Microwave laser composite radar cloud identification method and computer program product

Abstract

The invention discloses a microwave laser composite radar cloud identification method and a computer program product, and the method comprises the steps: removing ground clutters based on transfer learning and an R-FCN network, and eliminating the noise caused by the ground clutters of a microwave radar; a deep learning model is constructed by using transfer learning and R-FCN algorithms, microwave radar and laser radar echo images are input, an attention mechanism is optimized and used in the model to realize adaptive feature extraction, a multi-scale feature pyramid is introduced on the basis of an RPN network, and the recognition and detection capability of cloud regions with different scales is enhanced. Single-result joint inversion of microwave radar and laser radar cloud identification is realized by utilizing cloud boundary self-inspection, cloud feature marking, initial matching and joint inversion of cloud boundaries, and a composite identification cloud area is obtained. The method has a cloud identification function under complex atmospheric conditions such as multilayer clouds and equipment noise interference, is accurate in inversion result, is not sensitive to ground clutter and daytime background noise, and fully integrates the identification advantage of a laser radar on a thin-roll cloud area and a cloud bottom and the cloud top identification advantage of a microwave radar.

Description

Technical Field

[0001] The present invention belongs to the technical field of microwave laser composite radar cloud detection, and relates to a method for identifying microwave laser composite radar clouds based on transfer learning and R-FCN network. Background Technique

[0002] Clouds are visible polymers floating in the air, which are composed of small water droplets liquefied from water vapor in the atmosphere when cooled or small ice crystals sublimated. Clouds cover about 50% of the Earth's surface. They are meteorological factors that are of great concern in atmospheric science research and are also one of the key weather elements affecting people's production and life. In cloud physics research, the physical nature of clouds is often revealed from a certain aspect by studying the macroscopic structure of clouds. Among them, cloud height is one of the most basic but very important cloud physical parameters. The height of the cloud base, the shape of the cloud, and the structure of the cloud are considered to be the most direct and effective indicators for artificially distinguishing cloud types. Observing cloud macroscopic characteristics such as cloud height and cloud vertical structure is one of the most basic but very urgent observation requirements.

[0003] The macroscopic structure of clouds is closely related to cloud radiation characteristics, thus playing an important role in changing the thermal structure of the atmosphere and affecting atmospheric circulation. Before using radar observations to invert cloud macroscopic characteristics, it is first necessary to distinguish the noise in the original data from the cloud echo signal. In cloud detection, for strong echo signals significantly higher than the noise level, they can be easily identified by setting thresholds. However, when the cloud echo signal is weak and close to the noise level, it greatly increases the difficulty of accurate identification, which has always been a difficult problem in radar target recognition. These weak echo signals generally come from small-sized or thin ice cloud particles, often appearing at the top boundary of the cloud layer or high-altitude areas. The clouds in these positions have a strong greenhouse effect, which will significantly affect the radiation budget and is often related to the formation or disappearance of clouds. Accurately identifying these cloud droplet particles is of great significance for understanding the cloud formation and disappearance process and its role in the climate energy balance.

[0004] Conventional ground observations and satellite observations are important means for studying the climate characteristics of clouds, but they often cannot provide complete cloud macroscopic structure information. Ground observations are difficult to penetrate low clouds and accurately observe middle and high clouds; at the same time, they can only record the cloud base height and cannot provide cloud height information. Satellites, on the other hand, have difficulties in observing low clouds and determining the cloud base height. Although the temperature and humidity information obtained from radiosondes can accurately and completely invert the vertical distribution of clouds, radiosondes cannot continuously observe a specific area or type of cloud at fixed points. The all-sky imager can only obtain the all-sky cloud amount, cannot obtain cloud macroscopic parameter information, and only works during the day; the infrared imager can achieve day and night observations and can obtain cloud height and cloud amount information, but the inversion error is large and it is difficult to meet the business requirements. Summary of the Invention

[0005] The technical problem solved by this application is: overcoming the deficiencies of the prior art, providing a microwave-laser composite radar cloud recognition method based on transfer learning and R-FCN network, with cloud recognition function under complex atmospheric conditions such as multi-layer clouds and equipment noise interference, accurate inversion results, insensitive to ground clutter and daytime background noise, and fully integrating the recognition advantages of lidar for thin cirrus cloud areas and cloud bases and the recognition advantages of microwave radar for cloud tops. The method for jointly recognizing cloud boundaries is applicable to cloud recognition in the atmosphere and can effectively improve the accuracy and precision of cloud recognition.

[0006] The joint observation of lidar and microwave radar is the most important method for observing and studying local clouds at present. In order to more accurately determine the cloud boundaries under complex conditions, by using active detection means, combining the advantages of microwave radar and lidar, and complementing the shortcomings of the two detection methods, a microwave-laser composite radar cloud recognition method based on transfer learning and R-FCN network is proposed.

