Cloud radar ground object echo recognition method based on deep learning and physical statistics fusion

Abstract

The invention discloses a cloud radar ground object echo recognition method based on deep learning and physical statistics fusion, and belongs to the field of meteorological radar detection, and the method comprises the following steps: obtaining an observation variable of a millimeter wave cloud radar in a PPI scanning mode; determining a ground feature truth value based on the observation variable; and constructing a ground feature truth value sample based on the ground feature truth value, and performing accurate judgment of weather and ground feature echoes on the ground feature truth value sample by adopting a two-dimensional convolutional neural network to obtain a final recognition result. According to the method, more reliable data support is provided for severe convection weather early warning, extreme weather cause research and disastrous weather emergency plan making.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological radar detection, and in particular relates to a cloud radar ground object echo recognition method based on deep learning and physical statistics fusion. Background Art

[0002] The development of millimeter-wave cloud radar technology has provided more accurate data for meteorological detection. However, this technology faces a major challenge in practical applications: interference from echo signals of non-meteorological targets. These signals include ground objects, insects, birds, atmospheric turbulence, etc., which constitute clutter interference to atmospheric hydrometeors that meteorological researchers are concerned about, affecting the accurate judgment of weather processes such as precipitation areas and precipitation characteristics. Especially in scanning cloud radars, when the radar elevation angle is low or close to the horizontal, it will be interfered by surface objects including buildings, trees, and ships; when a weather system passes, it is difficult to distinguish the location of ground clutter.

[0003] In addition, when existing fuzzy logic algorithms process areas where the radial velocity is close to zero, they tend to misjudge these areas as ground object echoes. This is mainly because traditional algorithms rely on parameters such as velocity to judge ground object echoes, and these parameters may be close to zero at the edge of precipitation echoes and weak precipitation targets, leading to misjudgment.

[0004] The existing algorithms are not very effective in identifying ground object echoes in mixed precipitation and ground object areas, especially at the edges of precipitation echoes and weak precipitation targets. This is because the ground object echoes in these areas may show the characteristics of non-zero velocity characteristic parameters of precipitation echoes, resulting in the edges of precipitation echoes being misjudged as ground object echoes.

[0005] When using clutter maps and hardware methods (such as notch filters) for clutter suppression, precipitation echoes with zero radial velocity are often misidentified as ground clutter when filtering large areas of ground objects, and super-refractive echoes generated under abnormal propagation conditions cannot be identified and processed.

[0006] The main reasons for the above problems include: (1) Over-reliance on velocity parameters: The traditional algorithm's over-reliance on velocity parameters results in the inability to effectively distinguish between ground object echoes and precipitation echoes in areas where the velocity is close to zero, especially in areas where precipitation and ground objects are mixed. (2) Complex algorithm recognition process: The existing technology fails to fully utilize the observation variables provided by cloud radar, and the reliance on a single variable results in a relatively complex algorithm structure and poor portability.

[0007] In summary, the existing technology has obvious deficiencies in the identification and elimination of ground object echoes by millimeter wave cloud radar, and new technical solutions are needed to improve the accuracy and reliability of meteorological detection data. In view of this problem, the present invention proposes a ground object echo identification method based on computer vision technology and deep learning algorithm, in order to solve the problems existing in the existing technical solutions. Summary of the invention

[0008] In order to solve the above technical problems, the present invention proposes a cloud radar object echo recognition method based on deep learning and physical statistics fusion to solve the problems existing in the above-mentioned prior art.

[0009] To achieve the above object, the present invention provides a cloud radar object echo recognition method based on deep learning and physical statistics fusion, comprising:

[0010] Obtain the observed variables under the millimeter wave cloud radar PPI scanning mode;

[0011] Determining the true value of the ground feature based on the observed variables;

[0012] A ground object true value sample is constructed based on the ground object true value, and a two-dimensional convolutional neural network is used to accurately distinguish the meteorological and ground object echoes of the ground object true value sample to obtain a final recognition result.

[0013] Optionally, the observed variables include: radar reflectivity factor, radial velocity, velocity spectrum width and signal-to-noise ratio.

