Radar reflectivity inversion method based on satellite infrared data and lightning observation data

Abstract

The invention discloses a radar reflectivity inversion method based on satellite infrared data and lightning observation data, which belongs to the technical field of meteorological monitoring, and comprises the following steps: taking a three-channel input image formed by the lightning observation data and two infrared images from different infrared channels as the input of a scattering model; the lightning activity is closely related to severe convective weather, so that the scattering model can be helped to perform radar reflectivity inversion in areas lacking radar data, such as an open sea, a plateau and a desert, and the defect of insufficient accurate modeling of strong weather events in the prior art is overcome; more comprehensive and multi-dimensional input information is provided for radar data reconstruction, and the accuracy of radar reflectivity inversion is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological monitoring, and particularly relates to a method for retrieving radar reflectivity based on satellite infrared data and lightning observation data. Background Art

[0002] In the context of the deep integration of meteorology into economic and social development, frequent occurrence of disastrous weather poses a serious threat to property and life safety, attracting wide attention from the public and various industries. China has a vast territory, with certain differences in climate and terrain conditions, and a wide variety of disastrous weather. Disastrous weather generated by atmospheric convective motion, such as thunderstorms, squall lines, hail, etc., has a highly non-linear structure. This kind of weather has a short life history, strong suddenness, and great difficulty in forecasting, and is extremely likely to cause various meteorological disasters. Accurately monitoring and identifying the formation and development of disastrous weather is crucial for improving the ability to defend against meteorological disasters and reducing economic and life losses.

[0003] With the continuous increase of meteorological observation data and the rapid development of artificial intelligence technology, deep learning methods based on data-driven have become a research hotspot in the meteorological field. Compared with the artificial feature extraction of traditional machine learning, deep learning methods can automatically extract useful features from a large amount of data and learn the laws therein, thereby improving the prediction performance of the model. At present, many meteorological tasks can be completed based on deep learning methods, such as retrieving precipitation, temperature, and wind fields using multi-source meteorological data such as satellite data and radar data, discriminating and nowcasting severe convective weather. However, affected by factors such as radar wave attenuation, terrain obstruction, and incomplete radar network coverage in some areas, some radar data is missing, and its availability and referenceability are restricted.

[0004] Currently, existing radar reflectivity retrievals include: traditional radar reflectivity retrieval methods based on non-radar data and radar reflectivity retrieval methods based on deep learning. Among them, in the process of data processing and conversion, traditional radar reflectivity retrieval methods based on non-radar data are easily affected by problems such as noise and incomplete data, resulting in low accuracy of the reconstructed radar reflectivity data, which brings great uncertainty in weather disaster early warning. In addition, radar reflectivity retrieval methods based on deep learning generally reconstruct high-resolution images through downsampling and upsampling operations. However, due to the loss of spatial resolution during downsampling, it is often impossible to fully maintain global context information when restoring image details subsequently, resulting in insufficient spatial sensitivity of high-resolution images and low accuracy of the retrieved radar reflectivity. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a method for retrieving radar reflectivity based on satellite infrared data and lightning observation data. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] The present invention provides a method for retrieving radar reflectivity based on satellite infrared data and lightning observation data, including:

[0007] Obtain lightning observation data and two infrared images from different infrared channels to form a three-channel input image;

[0008] Input the input image into an inversion model, and make the inversion model perform the following steps:

[0009] Extract features from the input image, and use a mapping unit to map the extracted features into a one-dimensional feature vector;

[0010] Input the one-dimensional feature vector into an encoder, and make each encoding unit connected in sequence in the encoder use two Swin Transformer modules to extract features from its own input data in sequence to obtain a feature vector X En , and send the feature vector X En to at least one decoding unit in an upsampling module and a decoder, and further use the upsampling layer to perform upsampling processing on the feature vector X En to obtain a first feature map;

[0011] Input the first feature map output by the last-level encoding unit into a bottleneck layer for feature extraction to obtain a feature vector X Bo and then output it to each decoding unit in the decoder;

[0012] Input the feature vectors into each decoding unit connected in sequence in the decoder, and make each decoding unit perform cross-scale feature fusion based on the input feature vector X En and the feature vector X Bo to perform cross-scale feature fusion, and then use two Swin Transformer modules to extract features from the fused feature map in sequence to obtain a second feature map;

[0013] Use a downsampling unit to perform downsampling processing on the second feature map output by the last-level decoding unit to obtain a reconstructed image, and each pixel point in the reconstructed image respectively represents the radar reflectivity value at the corresponding radar site.