[0007] This method not only solves the problem of limited penetration of a single lidar into clouds, but also makes up for the deficiency of poor detection ability of a single microwave radar for thin clouds and water clouds with small particles, and can accurately obtain the cloud-aerosol vertical structure and cloud phase information, which is the best means for observing cloud macroscopic parameters at present. The application of deep learning, especially convolutional neural networks, in image processing provides a more accurate solution. Transfer learning improves performance on a small-scale specific domain set by using models pre-trained on large-scale data sets. The R-FCN network has advantages in object detection tasks by combining the Region Proposal Network (RPN) and the Fully Convolutional Network (FCN).

[0008] The present invention proposes a microwave-laser composite radar cloud recognition method based on transfer learning and R-FCN network, which uses an attention mechanism to achieve adaptive feature extraction, introduces a multi-scale feature pyramid on the basis of the RPN network, and enhances the recognition and detection ability for cloud areas of different scales. Realize the cloud area recognition of lidar and microwave radar, and complete the cloud area recognition task under composite conditions through cloud boundary self-check, cloud feature marking and the method of jointly inverting cloud boundaries.

[0009] The technical solutions provided by this application are as follows:

[0010] A microwave-laser composite radar cloud recognition method, including:

[0011] S1. For the reflectivity factor image and linear depolarization ratio LDR obtained by the operation of the microwave radar, remove the ground clutter from the reflectivity factor image and linear depolarization ratio LDR to obtain an effective reflectivity factor image and an effective linear depolarization ratio LDR;

[0012] S2. Use the lidar to collect and obtain a distance-corrected signal image after distance correction;

[0013] S3. Based on the distance-corrected signal image and the effective reflectivity factor image, use a cloud recognition algorithm based on transfer learning and the region-based fully convolutional R-FCN network to locate and identify cloud regions, and obtain the exact positions of cloud regions on the reflectivity factor image and the distance-corrected signal image;

[0014] S4. Cloud boundary self-check. In the reflectivity factor image, determine whether the cloud base and cloud top of the cloud region are paired, remove the bad values that do not exist in pairs, and retain the paired cloud top and cloud bottom boundaries. A pair of cloud top and cloud bottom boundaries is used as a cloud information; in the distance-corrected signal image, determine whether the cloud base and cloud top of the cloud region are paired, remove the bad values that do not exist in pairs, and retain the paired cloud top and cloud bottom boundaries. A pair of cloud top and cloud bottom boundaries is used as a cloud information;

[0015] S5. Cloud feature marking. According to the time and height of the reflectivity factor image and the time and height of the distance-corrected signal image, match the cloud information in the reflectivity factor image with the cloud information in the height-distance corrected signal image one by one, and determine the retained cloud information according to the matching situation to obtain the final cloud information.

[0016] In the above S3, the cloud recognition algorithm based on transfer learning and the region-based fully convolutional R-FCN network includes:

[0017] The distance-corrected signal image obtained in S2 and the effective reflectivity factor image obtained in S1 generate multiple position-sensitive score maps through the trained R-FCN network model, and classify and vote on each position-sensitive score map to obtain the exact positions of cloud regions on the distance-corrected signal image and the effective reflectivity factor image;

[0018] Perform morphological processing on the cloud regions on the distance-corrected signal image and the effective reflectivity factor image to obtain the accurate morphology of the cloud regions on the distance-corrected signal image and the effective reflectivity factor image.

[0019] The trained R-FCN network model is obtained through the following methods:

[0020] Use the reflectivity factor image and the distance-corrected signal image as the data set, manually locate and label the cloud regions in the image to obtain the data set;

[0021] Use the method of transfer learning to transfer the public data set to the data set;

[0022] Divide the data set into a training set and a test set;

[0023] Build an R-FCN network model, which includes a feature extraction network module, a region generation network module, and a region of interest subnet module; the feature extraction network module is used to generate corresponding convolutional feature maps according to the reflectivity factor image and / or the distance correction signal image; the region generation network module is used to generate cloud region localization in the convolutional feature map to obtain a convolutional feature map with cloud region localization; the region of interest subnet is used to perform operations on the convolutional feature map with cloud region localization and multiple convolutional kernels using fully convolutional layers to generate multiple position-sensitive score maps;

[0024] Configure training parameters for the R-FCN network model, where the training parameters include the number of iterations, the dataset category, and the initial learning rate;

[0025] Take the training set as the input, and the loss function is used according to and with the convolutional feature map with cloud region localization, the reflectivity factor image in the dataset corresponding to the convolutional feature map with cloud region localization, and the time or height distance correction signal image to obtain an evaluation result, and then update the training parameters of the R-FCN network model according to the evaluation result until the evaluation result meets the requirements to obtain a trained R-FCN network model.

[0026] The loss function is the cross-entropy loss function. Use the cross-entropy loss function to judge whether each convolutional feature map with cloud region localization contains a cloud region; according to whether it contains a cloud, give the model prediction probability Pi; if it does not contain a cloud region, Pi = 0; if it contains a cloud region, Pi = 1; according to the model prediction probability Pi, calculate the output value of the cross-entropy loss function. If the output value is less than the threshold, perform the next training; if the output value is greater than the threshold, adjust the initial learning rate until the output value of the cross-entropy loss function meets the requirements.