[0014] Optionally, the process of determining the true value of the ground object based on the observed variable includes:

[0015] Sorting the observed variables in chronological order to obtain changes of different radar variables in time series and constructing a data set;

[0016] The data set is shuffled in time order and a number of samples are extracted to obtain a sub-sample set;

[0017] Calculating the standard deviation distribution of different radar variables in the sub-sample set in time to obtain a standard deviation result;

[0018] Comparing the standard deviation result with the original observation result to obtain the radar variable that best fits the actual observation;

[0019] Based on computer vision technology, the ground object clutter is completely extracted from the radar variable that best fits the actual observation to obtain the ground object true value.

[0020] Optionally, the process of completely extracting ground clutter from the radar variable that best fits actual observation based on computer vision technology includes:

[0021] Based on the radar variable that best fits the actual observation, the edge gradient of the standard deviation of the longitudinal velocity is calculated using the Sobel gradient operator to obtain an edge detection result;

[0022] Based on the edge detection result, a watershed algorithm is used to automatically identify the boundary position to extract the ground clutter boundary and calibrate the true value.

[0023] Optionally, the process of using a two-dimensional convolutional neural network to accurately discriminate the meteorological and ground object echoes of the ground object true value samples to obtain a final recognition result includes:

[0024] Building a two-dimensional U-Net neural network, and using ground object true value samples and corresponding radar observation variables to train the two-dimensional U-Net neural network to obtain a trained two-dimensional U-Net neural network;

[0025] Based on the trained two-dimensional U-Net neural network, new radar data is intelligently identified and probability identification results of ground object echoes and weather echoes are output.

[0026] Optionally, the ground object echo recognition method further includes post-processing the probability recognition result output by the two-dimensional U-Net neural network; wherein the post-processing process includes: using a Sobel gradient operator and a watershed algorithm to determine the ground object clutter boundary.

[0027] The present invention also provides a computer terminal device, comprising:

[0028] one or more processors;

[0029] A memory, coupled to the processor, for storing one or more programs;

[0030] When the one or more programs are executed by the one or more processors, the one or more processors implement a cloud radar object echo recognition method based on deep learning and physical statistics fusion.

[0031] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, a cloud radar ground object echo recognition method based on deep learning and physical statistics fusion is implemented.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] The millimeter-wave cloud radar object echo recognition method of the present invention effectively determines the true value of the object by accurately obtaining the observed variables under the PPI scan, and uses a two-dimensional convolutional neural network to intelligently distinguish the true value samples of the object, thereby significantly improving the recognition accuracy of the object echo. This method reduces the dependence on a single velocity parameter and reduces the risk of misjudgment, especially in areas where the velocity is close to zero. In addition, the technical solution simplifies the algorithm structure, improves the portability and efficiency of the algorithm, and provides more reliable data support for meteorological services, especially in severe convective weather warnings and extreme weather causes. It has important application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 This is a flow chart of an algorithm for intelligently identifying ground object echoes according to an embodiment of the present invention;

[0036] Figure 2 Schematic diagram of the two-dimensional U-Net convolutional neural network structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0037] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0038] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] Embodiment 1

[0040] In view of the challenges encountered by meteorological radars in the detection process in the prior art, especially the problem that millimeter-wave cloud radars have difficulty distinguishing the echoes of meteorological targets and ground targets in actual detection, the purpose of the present invention is to provide an effective solution to improve the accuracy and reliability of meteorological detection data. Specifically, the present invention aims to achieve the following objectives:

[0041] (1) Improve the accuracy of ground object echo recognition:

[0042] By comprehensively analyzing the changing characteristics of different observation variables of long-time series millimeter-wave radar and utilizing the position stability of ground object echoes and the moving speed characteristics of meteorological targets, the present invention aims to improve the recognition accuracy of ground object echoes and reduce misjudgments and missed judgments.

[0043] (2) Development of new identification solutions:

[0044] Combining computer vision technology and deep learning algorithms, the present invention aims to develop a novel ground object echo recognition scheme to distinguish and exclude ground object echoes.

[0045] (3) Improving the reliability of weather radar data:

[0046] Through the above technical scheme, the present invention aims to improve the reliability of meteorological radar data, ensure the quality and accuracy of meteorological services, and thus provide more reliable data support for severe convective weather warnings, extreme weather cause research, and the formulation of disastrous weather emergency plans.

