Abnormal echo recognition method and system for radar mosaic data based on semantic segmentation

Abstract

The present invention provides a method and system for identifying abnormal echoes of radar jigsaw data based on semantic segmentation, and relates to the technical field of radar jigsaw data quality control. The method comprises the following steps: constructing a radar jigsaw image data set based on radar combined reflectivity images; marking abnormal echo images in the radar jigsaw image data set to construct a radar jigsaw abnormal echo training data set; introducing a middle-level feature extraction layer and a simple attention module into a semantic segmentation network to obtain an improved semantic segmentation network; constructing a radar jigsaw data abnormal echo recognition model based on the improved semantic segmentation network; training the radar jigsaw data abnormal echo recognition model based on the radar jigsaw abnormal echo training data set, performing abnormal echo recognition on the radar jigsaw data to be recognized, and obtaining an abnormal echo recognition result. The technical solution of the present invention provides technical support for further improving the data quality and application level of radar jigsaw data.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar mosaic data quality control, and in particular relates to a method and system for identifying abnormal echoes in radar mosaic data based on semantic segmentation. Background Art

[0002] In order to effectively carry out the monitoring, warning and forecasting of severe convective weather, radar jigsaw data is often used in meteorological services. Radar jigsaw data is the observation data of multiple weather radars in the area of ​​concern at the same time point, which is formed by splicing together radar echo data covering the entire area. It is the most frequently used and important data in weather forecasting services. With the help of radar jigsaw data, meteorological service personnel can track weather changes and disaster situation development in real time, comprehensively analyze and judge weather development trends, sort out weather changes or possible disaster situations, and accurately analyze and warn potential disaster weather in advance, thereby effectively reducing the probability of disasters and reducing the extent of disaster losses. Therefore, radar jigsaw data has become an important basis for improving disaster weather defense and accurate warning issuance.

[0003] However, affected by external electromagnetic interference, atmospheric environmental noise, transmitter-receiver system failures and other factors, weather radars often generate erroneous data during business operations. Since the quality control of radar-based data cannot guarantee that all erroneous information can be completely removed, abnormal echoes will also appear in radar jigsaw data. Abnormal echoes in radar jigsaw data will not only directly interfere with the business personnel's analysis and judgment of the weather system, but will also seriously affect the quantitative application effect of radar data. In addition, for meteorological business personnel, although abnormal echoes are not difficult to identify, it is a huge workload to manually identify them one by one from the massive radar jigsaw data, and it cannot be applied efficiently, which is not conducive to the stable operation of business systems such as numerical model assimilation systems and severe convective weather automatic identification systems. For this reason, it is of great significance to realize the automatic identification of abnormal echoes in radar jigsaw data.

[0004] Traditional image interpretation methods require manual feature setting to segment images. The parameter adjustment process requires strong professional knowledge, poor generalization and robustness, low recognition rate, and low degree of automation, making them unsuitable for large-scale business applications. In addition, the abnormal echoes have many details, the targets are flaky or scattered, the contours are irregular, and the features are obviously different, which also increases the difficulty of using traditional methods to identify abnormal echoes. With the vigorous development of deep learning and big data, deep learning algorithms based on big data have achieved breakthrough results in the field of computer vision, which has solved the problems that traditional image processing algorithms cannot solve. Among them, image segmentation technology plays a vital role in the field of computer vision. By segmenting the image, we can better understand the image content and conduct more accurate analysis and recognition. Among the many image segmentation technologies, semantic segmentation is one of the key technologies. To this end, it is an effective solution to try to train a semantic segmentation network using the radar puzzle data abnormal echo image dataset and use the trained semantic segmentation model to identify the radar puzzle data abnormal echo. Summary of the invention

[0005] In order to overcome the problems of low recognition rate and low automation in the traditional weather radar abnormal echo image interpretation method, the present invention improves the DeepLabv3+ semantic segmentation network, and combines the radar puzzle abnormal echo image data set to propose a radar puzzle data abnormal echo recognition method and system based on semantic segmentation, which realizes the automatic recognition of radar puzzle data abnormal echoes and provides technical support for further improving the data quality and application level of radar puzzle data. To achieve the above purpose, the present invention provides the following scheme:

[0006] The method for identifying abnormal echoes of radar mosaic data based on semantic segmentation includes the following steps:

[0007] Drawing a radar combined reflectivity image based on the collected radar mosaic data, and constructing a radar mosaic image data set based on the radar combined reflectivity image;

[0008] Annotating the abnormal echo images in the radar jigsaw image data set, and performing format conversion and expansion and enhancement processing on the abnormal echo images to construct a radar jigsaw abnormal echo training data set; wherein the annotated abnormal echo images include annular abnormal echo images and radial abnormal echo images;

[0009] Introduce the middle-level feature extraction layer and simple attention module into the DeepLabv3+ semantic segmentation network to obtain an improved DeepLabv3+ semantic segmentation network; Based on the improved DeepLabv3+ semantic segmentation network, build a radar puzzle data abnormal echo recognition model;

[0010] The radar jigsaw data abnormal echo recognition model is trained based on the radar jigsaw data abnormal echo training data set, and abnormal echo recognition is performed on the radar jigsaw data to be recognized based on the trained radar jigsaw data abnormal echo recognition model to obtain an abnormal echo recognition result.

