the front skeleton line.

Benefits of technology

It improves the accuracy and robustness of automatic front recognition, reduces the computational complexity, enhances the reliability and generalization ability of recognition results, and effectively solves the identification deviation and noise problems in existing methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115082791B_ABST
    Figure CN115082791B_ABST
Patent Text Reader

Abstract

The present invention provides a method for automatic identification of meteorological fronts based on a deep separable convolutional network, including: extending labels; calculating the gradient of each meteorological element; the encoding module performs a deep separable convolution operation to extract the gradient features of each meteorological element, and uses the maximum pooling to downsample the gradient features; fuses the gradient features; the decoding module upsamples the gradient features of the fused gradient abstract features; the Softmax function predicts the probability of the grid front; the network is trained by adding sample adaptive weights based on the multi-classification cross entropy loss function; the Max function is used to determine the front category of each grid point; the small front area is eliminated, and the skeleton line algorithm is used to determine the final front. Through the above method, the present invention can replace the manual identification of the front, and can automatically, efficiently and accurately identify the meteorological front.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of ground meteorological observation in atmospheric detection, and in particular relates to an automatic meteorological front recognition method based on a deep separable convolutional network. Background Art

[0002] In atmospheric science, a front is defined as the interface or transition zone between two air masses of different densities. A front is mainly formed by the meeting of two air masses with different thermodynamic properties in the convergence area of ​​airflow. The atmospheric baroclinicity near it is large, which is conducive to the development of vertical circulation and energy conversion. Therefore, there are often more drastic weather changes and the occurrence and development of air pressure systems near the front. For example, the area near the frontal system is often accompanied by severe weather phenomena such as heavy precipitation and storms, and is often the center of severe thunderstorms. Therefore, the positioning and analysis of the front helps forecasters to grasp the location, intensity and evolution of the frontal system, and accurately predict the changing trend of the frontal system and weather, which is of great significance to weather forecasting.

[0003] The existing automatic front analysis methods are mainly based on numerical front analysis (NFC), which some researchers define as grid processing without machine learning. Most numerical front analysis is based on manually set rules to analyze thermal parameters. For example, NOAA 2013, p.6 determines a threshold for rapid changes in the gradient of an input parameter for front analysis. The main disadvantage of the NFC method is that different researchers use different single input parameters and single rules (some also consider multiple quantities in more complex methods). The single input parameter does not consider the systematic nature of frontal weather; the single rule, such as determining the area with the largest parameter gradient change as the front, also has great defects. First, the gradient depends on the resolution of the grid data; second, there is noise when calculating the derivative on a finite grid; third, there is a strong systematic bias. These defects ultimately lead to a high misjudgment rate of its recognition results and weak generalization ability of the method.

[0004] In recent years, deep learning has also been applied to feature analysis in the meteorological field. For example, McGovern et al. 2019; Lagerquist et al. 2020; Biard et al. 2019 used fully convolutional neural networks (CFN) for automatic frontal recognition. Compared with the NFC method, deep learning integrates multiple meteorological elements and replaces the manually set rules with trained neural networks, which improves the accuracy of frontal recognition. However, there are some aspects that need to be improved in the deep learning methods used for automatic frontal recognition, such as: First, the structures of several FCN methods currently used are relatively simple, with a maximum of 8 layers of network structure, which cannot fit complex frontal weather systems well, and their recognition results and generalization capabilities are still poor; second, during network training, the class imbalance problem of labels was not scientifically handled, resulting in the recognition results being biased towards the no-front class.

[0005] At present, the intelligence level of automatic front identification is generally low, with poor reliability and weak robustness. Further research and improvement are needed to continuously improve its reliability, robustness and identification accuracy. Summary of the invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is that there are defects such as a single judgment rule in numerical front analysis and noise in the calculation of gradient derivatives; when using a fully convolutional neural network to identify the front, its network structure is too simple, and there are a large number of scattered small front areas in the results of identifying the front, and the class imbalance problem of the label is not scientifically handled during network training, resulting in the recognition result being biased towards the no-front class. In order to solve the above technical problems, a technical solution adopted by the present invention is: to provide a method for automatic identification of meteorological fronts based on a deep separable convolutional network, which includes the following steps:

[0007] Step 1, expand the tag;

[0008] Step 2, calculate the gradient of meteorological elements: the network input is 5 types of meteorological elements: potential temperature, specific humidity, U wind, V wind and sea level average pressure data; calculate the gradient of these 5 types of meteorological elements at each grid point respectively;