[0047] like Figure 1 As shown, this embodiment provides a cloud radar object echo recognition method based on deep learning and physical statistics fusion, including the following steps:

[0048] The present invention uses millimeter wave cloud radar PPI scanning data to construct a training set, which includes four key radar variables: radar reflectivity factor (Z, in dBZ units, measures the reflection ability of cloud precipitation particles, and is related to precipitation intensity), radial velocity (V, in m / s units, indicates the speed of meteorological targets relative to the radar, which is important for analyzing wind fields and precipitation motion characteristics), velocity spectrum width (SW, reflects the width of particle velocity distribution, and is related to particle size and shape) and signal-to-noise ratio (SNR, measures the ratio of target signal strength to background noise. A high signal-to-noise ratio means that the target signal is strong and easy to detect). The technical solution of the present invention participates in Figure 1 , specifically including: obtaining observation variables under the millimeter-wave cloud radar PPI scanning mode; determining the true value of the ground object based on the relevant observation variables; constructing a ground object true value sample based on the ground object true value, and using a two-dimensional convolutional neural network to accurately distinguish the meteorological and ground object echoes of the ground object true value sample to obtain the final recognition result.

[0049] As a specific implementation of this embodiment, it includes:

[0050] 1. Calibrate the true value of ground objects based on radar echo characteristic analysis. The present invention first analyzes the change characteristics of ground object echoes and meteorological echoes from a time series perspective. After clarifying these change characteristics, the radar variable with the most differential characterization variable is selected as the reference quantity, and computer vision technology is used to enhance the extraction of ground object features, thereby calibrating the true value of the ground object.

[0051] Furthermore, the process of determining the true value of the ground object based on the relevant observation variables includes: sorting the observation variables in chronological order to obtain changes in different radar variables in time series and constructing a data set; disrupting the time sequence of the data set and extracting a number of samples to obtain a sub-sample set; calculating the standard deviation distribution of different radar variables in the sub-sample set in time to obtain a standard deviation result; comparing the standard deviation result with the original observation result to obtain the radar variable that best fits the actual observation; and completely extracting ground object clutter from the radar variable that best fits the actual observation based on computer vision technology to obtain the true value of the ground object.

[0052] The specific steps are as follows: Step 1: Select the observation variables under the millimeter-wave cloud radar PPI scanning mode for a relatively long time and draw them. Then arrange these radar observation results in chronological order and analyze the changes of different radar variables in the time series. It can be found that compared with the meteorological echo, the boundary characteristics of the ground object echo are obvious, the volatility in the time series is more, and the band variation of the radar reflectivity factor inside the ground object is obvious, indicating that it is rough and has texture characteristics different from the meteorological echo.

[0053] Step 2: Random sampling and combining standard deviation to characterize the characteristics of ground objects. In order to further highlight the characteristics of ground object echoes, 40% of the samples are randomly extracted after the time sequence of the data set in step 1 is disrupted, thereby avoiding the interference of the stable precipitation system on the subsequent extraction of ground object echo signals. After that, the standard deviation distribution of different radar variables in time is calculated for the randomly sampled sub-sample set. The standard deviation results are compared with the original observation results, and the radar variables that best fit the actual observations are selected for subsequent ground object clutter extraction. Here, the standard deviation results of the meridional velocity are the best match.

[0054] Step 3: Extract the ground object true value based on computer vision technology. After processing in step 2, the location of the ground clutter can be basically determined, but how to completely extract the ground clutter is a new problem. Here, in order to avoid the problems of missed detection and false detection caused by the threshold method, consider using the Sobel gradient operator in computer vision to first calculate the edge gradient of the longitudinal velocity standard deviation to obtain the edge detection result. As mentioned earlier, the boundary features of the ground object echo are obvious, and this step can highlight the boundary information. Subsequently, the watershed algorithm is used to automatically identify the boundary position, thereby realizing the extraction of the ground clutter boundary and the calibration of the true value.