[0011] Preferably, the method for constructing a radar jigsaw abnormal echo training data set is:

[0012] Screening out abnormal echo images in the radar jigsaw image data set;

[0013] Using the Labelme software tool, label the screened abnormal echo images and generate a labeled data file in json format;

[0014] Convert the annotated json formatted labeled data file into a visual annotated image;

[0015] Convert the unlabeled original jpeg format abnormal echo image into a png format original image;

[0016] The visual annotation image and the original image in png format are respectively rotated, swapped left and right, and zoomed in and out to obtain the radar jigsaw image expansion and enhancement data set;

[0017] The original images in png format in the radar jigsaw image expansion and enhancement data set are converted into original images in jpeg format, and a training set and a validation set are divided to obtain the radar jigsaw abnormal echo training data set.

[0018] Preferably, the method for obtaining the improved DeepLabv3+ semantic segmentation network is:

[0019] Construct a simplified ResNet50 network as the backbone network for feature extraction; wherein the simplified ResNet50 network includes a deep feature extraction layer, a middle feature extraction layer, and a low feature extraction layer;

[0020] Adding a first simple attention module after the deep feature extraction layer to obtain deep features;

[0021] The deep features are processed by combining dilated convolutions with different dilation rates to form a dilated spatial convolution pooling pyramid to obtain multi-scale deep features;

[0022] Performing 1×1 convolution on the multi-scale deep features, and bilinearly upsampling the convolved multi-scale deep features by a factor of 4 to obtain upsampled multi-scale deep features;

[0023] Adding a second simple attention module after the low-level feature extraction layer to obtain low-level features; performing 1×1 convolution on the low-level features to obtain convolved low-level features;

[0024] Adding a third simple attention module after the middle-level feature extraction layer to obtain middle-level features; performing 1×1 convolution on the middle-level features, and bilinearly upsampling the convolved middle-level features by a factor of 2 to obtain upsampled middle-level features;

[0025] Connect the upsampled multi-scale deep features, the convolutional low-level features, and the upsampled middle-level features;

[0026] Use a 3×3 depth-separable convolutional layer to refine the connected features and obtain refined features;

[0027] Bilinearly upsampling the refined features by a factor of 4;

[0028] The upsampled refined features are convolved with a 1×1 convolution kernel that matches the number of abnormal echo categories to obtain the segmentation results, completing the improvement of the DeepLabv3+ semantic segmentation network.

[0029] Preferably, the simplified ResNet50 network structure has a total of fourteen layers; among them, the first layer is the conv1_x layer, the second to fourth layers are the conv2_x layers, the fifth to eighth layers are the conv3_x layers, and the ninth to fourteenth layers are the conv4_x layers.

[0030] Preferably, the application methods of the first simple attention module, the second simple attention module and the third simple attention module are all as follows:

[0031] Normalize the original features extracted by the backbone network;

[0032] Calculate the square of the difference between the normalized original feature and the mean of the normalized original feature to obtain a calculation result;

[0033] Based on the calculation result, obtaining an attention weight;

[0034] Attention weighting is performed on the original feature based on the attention weight to obtain the attention-weighted original feature.

[0035] The present invention also provides a radar mosaic data abnormal echo recognition system based on semantic segmentation, which is used to implement the method, including:

[0036] A data set construction module, used to draw a radar combined reflectivity image based on the collected radar mosaic data, and to construct a radar mosaic image data set based on the radar combined reflectivity image;

[0037] A labeling module is used to label the abnormal echo images in the radar jigsaw image data set, and perform format conversion and expansion and enhancement processing on the abnormal echo images to construct a radar jigsaw abnormal echo training data set; wherein the labeled abnormal echo images include annular abnormal echo images and radial abnormal echo images;

[0038] The network improvement module is used to introduce the middle-level feature extraction layer and the simple attention module into the DeepLabv3+ semantic segmentation network to obtain the improved DeepLabv3+ semantic segmentation network;

[0039] Model building module, used to build an abnormal echo recognition model for radar puzzle data based on the improved DeepLabv3+ semantic segmentation network;

[0040] The abnormal echo recognition module is used to train the radar puzzle data abnormal echo recognition model based on the radar puzzle data set, and perform abnormal echo recognition on the radar puzzle data to be recognized based on the trained radar puzzle data abnormal echo recognition model to obtain an abnormal echo recognition result.