[0014] In an embodiment of the present invention, the encoder includes a first encoding unit, a second encoding unit, and a third encoding unit connected in sequence, and the decoder includes a first decoding unit, a second decoding unit, and a third decoding unit connected in sequence; wherein,

[0015] Input the one-dimensional feature vector into the encoder, and enable each encoding unit connected in sequence in the encoder to perform feature extraction on its own input data using two Swin Transformer modules in sequence to obtain the feature vector X En , and send the feature vector X En to at least one decoding unit in the upsampling module and the decoder, and further use the upsampling layer to perform upsampling processing on the feature vector X En The steps for obtaining the first feature map include:

[0016] Input the one-dimensional feature vector into the first encoding unit, and use two Swin Transformer modules in the first encoding unit to perform feature extraction on the one-dimensional feature vector in sequence to obtain the feature vector and input them into the upsampling module, the first decoding unit, the second decoding unit, and the third decoding unit in the first encoding unit respectively;

[0017] Use the upsampling module in the first encoding unit to perform upsampling processing on the feature vector to obtain the first feature map F1 and input it into the second encoding unit;

[0018] Use two Swin Transformer modules in the second encoding unit to perform feature extraction on the first feature map F1 in sequence to obtain the feature vector and input them into the upsampling module, the first decoding unit, and the second decoding unit in the second encoding unit respectively;

[0019] Use the upsampling module in the second encoding unit to perform upsampling processing on the feature vector to obtain the first feature map F2 and input it into the third encoding unit;

[0020] Use two Swin Transformer modules in the third encoding unit to perform feature extraction on the first feature map F2 in sequence to obtain the feature vector and input it into the upsampling module in the third encoding unit and the first decoding unit;

[0021] Use the upsampling module in the third encoding unit to perform upsampling processing on the feature vector to obtain the first feature map F3 and input it into the bottleneck layer.

[0022] In an embodiment of the present invention, the bottleneck layer includes two Swin Transformer modules connected in sequence.

[0023] In one embodiment of the present invention, the feature vectors are respectively input into each decoding unit connected in sequence in the decoder, so that each decoding unit is based on the input feature vector X En and the feature vector X Bo perform cross-scale feature fusion, and then use two Swin Transformer modules to sequentially perform feature extraction on the fused feature map to obtain the second feature map. The steps include:

[0024] Input the feature vector X Bo into the first decoding unit, and use the first cross-scale feature fusion module in the first decoding unit to process the feature vector X Bo and the feature vector the feature vector and the feature vector After fusion, use the two Swin Transformer modules in the first decoding unit to sequentially perform feature extraction to obtain the second feature map F4 and input it into the second decoding unit and the third decoding unit;

[0025] Use the second cross-scale feature fusion module in the second decoding unit to process the feature vector X Bo the second feature map F4, the feature vector and the feature vector After fusion, use the two Swin Transformer modules in the second decoding unit to sequentially perform feature extraction to obtain the second feature map F5 and input it into the third decoding unit;

[0026] Use the third cross-scale feature fusion module in the third decoding unit to process the feature vector X Bo the obtained second feature map F4, the second feature map F5 and the feature vector After fusion, use the two Swin Transformer modules in the third decoding unit to sequentially perform feature extraction to obtain the second feature map F6.

[0027] In one embodiment of the present invention, the first cross-scale fusion module fuses the feature vector X Bo and the feature vector the feature vector and the feature vector as follows:

[0028] Upsample the feature vector twice, upsample the feature vector once, and upsample the feature vector X BoAfter performing one downsampling, it is concatenated with the said feature vector to obtain a first sub-feature map and send it to a fully connected layer to obtain the said feature vector X Bo and the said feature vector the said feature vector and the said feature vector to obtain a fused feature map.

[0029] In one embodiment of the present invention, the second cross-scale fusion module fuses the said feature vector X Bo , the second feature map F4, the said feature vector and the said feature vector according to the following steps:

[0030] Perform one downsampling on the said feature vector respectively, perform one upsampling on the said second feature map F4, perform two upsamplings on the said feature vector X Bo and then concatenate it with the said feature vector to obtain a second sub-feature map and send it to a fully connected layer to obtain the said feature vector X Bo , the second feature map F4, the said feature vector and the said feature vector to obtain a fused feature map.

[0031] In one embodiment of the present invention, the third cross-scale feature fusion module fuses the said feature vector X Bo , the obtained second feature map F4, the second feature map F5 and the feature vector according to the following steps:

[0032] Perform two downsamplings on the second feature map F4 respectively, perform one downsampling on the second feature map F5, perform three upsamplings on the said feature vector X Bo and then concatenate it with the said feature vector to obtain a third sub-feature map and send it to a fully connected layer to obtain the said feature vector X Bo , the obtained second feature map F4, the second feature map F5 and the feature vector to obtain a fused feature map.

[0033] In one embodiment of the present invention, the inversion model is trained in a supervised learning manner based on training samples, labels and a preset loss function; wherein each said training sample includes sample lightning observation data and two sample infrared images from different infrared channels, and the label is a radar composite reflectivity image.

[0034] In one embodiment of the present invention, the preset loss function is a weighted loss function L WSSIM = αL WMSE + βL mySSIM ;

[0035] Wherein,

[0036]

[0037] In the formula, N represents the total number of pixels of the current reconstructed image obtained in each round of training, y i represents the value of the position of the i-th pixel in the radar composite reflectivity image, w(y i ) represents the weight function dependent on y i , represents the value of the position of the i-th pixel in the current reconstructed image, SSIM represents the structural similarity index, which is used to measure the similarity between y i and , α represents the weight of the mean square error loss L WMSE , and β represents the weight of the structural similarity loss L mySSIM .

