Microwave-laser composite radar cloud identification method and computer program product

Abstract

The cloud recognition method and computer program product of the microwave laser composite radar are based on transfer learning and R-FCN network, remove the ground clutter, and remove the noise caused by the ground clutter of the microwave radar; a deep learning model is constructed by using transfer learning and R-FCN algorithm, microwave radar and laser radar echo images are input, the attention mechanism is used in the model to realize adaptive feature extraction, a multi-scale feature pyramid is introduced on the basis of the RPN network, and the recognition and detection capability for different scale cloud areas is enhanced. The single result of the microwave radar and the laser radar cloud recognition is realized by using cloud boundary self-checking, cloud feature marking, initial matching and joint inversion of the cloud boundary, and the composite recognition cloud area is obtained. The cloud recognition function is possessed under complex atmospheric conditions such as multi-layer clouds and under the interference of equipment noise, the inversion result is accurate, the ground clutter and the daytime background noise are not sensitive, and the recognition advantages of the laser radar for the thin roll cloud area and the cloud bottom and the recognition advantage of the microwave radar for the cloud top are fully integrated.

Description

Technical Field

[0001] This invention belongs to the field of microwave-laser composite radar cloud detection technology, and relates to a microwave-laser composite radar cloud identification method based on transfer learning and R-FCN network. Background Technology

[0002] Clouds are visible aggregates floating in the air, composed of tiny water droplets formed from the condensation of water vapor in the atmosphere upon cooling, or tiny ice crystals formed from its sublimation. Covering approximately 50% of the Earth's surface, clouds are a meteorological factor of great interest in atmospheric science research and a key weather element affecting people's lives and production. In cloud physics research, the macroscopic structure of clouds is often studied to reveal their physical nature from a particular perspective. Cloud height is one of the most fundamental yet crucial cloud physics parameters. The height of the cloud base, cloud morphology, and cloud structure are considered the most direct and effective indicators for artificially distinguishing cloud types. Observing macroscopic cloud characteristics such as cloud height and vertical structure is one of the most basic yet urgent observational needs.

[0003] The macroscopic structure and radiation characteristics of clouds are closely related, thus playing a crucial role in altering the thermal structure of the atmosphere and influencing atmospheric circulation. Before using radar observations to retrieve macroscopic cloud features, it is essential to distinguish between noise and cloud echo signals in the raw data. Strong echo signals significantly higher than the noise level can be easily identified by setting a threshold in cloud detection. However, when cloud echo signals are weak and close to the noise level, accurate identification becomes significantly more difficult, a persistent challenge in radar target identification. These weak echo signals generally originate from small-diameter or sparse ice cloud particles, often appearing at cloud top boundaries or high altitudes. Clouds in these locations exhibit a strong greenhouse effect, significantly impacting the radiation balance and frequently associated with cloud formation or dissipation. Accurate identification of these cloud droplet particles is crucial for understanding cloud formation and dissipation processes and their impact on climate energy balance.

[0004] Conventional ground-based and satellite observations are important tools for studying the climatological characteristics of clouds, but they often fail to provide complete information on the macroscopic structure of clouds. Ground-based observations struggle to penetrate low clouds and accurately observe mid- to high-level clouds; they can only record cloud base height, not cloud height information. Satellites, on the other hand, face difficulties in observing low clouds and determining their cloud base height. While temperature and humidity information obtained from radiosondes can accurately and completely invert the vertical distribution of clouds, radiosondes cannot perform continuous, fixed-point observations of clouds in a specific region or type. All-sky imagers can only obtain cloud cover over the entire sky, not macroscopic cloud parameters, and only operate during the day; infrared imagers can perform day and night observations and obtain cloud height and cloud cover information, but their inversion errors are large, making them difficult to meet operational needs. Summary of the Invention

[0005] The technical problem addressed in this application is to overcome the shortcomings of existing technologies and provide a microwave-lidar composite radar cloud identification method based on transfer learning and R-FCN networks. This method possesses cloud identification capabilities under complex atmospheric conditions such as multi-layered clouds and under equipment noise interference. The inversion results are accurate, and it is insensitive to ground clutter and daytime background noise. It fully integrates the advantages of lidar in identifying thin cirrus cloud regions and cloud bases with the advantages of microwave radar in identifying cloud tops. A joint cloud boundary identification method is also provided. This method is applicable to cloud identification in the atmosphere and can effectively improve the accuracy and precision of cloud identification.

