A method, device, equipment and medium for identifying marine weather systems using satellite cloud images

Abstract

The present application discloses a method, device, equipment and medium for identifying marine weather systems from satellite cloud images, and relates to the technical field of marine weather system identification. The method comprises: using historical satellite cloud images as input in advance, using the location information of marine weather systems in the historical satellite cloud images as labels, training a deep learning network model to obtain a trained marine weather system identification model. In practical applications, a satellite cloud image of an analysis area is obtained, and the analysis area includes an ocean area. The satellite cloud image is used as input, and the trained marine weather system identification model is used to label the marine weather systems in the satellite cloud image to identify the marine weather systems in the satellite cloud image. The present application can quickly, accurately and automatically identify marine weather systems in satellite cloud images.

Description

Technical Field

[0001] The present application relates to the technical field of marine weather system identification, and in particular to a method, device, equipment and medium for identifying marine weather systems using satellite cloud images. Background Art

[0002] With the development of meteorological satellite technology, satellite cloud images have become an indispensable reference for weather forecasting. Geostationary meteorological satellites, in particular, offer high observation frequency and a wide range of coverage. This makes satellite cloud images the primary basis for oceanographic and meteorological departments to understand marine weather developments and changes, given the scarcity of offshore observation data and information sources. However, my country's current application of meteorological satellite data is primarily focused on assisting in the assessment of marine weather systems, particularly those over land and nearshore areas. Assisting in the assessment of marine weather systems, particularly those in open ocean waters, remains insufficient.

[0003] A more obvious manifestation of the lack of application of satellite cloud images is that during the analysis and drawing of weather maps, some professionals are relatively accurate in marking the range and location of land weather systems due to the support of more observation data. However, the marking of offshore weather systems often has large errors. On the one hand, this is due to the lack of other observation data to assist, and on the other hand, there is a lack of understanding and in-depth understanding of the characteristics of offshore weather systems in satellite cloud images. In addition, when using satellite cloud images and weather maps, the two are two separate reference materials. Although meteorological departments now overlay infrared cloud images on weather maps for display, the weather systems that satellite cloud images can represent are relatively simple and cannot fully reflect the effects of satellite observations. In particular, offshore weather systems are often overlooked.

[0004] Because the understanding of satellite cloud images among my country's meteorological forecast professionals varies, and meteorological satellite professionals rarely participate in meteorological forecasting work, in the absence of other observation data to assist in the judgment of marine weather systems, the marine weather systems mapped manually and automatically by meteorological mapping software have large errors and often omissions. At present, there is no comprehensive analysis product for the mapping of marine weather systems based on satellite cloud images. Therefore, most professionals are unable to accurately identify some important marine weather systems when using satellite cloud images, which makes the functions of satellite cloud images not fully utilized. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, equipment and medium for identifying marine weather systems in satellite cloud images, which can quickly, accurately and automatically identify marine weather systems in satellite cloud images.

[0006] To achieve the above objectives, this application provides the following solutions.

[0007] In a first aspect, the present application provides a method for identifying marine weather systems using satellite cloud images, and the method includes the following steps.

[0008] Obtain a satellite cloud image of an analysis area; the analysis area includes an ocean area.

[0009] Taking the satellite cloud image as input, the trained marine weather system identification model is used to label the marine weather system in the satellite cloud image to identify the marine weather system in the satellite cloud image; wherein the trained marine weather system identification model is a model obtained by training a deep learning network model using historical satellite cloud images as input and the location information of the marine weather systems in the historical satellite cloud images as labels.

[0010] Optionally, the satellite cloud image includes a visible light cloud image and an infrared cloud image.

[0011] Optionally, the marine weather system includes a high-pressure center, a low-pressure center, a subtropical high-pressure area, a low-pressure trough line, a high-pressure ridge line and an equatorial convergence zone.

[0012] Optionally, the deep learning network model adopts an improved RT-DETR model, which is a model obtained by improving the RT-DETR model, and the improvement includes: replacing the convolution layer in the backbone network of the RT-DETR model with a spatial and channel reconstruction convolution layer, making the neck network of the RT-DETR model use a feature pyramid network, and replacing the convolution layer in the feature pyramid network with a spatial and channel reconstruction convolution layer.

[0013] Optionally, before taking the satellite cloud map as input and using a trained marine weather system identification model to label the marine weather system in the satellite cloud map, the satellite cloud map marine weather system identification method also includes: training to obtain a trained marine weather system identification model, specifically including the following steps.

[0014] Acquire multiple samples; the samples include historical satellite cloud images, historical ground weather analysis images, and historical 500hPa potential height images corresponding to the time of the analysis area.