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] As can be seen from the above embodiments, the beneficial effects of the present invention are as follows:

[0040] (1) In the radar reflectivity inversion method provided by the present invention, the three-channel input image formed by two infrared images and lightning observation data is used as the input of the scattering model. Since lightning activities are closely related to severe convective weather, it can help the scattering model to invert the radar reflectivity in areas lacking radar data such as the open sea, plateau, and desert. It not only makes up for the lack of accurate modeling of severe weather events in the prior art, but also provides more comprehensive and multi-dimensional input information for radar data reconstruction, significantly improving the accuracy of radar reflectivity inversion.

[0041] (2) For the scattering model, the present invention introduces the Swin Transformer module in each encoding unit, bottleneck layer of the encoder, and each decoding unit of the decoder, which better preserves the spatial details and global information of the image. The cross-scale feature fusion module can effectively fuse feature information at different levels. Therefore, the present invention enhances the ability of the scattering model to perceive features at different scales and levels, ensuring that the inversion model takes into account both local details and global context during the inversion process, breaking through the limitation of losing global perception during the improvement of spatial resolution in the prior art, and improving the accuracy and detail performance of the inversion results.

[0042] (3) For the training process of the inversion model, the present invention designs a weighted loss function L WSSIM , which assigns different weights to different regions and weather phenomena, thereby specifically optimizing the radar reflectivity inversion effect under complex weather conditions, especially performing well in the monitoring of extreme weather such as severe convective weather and typhoons.

[0043] The following will further elaborate on the present invention in conjunction with the drawings and embodiments. Description of the Drawings

[0044] Figure 1 is a flowchart of a method for radar reflectivity inversion based on satellite infrared data and lightning observation data provided by an embodiment of the present invention;

[0045] Figure 2 is a schematic structural diagram of an inversion model provided by an embodiment of the present invention;

[0046] Figure 3 is a schematic structural diagram of a first decoding unit provided by an embodiment of the present invention;

[0047] Figure 4 is a schematic structural diagram of a second decoding unit provided by an embodiment of the present invention;

[0048] Figure 5 is a schematic structural diagram of a third decoding unit provided by an embodiment of the present invention;

[0049] Figure 6 is an example diagram of a meteorological event in a dataset provided by an embodiment of the present invention. Detailed Embodiments

[0050] The following further describes the present invention in detail with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0051] Figure 1 is a flowchart of a method for radar reflectivity inversion based on satellite infrared data and lightning observation data provided by an embodiment of the present invention, Figure 2 is a schematic structural diagram of an inversion model provided by an embodiment of the present invention. Please refer to Figure 1-2 , an embodiment of the present invention provides a method for radar reflectivity inversion based on satellite infrared data and lightning observation data, including:

[0052] S1. Obtain lightning observation data and two infrared images from different infrared channels to form a three-channel input image;

[0053] S2. Input the input image into the inversion model, and make the inversion model perform the following steps:

[0054] S201. Extract features from the input image, and use the mapping unit to map the extracted features into a one-dimensional feature vector;

[0055] S202. Input the one-dimensional feature vector into the encoder, so that each encoding unit connected in sequence in the encoder uses two Swin Transformer modules (S-T modules) to extract features from its own input data in sequence, obtaining the feature vector X En , and send the feature vector X En to at least one decoding unit in the upsampling module and the decoder, and further use the upsampling layer to perform upsampling processing on the feature vector X En to obtain the first feature map;

[0056] S203. Input the first feature map output by the last-level encoding unit into the bottleneck layer for feature extraction, obtaining the feature vector X Bo and then output to each decoding unit in the decoder;

[0057] S204. Input the feature vector into each decoding unit connected in sequence in the decoder, so that each decoding unit performs cross-scale feature fusion based on the input feature vector X En and the feature vector X Bo , and then use two Swin Transformer modules to extract features from the fused feature map in sequence, obtaining the second feature map;

[0058] S205. Use the downsampling unit to perform downsampling processing on the second feature map output by the last-level decoding unit, obtaining the reconstructed image, and each pixel point in the reconstructed image represents the radar reflectivity value at the corresponding radar site.

[0059] It should be noted that the two infrared images obtained in step S1 come from two infrared channels respectively, namely: IR107 (10.7μm) and IR069 (6.9μm), which are two commonly used infrared channels in the field of meteorological satellites and can provide information on different heights and temperature levels. Among them, the IR107 channel mainly reflects the cloud top temperature and can be used to identify convective clouds, cloud top cooling, etc., while the IR069 channel mainly detects the middle layer water vapor and helps to analyze the humidity distribution and convective activities in the middle layer of the atmosphere.

[0060] These two infrared images are satellite observation data of the same time and the same area. The IR107 channel focuses on cloud top features, and the IR069 channel reflects the middle layer water vapor content. The combination of the two can improve the recognition ability of deep convection and precipitation systems and provide richer input features for radar reflectivity reconstruction. Therefore, in this embodiment, by combining the information of different channels, the atmospheric state can be more comprehensively characterized.

[0061] The present invention introduces lightning observation data into the radar reflectivity inversion process. Especially in severe convective weather and extreme climates, the introduction of lightning activity can significantly enhance the model's prediction ability for high-threshold weather events. Lightning observation data provides important clues about severe convective activities in meteorological systems, especially in areas where radar data cannot cover (such as the open sea, plateaus, deserts, etc.). Therefore, it can improve the inversion accuracy of radar reflectivity and contribute to enhancing the timeliness and accuracy of meteorological warnings.