[0006] Joint observation using lidar and microwave radar is currently the most common method for observing and studying local clouds. To more accurately determine cloud boundaries under complex conditions, this paper proposes a microwave-lidar hybrid radar cloud identification method based on transfer learning and an R-FCN network. This method combines the advantages of microwave radar and lidar, while mitigating the shortcomings of both approaches, and utilizes active detection techniques.

[0007] This method not only solves the limited penetration problem of single lidar into clouds, but also compensates for the poor detection capability of single microwave radar for thin clouds and water clouds with small particles. It can accurately obtain cloud-aerosol vertical structure and cloud phase information, making it the best means of observing macroscopic cloud parameters. Deep learning, especially convolutional neural networks, offers more accurate solutions in image processing. Transfer learning improves performance on small-scale, domain-specific datasets by utilizing models pre-trained on large-scale datasets. The R-FCN network, by combining a Region Proposal Network (RPN) and a Fully Convolutional Network (FCN), has advantages in object detection tasks.

[0008] This invention proposes a cloud identification method for microwave-LiDAR composite radar based on transfer learning and R-FCN networks. It utilizes an attention mechanism to achieve adaptive feature extraction and introduces a multi-scale feature pyramid on top of the RPN network to enhance the identification and detection capabilities of cloud regions at different scales. This method enables cloud region identification for both LiDAR and microwave radar, completing the cloud region identification task under composite conditions through cloud boundary self-checking, cloud feature labeling, and a joint cloud boundary inversion method.

[0009] The technical solution provided in this application is as follows:

[0010] Microwave-laser composite radar cloud identification method, including:

[0011] S1. Remove ground clutter from the reflectivity factor image and linear depolarization ratio (LDR) obtained from the microwave radar to obtain an effective reflectivity factor image and an effective LDR.

[0012] S2. Use lidar to collect and obtain distance-corrected signal images;

[0013] S3. Based on the distance-corrected signal image and the effective reflectivity factor image, a cloud recognition algorithm based on transfer learning and regional fully convolutional R-FCN network is adopted to locate and identify cloud areas and obtain the precise location of cloud areas 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 area are paired, remove bad values ​​that are not paired, and retain the paired cloud top and cloud base boundaries. A pair of paired cloud top and cloud base boundaries is considered as cloud information. In the distance correction signal image, determine whether the cloud base and cloud top of the cloud area are paired, remove bad values ​​that are not paired, and retain the paired cloud top and cloud base boundaries. A pair of paired cloud top and cloud base boundaries is considered as cloud information.

[0015] S5. Cloud feature labeling: Based on the reflectivity factor image and the time and height of the time and height correction signal image, match the cloud information in the reflectivity factor image with the cloud information in the height correction signal image one by one, and determine the cloud information to be retained based on the matching results to obtain the final cloud information.

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

[0017] The distance-corrected signal image obtained in S2 and the effective reflectivity factor image obtained in S1 are used to generate multiple location-sensitive score maps through a trained R-FCN network model. Each location-sensitive score map is classified and voted to obtain the precise location of the cloud area in the distance-corrected signal image and the effective reflectivity factor image.

[0018] Morphological processing is performed on the cloud regions in the distance-corrected signal image and the effective reflectivity factor image to obtain the accurate morphology of the cloud regions in the distance-corrected signal image and the effective reflectivity factor image.

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

[0020] Using reflectivity factor images and distance correction signal images as datasets, cloud areas in the images are manually located and labeled to obtain the dataset.

[0021] The public dataset is transferred to the dataset using transfer learning methods;

[0022] The dataset is divided into a training set and a test set;

[0023] An R-FCN network model is established, comprising a feature extraction network module, a region generation network module, and a region of interest (ROI) sub-network module. The feature extraction network module generates corresponding convolutional feature maps based on the reflectivity factor image and / or distance correction signal image, respectively. The region generation network module generates cloud region localization on the convolutional feature maps, obtaining convolutional feature maps with cloud region localization. The ROI sub-network uses fully convolutional layers to operate on the convolutional feature maps with cloud region localization and multiple convolutional kernels to generate multiple location-sensitive score maps.

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

[0025] The training set is used as input, and the loss function is applied based on the convolutional feature map with cloud localization, the reflectivity factor image and time or height distance correction signal image in the dataset corresponding to the convolutional feature map with cloud localization. The evaluation result is obtained, 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, and the trained R-FCN network model is obtained.

