Weather radar migration biological echo extraction method based on Deeplabv3 +

Abstract

The invention discloses a weather radar migratory organism echo extraction method based on Deeplabv3 +. The method comprises the following steps: (1) selecting a first elevation angle and a second elevation angle within a radar scanning range with a radius of 2.3 kilometers for monitoring; (2) preprocessing the collected data; (3) constructing a DeepLabv3 + semantic segmentation network based on Xception, and introducing an attention mechanism module and an edge extraction module into the Xception; (4) extracting biological echo information in the radar image; according to the invention, large-scale intelligent monitoring of aerial organisms is realized, the equipment deployment cost and difficulty are reduced, and the development of an aerial ecological intelligent monitoring technology is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of weather radar, and in particular to a method for extracting migratory organism echoes from weather radar based on Deeplabv3+. Background Art

[0002] Thousands of birds migrate between their breeding and wintering grounds around the world, and their migration routes overlap with areas of human activity. Studying animal migration behavior can help achieve two main goals: first, it can help rationally plan human activity areas and reduce interference with wildlife; second, it can provide important information to help better protect endangered birds, prevent the spread of diseases, respond to climate change, ensure food security, and reduce aviation safety risks.

[0003] However, compared with terrestrial organisms, it is more difficult to monitor aerial migratory organisms. A widely distributed weather radar network system makes it possible to monitor aerial biological migration on a large scale. At present, weather radars are mainly used for meteorological research, and biological echoes are regarded as interference that affects meteorological observations. Since single-polarization radars have fewer data types, directly applying data to machine learning algorithms for biological echo identification is not effective. At the same time, the existing deep learning models have a large number of parameters, making them difficult to promote and apply on a large scale. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for extracting migratory organism echoes from weather radar based on Deeplabv3+, so as to solve the problem that the data type collected by single-polarization radar is single and cannot be classified using existing data, and the existing deep learning model has a large number of parameters and is difficult to deploy.

[0005] Technical solution: The method for extracting migratory organism echoes from weather radar based on Deeplabv3+ of the present invention comprises the following steps: (1) Select the first elevation angle and the second elevation angle within the radar scanning range with a radius of 2.3 kilometers for monitoring; (2) Preprocess the collected data; (3) Construct a DeepLabv3+ semantic segmentation network based on Xception, and introduce an attention mechanism module and an edge extraction module to Xception; (4) Extract biological echo information from radar images.

[0006] Furthermore, step (2) includes the following steps: (21) Converting and rendering data from polar coordinates to Cartesian coordinates; (22) The converted Cartesian coordinate data is processed using the cubic spline interpolation method to generate a radar image with a resolution of 320 × 320; (23) Obtain the reflectivity and spectral width data at the first elevation angle of 230 km and the reflectivity data at the second elevation angle of 110 km as the red channel, blue channel, and green channel, respectively, and fuse them into a three-channel 320×320 color image; (24) Labelme was used to annotate images with bio-echoes.

[0007] Furthermore, step (21) is specifically as follows: The conversion formula is as follows: ; in, , and They represent the detection distance, elevation angle (0-90°) and azimuth angle (0-360°) respectively; x and y represent the position information in the Cartesian coordinate system respectively.

[0008] Furthermore, step (3) is specifically as follows: the Xception-based DeepLabv3+ semantic segmentation network includes: a hole convolution module, an attention mechanism module and an edge extraction module; wherein the attention mechanism module includes a spatial attention mechanism and a channel attention mechanism.

[0009] Furthermore, in step (3), the dilated convolution module is used to extract features from the original image and obtain multi-scale context information.

[0010] Furthermore, in step (3), the channel attention mechanism generates the importance weight of each channel by globally pooling the feature map in the channel dimension, while the spatial attention mechanism pools the spatial dimension of the feature map to generate the importance weight of each pixel position; after the two are fused, an attention-weighted intermediate feature map is generated.

[0011] Furthermore, in step (3), the edge extraction module applies Canny edge detection to the intermediate feature map to extract a preliminary edge feature map, denoted as edge_aux; then the channel and spatial attention mechanisms are applied to the intermediate feature map so that the generated attention-weighted feature map captures the significant areas and details in the image; the feature map enhanced by the attention mechanism is passed to the subsequent hole convolution layer and decoder part of DeepLabv3+ to gradually generate the final feature map; Canny edge detection is performed again on the final feature map to obtain a new edge feature map, denoted as edge; finally, edge_aux and edge are concatenated and the final edge feature map combined_edge is generated through a convolution layer; the edge feature map is directly input into the final feature fusion stage to be combined with the deep semantic features.