[0015] The marine weather systems in the historical ground weather analysis map and the historical 500hPa potential height map are marked to generate a historical marine weather system distribution map; the historical marine weather system distribution map contains the location information of the marine weather systems in the historical satellite cloud map.

[0016] The deep learning network model is trained using the historical satellite cloud image as input and the historical marine weather system distribution map as a label to obtain a trained marine weather system recognition model.

[0017] Optionally, obtaining multiple samples specifically includes the following steps.

[0018] For historical satellite cloud images, historical satellite cloud images of the analysis area are obtained; alternatively, L1 level full disk data of the analysis area are obtained, and the historical satellite cloud images are drawn based on the L1 level full disk data.

[0019] For historical surface weather analysis maps and historical 500hPa potential height maps, obtain historical surface weather analysis maps and global meteorological reanalysis data for the analysis area, and draw historical 500hPa potential height maps based on the global meteorological reanalysis data; or obtain global meteorological reanalysis data for the analysis area, and draw historical surface weather analysis maps and historical 500hPa potential height maps based on the global meteorological reanalysis data.

[0020] The time periods of the historical satellite cloud map, the historical ground weather analysis map and the historical 500hPa potential height map correspond to each other, and the map projections are all Mercator projections.

[0021] Optionally, when training the deep learning network model, the loss functions used include a classification loss function and a bounding box regression loss function, the classification loss function adopts a cross entropy function, and the bounding box regression loss function adopts a Smooth L1 function.

[0022] In a second aspect, the present application provides a satellite cloud image marine weather system identification device, which includes the following modules.

[0023] The acquisition module is used to obtain a satellite cloud image of an analysis area; the analysis area includes an ocean area.

[0024] The identification module is used to use the satellite cloud image as input and use the trained marine weather system identification model to label the marine weather system in the satellite cloud image to identify the marine weather system in the satellite cloud image; wherein the trained marine weather system identification model is a model obtained by training a deep learning network model with historical satellite cloud images as input and the location information of the marine weather systems in the historical satellite cloud images as labels.

[0025] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned method for identifying marine weather systems based on satellite cloud images.

[0026] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for identifying marine weather systems using satellite cloud images.

[0027] According to the specific embodiments provided in this application, this application has the following technical effects.

[0028] The present application provides a method, apparatus, device and medium for identifying marine weather systems from satellite cloud images. Historical satellite cloud images are used as input in advance, and the location information of marine weather systems in the historical satellite cloud images is used as labels to train a deep learning network model to obtain a trained marine weather system identification model. In practical applications, a satellite cloud image of an analysis area is obtained, and the analysis area includes an ocean area. The satellite cloud image is used as input, and the trained marine weather system identification model is used to label the marine weather systems in the satellite cloud image to identify the marine weather systems in the satellite cloud image. As a result, the marine weather systems in the satellite cloud image can be quickly, accurately and automatically identified through model recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 This is an application environment diagram of a method for identifying marine weather systems using satellite cloud images provided in Example 1 of the present application.

[0031] Figure 2 A flowchart of a method for identifying marine weather systems using satellite cloud images provided in Example 1 of the present application.

[0032] Figure 3 Schematic diagram of the network structure of the trained marine weather system identification model provided in Example 1 of the present application.

[0033] Figure 4 Schematic diagram of the fusion of multi-scale feature maps provided in Example 1 of the present application.

[0034] Figure 5 A schematic diagram of the functional modules of a satellite cloud image marine weather system identification device provided in Example 2 of the present application.

[0035] Figure 6 A schematic diagram of the structure of a computer device provided in Example 3 of the present application. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0037] Example 1.

[0038] The satellite cloud image marine weather system identification method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal communicates with the server via a network. The data storage system can store data that the server needs to process. The data storage system can be set up separately, integrated on the server, or placed on the cloud or other servers. The terminal can send the satellite cloud image to be processed to the server. After the server receives the satellite cloud image to be processed, it uses the satellite cloud image as input and uses the trained marine weather system identification model to annotate the marine weather systems in the satellite cloud image to identify the marine weather systems in the satellite cloud image. The server can then feedback the marine weather system identification results obtained for the satellite cloud image to the terminal.

[0039] In addition, in some embodiments, the satellite cloud image marine weather system identification method can also be implemented independently by a server or a terminal. For example, the terminal can directly process the satellite cloud image to be processed, or the server can obtain the satellite cloud image to be processed from the data storage system and process the satellite cloud image to be processed.

[0040] The terminals may include, but are not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart car devices. Portable wearable devices may include smart watches, smart bracelets, and head-mounted devices. The server may be implemented as a standalone server or a server cluster consisting of multiple servers, or as a cloud server.