[0026] The loss function is the cross-entropy loss function. The cross-entropy loss function is used to determine whether each convolutional feature map with cloud region localization contains a cloud region. Based on whether it contains a cloud, a model prediction probability Pi is given. If it does not contain a cloud region, Pi = 0; if it contains a cloud region, Pi = 1. Based on the model prediction probability Pi, the output value of the cross-entropy loss function is calculated. If the output value is less than a threshold, the next training iteration is performed. 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.

[0027] The distance-corrected signal image obtained in S2 and the effective reflectivity factor image obtained in S1 are used to generate multiple location-sensitive score maps through a trained R-FCN network model. Each location-sensitive score map is then classified and voted on to obtain the precise location of the cloud region in the distance-corrected signal image and the effective reflectivity factor image, including:

[0028] The predicted probabilities of multiple atmospheric objects included in each location sensitivity score map are weighted and averaged to obtain the specific category of the location sensitivity score map. The multiple atmospheric objects include clouds and others. If the category is clouds, the location sensitivity score map is retained as a cloud area; otherwise, it is not retained as a cloud area. Based on the location sensitivity score map that is retained as a cloud area, the accurate location of the cloud area is determined.

[0029] The morphological processing of 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 includes:

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

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

[0032] In step S5, based on the reflectivity factor image and the time and altitude of the time and altitude correction signal image, the cloud information in the reflectivity factor image is matched one-to-one with the cloud information in the altitude correction signal image. The cloud information to be retained is determined based on the matching results, resulting in the final cloud information, including:

[0033] S51. Based on the reflectivity factor image and the time and height of the time and height correction signal image, match the cloud information in the reflectivity factor image with the cloud information in the height correction signal image one by one.

[0034] S52. For the cloud information matched in S1 that corresponds to the same target cloud area, the following situations exist: there is no cloud information in either the reflectivity factor image or the height distance correction signal image, and therefore no corresponding cloud information is found; the cloud information is only present in the reflectivity factor image, and the final cloud information is based on the reflectivity factor image; the cloud information is only present in the height distance correction signal image, and the final cloud information is based on the height distance correction signal image; and cloud information is present in both the reflectivity factor image and the height distance correction signal image, so the relationship between the clouds in the reflectivity factor image and the height distance correction signal image needs to be considered comprehensively to obtain the final cloud information.

[0035] Both the reflectivity factor image and the height-distance corrected signal image contain cloud information. The relationship between the clouds in the reflectivity factor image and the height-distance corrected signal image needs to be comprehensively considered to obtain the final cloud information, including:

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

[0037] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of any of the microwave-laser composite radar cloud identification methods described above.

[0038] A cloud identification method for microwave lidar composite radar based on transfer learning and R-FCN network is proposed. The method includes: ground clutter removal, cloud identification algorithm based on transfer learning and regional fully convolutional network (R-FCN), cloud boundary self-inspection, cloud feature labeling, initial matching, and joint inversion of cloud boundaries.

[0039] 1) Ground clutter noise from microwave radar is removed using ground clutter removal methods to preprocess the dataset. 2) A deep learning model is constructed using transfer learning and the R-FCN algorithm. Microwave radar and lidar echo images are input, and the model optimizes adaptive feature extraction using an attention mechanism. A multi-scale feature pyramid is introduced on top of the RPN network to enhance the ability to identify and detect cloud areas at different scales. 3) Single-result cloud identification from microwave radar and lidar is jointly inverted using methods such as cloud boundary self-checking, cloud feature labeling, initial matching, and joint cloud boundary inversion to obtain a composite identified cloud area.

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

[0041] This invention proposes a cloud identification method for microwave-LiDAR composite radar based on transfer learning and R-FCN networks. It utilizes an attention mechanism for adaptive feature extraction and introduces a multi-scale feature pyramid on top of the RPN network to enhance the identification and detection capabilities of cloud regions at different scales. This method enables cloud region identification for both LiDAR and microwave radar. Through cloud boundary self-checking, cloud feature labeling, and joint cloud boundary retrieval methods, it completes cloud region identification tasks under composite conditions. It can be used for cloud boundary identification under various complex weather conditions, is less affected by ground clutter and background noise, has a high degree of automation, and yields more accurate results.