[0009] Step 3: Perform deep separable convolution to extract gradient features and downsample gradient features: Input the gradient data of the five types of meteorological elements into the corresponding five encoders, first perform deep separable convolution on the gradient of each meteorological element to extract gradient features, then perform maximum pooling downsampling operation, perform these two steps four times in a row, and output a highly abstract gradient feature map, whose size is 1 / 16 of the original;

[0010] Step 4, gradient feature fusion: first concatenate the output results of the five encoders in the third dimension, then input them into the fusion feature module for gradient feature fusion, and output the gradient fusion abstract feature map;

[0011] Step 5, upsampling gradient features: the fused abstract gradient feature map is input into the decoding module. After the decoding module performs four deconvolution upsampling on the input fused abstract gradient feature map, the output feature map has the same size as the original data. Finally, the convolution operation is performed on the convolution layer with a convolution kernel of 1 and a step size of 1.

[0012] Step 6, predict the front probability of the grid point: input the upsampled gradient feature map into the Softmax function to obtain a 3D probability map of each grid point in the recognition area for five types of fronts (no front, cold front, warm front, quasi-stationary front, and occluded front);

[0013] Step 7, network training: During network training, the network parameters are initialized first, and then the adaptive weights are calculated for each type of peak in the training sample. Then, the sample weighted cross entropy loss function is used to calculate the loss and backpropagate. Finally, the Adam optimizer is used to optimize the parameters and train the network. The learning rate of the stochastic gradient is set to 0.001.

[0014] Step 8, determine the frontal category of the grid point;

[0015] Step 9, determine the final front: first remove the small area fronts in the front area identified by the network, and then perform skeleton line extraction to determine the final front.

[0016] In step 1, the extension object is the non-frontal grid points that are adjacent to the frontal grid points in the eight directions; then, the non-frontal grid points to be extended in the label are divided into three categories: 1. Non-frontal grid points where the frontal grid points are not adjacent and the number of non-frontal grid points between the two grid points is greater than or equal to 2, 2. Non-frontal grid points where the frontal grid points are not adjacent and there is only one non-frontal grid point between the two grid points, 3. Non-frontal grid points where the frontal grid points are adjacent; finally, the non-frontal grid points in these three categories are extended according to the rules; specifically, the following steps are included:

[0017] Step 1-1, expanding the grid points without fronts where the grid points with fronts are not adjacent and the number of grid points without fronts between the two grid points is greater than or equal to 2: expanding the grid points without fronts adjacent to the grid points with fronts in 8 directions, vertically, horizontally and diagonally, to the same front type;

[0018] Step 1-2, expand the grid points with fronts that are not adjacent and there is only one grid point without fronts between the two grid points: expand the grid point without fronts between two grid points of different types of fronts into priority fronts, and the priority levels from high to low are cold front, warm front, quasi-stationary front, and occluded front;

[0019] Step 1-3, expand the non-frontal grid points adjacent to the frontal grid points: expand the non-frontal grid points adjacent to the frontal grid points on the same side to the same type of front.

[0020] Step 3 includes the following steps:

[0021] Step 3-1, depth-wise separable convolution to extract gradient features: In the encoder (the depth-wise separable convolution layer and the maximum pooling layer are collectively referred to as the encoder), every two depth-wise separable convolution layers and one maximum pooling layer form a group, and there are a total of 9 depth-wise separable convolution layers and 4 maximum pooling layers connected in series. Each depth-wise separable convolution layer is followed by a ReLU (rectified linear unit) activation operation; the depth-wise separable convolution operation is performed in two steps. The first step is a channel-by-channel convolution operation, which convolves the gradient data on each channel separately, with the convolution kernel size set to 3×3, the step size set to 2, and the padding mode set to 1; the second step is a point-by-point convolution operation, with the convolution kernel size set to 1×1, the step size set to 1, and the padding mode set to 0;