[0041] Preferably, the marking module includes:

[0042] A screening unit, used for screening out abnormal echo images in the radar jigsaw image data set;

[0043] A labeling unit, used to label the screened abnormal echo image using the Labelme software tool to generate a label data file in json format;

[0044] The format conversion unit is used to convert the annotated json format marked data file into a visual annotated image; and convert the unannotated original jpeg format abnormal echo image into a png format original image;

[0045] A sample expansion unit, used to rotate, swap left and right, and zoom in and out the visual annotation image and the original image in png format, respectively, to obtain an expanded and enhanced data set of the radar jigsaw image;

[0046] The data set division unit is used to convert the original images in png format in the radar jigsaw image expansion and enhancement data set into original images in jpeg format, and divide them into training sets and verification sets to obtain the radar jigsaw abnormal echo training data set.

[0047] Preferably, the network improvement module includes:

[0048] A backbone network construction unit, used to construct a simplified ResNet50 network as a backbone network for feature extraction; wherein the simplified ResNet50 network includes a deep feature extraction layer, a middle feature extraction layer, and a low feature extraction layer;

[0049] A deep feature processing unit is used to add a first simple attention module after the deep feature extraction layer to obtain deep features; process the deep features by combining dilated convolutions with different expansion rates to form a dilated spatial convolution pooling pyramid to obtain multi-scale deep features; perform 1×1 convolution on the multi-scale deep features, and perform bilinear upsampling on the convolved multi-scale deep features with a 4-fold coefficient to obtain upsampled multi-scale deep features;

[0050] A low-level feature processing unit, configured to add a second simple attention module after the low-level feature extraction layer to obtain low-level features; and perform 1×1 convolution on the low-level features to obtain convolved low-level features;

[0051] A middle-level feature processing unit, configured to add a third simple attention module after the middle-level feature extraction layer to obtain middle-level features; perform 1×1 convolution on the middle-level features, and perform bilinear upsampling on the convolved middle-level features by a factor of 2 to obtain upsampled middle-level features;

[0052] The feature connection unit is used to connect the upsampled multi-scale deep features, the convolutional low-level features, and the upsampled middle-level features;

[0053] A feature refinement unit is used to refine the connected features using a 3×3 depthwise separable convolutional layer to obtain refined features;

[0054] An upsampling unit, configured to perform bilinear upsampling on the refined features by a factor of 4;

[0055] The segmentation unit is used to convolve the upsampled refined features with a 1×1 convolution kernel that matches the number of abnormal echo categories to obtain the segmentation result and complete the improvement of the DeepLabv3+ semantic segmentation network.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows: the radar puzzle data abnormal echo recognition method and system based on semantic segmentation proposed by the present invention improves the DeepLabv3+ semantic segmentation network, the model backbone network adopts a simplified ResNet50, adds an intermediate layer and adds a simple attention mechanism module SimAM after each extracted backbone feature layer, and effectively introduces additional details and backgrounds by merging the intermediate layers, which greatly improves the segmentation performance of the model; the model is trained by using the sorted radar puzzle abnormal echo training data set to obtain a recognition model that can effectively recognize the two types of annular and radial abnormal echoes that often appear in radar puzzle data, solving the problem that traditional image processing algorithms cannot accurately segment radar puzzle abnormal echoes. The technical solution provided by the present invention has strong recognition performance, high degree of automation, and simple processing flow, and can be applied to the data processing of existing radar puzzle business systems, which has practical significance for improving the data quality and application level of radar puzzle data. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0058] Figure 1 It is a flow chart of a method for identifying abnormal echoes of radar mosaic data based on semantic segmentation according to an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of the architecture of the improved DeepLabv3+ semantic segmentation network of an embodiment of the present invention;

[0060] Figure 3 A schematic diagram of a simplified version of the ResNet50 model backbone network of an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Embodiment 1

[0064] like Figure 1 As shown, the radar mosaic data abnormal echo recognition method based on semantic segmentation includes the following steps:

[0065] S1: Draw a radar combined reflectivity (CR) image based on the collected radar mosaic data, and construct a radar mosaic image dataset based on the radar combined reflectivity image. In this embodiment, the radar mosaic data of area A during the main flood season is obtained.

[0066] S2: Annotate the abnormal echo images in the radar jigsaw image dataset, perform format conversion and expansion and enhancement processing on the abnormal echo images, and construct a radar jigsaw abnormal echo training dataset; wherein the annotated abnormal echo images include annular abnormal echo images and radial abnormal echo images.