[0042] This method possesses cloud identification capabilities under complex atmospheric conditions such as multi-layered clouds and equipment noise interference. The inversion results are accurate, and it is insensitive to ground clutter and daytime background noise. It fully integrates the advantages of lidar in identifying thin cirrus cloud regions and cloud bases with the advantages of microwave radar in identifying cloud tops. The joint cloud boundary identification method is applicable to cloud identification in the atmosphere and can effectively improve the accuracy and precision of cloud identification. This method can be transferred to spaceborne platforms, and the identified cloud regions can lay the foundation for inverting global three-dimensional wind field information using spaceborne microwave-liquid composite wind measurement radar. Attached Figure Description

[0043] Figure 1 The cloud boundary results are obtained after the LiDAR image is processed by the R-FCN network model and classification voting, where the horizontal axis represents events and the vertical axis represents height;

[0044] Figure 2 The cloud boundary results are obtained after applying the microwave radar reflectivity factor to the R-FCN network model and classification voting.

[0045] Figure 3 This is a cloud area identification result obtained by a microwave-liquid composite radar cloud identification method based on transfer learning and R-FCN network;

[0046] Figure 4 This is a technical roadmap for a microwave-laser composite radar cloud identification method based on transfer learning and R-FCN network;

[0047] Figure 5 The cloud boundary results obtained from lidar images under low cloud and ground clutter conditions after passing through the R-FCN network model and classification voting.

[0048] Figure 6 The cloud boundary results obtained from microwave radar images under low cloud and ground clutter conditions after passing through the R-FCN network model and classification voting.

[0049] Figure 7 The cloud region identification results map is obtained by using the method of this application under the conditions of low clouds and ground clutter. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments disclosed in the present invention will be described in further detail below with reference to the accompanying drawings.

[0051] This application discloses a microwave-laser composite radar cloud identification method based on transfer learning and R-FCN network, such as... Figure 4 As shown, it includes the following steps:

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

[0053] Non-meteorological clutter observed by microwave radar is usually caused by echoes of low-altitude suspended objects, including haze, dust, and insects in the atmosphere. In microwave radar cloud identification methods, ground clutter is more or less identified as clouds, which affects the results of complex cloud boundary extraction. Ground clutter must be removed before joint inversion with lidar.

[0054] Considering that these ground clutters generally exhibit weak echo intensity, low velocity, and high linear depolarization ratio (LDR), a method combining the echo intensity reflectivity factor Z and the LDR is adopted for clutter removal. The calculation formula is as follows:

[0055] noise = (LDR>-22)∩(Z<-5);

[0056] Here, "noise" refers to noise or clutter.

[0057] S2. Atmospheric echo signals are acquired using lidar. The atmospheric echo signals are then corrected to obtain distance correction signals. These distance correction signals are continuously acquired over a period of time to obtain distance correction signal images. Based on the distance correction signal images and reflectivity factor images, a cloud recognition algorithm based on transfer learning and a region fully convolutional R-FCN network is used to locate and identify cloud areas, obtaining the precise location of the cloud areas on the reflectivity factor images and the distance correction signal images.

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

[0059] (1) Data preprocessing: The reflectivity factor image and the distance correction signal image are used as datasets respectively. The cloud areas (i.e. cloud areas) in the images are manually located and labeled to complete the dataset production.

[0060] (2) Transfer learning: The public dataset is transferred to the dataset in step (1) using the transfer learning method; the public dataset includes multiple reflectance factor images and distance correction signal images labeled with cloud areas;

[0061] (3) Divide the dataset: Divide the dataset into training set and test set in a ratio of 8:2, with no overlap or duplicate 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 location-sensitive score map. The R-FCN network model includes a feature extraction network, a region generation network (RPN), and a region of interest (ROI) subnet.

[0064] Atmospheric echo signals are acquired using lidar, and these signals are then corrected for distance to obtain distance-corrected signals. These distance-corrected signals are continuously acquired over a period of time to obtain distance-corrected signal images. 1) The reflectivity factor images and distance-corrected signal images acquired by microwave radar and lidar, respectively, are input into a feature extraction network. The feature extraction network generates corresponding convolutional feature maps based on the microwave radar reflectivity factor image and the lidar distance-corrected signal image, respectively. These convolutional feature maps contain information related to the spatial structure, location distribution, and distance of clouds. 2) A region generation network (RPN) receives and analyzes the two types of convolutional feature maps, generating cloud region localization on the convolutional feature maps, obtaining convolutional feature maps with cloud region localization. 3) A region of interest (ROI) subnet uses fully convolutional layers to operate on the convolutional feature maps with cloud region localization and multiple convolutional kernels to generate location-sensitive score maps. 4) Finally, the location-sensitive score maps are classified and voted on to obtain the precise location of the cloud region on the microwave radar and lidar images, achieving cloud localization and identification. Figure 1 and Figure 2 As shown.