[0022] Step 3-2, downsampling gradient features: first set the encoder's pooling mode to the maximum pooling mode, set the kernel size to 2×2, the step size to 2, and the padding mode to 1; then input the gradient data of each meteorological element into the encoder, set the input channel of the first convolution layer to 1, the number of output channels to 16, and the data resolution to 256×512 for depthwise separable convolution operation; set the number of input and output channels of the second convolution layer to 16, and the data resolution to 256×512 for depthwise separable convolution operation; then input the data into the first pooling layer for the first downsampling, and the number of input channels is 256×512. The number of input channels is set to 16, the data resolution is set to 256×512, and the number of output channels is set to 16. The data resolution is set to 128×256; the number of input channels of the third convolutional layer is set to 16, the number of output channels is set to 32, and the data resolution is set to 128×256 for depthwise separable convolution operation; the number of input and output channels of the fourth convolutional layer is set to 32, and the data resolution is set to 128×256 for depthwise separable convolution operation, and then the data is input into the second pooling layer for the second downsampling, the number of input channels is set to 32, the data resolution is set to 128×256, and the number of output channels is set to 3 2, the data resolution is set to 64×128; the input channel number of the fifth convolutional layer is set to 32, the output channel number is set to 64, and the data resolution is set to 64×128 for depthwise separable convolution operation; the input and output channel numbers of the sixth convolutional layer are set to 64, and the data resolution is set to 64×128 for depthwise separable convolution operation, and then the data is input into the third pooling layer for the third downsampling, the input channel number is set to 64, the data resolution is set to 64×128, the output channel number is set to 64, and the data resolution is set to 32×64; the input channel number of the seventh convolutional layer is set to 64, The number of output channels is set to 128, and the data resolution is set to 32×64 for depth-wise separable convolution operation; the number of input and output channels of the eighth convolutional layer is set to 128, and the data resolution is set to 32×64 for depth-wise separable convolution operation, and then the data is input into the fourth pooling layer for the fourth downsampling, the number of input channels is set to 128, the data resolution is set to 32×64, the number of output channels is set to 128, and the data resolution is set to 16×32; the number of input and output channels of the ninth convolutional layer is set to 128, and the data resolution is set to 16×32×128 for depth-wise separable convolution operation;

[0023] Step 4 includes the following steps:

[0024] Step 4-1, gradient feature fusion: the gradient abstract feature data of the five types of meteorological elements after downsampling are input into two series-connected convolutional layers for fusion, and each convolutional layer performs full convolution operation and ReLU (rectified linear unit) activation operation; the number of input channels of the first convolutional layer is set to 640, the data resolution is set to 16×32, the output data channel is set to 512, and the resolution is set to 16×32; the number of input channels of the second convolutional layer is set to 512, the data resolution is set to 16×32, the number of output channels is set to 256, and the data resolution is set to 16×32.

[0025] Step 5 includes the following steps:

[0026] Step 5-1, decoder decoding: In the decoder (the convolution layer and deconvolution layer are collectively referred to as the decoder), each deconvolution layer and two convolution layers form a group, and a total of 9 convolution layers and 4 deconvolution layers are connected in series. ReLU activation operation is performed after each convolution layer. The convolution method of the convolution layer is the same as that of the encoder, and the depth-separable convolution operation method is also used. For specific methods, please refer to step 3-2. The convolution kernel size of the deconvolution layer is set to 2×2, the step size is set to 2, and the padding method is set to 1.

[0027] Step 5-2, upsampling gradient features: input the fused feature data output from step 4 into the decoder, set the input channel number of the first deconvolution to 256, the data resolution to 16×32, the output channel number to 256, and the data resolution to 32×64×256; then input to two consecutive convolutional layers, set the input and output channel numbers to 256, and the data resolution to 32×64; the input channel number of the second deconvolutional layer is set to 256, the data resolution is set to 32×64, the output channel number is set to 128, and the data resolution is set to 64×128; then input to two consecutive convolutional layers, set the input and output channel numbers to 128, and the data resolution is set to 64×128; The number of input channels of the convolutional layer is set to 128, the data resolution is set to 64×128, the number of output channels is set to 64, and the data resolution is set to 128×256; then it is input to two consecutive convolutional layers, the number of input and output channels is set to 64, and the data resolution is set to 128×256; the number of input channels of the fourth deconvolutional layer is set to 64, the data resolution is set to 128×256, the number of output channels is set to 32, and the data resolution is set to 256×512; then it is input to two consecutive convolutional layers, the number of input and output channels is set to 32, and the data resolution is set to 256×512; the number of input channels of the last convolutional layer is set to 32, the number of output channels is set to 5, and the data resolution is set to 256×512.

[0028] In step 6, the Softmax function calculation formula is:

[0029]

[0030] where z i is the output value of the ith channel, C is the number of output channels, that is, the number of classification categories, and e is a natural constant.