[0027] The distance correction signal image obtained in S2 and the effective reflectivity factor image obtained in S1 generate multiple position-sensitive score maps through the trained R-FCN network model, and classify and vote on each position-sensitive score map to obtain the accurate positions of the cloud regions on the distance correction signal image and the effective reflectivity factor image, including:

[0028] Weight-average the prediction probabilities of multiple types of atmospheric targets included in each position-sensitive score map to obtain the specific category of this position-sensitive score map. The multiple types of atmospheric targets include clouds and others. If the category is cloud, this position-sensitive score map is retained as a cloud region. If it is other, it is not retained as a cloud region; determine the accurate cloud region position according to the position-sensitive score maps retained as cloud regions.

[0029] Perform morphological processing on the cloud regions on the distance correction signal image and the effective reflectivity factor image to obtain the accurate morphology of the cloud regions on the distance correction signal image and the effective reflectivity factor image, including:

[0030] Morphological processing includes dilation and erosion. Erosion is used to remove noise, and dilation is used to merge adjacent cloud regions. In S1, removing ground clutter from the reflectivity factor image and the linear depolarization ratio LDR includes:

[0031] noise = (LDR > -22) ∩ (Z < -5); where, is the linear depolarization ratio; Z is the reflectivity factor.

[0032] In S5, according to the time and height of the reflectivity factor image and the time and height of the time and height distance corrected signal image, the cloud information in the reflectivity factor image is matched one by one with the cloud information in the time and height distance corrected signal image, and the cloud information to be retained is determined according to the matching situation to obtain the final cloud information, including:

[0033] S51. According to the time and height of the reflectivity factor image and the time and height of the time and height distance corrected signal image, the cloud information in the reflectivity factor image is matched one by one with the cloud information in the time and height distance corrected signal image;

[0034] S52. For the cloud information that is matched and corresponding in S1 and belongs to the same target cloud region, for the same target cloud region, there are the following situations: there is no cloud information in both the reflectivity factor image and the time and height distance corrected signal image, and there is no corresponding cloud information finally; there is only cloud information in the reflectivity factor image, and the final cloud information is based on the reflectivity factor image; there is only cloud information in the time and height distance corrected signal image, and the final cloud information is based on the time and height distance corrected signal image; there is cloud information in both the reflectivity factor image and the time and height distance corrected signal image, and it is necessary to comprehensively consider the mutual relationship of the clouds in the reflectivity factor image and the time and height distance corrected signal image to obtain the final cloud information.

[0035] When there is cloud information in both the reflectivity factor image and the time and height distance corrected signal image, it is necessary to comprehensively consider the mutual relationship of the clouds in the reflectivity factor image and the time and height distance corrected signal image to obtain the final cloud information, including:

[0036] For non-precipitating clouds, the lowest cloud base in the reflectivity factor image and the time and height distance corrected signal image is used as the final cloud boundary cloud base of the cloud region, and the highest cloud top in the reflectivity factor image and the time and height distance corrected signal image is used as the final cloud boundary cloud top of the cloud region.

[0037] A computer program product includes a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above-mentioned microwave laser composite radar cloud recognition method are implemented.

[0038] A cloud recognition method for microwave-laser combined radar based on transfer learning and R-FCN network, the method includes: removing ground clutter, cloud recognition algorithm based on transfer learning and region fully convolutional network (R-FCN), cloud boundary self-check, cloud feature marking, initial matching, and joint inversion of cloud boundary, etc.

[0039] 1) Remove the noise caused by ground clutter of the microwave radar through the ground clutter removal method to complete the preprocessing of the dataset; 2) Use transfer learning and R-FCN algorithm to construct a deep learning model, input the echo images of the microwave radar and lidar, and optimize the use of the attention mechanism in the model to achieve adaptive feature extraction. Introduce a multi-scale feature pyramid on the basis of the RPN network to enhance the recognition and detection ability of cloud regions at different scales. 3) Use methods such as cloud boundary self-check, cloud feature marking, initial matching, and joint inversion of cloud boundary to achieve the joint inversion of the single results of cloud recognition of the microwave radar and lidar, and obtain the composite recognition cloud region.

[0040] In summary, the present application at least includes the following beneficial technical effects:

[0041] The present invention proposes a cloud recognition method for microwave-laser combined radar based on transfer learning and R-FCN network, uses the attention mechanism to achieve adaptive feature extraction, introduces a multi-scale feature pyramid on the basis of the RPN network, and enhances the recognition and detection ability of cloud regions at different scales. Realize the cloud region recognition of lidar and microwave radar, and complete the cloud region recognition task under composite conditions through methods such as cloud boundary self-check, cloud feature marking, and joint inversion of cloud boundary. It can be used for cloud boundary recognition under various complex weather conditions, is less affected by ground clutter and background noise, has a high degree of automation, and can obtain more accurate results.