[0065] Specifically, the location-sensitive score map is categorized and voted on. The location-sensitive score map represents the predicted probability that each location on the convolutional feature map with cloud region localization belongs to a certain class of atmospheric objects. A certain class of atmospheric objects is either clouds or other, including background and aerosols. The predicted probabilities of multiple atmospheric object classes included in each location-sensitive score map are weighted and averaged to obtain the specific category of that location-sensitive score map. If the category is clouds, the location-sensitive score map is retained as a cloud region; otherwise, it is not retained. Based on the location-sensitive score maps retained as cloud regions, the accurate cloud region location is determined.

[0066] The convolutional feature map is processed by 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 was chosen as the feature extraction network. ResNet is a deep neural network architecture designed to address the gradient descent problem (difficulty in updating weights) that arises with increasing network layers. Transfer learning is a machine learning method that improves learning efficiency by transferring knowledge learned from existing domain models to cloud recognition. A ResNet model pre-trained on a large existing dataset was used for feature extraction, adapting it to the data features of LiDAR and microwave radar. An attention mechanism was employed to enhance the model's focus on cloud features in LiDAR and microwave radar images, ignoring noise and improving cloud recognition accuracy.

[0068] The residual blocks introduced by the ResNet model solve the gradient descent problem in neural networks. Based on the pre-trained model, transfer learning and attention mechanisms are used for fine-tuning to adapt it 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 training parameters, including: the number of iterations is set to 30,000, the number of dataset categories 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: The training set is used as input to train and optimize the R-FCN network model. The model is optimized using a loss function, which can be either the cross-entropy loss function or the bounding box regression loss function. The training parameters are adjusted based on 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 region localization, the reflectivity factor image and the time or height distance correction signal image in the dataset corresponding to the convolutional feature map with cloud region localization, and then updates the parameters of the R-FCN network model based on the evaluation result.

[0072] Specifically, taking the cross-entropy loss function as an example, the process includes: using the cross-entropy loss function to determine whether each convolutional feature map with cloud region localization contains a cloud region; and based on whether it contains a cloud, predicting a probability P for the given model. i Excluding cloud areas, P i =0; includes cloud regions, P i =1; Based on the model, the predicted probability P i The output value (i.e. the evaluation result) of the cross-entropy loss function is calculated. If the output value is less than the threshold, the next training iteration is performed; if the output value is greater than the threshold, the initial learning rate is adjusted.

[0073] (7) Cloud area post-processing: The test set is used as input to obtain the cloud area detection results. Morphological processing is performed on the extraction results, including operations such as dilation and erosion. Erosion removes noise, and dilation merges adjacent cloud areas to improve extraction accuracy.

[0074] The cloud recognition algorithm based on transfer learning and regional fully convolutional R-FCN networks is used to achieve cloud recognition for microwave radar and lidar. It fully leverages the similarity between data, tasks, and models, applying models learned in the old domain to the new domain, thus reducing the requirements for the amount of training data and the similarity of feature distributions between training and test data samples. A feature extraction module incorporating an attention mechanism is added to the design to 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 determine whether the cloud boundary results obtained from a single radar inversion exist in pairs. If they exist in pairs, the preliminary extraction results are accepted. If the cloud boundaries do not exist in pairs, it is necessary to determine whether the cloud top or cloud base has a bad value, and the bad value needs to be removed to obtain the paired cloud top and cloud base boundaries.

[0076] Specifically, this includes: in the reflectivity factor image, determining whether the cloud base and cloud top of the cloud area are paired, removing bad values ​​that are not paired, retaining the paired cloud top and cloud base boundaries, and taking a pair of paired cloud top and cloud base boundaries as cloud information; in the distance correction signal image, determining whether the cloud base and cloud top of the cloud area are paired, removing bad values ​​that are not paired, retaining the paired cloud top and cloud base boundaries, and taking a pair of paired cloud top and cloud base boundaries as cloud information;

[0077] S4. Cloud Feature Marking: After step S3, cloud boundaries can be considered to exist in pairs and are free from ground clutter. Based on this, the following steps are performed:

[0078] S41. Initial Cloud Boundary Matching: Before joint cloud boundary inversion, it is necessary to first determine that the cloud boundaries obtained by the two working modes are for the same target. After removing ground clutter, efforts will be made to ensure that the target information obtained by a single detection method exists in pairs and corresponds one-to-one. In reality, there may be cases where the result of one detection method is greater than that of another, but they represent the same target cloud body (cloud body, i.e., cloud region). Therefore, it is necessary to traverse the cloud boundaries of lidar and microwave radar, mark them simultaneously, perform logical calculations on the same target cloud body, and save the typical characteristics of different cloud bodies.