[0067] A further implementation method is that the method for constructing a radar jigsaw abnormal echo training data set is:

[0068] Filter out abnormal echo images in radar mosaic image dataset;

[0069] Use the Labelme software tool to label the screened abnormal echo images and generate a labeled data file in json format;

[0070] Convert the annotated json formatted labeled data file into a visual annotated image in png format;

[0071] Convert the unlabeled original jpeg format abnormal echo image into a png format original image;

[0072] The visual annotated images in png format and the original images in png format are rotated, swapped left and right, and zoomed in and out to obtain the radar puzzle image expansion and enhancement dataset, which includes 20,000 images.

[0073] The original images in PNG format in the radar puzzle image augmentation and enhancement dataset are converted into original images in JPEG format, and divided into training set and validation set to obtain the radar puzzle abnormal echo training dataset. Specifically, 80% of the dataset is divided into training dataset (containing 16,000 images) and 20% is divided into validation dataset (containing 4,000 images).

[0074] S3: Introduce the middle-level feature extraction layer and the simple attention module into the DeepLabv3+ semantic segmentation network to obtain an improved DeepLabv3+ semantic segmentation network; Based on the improved DeepLabv3+ semantic segmentation network, build a radar puzzle data abnormal echo recognition model; Figure 2 shown.

[0075] A further implementation method is to obtain an improved DeepLabv3+ semantic segmentation network by:

[0076] S31: Redesign the ResNet50 network as the backbone network, delete the 4th block of the ResNet50 network, use the simplified ResNet50, that is, the backbone network with a smaller depth, for feature extraction, and obtain better results on our data set; wherein, the simplified ResNet50 network includes a deep feature extraction layer, a middle feature extraction layer, and a low feature extraction layer; a further implementation method is that the simplified ResNet50 network structure has a total of fourteen layers; wherein, the first layer is the conv1_x layer, the second to fourth layers are the conv2_x layers, the fifth to eighth layers are the conv3_x layers, and the ninth to fourteenth layers are the conv4_x layers. Figure 3 shown.

[0077] In this embodiment, a simplified ResNet50 network conv1_x layer is constructed, and the first layer is a convolution operation and a maximum pooling operation, where: the convolution kernel size is 7×7, and the number of convolution kernels is 64;

[0078] Construct the conv2_x layer of the simplified ResNet50 network. The second to fourth layers are a residual block consisting of three convolution operations, where: the first convolution kernel size is 1×1 and the number of convolution kernels is 64; the second convolution kernel size is 3×3 and the number of convolution kernels is 64; the third convolution kernel size is 1×1 and the number of convolution kernels is 256;

[0079] Construct the conv3_x layer of the simplified ResNet50 network. The fifth to eighth layers are a residual block, which consists of three convolution operations, where: the first convolution kernel size is 1×1, the number of convolution kernels is 128; the second convolution kernel size is 3×3, the number of convolution kernels is 128; the third convolution kernel size is 1×1, the number of convolution kernels is 512;

[0080] Construct the conv4_x layer of the simplified ResNet50 network. The ninth to fourteenth layers are a residual block, which consists of three convolution operations, where: the first convolution kernel size is 1×1, and the number of convolution kernels is 256; the second convolution kernel size is 3×3, and the number of convolution kernels is 256; the third convolution kernel size is 1×1, and the number of convolution kernels is 1024.

[0081] S32: Add the first simple attention module (SimAM) after the deep feature extraction layer to obtain deep features;

[0082] S33: By combining dilated convolutions with different dilation rates to form a dilated spatial convolution pooling pyramid (ASPP), deep features are processed to obtain multi-scale deep features; the specific structure of ASPP is as follows:

[0083] The first branch is a depth-wise separable convolution with a kernel size of 1×1 and 256 kernels.

[0084] The second branch is a depth-wise separable convolution with a kernel size of 3×3, 256 kernels, and a dilation rate of 6.

[0085] The third branch is a depth-wise separable convolution with a kernel size of 3×3, 256 kernels, and a dilation rate of 12.

[0086] The fourth branch is a depth-wise separable convolution with a kernel size of 3×3, 256 kernels, and a dilation rate of 18.

[0087] The fifth branch is a convolution operation with a convolution kernel size of 1×1 and a number of convolution kernels of 256 after average pooling, followed by a bilinear upsampling with a coefficient of 2.

[0088] S34: performing 1×1 convolution on the multi-scale deep features of the encoder, and bilinearly upsampling the convolved multi-scale deep features by a factor of 4 to obtain upsampled multi-scale deep features;

[0089] S35: adding a second simple attention module after the low-level feature extraction layer to obtain low-level features; performing 1×1 convolution on the low-level features to obtain the convolved low-level features;

[0090] S36: Add a third simple attention module after adding an additional mid-level feature extraction layer to the encoder backbone network to obtain mid-level features; perform 1×1 convolution on the mid-level features, and bilinearly upsample the convolved mid-level features by a factor of 2 to obtain upsampled mid-level features to match the size of other reduced feature layers of the decoder.