[0042] This method has the function of cloud recognition under complex atmospheric conditions such as multi-layer clouds and under equipment noise interference. The inversion result is accurate, insensitive to ground clutter and daytime background noise, and fully integrates the recognition advantages of lidar for thin cirrus cloud regions and cloud bases and the recognition advantages of microwave radar for cloud tops. The joint recognition cloud boundary method is applicable to cloud recognition in the atmosphere and can effectively improve the accuracy and precision of cloud recognition. This method can be migrated to spaceborne platforms for application, and the recognized cloud regions can lay a foundation for the inversion of global three-dimensional wind field information by spaceborne microwave-laser combined wind measurement radar. Description of the Drawings

[0043] Figure 1 It is the cloud boundary result obtained after the lidar image passes through the R-FCN network model and classification voting. Among them, the horizontal axis is the event and the vertical axis is the height;

[0044] Figure 2 It is the cloud boundary result obtained after the microwave radar reflectivity factor passes through the R-FCN network model and classification voting;

[0045] Figure 3 It is the result map of cloud area recognition obtained by a microwave-laser combined radar cloud recognition method based on transfer learning and R-FCN network;

[0046] Figure 4 It is the technical roadmap of a microwave-laser combined radar cloud recognition method based on transfer learning and R-FCN network;

[0047] Figure 5 It is the cloud boundary result obtained after the lidar image passes through the R-FCN network model and classification voting under the condition of low cloud and ground clutter;

[0048] Figure 6 It is the cloud boundary result obtained after the microwave radar image passes through the R-FCN network model and classification voting under the condition of low cloud and ground clutter;

[0049] Figure 7 It is the result map of jointly identifying the cloud area by using the method of this application under the condition of low cloud and ground clutter. Detailed implementation mode

[0050] To make the objectives, technical solutions and advantages of the present invention clearer, the following will further describe in detail the disclosed implementation modes of the present invention in conjunction with the accompanying drawings.

[0051] This application embodiment discloses a microwave-laser combined radar cloud recognition method based on transfer learning and R-FCN network, as Figure 4 shown, including the following steps:

[0052] S1. Remove ground clutter: The microwave radar works to obtain the reflectivity factor Z and the linear depolarization ratio LDR. The reflectivity factor Z is continuously collected over a period of time to obtain a reflectivity factor image. Ground clutter is removed from the reflectivity factor Z and the linear depolarization ratio LDR to obtain an effective reflectivity factor Z and an effective linear depolarization ratio LDR;

[0053] The non-meteorological clutter observed when the microwave radar works is usually caused by the echo of low-altitude suspended matter, including haze, dust, insects, etc. in the atmosphere. In the microwave radar cloud recognition method, more or less, ground clutter will be recognized as clouds, which will affect the extraction result of the composite cloud boundary. Ground clutter must be removed before joint inversion with the lidar.

[0054] Considering that the general characteristics of these ground clutters are weak echo intensity, small movement speed, and very high linear depolarization ratio. Therefore, a method of combining the echo intensity reflectivity factor Z and the linear depolarization ratio LDR is selected to remove clutter, and the calculation formula is as follows:

[0055] noise = (LDR > -22) ∩ (Z < -5);

[0056] Among them, noise is noise, that is, clutter.

[0057] S2. Obtain the atmospheric echo signal by using lidar. The atmospheric echo signal is corrected by distance to obtain the distance-corrected signal. The distance-corrected signal is continuously collected within a period of time to obtain the distance-corrected signal image. According to the distance-corrected signal image and the reflectivity factor image, a cloud recognition algorithm based on transfer learning and region-based fully convolutional R-FCN network is used to locate and identify the cloud area, and obtain the accurate positions of the cloud area on the reflectivity factor image and the distance-corrected signal image.

[0058] The main process of the cloud recognition algorithm based on transfer learning and region-based fully convolutional R-FCN network is as follows:

[0059] (1) Data preprocessing: Respectively take the reflectivity factor image and the distance-corrected signal image as the data sets, and manually locate and label the cloud-containing areas (i.e., cloud areas) in the images to complete the production of the data sets.

[0060] (2) Transfer learning: Use the method of transfer learning to transfer the public data set to the data sets in step (1); the public data set includes multiple reflectivity factor images and distance-corrected signal images labeled with cloud areas.

[0061] (3) Divide the data set: Divide the data set into a training set and a test set according to the ratio of 8:2, and there is no cross-repeated data between the training set and the test set.

[0062] (4) R-FCN network model

[0063] The core of the R-FCN network model is the position-sensitive score map. The R-FCN network model includes a feature extraction network, a region proposal network (RPN), and a region of interest subnet (ROI).

[0064] The atmospheric echo signal is collected by lidar. The atmospheric echo signal is corrected for distance to obtain a distance-corrected signal. The distance-corrected signal is continuously collected over a period of time to obtain a distance-corrected signal image. 1) The reflectivity factor image and the distance-corrected signal image collected by the microwave radar and the lidar respectively are input into the feature extraction network. The feature extraction network generates corresponding convolutional feature maps based on the reflectivity factor image of the microwave radar and the distance-corrected signal image of the lidar. These convolutional feature maps contain information related to the spatial structure, position distribution, and distance of clouds, etc. 2) The Region Proposal Network (RPN) receives and analyzes the two convolutional feature maps, generates cloud region localization on the convolutional feature maps, and obtains a convolutional feature map with cloud region localization. 3) The Region of Interest subnet (ROI) uses a fully convolutional layer to operate on the convolutional feature map with cloud region localization and multiple convolutional kernels to generate position-sensitive score maps (Scores Maps). 4) Finally, the position-sensitive score maps are classified and voted to obtain the accurate positions of the cloud regions in the microwave radar and lidar images, realizing the localization and recognition of clouds. As Figure 1 and Figure 2 shown.