[0041] In an exemplary embodiment, Figure 2 As shown, a method for identifying marine weather systems in satellite cloud images is provided. The method is executed by a computer device, specifically a computer device such as a terminal or a server, or a terminal and a server. In the embodiment of the present application, the method is applied to Figure 1 The following steps are used as an example to illustrate the server.

[0042] Step S1, obtaining a satellite cloud image of an analysis area; the analysis area includes an ocean area.

[0043] Step S2, taking the satellite cloud map as input, using the trained marine weather system identification model to label the marine weather system in the satellite cloud map to identify the marine weather system in the satellite cloud map; wherein the trained marine weather system identification model is a model obtained by training a deep learning network model using historical satellite cloud maps as input and the location information of the marine weather systems in the historical satellite cloud maps as labels.

[0044] By implementing the above-mentioned steps S1 to S2, this embodiment provides a method for identifying marine weather systems in satellite cloud images based on deep learning. Historical satellite cloud images are used as input in advance, and the location information of marine weather systems in the historical satellite cloud images is used as labels to train the deep learning network model to obtain a trained marine weather system identification model. In actual application, it is only necessary to use the satellite cloud image of the analysis area as input, and the trained marine weather system identification model can be used to mark the marine weather systems in the satellite cloud image to identify the marine weather systems in the satellite cloud image, thereby completing the identification of the marine weather systems in the satellite cloud image quickly, accurately and automatically.

[0045] In this embodiment, the satellite cloud map is an image of the cloud cover and surface features on the earth observed from top to bottom by a geostationary meteorological satellite, which may include visible light cloud maps and infrared cloud maps. The marine weather system refers to various weather systems occurring over the ocean, which may include high-pressure centers, low-pressure centers, subtropical high-pressure areas, low-pressure troughs, high-pressure ridges and equatorial convergence zones.

[0046] This embodiment is based on a deep learning method, combining ground weather analysis maps, 500hPa (hectopascal) potential height maps and satellite cloud maps obtained by corresponding geostationary meteorological satellites, to learn the characteristics of marine weather systems such as high-pressure centers, low-pressure centers, subtropical high-pressure areas, low-pressure troughs, high-pressure ridges, and equatorial convergence zones in the satellite cloud maps, and realize automatic recognition and labeling of corresponding marine weather systems in the satellite cloud maps obtained by geostationary meteorological satellites.

[0047] In this embodiment, before using the satellite cloud image as input and using the trained marine weather system identification model to mark the marine weather system in the satellite cloud image, the satellite cloud image marine weather system identification method of this embodiment also includes: training to obtain the trained marine weather system identification model, which specifically includes the following steps.

[0048] (1) Obtain multiple samples, including historical satellite cloud images, historical surface weather analysis images, and historical 500hPa geopotential height images corresponding to the time of the analysis area.

[0049] This embodiment directly obtains historical surface weather analysis maps, or uses global meteorological reanalysis data (such as ERA5 data) to draw historical surface weather analysis maps, and then uses global meteorological reanalysis data to draw historical 500hPa potential height maps. The main marine weather systems in the historical surface weather analysis maps and the historical 500hPa potential height maps are analyzed and annotated, and a historical weather analysis map library is established. At the same time, a geostationary meteorological satellite cloud map library that matches the time series and spatial position of the historical weather analysis map library is established. The geostationary meteorological satellite cloud map library includes historical satellite cloud maps corresponding to the time and space of the historical surface weather analysis maps and the historical 500hPa potential height maps, thereby constructing a data set.

[0050] At this time, in this embodiment, multiple samples are obtained, specifically including: for historical satellite cloud images, obtaining historical satellite cloud images of the analysis area; or obtaining L1 full-disk data of the analysis area, and drawing the historical satellite cloud images based on the L1 full-disk data. For historical surface weather analysis images and historical 500hPa geopotential height images, obtaining historical surface weather analysis images and global meteorological reanalysis data of the analysis area, and drawing the historical 500hPa geopotential height images based on the global meteorological reanalysis data; or obtaining global meteorological reanalysis data of the analysis area, and drawing the historical surface weather analysis images and historical 500hPa geopotential height images based on the global meteorological reanalysis data.

[0051] Among them, the time of historical satellite cloud images, historical surface weather analysis images and historical 500hPa potential height images corresponds to each other, and the map projections are all Mercator projections.