[0031] Step 7 includes the following steps:

[0032] Step 7-1, use the Xavier initialization method to initialize all convolution kernel parameters. The initialization formula is as follows:

[0033] Define the input dimension of the layer where the parameter is located as n, and the output dimension as m. The parameters will be evenly distributed in Initialize within the range of and set all biases to 0;

[0034] Step 7-2, calculate the adaptive weight of the sample: according to the formula Calculate the weights of the five types of labels for each training. class represents the weight of each class in the sample, c is a hyperparameter, set to 1.02, p class is the proportion of samples of this type. Finally, w class Restricted to the range [1,50];

[0035] Step 7-3, calculate the loss: select the multi-classification cross entropy loss function, and integrate the calculated sample adaptive weights into the multi-classification cross entropy loss function to calculate the loss: The calculation formula of the multi-classification cross entropy loss function is:

[0036]

[0037] Where C represents the loss value, n is the number of samples, x is the dimension of the prediction vector, y is the true value after onehot encoding, corresponding to the label on the x dimension, which is 1 or 0, and a is the prediction label output in onehot format, which is a value between 0 and 1;

[0038] Finally, the Adam optimizer is used to optimize the parameters and train the network; the learning rate of the Adam optimizer is set to 0.001.

[0039] Step 7-4, select the optimal network: During network training, select a set of parameter networks with the highest accuracy on the test set and save it as the optimal network model.

[0040] Step 8 includes: during network prediction, the 3D probability map output by the selected optimal network is input into the Max function (the maximum front probability output by the Softmax function is selected at each grid point) to determine the front category of each grid point in the identification area.

[0041] Step 9 includes the following steps:

[0042] Step 9-1, remove small area fronts: When the network is applied to actual prediction, in the frontal area identified in step 8, the fronts of small areas are removed according to the set range threshold. The range threshold is set to 5×5 grid points when the range is 850 hPa, and the threshold is different for different vertical heights, such as 2×2 grid points when near the ground.

[0043] Step 9-2, extract skeleton lines: remove the redundant frontal pixels in the frontal area processed by step 9-1 by continuously iterating the "erosion" boundary method until only a front line with a width of 1 grid point remains. Finally, select the longest front line as the final front recognition result.

[0044] Step 9-2 includes the following steps:

[0045] Step 9-2-2, iterative "erosion" operation: judge the grid points that are frontal in the identified area one by one. First, regard the grid points of the front as 1 and the grid points that are not frontal as 0, marking the boundary of the frontal area; then take each frontal grid point that is 1 as the center, and iterate in the range of 3×3, and perform X iterations in each iteration. 1 (Generally the value is 6) checks, each cycle checks whether there are 3 adjacent points to the point with a value of 1 in the area, starting from whether there are 7 adjacent points. If there are more than 3 non-0 points adjacent, delete the point with a value of 1 in the area.

[0046] The beneficial effects of the present invention are: the present method can replace manual identification of fronts and has the advantage of automation; the convolution layer of the present method uses a depth-separable convolution operation to greatly reduce calculations and has the advantage of high efficiency; the present method is based on an analysis of the advantages and disadvantages of some current methods for automatic identification of fronts and has made many scientific technical improvements, and the accuracy of identifying fronts is relatively high. Therefore, the present method is very effective in the actual application of automatic identification of meteorological fronts and is worthy of popularization and application. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0048] Figure 1 It is a schematic flow chart of the method of the present invention.

[0049] Figure 2 It is a flow chart of downsampling and upsampling in the method of the present invention.

[0050] Figure 3 It is a schematic diagram of depth-wise separable convolution in the method of the present invention.

[0051] Figure 4It is a schematic diagram of the post-processing flow of the network output in the method of the present invention.

[0052] Figure 5 It is a schematic diagram showing that the front-free grid points in 8 directions are expanded into the same front.

[0053] Figure 6 It is a class diagram that expands the non-front grid points between two different types of front grid points into a preferred front.

[0054] Figure 7 It is a schematic diagram of expanding the non-frontal grid points adjacent to the frontal grid points on the same side to the same type of front. DETAILED DESCRIPTION

[0055] like Figure 1 As shown, the present invention provides a method for automatic identification of meteorological fronts based on a deep separable convolutional network, and the specific implementation includes the following steps:

[0056] Step 1, offline network training: divide the training sample set and the test sample set, expand the labels, input the meteorological elements and calculate the gradient of the meteorological elements at each grid point, initialize the network and training parameters, use the training sample set to train the network using the sample weighted cross entropy loss strategy, and then use the test sample set to select a set of parameter networks with the highest accuracy as the network for automatic identification of meteorological fronts;

[0057] Step 2, network online prediction: input the five types of meteorological element data in the area where front identification is required into the network obtained in step 1, and obtain the front category predicted for each grid point in the identification area. Finally, post-process the front results of each grid point identified by the network to determine the final front.