[0012] The weather radar migratory organism echo extraction system based on Deeplabv3+ of the present invention comprises: Monitoring module: used to select the first elevation angle and the second elevation angle within the radar scanning range with a radius of 2.3 kilometers for monitoring; Preprocessing module: used to preprocess the collected data; DeepLabv3+ semantic segmentation module: used to build a DeepLabv3+ semantic segmentation network based on Xception, and introduce an attention mechanism module and an edge extraction module to Xception; Extraction module: used to extract biological echo information from radar images.

[0013] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, it implements any one of the weather radar migratory organism echo extraction methods based on Deeplabv3+.

[0014] A storage medium described in the present invention stores a computer program, and when the computer program is executed by a processor, it implements any one of the weather radar migratory organism echo extraction methods based on Deeplabv3+.

[0015] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the DeepLabv3+ model uses lightweight Xception as the feature extraction network, and integrates the attention mechanism module and the edge extraction module to improve the model's ability to extract biological echoes while maintaining the lightweight characteristics of the model. This algorithm aims to achieve large-scale intelligent monitoring of aerial organisms through a widely distributed weather radar network, reduce the cost and difficulty of equipment deployment, and promote the development of intelligent aerial ecological monitoring technology. This method can also help rationally plan human activity areas, such as aviation route design and high-rise building construction site selection; in addition, it can also help to more effectively protect endangered birds, prevent virus transmission, respond to climate change, and ensure food security. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the present invention; Figure 2 It is the improved DeepLabv3+ network principle of the present invention. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0018] like Figure 1 As shown, an embodiment of the present invention provides a method for extracting migratory organism echoes from weather radar based on Deeplabv3+, comprising the following steps: (1) Select the first and second elevation angles within the radar scanning range with a radius of 2.3 kilometers for monitoring; the details are as follows: According to the flight behavior of migratory birds, they usually fly at a lower altitude, and the flight altitude usually does not exceed 2.5 kilometers. In the process of radar monitoring of migratory organisms, it is reasonable to select the first and second elevation angles within the radar scanning range with a radius of 2.3 kilometers for monitoring. This choice is based on reasonable speculation about the flight altitude of migratory birds and the limitations of radar technical parameters. By reasonably selecting the elevation angle, the radar can obtain effective echo data and capture the activity trajectory of migratory organisms in the air. This monitoring method can provide key information on migratory bird migration, such as the number of migrants, flight paths, and migration periods.

[0019] (2) Preprocess the collected data; including the following steps: (21) The data is converted and rendered from polar coordinates to Cartesian coordinates; the details are as follows: The conversion formula is as follows: ; in, , and They represent the detection distance, elevation angle (0-90°) and azimuth angle (0-360°) respectively; x and y represent the position information in the Cartesian coordinate system respectively.

[0020] (22) The converted Cartesian coordinate data is processed using the cubic spline interpolation method to generate a radar image with a resolution of 320 × 320; (23) Obtain the reflectivity and spectral width data at the first elevation angle of 230 km and the reflectivity data at the second elevation angle of 110 km as the red channel, blue channel and green channel respectively, and fuse them into a 3-channel 320×320 color image; specifically as follows: Since the biological echo is generally 10-35 dBz, while the rainfall echo can exceed 50 dBz. In terms of spectral width, since bird migration moves faster than rainfall, the spectral width of rainfall echo is usually less than 3 m / s, while the spectral width value of bird migration is usually greater than 8 m / s. Considering that the migration height of organisms usually does not exceed 2.5 km, the new generation of weather radar can reach the usual altitude of biological migration at the first and second elevation angles of 230 km and 110 km respectively. Therefore, the present invention selects the reflectivity and spectral width data at the first elevation angle of 230 km and the reflectivity data at the second elevation angle of 110 km as the red channel, blue channel and green channel respectively, and fuses them into a 3-channel 320×320 color image, so as to be more conducive to learning and extracting the characteristics of birds in the image using semantic segmentation algorithms.

[0021] (24) Labelme was used to annotate the images with biological echoes. Specifically, the scanning data of this time period was selected from the radar scans of Nanchang Station in 2019. After data preprocessing, the original radar data was screened and rendered to obtain 6660 radar rendering images, which were then annotated with biological echoes. Finally, they were randomly divided into a training set of 6160 images and a validation set of 500 images.