[0062] Optionally, the encoder includes a first encoding unit (Encoder1), a second encoding unit (Encoder2), and a third encoding unit (Encoder3) connected in sequence, and the decoder includes a first decoding unit (Decoder1), a second decoding unit (Decoder2), and a third decoding unit (Decoder3) connected in sequence; wherein,

[0063] In step S202, the one-dimensional feature vector is input into the encoder, and each encoding unit connected in sequence in the encoder uses two Swin Transformer modules to sequentially extract features from its own input data to obtain the feature vector X En , and the feature vector X En is sent to at least one decoding unit in the upsampling module and the decoder, and the upsampling layer is further used to perform upsampling processing on the feature vector X En to obtain the first feature map. The steps include:

[0064] Input the one-dimensional feature vector into the first encoding unit, and use two Swin Transformer modules in the first encoding unit to sequentially extract features from the one-dimensional feature vector to obtain the feature vector and input them into the upsampling module, the first decoding unit, the second decoding unit, and the third decoding unit in the first encoding unit respectively;

[0065] Use the upsampling module in the first encoding unit to perform upsampling processing on the feature vector to obtain the first feature map F1 and input it into the second encoding unit;

[0066] Use two Swin Transformer modules in the second encoding unit to sequentially extract features from the first feature map F1 to obtain the feature vector and input them into the upsampling module, the first decoding unit, and the second decoding unit in the second encoding unit respectively;

[0067] Use the upsampling module in the second encoding unit to perform upsampling processing on the feature vector to obtain the first feature map F2 and input it into the third encoding unit;

[0068] Use two Swin Transformer modules in the third encoding unit to sequentially perform feature extraction on the first feature map F2 to obtain a feature vector and input it into the upsampling module and the first decoding unit in the third encoding unit;

[0069] Use the upsampling module in the third encoding unit to perform upsampling on the feature vector to obtain the first feature map F3 and input it into the bottleneck layer.

[0070] In this embodiment, the bottleneck layer includes two sequentially connected Swin Transformer modules.

[0071] In step S204, input the feature vector into each decoding unit connected in sequence in the decoder, so that each decoding unit is based on the input feature vector X En and the feature vector X Bo perform cross-scale feature fusion, and then use two Swin Transformer modules to sequentially perform feature extraction on the fused feature map to obtain the second feature map. The steps include:

[0072] Input the feature vector X Bo into the first decoding unit, and use the first cross-scale feature fusion module (MSFF module 1) in the first decoding unit to fuse the feature vector X Bo and the feature vector feature vector and the feature vector After fusion, use two Swin Transformer modules in the first decoding unit to sequentially perform feature extraction to obtain the second feature map F4 and input it into the second decoding unit and the third decoding unit;

[0073] Use the second cross-scale feature fusion module (MSFF module 2) in the second decoding unit to fuse the feature vector X Bo 、the second feature map F4, the feature vector and the feature vector After fusion, use two Swin Transformer modules in the second decoding unit to sequentially perform feature extraction to obtain the second feature map F5 and input it into the third decoding unit;

[0074] Use the third cross-scale feature fusion module ((MSFF module 3) in the third decoding unit to fuse the feature vector X Bo 、obtain the second feature map F4, the second feature map F5 and the feature vector After fusion, use two Swin Transformer modules in the third decoding unit to sequentially perform feature extraction to obtain the second feature map F6.

[0075] Specifically, the first cross-scale fusion module fuses the feature vector X Bo and the feature vector feature vector and the feature vector as follows:

[0076] Perform two upsamplings on the feature vector respectively, one upsampling on the feature vector , and one downsampling on the feature vector X Bo . Then, concatenate with the feature vector to obtain the first sub-feature map and send it to the fully connected layer to obtain the feature vector X Bo , the feature vector feature vector and the feature vector of the fused feature map.

[0077] The second cross-scale fusion module fuses the feature vector X Bo , the second feature map F4, the feature vector and the feature vector as follows:

[0078] Perform one downsampling on the feature vector respectively, one upsampling on the second feature map F4, and two upsamplings on the feature vector X Bo . Then, concatenate with the feature vector to obtain the second sub-feature map and send it to the fully connected layer to obtain the feature vector X Bo , the second feature map F4, the feature vector and the feature vector of the fused feature map.

[0079] The third cross-scale feature fusion module fuses the feature vector X Bo , obtains the second feature map F4, the second feature map F5 and the feature vector as follows:

[0080] Perform two downsamplings on the second feature map F4, one downsampling on the second feature map F5, and three upsamplings on the feature vector X Bo . Then, concatenate with the feature vector to obtain the third sub-feature map and send it to the fully connected layer to obtain the feature vector X Bo , obtains the second feature map F4, the second feature map F5 and the feature vector of the fused feature map.

[0081] It should be noted that the above inversion model is trained in a supervised learning manner based on training samples, labels, and a preset loss function; among them, each training sample includes sample lightning observation data and two sample infrared images from different infrared channels, and the label is the radar composite reflectivity image.