[0065] Specifically, classifying and voting on the position-sensitive score maps includes: The position-sensitive score map represents the prediction probability that each position on the convolutional feature map with cloud region localization is classified as a certain type of atmospheric target. A certain type of atmospheric target is cloud or others, and others include background and aerosol. The prediction probabilities of multiple types of atmospheric targets included in each position-sensitive score map are weighted and averaged to obtain the specific category of this position-sensitive score map. If the category is cloud, this position-sensitive score map is retained as a cloud region. If it is others, it is not retained as a cloud region. According to the position-sensitive score maps retained as cloud regions, the accurate cloud region positions are determined.

[0066] The convolutional feature map passes through the RPN to generate a convolutional feature map with cloud region localization. A multi-scale feature pyramid design is introduced to obtain information from feature maps of different scales and identify cloud regions of different sizes.

[0067] ResNet is selected as the feature extraction network. ResNet is a deep neural network architecture designed to solve the problem of gradient descent (difficulty in updating weights) brought about by the increase in the number of network layers. Transfer learning is a machine learning method that improves learning efficiency by transferring the knowledge learned in an existing domain model to the cloud recognition domain. The pre-trained ResNet model using an existing large dataset is used for feature extraction to adapt it to the data characteristics of lidar and microwave radar. The attention mechanism is used to improve the model's attention to cloud region features in lidar and microwave radar images, ignore noise, and improve the recognition accuracy of clouds.

[0068] The residual blocks introduced by the ResNet model are used to solve the gradient descent problem in neural networks. Based on the pre-trained model, transfer learning and attention mechanism are used for fine-tuning to make it adapt to the specific features of lidar and microwave radar images, which can effectively improve the accuracy and efficiency of extracting cloud areas from lidar and microwave radar images.

[0069] (5) Configure the training parameters, including: the number of iterations is set to 30,000 times, the dataset category is set to 3, the initial learning rate is 0.001, and the number of data processed each time is 1.

[0070] (6) Training and optimization: Use the training set as the input to train and optimize the R-FCN network model. The loss function is used to optimize the model. The loss function is the cross-entropy loss function or the bounding box regression loss function. Adjust the training parameters according to the evaluation results to improve the classification accuracy.

[0071] That is: The loss function obtains the evaluation result based on the convolutional feature map with cloud area localization, the reflectivity factor image and the time or height-distance correction signal image in the dataset corresponding to the convolutional feature map of the cloud area localization, and then updates the parameters of the R-FCN network model according to the evaluation result.

[0072] Taking the cross-entropy loss function as an example specifically, it includes: using the cross-entropy loss function to judge whether each convolutional feature map with cloud area localization contains a cloud area; according to whether it contains a cloud, give the model prediction probability P i ; if it does not contain a cloud area, P i = 0; if it contains a cloud area, P i = 1; according to the model prediction probability P i , calculate the output value of the cross-entropy loss function (i.e., the evaluation result). If the output value is less than the threshold, perform the next training; if the output value is greater than the threshold, adjust the initial learning rate.

[0073] (7) Cloud area post-processing: Use the test set as the input to obtain the cloud area detection result. Perform morphological processing on the extraction result. The morphological processing includes operations such as dilation and erosion. Erosion realizes noise removal, and dilation realizes merging adjacent cloud areas to improve the extraction accuracy.

[0074] The cloud recognition algorithm based on transfer learning and region fully convolutional R-FCN network is used to realize cloud recognition of microwave radar and lidar. By making full use of the similarity between data, tasks and models, the model learned in the old field is applied in the new field, reducing the requirements for the amount of training data, the similarity of the feature distributions of training data samples and test data samples. In the design, a feature extraction module containing an attention mechanism is added, which can further enhance the extraction of cloud area features.

[0075] S3. Cloud boundary self-check: The cloud boundary self-check mainly determines whether the cloud base and cloud top exist in pairs. First, it is necessary to separately determine whether the cloud boundary results obtained by single radar inversion exist in pairs. If they exist in pairs, the preliminary extraction results are recognized. If the cloud boundaries do not exist in pairs, it is necessary to determine which of the cloud top or cloud base has a bad value and remove this bad value to obtain paired cloud top and cloud base boundaries.

[0076] Specifically, it includes: In the reflectivity factor image, determine whether the cloud base and cloud top in the cloud area exist in pairs, remove the bad values that do not exist in pairs, and retain the paired cloud top and cloud base boundaries. A set of paired cloud top and cloud base boundaries is used as a cloud information. In the range-corrected signal image, determine whether the cloud base and cloud top in the cloud area exist in pairs, remove the bad values that do not exist in pairs, and retain the paired cloud top and cloud base boundaries. A set of paired cloud top and cloud base boundaries is used as a cloud information.