[0052] Specifically, if the historical surface weather analysis map can be directly obtained, the historical surface weather analysis map of the North Pacific region published by the National Oceanic and Atmospheric Administration of the United States on the Internet can be obtained. The historical surface weather analysis map covers most of the North Pacific Ocean, with a longitude and latitude range of 0-60°N (north latitude), 110°E (east longitude)-150°W (west longitude), and the map projection is the Mercator projection. The historical surface weather analysis map has manually marked marine weather systems such as high-pressure centers, low-pressure centers, low-pressure troughs, high-pressure ridges, and equatorial convergence zones. At the same time, the historical 500hPa potential height map is drawn using global meteorological reanalysis data. The map projection selects the Mercator projection, and the drawing range is 0-60°N, 110°E-150°W. The coverage range of the 588hPa isohyetal lines in the historical 500hPa potential height map is the subtropical high pressure area. When drawing the historical 500hPa potential height map, the 588hPa isohyetal lines can be bolded to mark the subtropical high pressure area.

[0053] If the historical surface weather analysis map cannot be obtained directly, the global meteorological reanalysis data will be used to draw the historical surface weather analysis map and the historical 500hPa potential height map. The Mercator projection is selected as the map projection, and the drawing range is 0-60°N, 110°E-150°W. The high-pressure center and the low-pressure center can be automatically marked in the historical surface weather analysis map during the drawing process, and the low-pressure trough line, high-pressure ridge line and equatorial convergence zone can be manually marked. The coverage of the 588hPa isohyet in the historical 500hPa potential height map is the subtropical high-pressure area. When drawing the historical 500hPa potential height map, the 588hPa isohyet can be bolded to mark the subtropical high-pressure area.

[0054] If historical satellite cloud images can be obtained directly, the historical visible light cloud images and historical infrared cloud images of the Himawari-8 geostationary meteorological satellite (of course, other geostationary meteorological satellites can also be used instead) can be obtained from the official website of the Japan Meteorological Agency. The historical visible light cloud images and historical infrared cloud images constitute the historical satellite cloud images. The cloud image range is selected from 0-60°N, 110°E-150°W, and the map projection of the cloud image is the Mercator projection.

[0055] If historical satellite cloud images cannot be obtained directly, obtain the L1 full-disk data of the Himawari-8 geostationary meteorological satellite from the official website of the Japan Meteorological Agency. Use the NetCDF (Network Common Data Format) data file format, with the file name NC_H08_YYYYMDD_hhmm_Rbb_FLDK.xxxxx_yyyyy.nc, where YYYYMDD_hhmm is UTC (Coordinated Universal Time) and bb represents the satellite channel. The time of the L1 full-disk data should match the time of the historical surface weather analysis charts and historical 500hPa geopotential height charts in the Historical Weather Analysis Chart Library. After obtaining the L1 full-disk data, select data from any channel 1-3 to draw historical visible light cloud images, and select data from any channel 10-13 to draw historical infrared cloud images. Set the map projection for drawing cloud images to Mercator projection.

[0056] (2) Label the marine weather systems in the historical surface weather analysis map and the historical 500hPa geopotential height map to generate a historical marine weather system distribution map, which contains the location information of the marine weather systems in the historical satellite cloud map.

[0057] The historical surface weather analysis map includes the location information of marine weather systems such as high-pressure centers, low-pressure centers, low-pressure troughs, high-pressure ridges, and equatorial convergence zones. The historical 500hPa geopotential height map includes the location information of marine weather systems such as subtropical high-pressure areas. Based on the location information of marine weather systems, the marine weather systems in the historical surface weather analysis map and the historical 500hPa geopotential height map are further labeled to generate a historical marine weather system distribution map to create a label file. Specifically, the LabelImg tool is used for labeling based on the location information of the marine weather system. The LabelImg tool supports rectangular box labeling and can create label files in XML (eXtensible Markup Language) or JSON (JavaScript Object Notation) format.

[0058] Taking the low-pressure trough line as an example, the specific labeling steps are: collect the weather system distribution map of the low-pressure trough line category (that is, the historical ground weather analysis map with the location information of the low-pressure trough line as a marine weather system), start the LabelImg tool, open the folder containing the weather system distribution map of the low-pressure trough line category, and in the image preview window, select an image (that is, the weather system distribution map of the low-pressure trough line category) for labeling. Use the rectangular frame tool to drag the mouse on the image to draw a labeling frame around the low-pressure trough line, and enter "trough_line (that is, low-pressure trough line)" as the label in the label input box. If there are multiple low-pressure trough lines in the image, repeat the above steps to create a labeling frame and label for each low-pressure trough line. After completing the labeling, save the labeling information as an XML file to obtain the historical marine weather system distribution map. Other categories can be labeled by referring to the same steps.

[0059] (3) Using historical satellite cloud images as input and historical marine weather system distribution maps as labels, the deep learning network model is trained to obtain a trained marine weather system recognition model.