[0091] A further implementation method is that the application methods of the first simple attention module, the second simple attention module and the third simple attention module are all as follows:

[0092] The SimAM attention mechanism seeks the linear separability between important neurons and other neurons by establishing an energy function, which is defined as:

[0093]

[0094] In the formula, and The target neuron and other neurons for a single channel input feature, is the spatial dimension coefficient, and for and The linear transformation of is the number of neurons in a single channel (H, W are characteristic height and width), is a hyperparameter;

[0095] The analytical solution of the above formula is:

[0096]

[0097]

[0098] In the formula, and For channel The mean and variance of all neurons except , so the minimum energy formula is:

[0099]

[0100] The above formula shows that the energy The lower the neuron The greater the difference between a neuron and other surrounding neurons, the more important it is to visual processing.

[0101] Normalize the original features extracted by the backbone network; specifically, obtain the number of dimension feature channels C, feature height H, and feature width W of the input feature X, and calculate the number of elements of feature X minus one, that is, M=W*H-1, for normalization;

[0102] Calculate the square of the difference between the normalized original feature X and the mean of the normalized original feature X to obtain a calculation result;

[0103] Based on the calculation results, the attention weight is obtained;

[0104] Based on the attention weight, the original features are weighted to obtain the original features after attention weighting. Specifically, the attention weight is calculated through the core formula of SimAM and the input features after attention weighting are returned, that is: , Represents all minimum energy functions sum.

[0105] S37: Connect the upsampled multi-scale deep features, the convolutional low-level features, and the upsampled middle-level features to the encoder output;

[0106] S38: Use a 3×3 depthwise separable convolutional layer (SepConv) to refine the connected features to obtain refined features;

[0107] S39: Bilinear upsampling of refined features by a factor of 4;

[0108] S310: The upsampled refined features are convolved with a 1×1 convolution kernel that matches the number of abnormal echo categories to obtain mask image segmentation results of two types of abnormal echoes, radial and annular, completing the improvement of the DeepLabv3+ semantic segmentation network.

[0109] S4: training a radar jigsaw data abnormal echo recognition model based on the radar jigsaw data abnormal echo training data set, and performing abnormal echo recognition on the radar jigsaw data to be recognized based on the trained radar jigsaw data abnormal echo recognition model to obtain an abnormal echo recognition result.

[0110] In this embodiment, S41: the specific method of training the model is:

[0111] An improved DeepLabv3+ model network was built based on the PyTorch deep learning framework, and the improved DeepLabv3+ semantic segmentation network was trained. During the training process, the standard Adam optimizer and classification cross entropy loss were used, and appropriate hyperparameters were set, where the learning rate was initialized to 0.007 and the batch size was 100. GPU was used to accelerate training and improve computing efficiency. The training process and effect of the model were observed through visualization technology, and hyperparameters and optimization strategies were adjusted in a timely manner.

[0112] Determine whether the model has converged and end training based on the size of the loss value.

[0113] In this embodiment, S42: the method for obtaining the abnormal echo identification result is:

[0114] Solidify the trained radar jigsaw data abnormal echo recognition model, that is, generate a pb model file;

[0115] For the solidified radar puzzle data abnormal echo recognition model, the radar puzzle data to be recognized is input, and the corresponding image with the abnormal echo type label is output, so as to obtain the abnormal echo recognition result.

[0116] Embodiment 2

[0117] The radar mosaic data selected in this embodiment comes from the radar combined reflectivity mosaic (MCR) product of the Severe Weather Short-term Nowcasting Operation System (SWAN) operated by the meteorological station in area A. The product covers area A and surrounding areas, with a grid horizontal resolution of 1 km and a time resolution of 6 minutes.

[0118] The following analyzes the recognition effect of the radar puzzle data abnormal echo recognition method based on semantic segmentation involved in the present invention in combination with the actual situation of the radar puzzle data.

[0119] In order to verify the actual effect of the abnormal echo identification method of radar mosaic data based on semantic segmentation, the radar mosaic data of the study area was used for evaluation.

[0120] According to the radar mosaic at time X in area A and the corresponding abnormal echo identification results, it can be found that a significant abnormal echo appeared at the junction of area A and area B. The abnormal echo showed a ring-shaped distribution as a whole, and the intensity ranged from 20 dBZ to 40 dBZ. This abnormal echo not only seriously interferes with the forecaster's analysis and judgment of the weather system, but also produces obvious deviations when the radar data is quantitatively applied, resulting in large-scale false echoes. In the output results of the radar mosaic data abnormal echo identification model, the complete outline of the abnormal echo was effectively identified.