[0082] Optionally, the preset loss function is the weighted loss function L WSSIM = αL WMSE + βL mySSIM ;

[0083] Wherein,

[0084]

[0085] In the formula, y i represents the value of the position of the i-th pixel in the radar composite reflectivity image, w(y i ) represents the weight function dependent on y i , represents the value of the position of the i-th pixel in the current reconstructed image, SSIM represents the structural similarity index, which is used to measure the similarity between y i and , α represents the weight of the mean square error loss L WMSE , β represents the weight of the structural similarity loss L mySSIM , N represents the total number of pixels of the current reconstructed image obtained in each round of training. It should be understood that the total number of pixels of the two sample infrared images and the radar composite reflectivity image is also N.

[0086] The training process of the inversion model is introduced below.

[0087] Step 1: Select representative storm event images from the Storm EVent ImagRy (SEVIR) dataset. Select two infrared images and lightning observation data as the input sample images for each round of training of the evolution model. Among them, the infrared contains the infrared radiation information during the storm occurrence process, and the lightning observation data provides the spatio-temporal information related to the radar observation data. The two can effectively supplement the deficiencies of the radar data. In addition, select the radar composite reflectivity image as the label for the target of supervised learning in the evolution model training. The composite reflectivity image provides the reflectivity information observed by the radar and can be used as high-precision ground observation data for training the deep learning model. When processing the data, the input infrared images and lightning observation data are corresponded one by one with the radar composite reflectivity image to ensure the accurate pairing of the input and the label. At the same time, for factors such as different meteorological conditions and geographical locations, the dataset is preprocessed such as normalization and denoising to ensure that the model can learn effective features.

[0088] Step 1 includes the following steps:

[0089] (1) Save the SEVIR data in the HDF file in integer type. Decode the SEVIR data using linear scaling or exponential transformation, and the processing process is as follows: The infrared images from the two channels of IR107 and IR069 are converted to floating-point type using linear scaling; the lightning observation data lght is stored in matrix form and converted to the standard format after preprocessing; the radar composite reflectivity image is used as a label and stored in integer form, with the numerical range of 0 - 254, and 255 is used to represent the area where data is missing. Figure 6 It is an example diagram of a meteorological event in the dataset provided by the embodiment of the present invention.

[0090] (2) Select samples from the dataset with consistent time and region containing two infrared images from the IR069 and IR107 channels and the lightning observation data lght, and the radar composite reflectivity image is used as the label for model training. Ensure the consistency of sample time and space to ensure the matching degree of data.

[0091] (3) Flatten the two infrared images, the lightning observation data lght, and the radar composite reflectivity image into one-dimensional vectors for unified formatting processing.

[0092] (4) Read the radar composite reflectivity image, calculate the number of grid points in the radar composite reflectivity image that are all greater than 5 dbz, and retain the radar composite reflectivity image where the number of grid points meeting the conditions accounts for 1 / 20 of the total number of grid points.

[0093] (5) Perform data screening on the radar reflectivity, set the threshold to 0 - 75, set the grid point values less than 0 to 0, set the grid point values greater than 75 to 75, and delete the samples containing nan values in the data.

[0094] (6) Align the radar composite reflectivity image, the two infrared images, and the lightning observation data lght in time and space to ensure the consistency of different data sources during training.

[0095] (7) Divide the radar reflectivity inversion dataset after spatio-temporal alignment processing into training set, test set, and validation set to ensure the stability and generalization ability of model training and validation.

[0096] By performing unified formatting and spatio-temporal alignment processing on the composite reflectivity image, the two infrared images, and the lightning observation data lght, the consistency of multi-source data during training is ensured, the generalization ability and training efficiency of the model are improved, and the problems brought by data source heterogeneity are solved.

[0097] Step 2: Build an inversion model based on satellite infrared data and lightning observation data. The architecture of the inversion model consists of an S-T module and a multi-scale feature fusion module. Designing the multi-scale feature fusion module can fuse low-level and high-level feature information during the decoding process, thereby improving the model's learning ability for details and global information. Design the S-T module to extract multi-scale context information from the input satellite images and enhance the inversion model's perception ability for features at different scales.

[0098] For the radar reflectivity inversion task, the present invention also designs a weighted loss function L WSSIM = αL WMSE + βL mySSIM , where the weighted mean square error loss function L WMSE weights the errors in different regions to ensure that the scattering model can pay more attention to the inversion accuracy of important regions during training; the structural similarity loss L mySSIM is used to optimize the structural similarity of the inverted image and ensure that the inversion result is consistent with the real radar reflectivity image in structure.

[0099] Step 2 includes the following steps:

[0100] (1) Build a deep learning model based on the S-T module and the multi-scale feature fusion module. The inversion model consists of an encoder, a bottleneck layer, and a decoder. The encoder is responsible for extracting the features of the input image, the bottleneck layer is used to enhance the global information interaction of the features, and the decoder restores the spatial resolution of the image. The inversion model introduces the S-T module, which is a deep learning architecture that can effectively capture local and global features. The inversion model realizes the feature recovery from low resolution to high resolution through step-by-step downsampling and upsampling. Between the decoder and the encoder, a multi-scale feature fusion module is adopted to retain semantic information at different scales and ensure the effective fusion of multi-layer feature information.