[0058] Step 1 includes the following steps:

[0059] Step 1-1, expand the label: First, expand the non-frontal grid points adjacent to the frontal grid points in 8 directions; then, divide the non-frontal grid points to be expanded in the label into 3 categories for expansion:

[0060] 1. The frontal grid points are not adjacent and the number of non-frontal grid points between the two grid points is ≥ 2, such as Figure 5 As shown, the front-free grid points in the eight directions are all expanded to the same front;

[0061] 2. If the grid points with fronts are not adjacent and there is only one grid point without fronts between the two grid points, the grid point without fronts between the two grid points of different types will be expanded into priority fronts. The priority level is cold front > warm front > quasi-stationary front > occluded front, such as Figure 6 As shown, it is expanded into the priority front class;

[0062] 3. When there are grid points with fronts adjacent to grid points without fronts, the grid points without fronts adjacent to the grid points with fronts on the same side are expanded to the same type of fronts, as follows Figure 7 Shown extensions;

[0063] Step 1-2, calculate the horizontal gradient of meteorological elements: the network input is potential temperature, specific humidity, U wind, V wind and sea level average pressure data; the horizontal gradient of these five types of meteorological elements at each grid point is calculated respectively, and the calculation formula is:

[0064]

[0065]

[0066] Where τ is the input meteorological element data;

[0067] Step 1-3, initialization of training parameters: the network is trained by the Adam optimizer, the learning rate is set to 0.001, the number of samples input each time during the training phase is set to 10, and the maximum number of iterations of network training is set to 30;

[0068] Step 1-4, read gradient data: adopt batch training method, read 10 groups of 5 types of meteorological element gradient data from step 1-2 each time;

[0069] Step 1-5, forward propagation: input the 10 groups of 5 types of meteorological element gradient data read into Figure 2 The encoder shown in A performs deep separable convolution to extract gradient features and downsamples the gradient features. Then the downsampled gradient feature data is input into the fusion module for fusion operation. The output data is input into the decoder for upsampling operation. After the Softmax function, the forward propagation result is 10 groups of 5 types of 3D probability scalars. The calculation formula of the Softmax function is:

[0070] where z i is the output value of the i-th channel, C is the number of output channels, that is, the number of classification categories;

[0071] Steps 1-6, network training strategy: During network training, first perform adaptive weight calculation on each type of peak in the training sample, then use the sample weighted cross entropy loss function to calculate the loss of both the forward propagation output result and the expanded label, then calculate the gradient of the trained network parameters according to the loss function, and finally use the stochastic gradient descent method to update the parameters;

[0072] Steps 1-5 include the following steps:

[0073] Step 1-5-1, use the Xavier initialization method to initialize all convolution kernel parameters. The initialization formula is as follows:

[0074] Define the input dimension of the layer where the parameter is located as n, and the output dimension as m. The parameters will be evenly distributed in Initialize within the range of and set all biases to 0;

[0075] Step 1-5-2, perform deep separable convolution to extract gradient features and downsample gradient features: input the gradient data of the five types of meteorological elements into the corresponding five encoders, first perform deep separable convolution on the gradient of each meteorological element in the encoder to extract gradient features, then perform maximum pooling downsampling operation, these two steps are performed four times in a row, and a highly abstract gradient feature map is output, whose size is 1 / 16 of the original;

[0076] Step 1-5-3, fusion gradient features: first concatenate the output results of the five encoders in the third dimension, then input them into the fusion feature module for gradient feature fusion, and output the gradient fusion abstract feature map;

[0077] Step 1-5-4, upsampling gradient features: the fused abstract gradient feature map is input into the decoding module. After the decoder performs four deconvolution upsampling on the input fused feature map, the output feature map has the same size as the original data. Finally, the convolution operation is performed on the convolution layer with a convolution kernel of 1 and a step size of 1.