[0060] Among them, the deep learning network model adopts the improved RT-DETR (Real-Time Detection Transformer) model. The improved RT-DETR model is a model obtained by improving the RT-DETR model. The improvements include: replacing the convolutional layers in the backbone network of the RT-DETR model with spatial and channel reconstruction convolutional layers, making the neck network of the RT-DETR model use a feature pyramid network, and replacing the convolutional layers in the feature pyramid network with spatial and channel reconstruction convolutional layers.

[0061] Specifically, such as Figure 3As shown in the figure, the deep learning network model adopts the improved RT-DETR model. The improved RT-DETR model is a model obtained by improving the RT-DETR model. The RT-DETR model includes a backbone network (Backbone), a neck network (Neck), an efficient hybrid encoder (Efficient Hybrid Encoder, containing a Transformer encoder), query selection and decoder based on IOU (Intersection over Union, intersection ratio) perception and a prediction head (Detection&Head). The backbone network is the HGNetV2 (High Performance GPU Net V2) model. The HGNetV2 model is mainly composed of multiple HG-Blocks (Hierarchical Graph-Blocks). The hierarchical blocks are composed of a series of 3 3 traditional convolutional layers stacked together. That is, in the RT-DETR model, the backbone network contains a large number of traditional convolutional layers. To improve feature extraction efficiency and network performance, this embodiment uses spatial and channel reconstruction convolution layers (SCConv) in the backbone network to replace these traditional convolutional layers, thereby optimizing the feature extraction process and reducing computing resource consumption. In addition, in the neck network, this embodiment introduces SCConv-enhanced Feature Pyramid Networks (FPN). SCConv-enhanced Feature Pyramid Networks replace the convolutional layers in the Feature Pyramid Networks with spatial and channel reconstruction convolution layers. This SCConv-enhanced Feature Pyramid Network can generate feature representations rich in multi-scale information, significantly enhancing the model's ability to represent image features, thereby improving overall network performance.

[0062] At this time, the workflow of the deep learning network model is as follows: first, the backbone network of the improved RT-DETR model is used to generate multi-scale feature maps. The backbone network uses HGNetV2 as the basic structure and replaces all traditional convolutional layers with SCConv to enhance feature extraction capabilities; the multi-scale feature maps generated by the backbone network are input into the FPN (i.e., the neck network) enhanced by SCConv, and further reconstructed using spatial and channel information to improve the feature expression capability and output multi-scale features; using an efficient hybrid encoder, by decoupling the attention-based intra-scale feature interaction (AIFI) and cross-scale feature fusion (Cross-scale Feature Fusion (CCFF) efficiently processes the multi-scale features output by FPN, reduces computational redundancy, improves feature extraction efficiency, and obtains a feature fusion map. The feature fusion map output by CCFF enters the query selection based on IOU perception. Finally, the decoder and prediction head process the position embedding information, image feature information, and target query information to determine the category, position, and bounding box size of the target prediction box. That is, the target prediction box is used to select the marine weather system, realizing the learning of the corresponding cloud map features of the marine weather system and the accurate identification of the marine weather system.

[0063] After constructing the deep learning network model, historical satellite cloud images can be used as input, and historical marine weather system distribution maps can be used as labels to train the deep learning network model, resulting in a trained marine weather system identification model. It should be noted that when training the deep learning network model, a corresponding marine weather system identification model is obtained for each marine weather system, that is, a trained marine weather system identification model corresponding to each marine weather system is obtained.

[0064] During training, the input data consists of geostationary satellite imagery, specifically historical satellite cloud images, including visible and infrared images. Since both images are single-channel, different images are used as input channels, resulting in two-channel input data. The labeled data consists of historical maps of marine weather systems. The dataset, consisting of multiple samples, is divided into training, validation, and test sets, maintaining a roughly 8:1:1 ratio. This partitioning aims to train the model using a training set containing sufficient training data, fine-tune model parameters using the validation set, and ultimately evaluate the model's generalization ability using the test set, ensuring its stability and reliability in practical applications. During training, two loss functions are used: a classification loss and a bounding box regression loss, which together constitute the overall loss function. The classification loss can be implemented using the cross-entropy function, while the bounding box regression loss can be implemented using the Smooth L1 function.

[0065] When training a model, the usual practice is to perform a weighted summation of the classification loss function and the bounding box regression loss function to form a total loss function. This approach allows the model to consider both the accuracy of object classification and the accuracy of bounding box positioning during the optimization process. Specifically, the total loss It can be calculated by the following formula (1).

[0066] (1).

[0067] In formula (1), is the total loss; is a hyperparameter used to balance the weights of classification loss and bounding box regression loss; is the classification loss; is the bounding box regression loss.

[0068] The calculation formula of classification loss is as follows (2).

[0069] (2).