[0101] (2) The input image is divided into non-overlapping image patches of a fixed size (for example, 16×16 or 32×32 pixels), and these image patches are independently fed into the encoder part for processing. This segmentation method helps to reduce memory consumption and can process each image patch in parallel, improving the calculation efficiency. In addition, the segmented image patches can reduce the redundancy of spatial information in the image, making the model more focused on the extraction of detail features.

[0102] (3) The encoder part uses three groups of S-T modules. Each group has a different window size and is used to extract context features of different scales of image patches, which are then sent to the multi-scale feature fusion module for processing. The window size and stride of each group are carefully designed to enhance the model's ability to perceive features at different scales. Through progressive downsampling, the encoder can not only extract local features but also capture global context information at a larger scale. This enables the model to flexibly handle weather systems of different sizes and complex meteorological phenomena;

[0103] (4) Each layer of the encoder's feature map is saved so that it can be used later in the decoder part. During the multi-scale feature fusion process, the low-level feature maps of the encoder can provide detailed information, while the high-level feature maps contain more global information. Saving each layer of the feature map helps combine this information during the decoder stage to ensure the consistency of the inverted image in terms of spatial resolution and semantic content;

[0104] (5) At the bottleneck layer, a group of S-T modules with a larger window size is used. The function of the bottleneck layer is to enhance the global information interaction between features. At this layer, a group of S-T modules with a larger window size is used to improve the ability to capture global context information by expanding the receptive field of the window. The larger window can span a wider area and capture the associated information of the global weather system, thereby improving the model's ability to handle large-scale meteorological events;

[0105] (6) The decoder part upsamples the image through three groups of symmetric S-T modules to restore the spatial resolution of the image. Corresponding to the downsampling process of the encoder, the upsampling process of the decoder gradually restores the resolution of the feature map to the size of the input image. Each S-T module restores higher-resolution details by refining the feature map while reducing information loss.

[0106] (7) To ensure that information at different scales can be fully fused in the decoder, two methods are used to adjust the sizes of outputs at different levels: merging adjacent image patches and the patch expansion layer. The purpose of merging adjacent image patches is to combine adjacent image patches of low-scale outputs, while the patch expansion layer is used to expand the patches of high-scale outputs to ensure the consistency of the input and output sizes of the decoder. Through these operations, it is ensured that the decoder can make full use of the encoder's information when processing high-resolution images.

[0107] (8) In the decoder, the outputs of the low scale and the high scale are merged to ensure the effective fusion of information at different scales. First, the outputs of the merged adjacent image patches and the patch expansion layer are adjusted to the same size as the decoder, and then they are merged with the encoder output of the same scale as the decoder. Through this merging operation, the model can capture different features at multiple scales, improving the accuracy of the inverted image;

[0108] (9) After the size adjustment is completed, the decoder merges the feature map of the encoder with the output map of the decoder to form the final skip connection. Through the skip connection, the detailed features of the low scale and the global information of the high scale are effectively retained, thus improving the reconstruction accuracy of the image. This step ensures the comprehensive fusion of multi-scale features, enabling the full utilization of semantic information at different levels;

[0109] (10) During the training process, we calculate the probability distribution based on the values of the tensor elements in the dataset and assign different weights to different regions. The method of setting the weights is based on the reciprocal of the region ratio, which means that elements in larger regions are given smaller weights, while elements in smaller regions are given larger weights. This strategy can help the model better focus on key regions when calculating the loss, improve the adaptability to data in different regions during the training process, and ultimately enhance the prediction accuracy of the model. After calculation, the weight sizes of the data in each region are shown in Table 1 below:

[0110] Table 1

[0111] Threshold interval 0-16 16-74 74-133 133-160 160-181 181-219 >219 Weight 1.3363 8.4795 10.7041 45.0041 117.9627 133.3007 476.5719

[0112] Step 3: Train on the unordered SEVIR storm event dataset to optimize the parameters of the scattering model. During the training process, the input data is processed in batches to ensure that each update can effectively learn the features in the data. During the training process, hyperparameters such as the learning rate, batch size, and number of training epochs are adjusted to optimize the performance of the model. A validation set can be used to monitor the performance of the model to avoid overfitting. Through the trained model, high-resolution radar reflectivity inversion products are generated. These inversion results will be used as the model output, which can provide higher spatio-temporal resolution than traditional radar observations and have important application value especially in areas with radar coverage gaps (such as the open sea, deserts, and plateau regions). The inversion results are evaluated using real radar reflectivity images, and the errors between the model inversion results and the real data are compared to further optimize the model and ensure its reliability and accuracy in practical applications.

[0113] Specifically, Step 3 includes:

[0114] (1) Initialize the parameters of the inversion model to ensure that the model can converge stably at the beginning of training.

[0115] (2) Set the batch size to 32, which means that 32 samples will be randomly selected from the training dataset for processing each time during training, thereby improving the computational efficiency and preventing overfitting.

[0116] (3) Set the learning rate lr to 0.0001, which is used to control the magnitude of each gradient update and avoid unstable parameter updates during the model training process due to overly large updates.