[0078] Step 1-5-2 includes the following steps:

[0079] Step 1-5-2-1, depthwise separable convolution: In the encoder, every two depthwise separable convolution layers and one maximum pooling layer form a group, and there are a total of 9 depthwise separable convolution layers and 4 maximum pooling layers in series. Each depthwise separable convolution layer is followed by a ReLU activation operation; the depthwise separable convolution operation is performed in two steps. The first step is as follows Figure 3 As shown in the left figure, the convolution operation is performed channel by channel first, and the gradient data on each channel is convolved separately. The convolution kernel size is set to 3×3, the step size is set to 2, and the padding mode is set to 1; the second step is as follows Figure 3 As shown in the right figure, a point-by-point convolution operation is performed, the convolution kernel size is set to 1×1, the step size is set to 1, and the padding mode is set to 0;

[0080] Step 1-5-2-2, downsample gradient features: Figure 2In A, first set the pooling mode of the encoder to the maximum pooling mode, set the kernel size to 2×2, the step size to 2, and the padding mode to 1; then input the gradient data of each meteorological element into the encoder, set the input channel of the first convolution layer to 1, the number of output channels to 16, and the data resolution to 256×512 for depthwise separable convolution operation; set the input and output channel numbers of the second convolution layer to 16, and the data resolution to 256×512 for depthwise separable convolution operation; then input the data into the first pooling layer for the first downsampling, the number of input channels is set to 16, and the data resolution is set to 256×512. The resolution is set to 256×512, and the number of output channels is set to 16. The data resolution is set to 128×256; the number of input channels of the third convolutional layer is set to 16, the number of output channels is set to 32, and the data resolution is set to 128×256 for depthwise separable convolution operation; the number of input and output channels of the fourth convolutional layer is set to 32, and the data resolution is set to 128×256 for depthwise separable convolution operation, and then the data is input into the second pooling layer for the second downsampling, the number of input channels is set to 32, the data resolution is set to 128×256, the number of output channels is set to 32, and the data The resolution is set to 64×128; the input channel number of the fifth convolutional layer is set to 32, the output channel number is set to 64, and the data resolution is set to 64×128 for depthwise separable convolution operation; the input and output channel number of the sixth convolutional layer is set to 64, and the data resolution is set to 64×128 for depthwise separable convolution operation, and then the data is input to the third pooling layer for the third downsampling, the input channel number is set to 64, the data resolution is set to 64×128, the output channel number is set to 64, and the data resolution is set to 32×64; the input channel number of the seventh convolutional layer is set to 64, the input channel number is set to 64, and the output channel number is set to 64. The number of output channels is set to 128, and the data resolution is set to 32×64 for depth-wise separable convolution operation; the number of input and output channels of the eighth convolutional layer is set to 128, and the data resolution is set to 32×64 for depth-wise separable convolution operation, and then the data is input into the fourth pooling layer for the fourth downsampling, the number of input channels is set to 128, the data resolution is set to 32×64, the number of output channels is set to 128, and the data resolution is set to 16×32; the number of input and output channels of the ninth convolutional layer is set to 128 and the data resolution is set to 16×32×128 for depth-wise separable convolution operation;

[0081] Step 1-5-3 includes the following steps:

[0082] Step 1-5-3-1, fusion of gradient features: the gradient abstract feature data of the five types of meteorological elements after downsampling are input into two series-connected convolutional layers for fusion, and each convolutional layer performs full convolution operation and ReLU activation operation; the number of input channels of the first convolutional layer is set to 640, the data resolution is set to 16×32, the output data channel is set to 512, and the resolution is set to 16×32; the number of input channels of the second convolutional layer is set to 512, the data resolution is set to 16×32, the number of output channels is set to 256, and the data resolution is set to 16×32.

[0083] Step 1-5-4 includes the following steps:

[0084] Step 1-5-4-1, decoder decoding: Figure 2 As shown in B in the figure, each deconvolution layer in the decoder is grouped with two convolution layers. There are a total of 9 convolution layers and 4 deconvolution layers connected in series. ReLU activation operation is performed after each convolution layer. The convolution method of the convolution layer is the same as that of the encoder, which also uses the depth-separable convolution operation method, as shown in Figure 2 As shown in the figure, the convolution kernel size of the deconvolution layer is set to 2×2, the step size is set to 2, and the padding mode is set to 1;

[0085] Step 1-5-4-2, upsampling gradient features: Figure 3 As shown in the figure, the output fusion feature data is input into the decoder, the input channel number of the first deconvolution is set to 256, the data resolution is set to 16×32, the output channel number is set to 256, and the data resolution is set to 32×64×256; then it is input into two consecutive convolutional layers, the input and output channel numbers are set to 256, and the data resolution is set to 32×64; the input channel number of the second deconvolutional layer is set to 256, the data resolution is set to 32×64, the output channel number is set to 128, and the data resolution is set to 64×128; then it is input into two consecutive convolutional layers, the input and output channel numbers are set to 128, and the data resolution is set to 64×128; the input channel number of the third deconvolutional layer is set to 128, and the data resolution is set to 64×128. The number of channels is set to 128, the data resolution is set to 64×128, the number of output channels is set to 64, and the data resolution is set to 128×256; then it is input to two consecutive convolutional layers, the number of input and output channels is set to 64, and the data resolution is set to 128×256; the number of input channels of the fourth deconvolutional layer is set to 64, the data resolution is set to 128×256, the number of output channels is set to 32, and the data resolution is set to 256×512; then it is input to two consecutive convolutional layers, the number of input and output channels is set to 32, and the data resolution is set to 256×512; the number of input channels of the last convolutional layer is set to 32, the number of output channels is set to 5, and the data resolution is set to 256×512.