[0077] S4. Cloud feature marking: After the cloud boundary in step S3, it can be considered that the cloud boundaries exist in pairs and are not affected by ground clutter. On this basis, proceed with:

[0078] S41. Initial matching of cloud boundaries: Before the joint inversion of cloud boundaries, it is first necessary to determine that the cloud boundaries obtained by the two working systems are for the same target. After removing ground clutter, it is necessary to ensure as much as possible that the target information obtained by a single detection method exists in pairs and corresponds one by one. In actual situations, there will be cases where the result of one detection method is greater than that of another detection method, but it represents the same target cloud body (the cloud body is the cloud area). Therefore, it is necessary to traverse and mark the cloud boundaries of lidar and microwave radar simultaneously, perform logical calculations on the same target cloud body, and save the typical features of different cloud bodies.

[0079] S42. Joint inversion of cloud boundaries: For non-precipitating clouds, the lowest cloud base is used as the cloud base of the final cloud boundary; the highest cloud top is used as the cloud top of the final cloud boundary. If there is only the result of a certain detection method, the final cloud boundary is assigned with this value. This situation corresponds to the case where there is only the water cloud result detected by lidar or the high and thick cloud detected by microwave radar in actual detection, and lidar does not detect it.

[0080] Specifically, it includes: The cloud boundaries extracted based on microwave radar and lidar are mainly divided into the following four situations: 1) If neither type of radar has cloud information, there is no corresponding cloud information in the final inversion result; 2) If lidar has no cloud information and only microwave radar cloud information exists, then the final cloud information is based on microwave radar; 3) If lidar has cloud information and microwave radar has no cloud information, the final result is based on lidar; 4) If both radars have cloud information, which is also the most complex situation, it is necessary to comprehensively consider the mutual relationship between the clouds obtained by the two radars to obtain the final cloud information.

[0081] In the above four cases, cloud boundary initial matching needs to be considered in each case to determine the consistency of the inversion target. Otherwise, the value substituted into the logic is an invalid value, which will affect the final cloud boundary extraction result.

[0082] The result map of cloud area recognition obtained by a microwave laser composite radar cloud recognition method based on transfer learning and R-FCN network according to the present application is as Figure 3 shown, while a single detection means cannot complete the recognition of the cloud area.

[0083] For different times, under the condition of low clouds and ground clutter, the cloud boundary results obtained after the lidar image passes through the R-FCN network model and classification voting are as Figure 5 shown, the cloud boundary results obtained after the microwave radar image passes through the R-FCN network model and classification voting are as Figure 6 shown, and the result of jointly recognizing the cloud area by using the method of the present application is as Figure 7 shown.

[0084] The present application fully integrates the recognition advantages of lidar for cirrus cloud areas and cloud bases and the recognition advantages of microwave radar for cloud tops. This method of jointly recognizing cloud boundaries is applicable to cloud recognition in the atmosphere and can effectively improve the accuracy and precision of cloud recognition.

[0085] Specifically, step S2 includes:

[0086] S21. The reflectivity factor image and the range correction signal image respectively generate corresponding convolutional feature maps through the feature extraction network;

[0087] The deep residual neural network (ResNet) is a network formed by repeatedly stacking residual learning modules, which solves the problem of gradient dispersion during model training and thus improves the learning and recognition ability of the model. In this method, ResNet-101 is selected as the feature extraction network;

[0088] Assume the input is x, and after the operations of convolutional layers W 1 and W 2 , the output is F(x, W 1 , W 2 ). The activation function is ReLU, and the output y of the residual module is expressed as:

[0089] y = F(x, W 1 , W 2 ) + W s x

[0090] where W 1 , W 2 are the learning weights corresponding to the first and second layers of the convolutional neural network respectively; F(x, W 1 , W 2) is a mapping function composed of convolution activation functions, representing the output after calculation by the convolution layer; the reflectivity factor or distance correction signal of each point in the reflectivity factor image or the time and height distance correction signal image is brought into the mapping function F(x, W 1 , W 2 ), and multiple y values are calculated to form the convolution feature map; W s is the weight of the variable x from input to output. The input x is the reflectivity factor image and the distance correction signal image;

[0091] S22. Position score sensitive map and pooling; In order to accurately obtain the cloud area position on the convolution feature map with cloud area localization, the convolution layer obtained after performing convolution operation on the original image is used to generate the position sensitive score map. "The convolution feature map with cloud area localization and multiple convolution kernels are operated to generate the position sensitive score map" - this process generates k 2 position sensitive score maps. There are c types of atmospheric targets (clouds, aerosols) and 1 background, corresponding to the number of channels k 2 (c + 1). In the algorithm model of this paper, k = 3 is set. The meaning of k is: the length and width of the convolution feature map are sliced into k equal parts. The detection categories are clouds and aerosols, and the number of categories c = 2. After the RPN gives the ROIs of interest, by performing position sensitive pooling on each convolution feature map with cloud area localization, the position sensitive score map with the number of channels k 2 (c + 1) is obtained. When k = 3, the convolution feature map with cloud area localization can be divided into 9 rectangular blocks, and then the corresponding position sensitive score map is also divided into 9 blocks, namely (top - left, top - center, top - right,..., bottom - right). Then, mean pooling is performed on the divided position sensitive score map to obtain a vector with a length of c + 1, and softmax classification (i.e., classification voting) is performed on this vector to finally obtain the detection category of the cloud area in the convolution feature map with cloud area localization.