[0055] 2. Intelligent recognition algorithm based on two-dimensional U-Net neural network training

[0056] It should be emphasized that the accurate calibration of ground clutter is based on the existence of a certain number of observation samples in advance, which is obviously not conducive to the real-time identification of ground object echoes. Therefore, after obtaining reasonably accurate meteorological and ground object true value samples, it is necessary to design an objective and intelligent meteorological and ground object clutter classification method. Here, the U-structure Neural Network (U-Net) convolutional neural network architecture in the semantic segmentation model is used to extract the texture and boundary information of the ground object echo, thereby completing the accurate discrimination of meteorological and ground object echoes. The specific steps are as follows:

[0057] Step 1: Construction of a 2D convolutional neural network and algorithm training scheme. Based on the above high-precision label information, the classic U-structure Neural Network (U-Net) convolutional neural network architecture skip connection idea is retained to reduce the depth of the model (see the model structure for details). Figure 2 ). The four radar observation variables in the training data set were cut into 30X50 size and normalized as multi-channel values ​​and input into the model. By setting a decreasing learning rate, using the cross entropy loss function, and setting the batch size to 64, the algorithm was built after 50 iterations on a 160-core CPU. Finally, the recognition algorithm outputs the corresponding probability recognition results in units of 30X50 size, and the recognition results are spliced ​​into the size of the original radar observation variables.

[0058] Step 2: Computer vision-based method for determining the boundary of ground clutter. Based on the constructed two-dimensional U-Net meteorological and ground object echo intelligent classification model, in order to avoid the potential interference of output size on the determination of ground object boundary, the probability output result of the model is further post-processed. The edge information is obtained by calculating the Sobel gradient operator on the probability discrimination result, and then the watershed algorithm is used to extract the ground object edge to complete the ground object and meteorological echo recognition in the PPI observation mode.

[0059] Through the above detailed technical solutions, the present invention can effectively realize high-precision recognition of ground objects, and provide strong technical support for research and application in related fields.

[0060] Based on the specific implementation of the above embodiments, the present invention has the following significant advantages and positive effects:

[0061] (1) Improve recognition accuracy: By comprehensively analyzing the changing characteristics of different observation variables of long-term millimeter-wave radar, the present invention can effectively distinguish ground object echoes from meteorological echoes, reducing misjudgments and missed judgments. The position stability of ground object echoes and the moving speed characteristics of meteorological targets are used to improve the accuracy of ground object echo recognition.

[0062] (2) Development of a new recognition scheme: Combining computer vision technology and deep learning algorithms, the present invention has developed a new ground object echo recognition scheme. This scheme can effectively solve the problem that the traditional method has poor effect in identifying ground object echoes in mixed precipitation and ground object areas, especially at the edge of precipitation echoes and weak precipitation targets.

[0063] (3) Improving the reliability of weather radar data: Through the above technical solution, the present invention can improve the reliability of weather radar data and ensure the quality of meteorological services. This provides more reliable data support for severe convective weather warnings, extreme weather cause research, and the formulation of disastrous weather emergency plans.

[0064] (4) Reducing dependence on velocity parameters: The present invention reduces dependence on velocity parameters. By comprehensively analyzing multiple radar variables, the risk of not being able to effectively distinguish between ground object echoes and precipitation echoes in areas where the velocity is close to zero is reduced, especially in areas where precipitation and ground objects are mixed.

[0065] (5) Simplify the algorithm recognition process: The present invention makes full use of the observation variables provided by the cloud radar, simplifies the algorithm structure, and improves the portability of the algorithm. Through the application of the deep learning model U-Net, the present invention can automatically extract the texture and boundary information of the ground object echo, reduce the dependence on a single variable, and improve the efficiency and accuracy of the algorithm.

[0066] (6) Real-time recognition capability: The technical solution of the present invention is not only applicable to post-analysis, but also can realize real-time recognition of ground object echoes. By training the intelligent recognition algorithm, the present invention can quickly and accurately identify ground object echoes in practical applications, thereby improving the processing speed and response capability of meteorological radar data.

[0067] (7) Improved data processing efficiency: Through automated object echo recognition and elimination, the present invention reduces the need for manual intervention and improves data processing efficiency, which is particularly important for meteorological services that require rapid response.

[0068] Furthermore, the core of the present invention lies in its unique ground object echo recognition method, which improves the accuracy and efficiency of millimeter wave cloud radar ground object echo recognition through precise true value calibration process and advanced model architecture. The following are the key technologies claimed by the present invention:

[0069] 1. True value calibration process: The true value calibration method involved in the present invention includes randomly sampling radial velocity variables under the millimeter-wave cloud radar PPI scanning mode for a relatively long time and calculating the standard deviation distribution, and then extracting the true value of the ground object based on computer vision technology, using the Sobel gradient operator to calculate the edge gradient of the longitudinal velocity standard deviation, and using the watershed algorithm to automatically identify the boundary position, so as to realize the extraction of ground clutter boundaries and true value calibration.