[0079] S42. Joint Cloud Boundary Retrieval: For non-precipitating clouds, the lowest cloud base is used as the final cloud boundary base; the highest cloud top is used as the final cloud boundary top. If only one detection method's result exists, the final cloud boundary is assigned that value. This situation corresponds to the actual detection where only water cloud results detected by lidar or high-thick clouds detected by microwave radar are available, but not detected by lidar.

[0080] Specifically, the cloud boundaries extracted by microwave radar and lidar are mainly divided into the following four cases: 1) If neither type of radar has cloud information, the final inversion result will not have corresponding cloud information; 2) If lidar has no cloud information, but only microwave radar has cloud information, then the final cloud information will be based on microwave radar; 3) If lidar has cloud information, but microwave radar has no cloud information, then the final result will be based on lidar; 4) If both radars have cloud information, which is the most complex case, the relationship between the clouds obtained by the two radars needs to be comprehensively considered to obtain the final cloud information.

[0081] In each of the four scenarios above, initial cloud boundary matching needs to be considered to determine the consistency of the inversion target; otherwise, the value substituted into the logic will be invalid and will affect the final cloud boundary extraction result.

[0082] The cloud area identification result obtained according to the microwave lidar cloud identification method based on transfer learning and R-FCN network of this application is shown in the figure below. Figure 3 As shown, no single detection method can identify the cloud area.

[0083] For different times, under low cloud and ground clutter conditions, the cloud boundary results obtained from the LiDAR images after processing through the R-FCN network model and classification voting are as follows: Figure 5 As shown, the cloud boundary results obtained after processing the microwave radar image through the R-FCN network model and classification voting are as follows: Figure 6 As shown, the results of joint identification of cloud areas using the method of this application are as follows: Figure 7 As shown.

[0084] This application fully integrates the advantages of lidar in identifying cirrus cloud regions and cloud bases with the advantages of microwave radar in identifying cloud tops. This joint method for identifying cloud boundaries is applicable to cloud identification in the atmosphere and can effectively improve the accuracy and precision of cloud identification.

[0085] Specifically, step S2 includes:

[0086] S21, the reflectivity factor image, and the distance correction signal image are processed by a feature extraction network to generate corresponding convolutional feature maps;

[0087] Deep residual neural networks (ResNet) are networks formed by repeatedly stacking residual learning modules. They solve the gradient vanishing problem during model training, thereby improving the model's learning and recognition capabilities. This method selects ResNet-101 as the feature extraction network.

[0088] Assuming the input is x, after passing through convolutional layers W1 and W2, the output is F(x,W1,W2), the activation function is ReLU, and the residual module output y is represented as:

[0089] y = F(x, W1, W2) + W s x

[0090] Where W1 and W2 are the learning weights corresponding to layers 1 and 2 of the convolutional neural network, respectively; F(x,W1,W2) is a mapping function composed of convolutional activation functions, representing the output calculated by the convolutional layer; the reflectivity factor or distance correction signal of each point in the reflectivity factor image or time and height distance correction signal image is fed into the mapping function F(x,W1,W2) to calculate multiple y values, which constitute the convolutional feature map; W sThese are the weights of variable x from input to output. Input x consists of the reflectance factor image and the distance correction signal image;

[0091] S22, Location-Sensitive Fraction Map and Pooling; To accurately obtain the cloud region location on the convolutional feature map with cloud region localization, the convolutional layer obtained after convolutional operation on the original image is used to generate the location-sensitive fraction map. "The convolutional feature map with cloud region localization is processed with multiple convolutional kernels to generate the location-sensitive fraction map"—this process generates k. 2 There are k location-sensitive score maps, containing class C atmospheric targets (clouds, aerosols) and one background, corresponding to k channels. 2 (c+1). The algorithm model in this paper sets k=3, where k means that the length and width of the convolutional feature map are divided into k equal slices. The detection categories are clouds and aerosols, with c=2 categories. After RPN provides the ROIs of interest, it performs position-sensitive pooling on each convolutional feature map with cloud region localization to obtain k channels. 2 (c+1) Position-sensitive score map. When k=3, the convolutional feature map with cloud region localization can be divided into 9 rectangular blocks. Therefore, the corresponding position-sensitive score map is also divided into 9 blocks: (top-left, top-center, top-right, ..., bottom-right). The segmented position-sensitive score map is then subjected to mean pooling to obtain a vector of length c+1. ​​This vector is then used for softmax classification (i.e., classification voting) to finally determine the detection category of the cloud region in the convolutional feature map with cloud region localization.