[0121] According to the radar mosaic at time Y in area A and the corresponding abnormal echo identification results, it can be found that multiple radars in area A have abnormal echoes in multiple radial directions. The abnormal echoes are mainly characterized by continuous or intermittent radial divergence distribution characteristics, and the echo intensity is concentrated in the range of 20–40 dBZ. This abnormal echo causes false echoes to appear in multiple radial directions in the radar mosaic data, which obviously interferes with the business application of radar mosaic data. In the output results of the radar mosaic data abnormal echo identification model, multiple radial abnormal echoes can be identified, and the identification effect is good.

[0122] This example focuses on two cases of heavy rainfall in region A at time X and time Y. First, the radar mosaic data is read and plotted as a radar echo image output; then, the trained radar mosaic data abnormal echo recognition model is input; finally, the corresponding image with abnormal echo type label is output. The following conclusions are initially obtained:

[0123] (1) Affected by external electromagnetic interference, atmospheric environmental noise, transmitter-receiver system failures and other factors, radar mosaic data often show abnormal echoes on the image, mainly in the form of rings or radial divergence, which seriously affects the analysis and judgment of weather systems by business personnel, as well as the quantitative application of radar mosaic data.

[0124] (2) The radar mosaic data abnormal echo recognition method based on semantic segmentation involved in the present invention takes the image features of abnormal echoes as the key, improves the DeepLabv3+ semantic segmentation network, and trains and builds the radar mosaic data abnormal echo recognition model. From the recognition results of the radar mosaic data example, this method can effectively identify annular and radially divergent radar abnormal echoes, providing strong technical support for further improving the quality of radar mosaic data and its quantitative application effect.

[0125] From the above conclusions, it can be seen that the present invention can convert radar puzzle data into image files and manually annotate radar abnormal echoes; then construct a training data set through data processing and enhancement; then combine the improved DeepLabv3+ semantic segmentation network to carry out the training of radar puzzle data abnormal echo recognition model; finally, input the radar puzzle data image into the trained recognition model, and output the corresponding image with the abnormal echo type label, thereby realizing radar puzzle data abnormal echo recognition based on semantic segmentation.

[0126] Embodiment 3

[0127] The present invention also provides a radar mosaic data abnormal echo recognition system based on semantic segmentation, which is used to implement the method, including:

[0128] A data set construction module, used for drawing a radar combined reflectivity image based on the collected radar mosaic data, and constructing a radar mosaic image data set based on the radar combined reflectivity image;

[0129] The annotation module is used to annotate the abnormal echo images in the radar jigsaw image data set, and perform format conversion and expansion and enhancement processing on the abnormal echo images to construct a radar jigsaw abnormal echo training data set; wherein the annotated abnormal echo images include annular abnormal echo images and radial abnormal echo images;

[0130] The network improvement module is used to introduce the middle-level feature extraction layer and the simple attention module into the DeepLabv3+ semantic segmentation network to obtain the improved DeepLabv3+ semantic segmentation network;

[0131] Model building module, used to build an abnormal echo recognition model for radar puzzle data based on the improved DeepLabv3+ semantic segmentation network;

[0132] The abnormal echo recognition module is used to train a radar puzzle data abnormal echo recognition model based on a radar puzzle data set of abnormal echoes, and to perform abnormal echo recognition on the radar puzzle data to be recognized based on the trained radar puzzle data abnormal echo recognition model to obtain an abnormal echo recognition result.

[0133] A further implementation is that the marking module includes:

[0134] A screening unit, used for screening out abnormal echo images in the radar jigsaw image data set;

[0135] The labeling unit is used to label the screened abnormal echo images using the Labelme software tool and generate a label data file in json format;

[0136] The format conversion unit is used to convert the annotated json format marked data file into a visual annotated image; and convert the unannotated original jpeg format abnormal echo image into a png format original image;

[0137] The sample expansion unit is used to rotate, swap left and right, and zoom in and out the visual annotation image and the original image in PNG format to obtain an expanded and enhanced dataset of radar jigsaw images;

[0138] The data set division unit is used to convert the original images in png format in the radar jigsaw image expansion and enhancement data set into original images in jpeg format, and divide them into training sets and verification sets to obtain the radar jigsaw abnormal echo training data set.