[0070] 2. Model post-processing solution: Perform Sobel gradient operator calculation on the probability output results of the model to obtain edge information, use the watershed algorithm to extract the edge of the ground object, and complete the identification of ground objects and meteorological echoes under the PPI observation mode.

[0071] These key technologies of the present invention not only improve the accuracy of ground object echo recognition, but also significantly improve the efficiency and reliability of weather radar data processing through automated and intelligent methods.

[0072] Embodiment 2

[0073] This embodiment also provides a computer terminal device, including:

[0074] one or more processors;

[0075] A memory, coupled to the processor, for storing one or more programs;

[0076] When the one or more programs are executed by the one or more processors, the one or more processors implement a cloud radar object echo recognition method based on deep learning and physical statistics fusion.

[0077] This embodiment also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, a cloud radar object echo recognition method based on deep learning and physical statistics fusion is implemented.

[0078] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cloud radar object echo recognition method based on deep learning and physical statistics fusion, characterized in that: The following steps are involved: Obtain the observed variables under the millimeter wave cloud radar PPI scanning mode; Determining the true value of the ground feature based on the observed variables; A ground object true value sample is constructed based on the ground object true value, and a two-dimensional convolutional neural network is used to accurately distinguish the meteorological and ground object echoes of the ground object true value sample to obtain a final recognition result.

2. The cloud radar object echo recognition method based on deep learning and physical statistics fusion according to claim 1 is characterized in that: The observed variables include: radar reflectivity factor, radial velocity, velocity spectrum width and signal-to-noise ratio.

3. The cloud radar object echo recognition method based on deep learning and physical statistics fusion according to claim 1 is characterized in that: The process of determining the true value of the ground object based on the observed variables includes: Sorting the observed variables in chronological order to obtain changes of different radar variables in time series and constructing a data set; The data set is shuffled in time order and a number of samples are extracted to obtain a sub-sample set; Calculating the standard deviation distribution of different radar variables in the sub-sample set in time to obtain a standard deviation result; Comparing the standard deviation result with the original observation result to obtain the radar variable that best fits the actual observation; Based on computer vision technology, the ground object clutter is completely extracted from the radar variable that best fits the actual observation to obtain the ground object true value.

4. The cloud radar object echo recognition method based on deep learning and physical statistics fusion according to claim 3 is characterized in that: The process of completely extracting ground clutter from the radar variables that best fit actual observations based on computer vision technology includes: Based on the radar variable that best fits the actual observation, the edge gradient of the standard deviation of the longitudinal velocity is calculated using the Sobel gradient operator to obtain an edge detection result; Based on the edge detection result, a watershed algorithm is used to automatically identify the boundary position to extract the ground clutter boundary and calibrate the true value.

5. The cloud radar object echo recognition method based on deep learning and physical statistics fusion according to claim 1 is characterized in that: The process of using a two-dimensional convolutional neural network to accurately discriminate the meteorological and ground object echoes of the ground object true value samples to obtain the final recognition result includes: Building a two-dimensional U-Net neural network, and using ground object true value samples and corresponding radar observation variables to train the two-dimensional U-Net neural network to obtain a trained two-dimensional U-Net neural network; Based on the trained two-dimensional U-Net neural network, new radar data is intelligently identified and probability identification results of ground object echoes and weather echoes are output.

6. The cloud radar object echo recognition method based on deep learning and physical statistics fusion according to claim 5 is characterized in that: The ground object echo recognition method also includes post-processing the probability recognition result output by the two-dimensional U-Net neural network; wherein the post-processing process includes: using the Sobel gradient operator and the watershed algorithm to determine the ground object clutter boundary.

7. A computer terminal device, characterized in that: include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the cloud radar object echo recognition method based on deep learning and physical statistics fusion as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the cloud radar object echo recognition method based on deep learning and physical statistics fusion as described in any one of claims 1 to 6 is implemented.

Images