[0092] Position-sensitive pooling specifically involves dividing a convolutional feature map of size ω×h, containing cloud region localization, into k... 2 There are rectangular blocks of size (ω / k)×(h / k), and (i,j) represents the position of each block in the convolutional feature map and the position-sensitive fractional map with cloud localization, where the value of (i,j) is in the range of (0≤i≤k-1,0≤j≤k-1).

[0093] Set the number of channels pointed to by the top-left arrow to k. 2 The (0,0) square in the convolutional feature map corresponding to the feature map with cloud region localization is placed at position (0,0) in the position-sensitive score map; the number of channels pointed to by the top-center arrow is k. 2 The (1,0)th square in the convolutional feature map corresponding to the feature map with cloud region localization is placed at position (1,0) in 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 as follows:

[0094]

[0095] Where, r c (i,j|θ) represents the c-th position sensitivity score map k. 2 The (i,j)th block in the Z blocks; i.j.c is the first feature map; bin(i,j) is the set of locations in the ROI and the location-sensitive score map; (x0,y0) is the coordinate of the top left corner of the ROI; (x,y) is the range of coordinate values ​​for each element in the ROI with (x0,y0) as 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 from the position-sensitive fractional map is as follows:

[0097]

[0098] S23, Loss Function; When training using gradient descent, both classification loss and location loss need to be considered. The loss function for each ROI is the sum of the cross-entropy loss and the boundary regression loss, expressed as:

[0099]

[0100] Where λ=1; c * The class label of ROI (c * When = 0, it represents the background class of the ROI; L is the cross-entropy loss function used for classification; reg In t represents the parameters of the true regression box; t represents the parameters of the regression box obtained from RPN. t = (t x ,t y ,t w ,t h ).

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

[0102] The present application has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present application. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and implementation methods of the present application without departing from the spirit and scope of the present application, and all such modifications and improvements fall within the scope of the present application. The scope of protection of the present application is determined by the appended claims.

Claims

1. A microwave laser complex radar cloud recognition method, characterized by, The method comprises the following steps: S1, removing ground clutter from the reflectivity factor image and the linear depolarization ratio , removing ground clutter from the reflectivity factor image and the linear depolarization ratio , removing ground clutter from the reflectivity factor image and the linear depolarization ratio ; The reflectivity factor image and the linear depolarization ratio Removing ground clutter includes: wherein, Z is the reflectivity factor; and Z is the reflectivity factor; and S2, collecting and obtaining a range-corrected signal image by range correction through a laser radar; S3, according to the range-corrected signal image and the effective reflectivity factor image, a cloud recognition algorithm based on transfer learning and a regional full convolutional R-FCN network is used to locate and identify the cloud area, and the accurate position of the cloud area on the reflectivity factor image and the range-corrected signal image is obtained; The cloud recognition algorithm based on transfer learning and the regional full convolutional R-FCN network comprises the following steps: the range-corrected signal image obtained in S2 and the effective reflectivity factor image obtained in S1 are input into a trained R-FCN network model to generate a plurality of position-sensitive score maps, each position-sensitive score map is classified and voted to obtain the accurate position of the cloud area on the range-corrected signal image and the effective reflectivity factor image; and morphological processing is performed on the cloud area on the range-corrected signal image and the effective reflectivity factor image to obtain the accurate shape of the cloud area on the range-corrected signal image and the effective reflectivity factor image; The trained R-FCN network model is obtained by the following method: The reflectivity factor image and the range-corrected signal image are used as a data set, and the cloud area in the image is manually located and labeled to obtain the data set; The public data set is transferred to the data set by using the transfer learning method; The data set is divided into a training set and a test set; An R-FCN network model is established, which comprises 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 convolution feature maps according to the reflectivity factor image and / or the range-corrected signal image; the region generation network module is used to generate cloud area positioning on the convolution feature maps to obtain convolution feature maps with cloud area positioning; and the region of interest subnet is used to generate a plurality of position-sensitive score maps by using a full convolution layer to operate the convolution feature maps with cloud area positioning and a plurality of convolution kernels; Training parameters of the R-FCN network model are configured, and the training parameters comprise the number of iterations, the data set category and the initial learning rate; The training set is used as an input, a loss function is used according to and uses the convolution feature maps with cloud area positioning, the reflectivity factor image and the time or height range-corrected signal image in the data set corresponding to the convolution feature maps with cloud area positioning 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, so that the trained R-FCN network model is obtained; S4, cloud boundary self-checking, in the reflectivity factor image, whether the cloud bottom and the cloud top of the cloud area are paired is judged, the bad values existing in the unpaired state are removed, the paired cloud top and cloud bottom boundaries are retained, and a group of paired cloud top and cloud bottom boundaries is used as a cloud information; in the range-corrected signal image, whether the cloud bottom and the cloud top of the cloud area are paired is judged, the bad values existing in the unpaired state are removed, the paired cloud top and cloud bottom boundaries are retained, and a group of paired cloud top and cloud bottom boundaries is used as a cloud information; S5, cloud feature marking, according to the reflectivity factor image and the time and height of the range correction signal image, the cloud information in the reflectivity factor image is matched with the cloud information in the height range correction signal image one by one, the reserved cloud information is determined according to the matching correspondence, and the final cloud information is obtained.