[0117] (4) During the training process, two loss functions are used: WMSE (Weighted Mean Square Error Loss) and WSSIM (Weighted Structural Similarity Loss) to optimize the model. WMSE can handle the distribution differences between regions by weighting the errors in different regions, while WSSIM helps to improve the structural similarity of the inverted images. The training process will be carried out for 200 rounds, and the model parameters will be updated after each round of training. Finally, the trained model parameters will be saved for subsequent use.

[0118] (5) Load the model parameters obtained from the previous training. Loading the model parameters means restoring the weights and biases of the model from the saved file to ensure that the inversion model can continue to use the knowledge during the training process in the subsequent stage. The loading process includes reading the saved model file and applying it to the current model architecture.

[0119] (6) Load the validation set data and input it into the trained inversion model to obtain the inversion results. In this stage, we do not update the model parameters but use the validation set to test the performance of the current model to ensure that the model can effectively invert the radar data. The results of the validation set will help evaluate the performance of the model under different meteorological scenarios.

[0120] (7) Adjust the hyperparameters according to the performance of the validation set to optimize the model performance. By adjusting the hyperparameters, we can improve the performance of the model on different data and reduce the risk of overfitting or underfitting. Repeat step 3 until the inversion values of the model reach a relatively stable state and save the final model parameters. If the inversion results of the model are relatively stable and the loss of the validation set gradually converges, it indicates that the hyperparameters have been tuned to an appropriate state. After that, we save the final model parameters for subsequent use.

[0121] (8) After completing the hyperparameter adjustment, we load the best model parameters obtained previously. This is to ensure that the optimized and better-performing model is used in the testing stage and to avoid using untuned model parameters.

[0122] (9)Finally, load the test set and input it into the tuned and validated inversion model. The test set is used to evaluate the final performance of the model and usually does not participate in the model training or validation process. By performing inference on the test set, we can evaluate the performance of the model in actual scenarios and ensure that it can provide stable and accurate inversion results under different data and different environments. The performance on the test set ultimately determines whether the model can be applied to an actual meteorological disaster warning system.

[0123] As can be seen from the above embodiments, the beneficial effects of the present invention are as follows:

[0124] (1) In the radar reflectivity inversion method provided by the present invention, a three-channel input image formed by two infrared images and lightning observation data is used as the input of the scattering model. Since lightning activity is closely related to severe convective weather, it can help the scattering model invert the radar reflectivity in areas lacking radar data, such as the open sea, plateau, and desert. This not only makes up for the lack of accurate modeling of severe weather events in the prior art but also provides more comprehensive and multi-dimensional input information for radar data reconstruction, significantly improving the accuracy of radar reflectivity inversion.

[0125] (2) For the scattering model, the present invention introduces the Swin Transformer module in each encoding unit of the encoder, the bottleneck layer, and each decoding unit of the decoder, which better preserves the spatial details and global information of the image. The cross-scale feature fusion module can effectively fuse feature information at different levels. Therefore, the present invention enhances the ability of the scattering model to perceive features at different scales and levels, ensuring that the inversion model takes into account both local details and global context during the inversion process, breaking through the limitation of losing global perception during the improvement of spatial resolution in the prior art, and improving the accuracy and detail performance of the inversion results.

[0126] (3) For the training process of the inversion model, the present invention designs a weighted loss function L WSSIM , assigns different weights to different regions and weather phenomena, thereby specifically optimizing the radar reflectivity inversion effect under complex weather conditions, especially performing well in the monitoring of extreme weather such as severe convective weather and typhoons.

[0127] In the description of the present invention, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0128] The description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0129] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A radar reflectivity inversion method based on satellite infrared data and lightning observation data, characterized in that: include: Obtain lightning observation data and two infrared images from different infrared channels to form a three-channel input image; The input image is input into the inversion model, so that the inversion model performs the following steps: Extracting features from the input image, and mapping the extracted features into a one-dimensional feature vector using a mapping unit; The one-dimensional feature vector is input into the encoder, so that each encoding unit connected in sequence in the encoder uses two Swin Transformer modules to extract features from its own input data in sequence, and obtains the feature vector X En , and the feature vector X En Send to at least one decoding unit in the upsampling module and the decoder, and further use the upsampling layer to process the feature vector X En Perform upsampling processing to obtain a first feature map; The first feature map output by the last level encoding unit is input into the bottleneck layer for feature extraction to obtain the feature vector X Bo Then output each decoding unit in the decoder; The feature vectors are respectively input into the decoding units connected in sequence in the decoder, so that each of the decoding units can generate a feature vector X based on the input feature vector X. En and the eigenvector X Bo Perform cross-scale feature fusion, and then use two SwinTransformer modules to extract features from the fused feature maps in turn to obtain the second feature map; The second feature map output by the last level decoding unit is downsampled by a downsampling unit to obtain a reconstructed image, wherein each pixel in the reconstructed image represents a radar reflectivity value at a corresponding radar site.

2. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 1 is characterized in that: The encoder comprises a first encoding unit, a second encoding unit and a third encoding unit connected in sequence, and the decoder comprises a first decoding unit, a second decoding unit and a third decoding unit connected in sequence; wherein, The one-dimensional feature vector is input into the encoder, so that each encoding unit connected in sequence in the encoder uses two Swin Transformer modules to extract features from its own input data in sequence, and obtains the feature vector X En , and the feature vector X En Send to at least one decoding unit in the upsampling module and the decoder, and further use the upsampling layer to process the feature vector X En The step of performing upsampling processing to obtain a first feature map includes: The one-dimensional feature vector is input into the first coding unit, and the two SwinTransformer modules in the first coding unit are used to extract features of the one-dimensional feature vector in turn to obtain a feature vector and input the up-sampling module in the first encoding unit, the first decoding unit, the second decoding unit and the third decoding unit respectively; The feature vector is coded by using an upsampling module in the first coding unit. Perform upsampling processing to obtain a first feature map F1 and input it into the second encoding unit; The two Swin Transformer modules in the second encoding unit are used to extract features from the first feature map F1 in turn to obtain a feature vector and input the up-sampling module in the second encoding unit, the first decoding unit and the second decoding unit respectively; The feature vector is coded by using an upsampling module in the second coding unit. Perform upsampling processing to obtain a first feature map F2 and input it into the third encoding unit; The two Swin Transformer modules in the third encoding unit are used to extract features from the first feature map F2 in turn to obtain a feature vector and input into the up-sampling module in the third encoding unit and the first decoding unit; The feature vector is encoded by using an upsampling module in the third encoding unit. An upsampling process is performed to obtain a first feature map F3 and input it into the bottleneck layer.

3. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 2 is characterized in that: The bottleneck layer includes two Swin Transformer modules connected in sequence.

4. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 3 is characterized in that: The feature vectors are respectively input into the decoding units connected in sequence in the decoder, so that each of the decoding units can generate a feature vector X based on the input feature vector X. En and the eigenvector X Bo The step of performing cross-scale feature fusion and then extracting features from the fused feature maps in sequence using two Swin Transformer modules to obtain a second feature map includes: The feature vector X Bo Input the first decoding unit, and use the first cross-scale feature fusion module in the first decoding unit to fusion the feature vector X Bo And the feature vector The feature vector and the eigenvector After fusion, the two Swin Transformer modules in the first decoding unit are used to extract features in sequence to obtain a second feature map F4 and input it into the second decoding unit and the third decoding unit; The feature vector X is fused using the second cross-scale feature fusion module in the second decoding unit. Bo , the second feature map F4, the feature vector and the eigenvector After fusion, the two Swin Transformer modules in the second decoding unit are used to extract features in sequence to obtain a second feature map F5 and input it into the third decoding unit; The feature vector X is fused using the third cross-scale feature fusion module in the third decoding unit. Bo , the second feature map F4, the second feature map F5 and the feature vector After fusion, the two Swin Transformer modules in the third decoding unit are used to extract features in sequence to obtain a second feature map F6.

5. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 4 is characterized in that: The first cross-scale fusion module performs the following steps on the feature vector X Bo And the feature vector The feature vector and the eigenvector To perform the fusion: The feature vectors Perform two upsamplings and transform the feature vector Perform an upsampling and resample the feature vector X Bo After downsampling once, the feature vector Concatenate to get the first sub-feature map and send it to the fully connected layer to get the feature vector X Bo , the feature vector The feature vector and the eigenvector The fused feature map.

6. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 4 is characterized in that: The second cross-scale fusion module performs the following steps on the feature vector X Bo , the second feature map F4, the feature vector and the eigenvector To perform the fusion: The feature vectors Down-sampling is performed once, up-sampling is performed once on the second feature map F4, and the feature vector X Bo After two upsamplings, the feature vector Concatenate to get the second sub-feature map and send it to the fully connected layer to get the feature vector X Bo , the second feature map F4, the feature vector and the eigenvector The fused feature map.

7. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 4 is characterized in that: The third cross-scale feature fusion module performs the following steps on the feature vector X Bo , the second feature map F4, the second feature map F5 and the feature vector To perform the fusion: The second feature map F4 is downsampled twice, the second feature map F5 is downsampled once, and the feature vector X is downsampled once. Bo After three upsamplings, the feature vector Concatenate to get the third sub-feature map and send it to the fully connected layer to get the feature vector X Bo , the second feature map F4, the second feature map F5 and the feature vector The fused feature map.

8. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 1 is characterized in that: The inversion model is trained and obtained in a supervised learning manner based on training samples, labels and a preset loss function; wherein each of the training samples includes sample lightning observation data and two sample infrared images from different infrared channels, and the label is a radar composite reflectivity image.

9. The radar reflectivity inversion method based on satellite infrared data and lightning observation data according to claim 8, characterized in that: The preset loss function is a weighted loss function L WSSIM =αL WMSE +βL mySSIM ; in, Where N represents the total number of pixels of the current reconstructed image obtained in each round of training, and y i represents the value of the position of the i-th pixel in the radar composite reflectivity image, w(y i ) indicates that it depends on y i The weight function of represents the value of the i-th pixel position in the current reconstructed image, and SSIM represents the structural similarity index, which is used to measure y i and The similarity between them, α represents the mean square error loss L WMSE The weight of β represents the structural similarity loss L mySSIM The weight of .

Images