[0070] In formula (2), is the number of samples; is the number of categories; is an indicator variable, if the sample The true category is ,but ,otherwise, ; Predict samples for the model Belong to category probability.

[0071] The calculation formula of the bounding box regression loss is as follows (3).

[0072] (3).

[0073] In formula (3), is the number of samples; The bounding box parameters predicted by the model (usually the center point coordinates ,width ,high ); are the actual bounding box parameters.

[0074] is the Smooth L1 function, which is defined as the following formula (4).

[0075] (4).

[0076] In formula (4), For input.

[0077] At this time, in this embodiment, when training the deep learning network model, the loss functions used include a classification loss function and a bounding box regression loss function. The classification loss function adopts a cross entropy function, and the bounding box regression loss function adopts a Smooth L1 function.

[0078] The processing process of the deep learning network model on the historical satellite cloud image is as follows: Feature extraction: The historical satellite cloud image in the dataset is input into the backbone network of the improved RT-DETR model. The backbone network outputs a feature map with three dimensions [C, H, W], where C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map; taking the third, fourth, and fifth layers of feature maps output by the backbone network as an example (here we take the backbone network as an example to generate five layers of feature maps), they are respectively recorded as FP3, FP4, and FP5, as shown in Figure 2. Figure 4As shown in the figure, first, the number of channels of these feature maps is uniformly adjusted to 256 to obtain FP3_Later, FP4_Later, and FP5_Later. Then, the nearest neighbor interpolation method is used to upsample FP5_Later to obtain upsampled_FP5, and it is horizontally connected with FP4_Later, that is, the corresponding element values ​​are added to obtain the fused merged_FP4. The merged_FP4 is upsampled and horizontally connected with FP3_Later to obtain the fused merged_FP3. Finally, these fused multi-scale feature maps are passed to the next processing stage, that is, the multi-scale feature map is output; Multi-scale feature map fusion: In FPN, the semantic information of the high-level feature map is fused with the spatial information of the low-level feature map to generate a feature representation with rich multi-scale information and output multi-scale features; Capture global context information: The output multi-scale feature map The degree features are input into the Transformer encoder of the model. In this process, the position encoding, query selection, self-attention mechanism, multi-head attention and multiple calculation steps of the encoder and decoder are combined to generate a feature vector with stronger global context information capture ability, and output a feature fusion map; detection head processing: IOU-aware query selection receives the feature fusion map from the Transformer encoder, and after further processing, it is passed to the decoder and prediction head. First, the decoder and prediction head perform a linear transformation on these feature vectors to generate class classification logits (referring to the original prediction value of the network output layer), and apply the Softmax activation function to predict the probability of each category. Then, another linear transformation is performed on the feature vector to predict the regression value of the bounding box. According to the classification results of the marine weather system, the position and size of the target prediction box are adjusted to realize the inspection and recognition of cloud map features.

[0079] After obtaining the trained marine weather system identification model, obtain the satellite cloud image of the geostationary meteorological satellite in the same area as the historical satellite cloud image in the data set, input the satellite cloud image into the trained marine weather system identification model, mark the marine weather system in the satellite cloud image, and output the new cloud image product to complete the identification of the marine weather system.

[0080] At this time, a satellite cloud image of the analysis area is obtained. The analysis area includes the ocean area. The satellite cloud image is used as input, and the trained marine weather system recognition model is used to mark the marine weather system in the satellite cloud image to identify the marine weather system in the satellite cloud image and output a new cloud image product. In the new cloud image product, the marine weather system in the satellite cloud image is selected using a marking box.

[0081] At present, the main purpose of satellite cloud images of geostationary meteorological satellites is to assist in judging the intensity and location of disastrous weather systems or processes such as tropical cyclones, severe convection, and sea fog. These disastrous weather systems are only part of the marine weather systems displayed in satellite cloud images. On the one hand, they must be manually interpreted in combination with weather maps. On the other hand, these application scenarios also have certain limitations and fail to give full play to the functions of satellite cloud images. In order to solve this problem, this embodiment provides a method for identifying marine weather systems in satellite cloud images based on deep learning. It can correspond and learn the characteristics of marine weather systems such as high-pressure centers, low-pressure centers, subtropical high-pressure areas, low-pressure troughs, high-pressure ridges, and equatorial convergence zones in weather maps and satellite cloud images of corresponding geostationary meteorological satellites, and realize automatic identification and labeling of corresponding marine weather systems in satellite cloud images, replacing manual judgment and providing a reference basis for weather system analysis and mapping in marine areas.