[0022] (3) If Figure 2 As shown in the figure, a DeepLabv3+ semantic segmentation network based on Xception is constructed, and an attention mechanism module and an edge extraction module are introduced to Xception; specifically, the DeepLabv3+ semantic segmentation network based on Xception includes: a hole convolution module, an attention mechanism module and an edge extraction module; among them, the attention mechanism module includes a spatial attention mechanism and a channel attention mechanism.

[0023] Among them, the hole convolution module is used to extract features from the original image and obtain multi-scale context information. The formula is as follows: ; in, Represents a set of N dilated convolutions; It means The pixel values ​​of the input feature map; Represents the dilation rate of the nth dilated convolution; is the convolution kernel weight of the nth dilated convolution; N is the number of different dilation rates.

[0024] The channel attention mechanism generates the importance weight of each channel by globally pooling the feature map in the channel dimension, while the spatial attention mechanism pools the spatial dimension of the feature map to generate the importance weight of each pixel position; after the two are combined, an attention-weighted intermediate feature map is generated. The channel attention formula is as follows: ; Among them, F represents the input feature map, the shape is (C, H, W); C is the number of channels; H is the height; W is the width; and They represent the global average pooling and global maximum pooling operations of the feature map, respectively, for generating features on two channel dimensions (C, 1, 1); MLP represents a multi-layer perceptron that compresses or expands the channel dimension; is the Sigmoid activation function, which is used to limit the output to the range [0,1]. The spatial attention can be expressed as: ; Among them, F represents the input feature map, the shape is (C, H, W); C is the number of channels; H is the height; W is the width; FC represents the fully connected layer, and They represent the global average pooling and global maximum pooling operations of the feature map, respectively, and two feature maps with shapes of (1, H, W) are obtained. represents the convolution operation used to generate the spatial attention map, is the Sigmoid activation function, which is used to limit the output to the range [0,1].

[0025] The edge extraction module applies Canny edge detection to the intermediate feature map to extract a preliminary edge feature map, denoted as edge_aux; then the channel and spatial attention mechanisms are applied to the intermediate feature map so that the generated attention-weighted feature map captures the significant areas and details in the image; the feature map enhanced by the attention mechanism is passed to the subsequent hole convolution layer and decoder part of DeepLabv3+ to gradually generate the final feature map; Canny edge detection is performed again on the final feature map to obtain a new edge feature map, denoted as edge; finally, edge_aux and edge are concatenated and the final edge feature map combined_edge is generated through a convolution layer; the edge feature map is directly input into the final feature fusion stage to be combined with deep semantic features.

[0026] Among them, the edge extraction module includes: Gaussian smoothing, gradient calculation, non-maximum suppression, dual threshold detection and edge connection. The process is as follows: First, the image is Gaussian filtered to reduce the impact of noise. The image is smoothed by the Gaussian convolution kernel to effectively suppress the noise in the image while retaining the edge information. The Gaussian filter is defined as follows: ; in, represents the weight of the Gaussian filter; x and y are the horizontal and vertical distances of the pixel relative to the center of the filter; is the standard deviation of the Gaussian filter, which controls the smoothness of the filter, The larger the value, the stronger the smoothing effect. After smoothing, the edge is extracted by calculating the horizontal and vertical gradients of the image. Similar to the Sobel operator, the horizontal gradient is obtained through convolution operation. and vertical gradient The formula is as follows: ; Where I is the input image, and are the convolution kernels in the horizontal and vertical directions respectively. The gradient intensity of the pixel and the gradient direction It can be calculated as follows: ; ; Among them, G represents the gradient strength of the pixel point, Indicates the gradient direction. is the horizontal gradient; is the vertical gradient. Again, non-maximum suppression is entered, by checking the neighbors of each pixel along its gradient direction, retaining the local maximum and suppressing other pixels. This ensures the refinement and accuracy of the edge. Double threshold detection is another important step of the Canny module. By setting a high threshold and low threshold , higher than Pixels with a value lower than The pixels between the two are suppressed, while the pixels between them are considered as weak edge points. Finally, the edge coherence is ensured by edge connection, i.e. hysteresis threshold processing. Weak edge points are retained only when they are connected to strong edge points, which further improves the accuracy of edge detection.

[0027] (4) Extract biological echo information from radar images.