[0092] The specific operation of position sensitive pooling is to divide a convolution feature map with cloud area localization of size ω×h into k 2 rectangular blocks of size (ω / k)×(h / k), and use (i, j) to represent the position of each block in the convolution feature map with cloud area localization and the position sensitive score map, where the value range of (i, j) is (0 ≤ i ≤ k - 1, 0 ≤ j ≤ k - 1).

[0093] Place the block at (0, 0) of the convolution feature map with cloud area localization corresponding to the feature map with the number of channels k 2 pointed by the top - left arrow to the (0, 0) position of the position sensitive score map; the block at (0, 0) of the convolution feature map with cloud area localization corresponding to the feature map with the number of channels k pointed by the top - center arrow;2 The (1,0)th square in the convolutional feature map with cloud area positioning corresponding to the feature map is placed at the (1,0) position of the position sensitive score map, and so on. The specific operation to obtain the (i,j)th rectangular square in the position sensitive score map is:

[0094]

[0095] Among them, r c (i,j|θ) is the position-sensitive score map k of the cth class 2 The (i,j)th block among the blocks; Z i.j.c is the first feature map; bin(i,j) is the location set in the ROI and position-sensitive score map; (x 0 ,y 0 ) is the coordinate of the upper left corner of ROI; (x, y) is the coordinate of the upper left corner of ROI. 0 ,y 0 ) is the origin; n is the total number of pixels in the (i, j)th block; θ is the network parameter.

[0096] The mean pooling operation for generating a vector of length c+1 for the position-sensitive score map is:

[0097]

[0098] S23, loss function; using gradient descent for training, it is necessary to consider both classification loss and position loss. The loss function of each ROI is the sum of the cross entropy loss and the boundary regression loss, expressed as:

[0099]

[0100] Where λ = 1; c * is the class label of ROI (c * =0, indicating the background class of ROI); is the cross entropy loss function used for classification; L reg In is the parameter of the real regression box; t is the parameter of the regression box obtained by RPN t = (t x ,t y ,t w ,t h ).

[0101] The contents not described in detail in this application specification belong to the common knowledge of those skilled in the art.

[0102] The present application has been described in detail above in connection with specific embodiments and exemplary examples. However, these descriptions should not be construed as limiting the present application. Those skilled in the art understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and their implementation manners of the present application without departing from the spirit and scope of the present application, and all of these fall within the scope of the present application. The protection scope of the present application shall be subject to the appended claims.

Claims

1. A microwave laser composite radar cloud recognition method, characterized in that: include: S1. Remove ground clutter from the reflectivity factor image and the linear depolarization ratio LDR obtained by microwave radar operation to obtain an effective reflectivity factor image and an effective linear depolarization ratio LDR; S2, using laser radar to collect and obtain a distance correction signal image after distance correction; S3. According to the distance correction signal image and the effective reflectivity factor image, a cloud recognition algorithm based on transfer learning and regional fully convolutional R-FCN network is used to locate and identify the cloud area, and obtain the precise position of the cloud area on the reflectivity factor image and the distance correction signal image; S4. Cloud boundary self-check: In the reflectivity factor image, determine whether the cloud base and cloud top of the cloud area are paired, remove the bad values ​​that are not paired, retain the paired cloud top and cloud bottom boundaries, and a group of paired cloud top and cloud bottom boundaries are taken as a cloud information; In the distance correction signal image, determine whether the cloud base and cloud top of the cloud area are paired, remove the bad values ​​that are not paired, retain the paired cloud top and cloud bottom boundaries, and a group of paired cloud top and cloud bottom boundaries are taken as a cloud information; S5. Cloud feature marking. According to the time and height of the reflectivity factor image and the time and height distance correction signal image, the cloud information in the reflectivity factor image is matched with the cloud information in the height distance correction signal image one by one. The retained cloud information is determined based on the matching situation to obtain the final cloud information.

2. The microwave laser composite radar cloud recognition method according to claim 1, characterized in that: In S3, the cloud recognition algorithm based on transfer learning and regional full convolution R-FCN network includes: The distance-corrected signal image obtained by S2 and the effective reflectivity factor image obtained by S1 generate multiple position-sensitive score maps through the trained R-FCN network model, and classify and vote on each position-sensitive score map to obtain the precise position of the cloud area on the distance-corrected signal image and the effective reflectivity factor image; Morphological processing is performed on the cloud area on the distance correction signal image and the effective reflectivity factor image to obtain the accurate morphology of the cloud area on the distance correction signal image and the effective reflectivity factor image.