[0082] This application also provides an application scenario that applies the above-mentioned method for identifying marine weather systems in satellite cloud images. Specifically, the method for identifying marine weather systems in satellite cloud images provided in this embodiment can be applied in a marine weather system identification scenario. The marine weather system identification scenario includes an identification phase and a display phase. The identification phase is used to use a trained marine weather system identification model to annotate marine weather systems in satellite cloud images to identify marine weather systems in the satellite cloud images, and the display phase is used to display marine weather systems in the satellite cloud images. The method for identifying marine weather systems in satellite cloud images provided in this embodiment belongs to the identification phase.

[0083] Example 2.

[0084] Based on the same inventive concept, embodiments of the present application also provide a device for identifying marine weather systems using satellite cloud images, for implementing the aforementioned method for identifying marine weather systems using satellite cloud images. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for identifying marine weather systems using satellite cloud images provided below can be found in the aforementioned limitations of the method for identifying marine weather systems using satellite cloud images, and will not be further elaborated here.

[0085] In an exemplary embodiment, Figure 5 As shown, a satellite cloud image marine weather system identification device is provided, and the satellite cloud image marine weather system identification device includes the following modules.

[0086] The acquisition module M1 is used to acquire a satellite cloud image of an analysis area; the analysis area includes an ocean area.

[0087] The identification module M2 is used to use the satellite cloud image as input and use the trained marine weather system identification model to label the marine weather system in the satellite cloud image to identify the marine weather system in the satellite cloud image; wherein the trained marine weather system identification model is a model obtained by training a deep learning network model with historical satellite cloud images as input and the location information of the marine weather systems in the historical satellite cloud images as labels.

[0088] Example 3.

[0089] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, memory and input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for identifying marine weather systems using satellite cloud images is implemented.

[0090] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0091] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the method for identifying marine weather systems based on satellite cloud images in Example 1 is implemented.

[0092] Example 4.

[0093] In an exemplary embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the method for identifying marine weather systems using satellite cloud images in Example 1 is implemented.

[0094] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0095] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for identifying marine weather systems from satellite cloud images, characterized in that: The satellite cloud image marine weather system identification method comprises: Obtaining a satellite cloud image of an analysis area; the analysis area includes an ocean area; the satellite cloud image includes a visible light cloud image and an infrared cloud image; The satellite cloud image is used as input, and the trained marine weather system identification model is used to annotate the marine weather systems in the satellite cloud image to identify the marine weather systems in the satellite cloud image; wherein the trained marine weather system identification model is a model obtained by training a deep learning network model using historical satellite cloud images as input and the location information of marine weather systems in the historical satellite cloud images as labels; the location information of marine weather systems in the historical satellite cloud images is represented by a historical marine weather system distribution map, and the historical marine weather system distribution map is generated by annotating marine weather systems in historical surface weather analysis maps and historical 500hPa geopotential height maps; the marine weather system includes a high-pressure center, a low-pressure center, a subtropical high-pressure area, a low-pressure trough line, a high-pressure ridge line, and an equatorial convergence zone; The deep learning network model adopts the improved RT-DETR model, which is a model obtained by improving the RT-DETR model. The improvements include: replacing the convolutional layer in the backbone network of the RT-DETR model with a spatial and channel reconstruction convolutional layer, making the neck network of the RT-DETR model use a feature pyramid network, and replacing the convolutional layer in the feature pyramid network with a spatial and channel reconstruction convolutional layer; the third, fourth, and fifth feature maps output by the backbone network are respectively recorded as FP3, FP4, and FP5, and the number of channels of these feature maps is uniformly adjusted to 256 to obtain FP3_Later, FP4_Later, and FP5_Later, and FP5_Later is upsampled using the nearest neighbor interpolation method to obtain upsampled_FP5, and is horizontally connected with FP4_Later to obtain a fused merged_FP4, and the merged_FP4 is upsampled and horizontally connected with FP3_Later to obtain a fused merged_FP3, outputting a multi-scale feature map, inputting the multi-scale feature map into the neck network, and outputting multi-scale features; When training a deep learning network model, the loss functions used include a classification loss function and a bounding box regression loss function. Specifically, a weighted sum of the classification loss function and the bounding box regression loss function is performed to obtain a total loss. The classification loss function adopts a cross entropy function, and the bounding box regression loss function adopts a Smooth L1 function. The total loss is calculated as: ； in, is the total loss; is a hyperparameter; is the classification loss; is the bounding box regression loss; The classification loss is calculated as: ； in, is the number of samples; is the number of categories; is an indicator variable, if the sample The true category is ,but ,otherwise, ; Predict samples for the model Belong to category probability; The bounding box regression loss is calculated as: ； in, is the number of samples; is the center point coordinate, is the width, is the height; is the SmoothL1 function; The bounding box parameters predicted by the model; are the actual bounding box parameters.