[0139] A further implementation method is that the network improvement module includes:

[0140] A backbone network construction unit, used to construct a simplified ResNet50 network as a backbone network for feature extraction; wherein the simplified ResNet50 network includes a deep feature extraction layer, a middle feature extraction layer, and a low feature extraction layer;

[0141] A deep feature processing unit is used to add a first simple attention module after the deep feature extraction layer to obtain deep features; process deep features by combining dilated convolutions with different expansion rates to form a dilated spatial convolution pooling pyramid to obtain multi-scale deep features; perform 1×1 convolution on the multi-scale deep features, and perform bilinear upsampling on the convolved multi-scale deep features with a factor of 4 to obtain upsampled multi-scale deep features;

[0142] A low-level feature processing unit, used for adding a second simple attention module after the low-level feature extraction layer to obtain low-level features; performing 1×1 convolution on the low-level features to obtain the convolved low-level features;

[0143] A middle-level feature processing unit is used to add a third simple attention module after the middle-level feature extraction layer to obtain middle-level features; perform 1×1 convolution on the middle-level features, and perform bilinear upsampling on the convolved middle-level features by a factor of 2 to obtain upsampled middle-level features;

[0144] The feature connection unit is used to connect the upsampled multi-scale deep features, the convolutional low-level features, and the upsampled middle-level features;

[0145] A feature refinement unit is used to refine the connected features using a 3×3 depthwise separable convolutional layer to obtain refined features;

[0146] An upsampling unit for bilinear upsampling of refined features by a factor of 4;

[0147] The segmentation unit is used to convolve the upsampled refined features with a 1×1 convolution kernel that matches the number of abnormal echo categories to obtain the segmentation result and complete the improvement of the DeepLabv3+ semantic segmentation network.

[0148] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for identifying abnormal echoes in radar mosaic data based on semantic segmentation, characterized in that: The following steps are involved: Drawing a radar combined reflectivity image based on the collected radar mosaic data, and constructing a radar mosaic image data set based on the radar combined reflectivity image; Annotating the abnormal echo images in the radar jigsaw image data set, and performing format conversion and expansion and enhancement processing on the abnormal echo images to construct a radar jigsaw abnormal echo training data set; wherein the annotated abnormal echo images include annular abnormal echo images and radial abnormal echo images; A middle-level feature extraction layer and a simple attention module are introduced into the DeepLabv3+ semantic segmentation network to obtain an improved DeepLabv3+ semantic segmentation network. Based on the improved DeepLabv3+ semantic segmentation network, a radar puzzle data abnormal echo recognition model is constructed. A simplified ResNet50 network is constructed as the backbone network for feature extraction. The simplified ResNet50 network includes a deep feature extraction layer, a middle-level feature extraction layer, and a low-level feature extraction layer. Adding a third simple attention module after the middle-level feature extraction layer to obtain middle-level features; performing 1×1 convolution on the middle-level features, and bilinearly upsampling the convolved middle-level features by a factor of 2 to obtain upsampled middle-level features; Connect the upsampled multi-scale deep features, the convolutional low-level features, and the upsampled middle-level features; Use a 3×3 depth-separable convolutional layer to refine the connected features and obtain refined features; Bilinearly upsampling the refined features by a factor of 4; The upsampled refined features are convolved with a 1×1 convolution kernel that matches the number of abnormal echo categories to obtain the segmentation results, completing the improvement of the DeepLabv3+ semantic segmentation network; The radar jigsaw data abnormal echo recognition model is trained based on the radar jigsaw data abnormal echo training data set, and abnormal echo recognition is performed on the radar jigsaw data to be recognized based on the trained radar jigsaw data abnormal echo recognition model to obtain an abnormal echo recognition result.

2. The method for identifying abnormal echoes of radar mosaic data based on semantic segmentation according to claim 1, characterized in that: The method to construct the radar puzzle abnormal echo training data set is: Screening out abnormal echo images in the radar jigsaw image data set; Using the Labelme software tool, label the screened abnormal echo images and generate a labeled data file in json format; Convert the annotated json formatted labeled data file into a visual annotated image; Convert the unlabeled original jpeg format abnormal echo image into the original png format image; The visual annotation image and the original image in png format are respectively rotated, swapped left and right, and zoomed in and out to obtain the radar jigsaw image expansion and enhancement data set; The original images in png format in the radar jigsaw image expansion and enhancement data set are converted into original images in jpeg format, and a training set and a validation set are divided to obtain the radar jigsaw abnormal echo training data set.

3. The method for identifying abnormal echoes of radar mosaic data based on semantic segmentation according to claim 1, characterized in that: The method to obtain the improved DeepLabv3+ semantic segmentation network includes: Adding a first simple attention module after the deep feature extraction layer to obtain deep features; The deep features are processed by combining dilated convolutions with different dilation rates to form a dilated spatial convolution pooling pyramid to obtain multi-scale deep features; Performing 1×1 convolution on the multi-scale deep features, and bilinearly upsampling the convolved multi-scale deep features by a factor of 4 to obtain upsampled multi-scale deep features; A second simple attention module is added after the low-level feature extraction layer to obtain low-level features; and a 1×1 convolution is performed on the low-level features to obtain the convolved low-level features.