3. The microwave laser composite radar cloud recognition method according to claim 2, characterized in that: The trained R-FCN network model is obtained by: The reflectivity factor image and the distance correction signal image are used as data sets, and the cloud areas in the images are manually located and annotated to obtain the data sets; Use transfer learning methods to migrate public datasets into the dataset; Divide the dataset into training and testing sets; An R-FCN network model is established. The R-FCN network model includes a feature extraction network module, a region generation network module and an area of ​​interest subnetwork module; the feature extraction network module is used to generate corresponding convolution feature maps according to the reflectivity factor image and / or the distance correction signal image; the region generation network module is used to generate cloud area positioning in the convolution feature map to obtain a convolution feature map with cloud area positioning; the area of ​​interest subnetwork is used to use the full convolution layer to operate the convolution feature map with cloud area positioning and the multi-layer convolution kernel to generate multiple position sensitive score maps; Configure training parameters for the R-FCN network model, including the number of iterations, data set category, and initial learning rate; The training set is taken as input, and the loss function is based on and uses the convolutional feature map with cloud area positioning, the reflectivity factor image in the data set corresponding to the convolutional feature map with cloud area positioning, and the time or height distance correction signal image to obtain an evaluation result, and then the training parameters of the R-FCN network model are updated according to the evaluation result until the evaluation result meets the requirements to obtain a trained R-FCN network model.

4. The microwave laser composite radar cloud recognition method according to claim 3 is characterized by: The loss function is a cross entropy loss function, which is used to determine whether each convolution feature map with cloud area positioning contains a cloud area; according to whether it contains clouds, the model prediction probability P is given. i ; does not include cloud area, P i =0; including cloud area, P i =1; According to the model prediction probability P i , calculate the output value of the cross entropy loss function, if the output value is less than the threshold, perform the next training; If the output value is greater than the threshold, the initial learning rate is adjusted until the output value of the cross entropy loss function meets the requirements.

5. The microwave laser composite radar cloud recognition method according to claim 2, characterized in that: The distance correction signal image obtained by S2 and the effective reflectivity factor image obtained by S1 generate multiple position sensitive score maps through the trained R-FCN network model, and classify and vote on each position sensitive score map to obtain the precise position of the cloud area on the distance correction signal image and the effective reflectivity factor image, including: The predicted probabilities of multiple categories of atmospheric targets included in each position sensitive score map are weighted averaged to obtain the specific category of the position sensitive score map. The multiple categories of atmospheric targets include clouds and others. If the category is cloud, the position sensitive score map is retained as a cloud area. If it is others, it is not retained as a cloud area. According to the position sensitive score map retained as a cloud area, the accurate cloud area position is determined.

6. The microwave laser composite radar cloud recognition method according to claim 2, characterized in that: The step of performing morphological processing on the cloud area on the distance correction signal image and the effective reflectivity factor image to obtain accurate morphology of the cloud area on the distance correction signal image and the effective reflectivity factor image includes: Morphological processing includes dilation and erosion. Erosion is used to remove noise, and dilation is used to merge adjacent cloud areas.

7. The microwave laser composite radar cloud recognition method according to claim 1, characterized in that: In S1, the reflectivity factor image and the linear depolarization ratio LDR are used to remove ground clutter, including: noise=(LDR>-22)∩(Z<-5) Wherein, LDR is the linear depolarization ratio; Z is the reflectivity factor.

8. The microwave laser composite radar cloud recognition method according to claim 1, characterized in that: In S5, according to the time and height of the reflectivity factor image and the time and height distance correction signal image, the cloud information in the reflectivity factor image is matched with the cloud information in the height distance correction signal image one by one, and the retained cloud information is determined according to the matching situation to obtain the final cloud information, including: S51, according to the reflectivity factor image and the time and height of the time and height distance correction signal image, the cloud information in the reflectivity factor image is matched with the cloud information in the height distance correction signal image one by one; S52. For the cloud information matched in S1, which is the same target cloud area, for the same target cloud area, the following situations exist: there is no cloud information in both the reflectivity factor image and the height distance correction signal image, and finally there is no corresponding cloud information; the cloud information exists only in the reflectivity factor image, and the final cloud information is based on the reflectivity factor image; the cloud information exists only in the height distance correction signal image, and the final cloud information is based on the height distance correction signal image; there is cloud information in both the reflectivity factor image and the height distance correction signal image, and it is necessary to comprehensively consider the relationship between the clouds in the reflectivity factor image and the height distance correction signal image to obtain the final cloud information.

9. The microwave laser composite radar cloud recognition method according to claim 8, characterized in that: The reflectivity factor image and the height distance correction signal image both contain cloud information. It is necessary to comprehensively consider the relationship between the clouds in the reflectivity factor image and the height distance correction signal image to obtain the final cloud information, including: For non-precipitating clouds, the lowest cloud base in the reflectivity factor image and the height distance correction signal image is taken as the final cloud boundary cloud base of the cloud area, and the highest cloud top in the reflectivity factor image and the height distance correction signal image is taken as the final cloud boundary cloud top of the cloud area.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the microwave laser composite radar cloud identification method described in any one of claims 1-9 are implemented.

Images