2. The method for identifying marine weather systems from satellite cloud images according to claim 1, characterized in that: Before labeling the marine weather system in the satellite cloud image using the trained marine weather system identification model using the satellite cloud image as input, the method for identifying marine weather systems in the satellite cloud image further includes: training to obtain the trained marine weather system identification model, specifically including: Acquire multiple samples; the samples include historical satellite cloud images, historical surface weather analysis images, and historical 500hPa geopotential height images corresponding to the time of the analysis area; Annotating the marine weather systems in the historical surface weather analysis map and the historical 500hPa geopotential height map to generate a historical marine weather system distribution map; the historical marine weather system distribution map contains location information of marine weather systems in historical satellite cloud maps; The deep learning network model is trained using the historical satellite cloud image as input and the historical marine weather system distribution map as a label to obtain a trained marine weather system recognition model.

3. The method for identifying marine weather systems from satellite cloud images according to claim 2, characterized in that: Get multiple samples, including: For historical satellite cloud images, obtain historical satellite cloud images of the analysis area; or obtain L1 level full disk data of the analysis area, and draw a historical satellite cloud image based on the L1 level full disk data; For historical surface weather analysis maps and historical 500 hPa potential height maps, obtain historical surface weather analysis maps and global meteorological reanalysis data for the analysis area, and draw historical 500 hPa potential height maps based on the global meteorological reanalysis data; or obtain global meteorological reanalysis data for the analysis area, and draw historical surface weather analysis maps and historical 500 hPa potential height maps based on the global meteorological reanalysis data; The time periods of the historical satellite cloud map, the historical ground weather analysis map and the historical 500hPa potential height map correspond to each other, and the map projections are all Mercator projections.

4. A satellite cloud image marine weather system identification device, characterized in that: The satellite cloud image marine weather system identification device comprises: An acquisition module is used to acquire a satellite cloud image of an analysis area; the analysis area includes an ocean area; the satellite cloud image includes a visible light cloud image and an infrared cloud image; An identification module is configured to use the satellite cloud image as input and utilize a trained marine weather system identification model to annotate the marine weather systems in the satellite cloud image, so as to identify the marine weather systems in the satellite cloud image; wherein the trained marine weather system identification model is a model obtained by training a deep learning network model using historical satellite cloud images as input and using the location information of marine weather systems in the historical satellite cloud images as labels; the location information of marine weather systems in the historical satellite cloud images is represented by a historical marine weather system distribution map, which is generated by annotating marine weather systems in historical surface weather analysis maps and historical 500hPa geopotential height maps; the marine weather system includes a high-pressure center, a low-pressure center, a subtropical high-pressure area, a low-pressure trough, a high-pressure ridge, and an equatorial convergence zone; The deep learning network model adopts the improved RT-DETR model, which is a model obtained by improving the RT-DETR model. The improvements include: replacing the convolutional layer in the backbone network of the RT-DETR model with a spatial and channel reconstruction convolutional layer, making the neck network of the RT-DETR model use a feature pyramid network, and replacing the convolutional layer in the feature pyramid network with a spatial and channel reconstruction convolutional layer; the third, fourth, and fifth feature maps output by the backbone network are respectively recorded as FP3, FP4, and FP5, and the number of channels of these feature maps is uniformly adjusted to 256 to obtain FP3_Later, FP4_Later, and FP5_Later, and FP5_Later is upsampled using the nearest neighbor interpolation method to obtain upsampled_FP5, and is horizontally connected with FP4_Later to obtain a fused merged_FP4, and the merged_FP4 is upsampled and horizontally connected with FP3_Later to obtain a fused merged_FP3, outputting a multi-scale feature map, inputting the multi-scale feature map into the neck network, and outputting multi-scale features; When training a deep learning network model, the loss functions used include a classification loss function and a bounding box regression loss function. Specifically, a weighted sum of the classification loss function and the bounding box regression loss function is performed to obtain a total loss. The classification loss function adopts a cross entropy function, and the bounding box regression loss function adopts a Smooth L1 function. The total loss is calculated as: ； in, is the total loss; is a hyperparameter; is the classification loss; is the bounding box regression loss; The classification loss is calculated as: ； in, is the number of samples; is the number of categories; is an indicator variable, if the sample The true category is ,but ,otherwise, ; Predict samples for the model Belong to category probability; The bounding box regression loss is calculated as: ； in, is the number of samples; is the center point coordinate, is the width, is the height; is the SmoothL1 function; The bounding box parameters predicted by the model; are the actual bounding box parameters.

5. A computer device comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the satellite cloud image marine weather system identification method according to any one of claims 1-3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for identifying marine weather systems in satellite cloud images according to any one of claims 1 to 3 is implemented.

Images