[0028] The embodiment of the present invention also provides a weather radar migratory organism echo extraction system based on Deeplabv3+, comprising: Monitoring module: used to select the first elevation angle and the second elevation angle within the radar scanning range with a radius of 2.3 kilometers for monitoring; Preprocessing module: used to preprocess the collected data; DeepLabv3+ semantic segmentation module: used to build a DeepLabv3+ semantic segmentation network based on Xception, and introduce an attention mechanism module and an edge extraction module to Xception; Extraction module: used to extract biological echo information from radar images.

[0029] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, the method for extracting migratory organism echoes from weather radar based on Deeplabv3+ is implemented.

[0030] An embodiment of the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements any one of the weather radar migratory organism echo extraction methods based on Deeplabv3+.

Claims

1. A method for extracting migratory organism echoes from weather radar based on Deeplabv3+, characterized in that: The following steps are involved: (1) Select the first elevation angle and the second elevation angle within the radar scanning range with a radius of 2.3 kilometers for monitoring; (2) Preprocess the collected data; (3) Construct a DeepLabv3+ semantic segmentation network based on Xception, and introduce an attention mechanism module and an edge extraction module to Xception; (4) Extract biological echo information from radar images.

2. The method for extracting migratory organism echoes from weather radar based on Deeplabv3+ according to claim 1, characterized in that: Step (2) includes the following steps: (21) Converting and rendering data from polar coordinates to Cartesian coordinates; (22) The converted Cartesian coordinate data is processed using the cubic spline interpolation method to generate a radar image with a resolution of 320 × 320; (23) Obtain the reflectivity and spectral width data at the first elevation angle of 230 km and the reflectivity data at the second elevation angle of 110 km as the red channel, blue channel, and green channel, respectively, and fuse them into a three-channel 320×320 color image; (24) Labelme was used to annotate images with bio-echoes.

3. The method for extracting migratory organism echoes from weather radar based on Deeplabv3+ according to claim 2, characterized in that: Step (21) is as follows: The conversion formula is as follows: ； in, , and They represent the detection distance, elevation angle and azimuth respectively; x and y represent the position information in the Cartesian coordinate system respectively.

4. The method for extracting migratory organism echoes from weather radar based on Deeplabv3+ according to claim 1, characterized in that: Step (3) is as follows: The Xception-based DeepLabv3+ semantic segmentation network includes: a hole convolution module, an attention mechanism module and an edge extraction module; among them, the attention mechanism module includes a spatial attention mechanism and a channel attention mechanism.

5. The method for extracting migratory organism echoes from weather radar based on Deeplabv3+ according to claim 4, characterized in that: In step (3), the dilated convolution module is used to extract features from the original image and obtain multi-scale context information.

6. The method for extracting migratory organism echoes from weather radar based on Deeplabv3+ according to claim 4, characterized in that: In step (3), the channel attention mechanism generates the importance weight of each channel by globally pooling the feature map in the channel dimension, while the spatial attention mechanism pools the spatial dimension of the feature map to generate the importance weight of each pixel position; After the two are fused, an attention-weighted intermediate feature map is generated.

7. The method for extracting migratory organism echoes from weather radar based on Deeplabv3+ according to claim 4, characterized in that: In step (3), the edge extraction module applies Canny edge detection to the intermediate feature map to extract a preliminary edge feature map, denoted as edge_aux; then the channel and spatial attention mechanisms are applied to the intermediate feature map so that the generated attention-weighted feature map captures the salient areas and details in the image; The feature map enhanced by the attention mechanism is passed to the subsequent atrous convolution layer and decoder part of DeepLabv3+ to gradually generate the final feature map; Canny edge detection is performed again on the final feature map to obtain a new edge feature map, recorded as edge; finally, edge_aux and edge are concatenated and the final edge feature map combined_edge is generated through a convolution layer; the edge feature map is directly input into the final feature fusion stage to be combined with the deep semantic features.

8. A weather radar migratory organism echo extraction system based on Deeplabv3+, characterized in that: include: Monitoring module: used to select the first elevation angle and the second elevation angle within the radar scanning range with a radius of 2.3 kilometers for monitoring; Preprocessing module: used to preprocess the collected data; DeepLabv3+ semantic segmentation module: used to build a DeepLabv3+ semantic segmentation network based on Xception, and introduce an attention mechanism module and an edge extraction module to Xception; Extraction module: used to extract biological echo information from radar images.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the method for extracting weather radar migratory organism echoes based on Deeplabv3+ according to any one of claims 1 to 7 is implemented.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a method for extracting weather radar migratory organism echoes based on Deeplabv3+ according to any one of claims 1 to 7 is implemented.

Images