4. The method for identifying abnormal echoes of radar mosaic data based on semantic segmentation according to claim 3 is characterized in that: The simplified ResNet50 network structure has a total of fourteen layers; among them, the first layer is the conv1_x layer, the second to fourth layers are the conv2_x layers, the fifth to eighth layers are the conv3_x layers, and the ninth to fourteenth layers are the conv4_x layers.

5. The method for identifying abnormal echoes of radar mosaic data based on semantic segmentation according to claim 3 is characterized in that: The application methods of the first simple attention module, the second simple attention module and the third simple attention module are all as follows: Normalize the original features extracted by the backbone network; Calculate the square of the difference between the normalized original feature and the mean of the normalized original feature to obtain a calculation result; Based on the calculation result, obtaining an attention weight; Attention weighting is performed on the original feature based on the attention weight to obtain the attention-weighted original feature.

6. A radar mosaic data abnormal echo recognition system based on semantic segmentation, used to implement the method described in any one of claims 1 to 5, characterized in that: include: A data set construction module, used to draw a radar combined reflectivity image based on the collected radar mosaic data, and to construct a radar mosaic image data set based on the radar combined reflectivity image; A labeling module is used to label the abnormal echo images in the radar jigsaw image data set, and perform format conversion and expansion and enhancement processing on the abnormal echo images to construct a radar jigsaw abnormal echo training data set; wherein the labeled abnormal echo images include annular abnormal echo images and radial abnormal echo images; The network improvement module is used to introduce the middle-level feature extraction layer and the simple attention module into the DeepLabv3+ semantic segmentation network to obtain the improved DeepLabv3+ semantic segmentation network; Model building module, used to build an abnormal echo recognition model for radar puzzle data based on the improved DeepLabv3+ semantic segmentation network; The abnormal echo recognition module is used to train the radar puzzle data abnormal echo recognition model based on the radar puzzle data set, and perform abnormal echo recognition on the radar puzzle data to be recognized based on the trained radar puzzle data abnormal echo recognition model to obtain an abnormal echo recognition result.

7. The radar mosaic data abnormal echo recognition system based on semantic segmentation according to claim 6 is characterized in that: The marking module comprises: A screening unit, used for screening out abnormal echo images in the radar jigsaw image data set; A labeling unit, used to label the screened abnormal echo image using the Labelme software tool to generate a label data file in json format; The format conversion unit is used to convert the annotated json format marked data file into a visual annotated image; and convert the unannotated original jpeg format abnormal echo image into a png format original image; A sample expansion unit, used to rotate, swap left and right, and zoom in and out the visual annotation image and the original image in png format, respectively, to obtain an expanded and enhanced data set of the radar jigsaw image; The data set division unit is used to convert the original images in png format in the radar jigsaw image expansion and enhancement data set into original images in jpeg format, and divide them into training sets and verification sets to obtain the radar jigsaw abnormal echo training data set.

8. The radar mosaic data abnormal echo recognition system based on semantic segmentation according to claim 6, characterized in that: The network improvement module includes: A backbone network construction unit, used to construct a simplified ResNet50 network as a backbone network for feature extraction; wherein the simplified ResNet50 network includes a deep feature extraction layer, a middle feature extraction layer, and a low feature extraction layer; A deep feature processing unit is used to add a first simple attention module after the deep feature extraction layer to obtain deep features; process the deep features by combining dilated convolutions with different expansion rates to form a dilated spatial convolution pooling pyramid to obtain multi-scale deep features; perform 1×1 convolution on the multi-scale deep features, and perform bilinear upsampling on the convolved multi-scale deep features with a 4-fold coefficient to obtain upsampled multi-scale deep features; A low-level feature processing unit, configured to add a second simple attention module after the low-level feature extraction layer to obtain low-level features; and perform 1×1 convolution on the low-level features to obtain convolved low-level features; A middle-level feature processing unit, configured to add a third simple attention module after the middle-level feature extraction layer to obtain middle-level features; perform 1×1 convolution on the middle-level features, and perform bilinear upsampling on the convolved middle-level features by a factor of 2 to obtain upsampled middle-level features; The feature connection unit is used to connect the upsampled multi-scale deep features, the convolutional low-level features, and the upsampled middle-level features; A feature refinement unit is used to refine the connected features using a 3×3 depthwise separable convolutional layer to obtain refined features; An upsampling unit, configured to perform bilinear upsampling on the refined features by a factor of 4; The segmentation unit is used to convolve the upsampled refined features with a 1×1 convolution kernel that matches the number of abnormal echo categories to obtain the segmentation result and complete the improvement of the DeepLabv3+ semantic segmentation network.

Images