[0086] Steps 1-6 include the following steps:

[0087] Step 1-6-1, calculate the adaptive weight of the sample: according to the formula Calculate the weights of the five categories of labels for each training. Among them, c is a hyperparameter, set to 1.02, p class is the proportion of samples of this type. Finally, w class Restricted to the range [1,50];

[0088] Step 1-6-2, calculate the loss: select the multi-classification cross entropy loss function, and integrate the calculated sample adaptive weights into the multi-classification cross entropy loss function to calculate the loss. The calculation formula of the multi-classification cross entropy loss function is: Among them, C represents the loss value, n is the number of samples, x is the prediction vector dimension, y is the true value after onehot encoding, corresponding to the label on the x dimension, which is 1 or 0, and a is the prediction label output in onehot format, which is a value from 0 to 1.

[0089] Step 2 includes the following steps:

[0090] Step 2-1, read the five types of meteorological element data of each grid point in the identification area: input the five types of meteorological element data of the identification area into the optimal network trained in step 1;

[0091] Step 2-2, forward propagation: The five types of meteorological element data in the identification area are forward propagated through the optimal network to obtain a three-dimensional probability map of the five types of fronts in the identification area;

[0092] Step 2-3, determine the frontal category of each grid point: input the 3D probability map output by the network into the Max function to determine the frontal category of each grid point in the identification area;

[0093] Step 2-4, determine the final front: first remove the small area fronts in the front area identified by the network, and then perform skeleton line extraction to determine the final front.

[0094] Steps 2-4 include the following steps:

[0095] Step 2-4-1, remove small area fronts: In the frontal area identified in step 2-3, remove small area fronts according to the set range threshold. The range threshold is set to 5×5 grid points at 850 hPa, and the threshold is different for different vertical heights. For example, it can be set to 2×2 grid points when near the ground.

[0096] Step 2-4-2, extract skeleton line: Figure 4As shown, the redundant frontal pixels in the frontal area after the processing of step 2-4-1 are removed by continuously iterating the "erosion" boundary method until a front line with a width of 1 grid point remains. The specific implementation method of the iterative "erosion" operation is: judge the grid points that are frontal in the identified area one by one. First, the grid points of the front are regarded as 1, and the grid points that are not frontal are regarded as 0, marking the boundary of the frontal area; then, with each frontal grid point that is 1 as the center, the domain range is 3×3, and it is iterated. Each iteration performs 6 checks. Each cycle checks whether there are 3 adjacent points to the point that is 1 in the domain, starting from whether there are 7 adjacent points. If there are more than 3 non-0 adjacent points, the point is deleted. Finally, as Figure 4 The figure below (find the main skeleton line) shows that the longest front line is selected as the final front identification result.

[0097] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of a method for automatic identification of meteorological fronts based on a deep separable convolutional network provided by the present invention and some or all of the steps in each embodiment can be executed. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0098] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of a computer program, i.e., a software product, which can be stored in a storage medium and includes several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, a MUU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0099] The present invention provides a method for automatic identification of weather fronts based on a deep separable convolutional network. There are many methods and ways to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented using existing technologies.