2. The microwave laser compound radar cloud discrimination method of claim 1, wherein: The loss function is a cross-entropy loss function, and the cross-entropy loss function is used to judge whether each convolution feature map with cloud area positioning contains a cloud area; according to whether the cloud is contained, the model prediction probability P is given i ; no cloud area is contained, P i =0; a cloud area is contained, P i =1; according to the model prediction probability P i , the output value of the cross-entropy loss function is calculated, and if the output value is less than a threshold value, the next training is performed; If the output value is greater than the threshold value, the initial learning rate is adjusted until the output value of the cross-entropy loss function meets the requirements.

3. The microwave laser compound radar cloud discrimination method of claim 1, wherein, The distance correction signal image obtained in S2 and the effective reflectivity factor image obtained in S1 are input into the trained R-FCN network model to generate a plurality of position-sensitive score maps, each position-sensitive score map is classified and voted to obtain the accurate position of the cloud area on the distance correction signal image and the effective reflectivity factor image, including: The prediction probabilities of the plurality of atmospheric target objects included in each position-sensitive score map are weighted and averaged to obtain the specific category of the position-sensitive score map, the plurality of atmospheric target objects include clouds and others, if the category is cloud, the position-sensitive score map is retained as cloud area, and if it is other, it is not retained as cloud area; the accurate cloud area position is determined according to the position-sensitive score map retained as cloud area.

4. The microwave laser compound radar cloud discrimination method of claim 1, wherein: The cloud area on the distance correction signal image and the effective reflectivity factor image is morphologically processed to obtain the accurate morphology of the cloud area on the distance correction signal image and the effective reflectivity factor image, including: The morphological processing includes dilation and erosion, the erosion is used to remove noise, and the dilation is used to merge adjacent cloud areas.

5. The microwave laser compound radar cloud discrimination method of claim 1, wherein: In S5, according to the reflectivity factor image and the time and height of the range correction signal image, the cloud information in the reflectivity factor image is matched with the cloud information in the height range correction signal image one by one, the reserved cloud information is determined according to the matching correspondence, and the final cloud information is obtained, including: S51, according to the reflectivity factor image and the time and height of the range correction signal image, the cloud information in the reflectivity factor image is matched with the cloud information in the height range 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, there are the following situations: there is no cloud information in the reflectivity factor image and the height range correction signal image, and finally there is no corresponding cloud information; only the cloud information exists in the reflectivity factor image, and finally the cloud information is based on the reflectivity factor image; only the cloud information exists in the height range correction signal image, and finally the cloud information is based on the height range correction signal image; there is cloud information in the reflectivity factor image and the height range correction signal image, and the mutual relationship of the cloud information in the reflectivity factor image and the height range correction signal image needs to be considered comprehensively to obtain the final cloud information.

6. The microwave laser compound radar cloud discrimination method of claim 5, wherein: The reflectivity factor image and the height range correction signal image have cloud information, and the mutual relationship of the cloud information in the reflectivity factor image and the height range correction signal image needs to be considered comprehensively to obtain the final cloud information, including: For non-precipitation clouds, the lowest cloud base in the reflectivity factor image and the height-distance-corrected signal image is the final cloud boundary cloud base for the cloud region, and the highest cloud top in the reflectivity factor image and the height-distance-corrected signal image is the final cloud boundary cloud top for the cloud region.

7. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the microwave laser composite radar cloud identification method of any one of claims 1-6.

Images