Claims

1. A method for automatic identification of meteorological fronts based on deep separable convolutional networks, characterized in that: The following steps are involved: Step 1, expand the tag; Step 2, calculate the gradient of meteorological elements: the network input is 5 types of meteorological elements: potential temperature, specific humidity, U wind, V wind and sea level average pressure data; calculate the gradient of the 5 types of meteorological elements at each grid point respectively; Step 3, perform depth-separable convolution to extract gradient features and downsample the gradient features; Step 4, fusion gradient features; Step 5, upsampling gradient features; Step 6, predict the front probability of the grid point: input the upsampled gradient feature map into the Softmax function to obtain a 3D probability map of each grid point in the recognition area being one of the five types of fronts; Step 7, network training; Step 8, determine the frontal category of the grid point; Step 9, determine the final front; Step 7 includes the following steps: Step 7-1, use the Xavier initialization method to initialize all convolution kernel parameters. The initialization formula is as follows: Define the input dimension of the layer where the parameter is located as n, and the output dimension as m. The parameters will be evenly distributed in Initialize within the range of and set all biases to 0; Step 7-2, calculate the adaptive weight of the sample: according to the formula Calculate the weights of the five types of labels for each training, where w class represents the weight of each class in the sample, c is a hyperparameter, p class is the proportion of samples; finally, w class Restricted to the range [1,50]; Step 7-3, calculate the loss: select the multi-classification cross entropy loss function, and integrate the calculated sample adaptive weights into the multi-classification cross entropy loss function to calculate the loss: The calculation formula of the multi-classification cross entropy loss function is: Where C represents the loss value, n is the number of samples, x is the dimension of the prediction vector, y is the true value after onehot encoding, corresponding to the label on the x dimension, which is 1 or 0, and a is the prediction label output in onehot format, which is a value between 0 and 1; Finally, use the Adam optimizer to optimize the parameters and train the network; Step 7-4, select the optimal network: during network training, select a set of parameter networks with the highest accuracy on the test set and save it as the optimal network model; Step 8 includes: during network prediction, inputting the 3D probability map output by the selected optimal network into the Max function to determine the front category of each grid point in the identification area; Step 9 includes the following steps: Step 9-1, removing small area fronts: when applied to actual prediction, in the frontal area identified in step 8, the fronts of small areas are removed according to the set range threshold; Step 9-2, extract skeleton line: remove the redundant frontal pixels in the frontal area processed by step 9-1 by continuously iteratively corroding the boundary until only a frontal line with a width of one grid point remains. Finally, select the longest frontal line as the final frontal recognition result; Step 9-2 includes the following steps: Step 9-2-2, iterative corrosion operation: judge the grid points that are fronts in the identified area one by one. First, regard the grid points of the front as 1 and the grid points that are not fronts as 0, marking the boundary of the front area; then, take each front grid point that is 1 as the center, and iterate in the field range of 3×3. Each iteration performs X1 checks. Each cycle checks whether there are 3 adjacent points to 7 adjacent points for the points that are 1 in the field. If there are more than 3 non-0 adjacent points, delete the points that are 1 in the field.

2. The method according to claim 1, characterized in that Step 1 includes the following steps: Step 1-1, expanding the grid points without fronts where the grid points with fronts are not adjacent and the number of grid points without fronts between the two grid points is greater than or equal to 2: expanding the grid points without fronts adjacent to the grid points with fronts in 8 directions, vertically, horizontally and diagonally, to the same front type; Step 1-2, expand the grid points with fronts that are not adjacent and there is only one grid point without fronts between the two grid points: expand the grid point without fronts between two grid points of different types of fronts into priority fronts, and the priority levels from high to low are cold front, warm front, quasi-stationary front, and occluded front; Step 1-3, expand the non-frontal grid points adjacent to the frontal grid points: expand the non-frontal grid points adjacent to the frontal grid points on the same side to the same type of front.

3. The method according to claim 2, characterized in that Step 3 includes the following steps: Step 3-1, depthwise separable convolution to extract gradient features: The depthwise separable convolution layer and the maximum pooling layer are collectively called an encoder, and corresponding operations are performed in the encoder; Step 3-2, downsampling gradient features: first set the encoder's pooling method to the maximum pooling method, and then perform the corresponding operation.

4. The method according to claim 3, characterized in that Step 4 includes the following steps: In step 4-1, the gradient abstract feature data of the five types of meteorological elements after downsampling are input into two serially connected convolutional layers for fusion. Each convolutional layer performs full convolution operation and ReLU activation operation. The number of input channels of the first convolutional layer is set to 640, the data resolution is set to 16×32, the output data channel is set to 512, and the resolution is set to 16×32. The number of input channels of the second convolutional layer is set to 512, the data resolution is set to 16×32, the output channel number is set to 256, and the data resolution is set to 16×32.

5. The method according to claim 4, characterized in that Step 5 includes the following steps: Step 5-1, decoder decoding: The convolution layer and deconvolution layer are collectively called a decoder. In the decoder, each deconvolution layer and two convolution layers form a group, and then perform corresponding operations; Step 5-2, upsampling gradient features: input the fused feature data output from step 4 into the decoder for corresponding operations.

6. The method according to claim 5, characterized in that In step 6, the Softmax function calculation formula is: where z i is the output value of the i-th channel, and C is the number of output channels.