A radar reflectivity inversion method based on satellite infrared data and lightning observation data

Abstract

The application discloses a radar reflectivity inversion method based on satellite infrared data and lightning observation data, and belongs to the technical field of meteorological monitoring. The three-channel input image formed by lightning observation data and two infrared images from different infrared channels is taken as the input of the scattering model. Since lightning activity is closely related to severe convective weather, the scattering model can help to invert radar reflectivity in areas where radar data is missing, such as the open sea, plateau and desert. The application not only makes up for the insufficient accurate modeling of severe weather events in the prior art, but also provides more comprehensive and more dimensional input information for radar data reconstruction, and significantly improves the accuracy of radar reflectivity inversion.

Description

Technical Field

[0001] This invention belongs to the field of meteorological monitoring technology, specifically relating to a radar reflectivity inversion method based on satellite infrared data and lightning observation data. Background Technology

[0002] With meteorology deeply integrated into economic and social development, frequent severe weather events pose a serious threat to property and life, drawing widespread attention from the public and various industries. China has a vast territory with diverse climates and topography, resulting in a wide variety of severe weather events. Severe weather events generated by atmospheric convection, such as thunderstorms, squall lines, and hail, exhibit highly nonlinear structures. These events are short-lived, sudden, and difficult to forecast, easily triggering various meteorological disasters. Accurate monitoring and identification of the formation and development of severe weather events are crucial for improving meteorological disaster prevention capabilities and reducing economic and loss of life.

[0003] With the continuous increase in meteorological observation data and the rapid development of artificial intelligence technology, data-driven deep learning methods have become a research hotspot in the meteorological field. Compared with the manual feature extraction of traditional machine learning, deep learning methods can automatically extract useful features from large amounts of data and learn the patterns within them, thereby improving the predictive performance of the model. Currently, deep learning methods can accomplish many tasks in the meteorological field, such as using multi-source meteorological data such as satellite data and radar data to invert precipitation, temperature, and wind fields, and to identify and forecast severe convective weather. However, due to factors such as radar wave attenuation, terrain obstruction, and incomplete radar network coverage in some areas, some radar data is missing, limiting its usability and reference value.

[0004] Currently, existing radar reflectivity inversion methods include traditional methods based on non-radar data and deep learning-based methods. Traditional methods based on non-radar data are susceptible to noise and incomplete data during data processing and conversion, resulting in low accuracy of the reconstructed radar reflectivity data, which introduces significant uncertainty in weather disaster early warning. Furthermore, deep learning-based radar reflectivity inversion methods typically reconstruct high-resolution images through downsampling and upsampling operations. However, because spatial resolution is lost during downsampling, subsequent image detail recovery often fails to fully preserve global contextual information, leading to insufficient spatial sensitivity of the high-resolution image and consequently low accuracy of the inverted radar reflectivity. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a radar reflectivity inversion method based on satellite infrared data and lightning observation data. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] This invention provides a radar reflectivity inversion method based on satellite infrared data and lightning observation data, comprising:

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

[0008] The input image is input into the inversion model, causing the inversion model to perform the following steps:

[0009] Feature extraction is performed on the input image, and the extracted features are mapped into a one-dimensional feature vector using a mapping unit;

[0010] The one-dimensional feature vector is input into the encoder, and each encoding unit connected in sequence in the encoder uses two Swing Transformer modules to extract features from its own input data to obtain the feature vector X. En and the feature vector X En The feature vector X is sent to at least one decoding unit in the upsampling module and the decoder, and further processed by the upsampling layer. En Upsampling is performed to obtain the first feature map;

[0011] The first feature map output from the last-stage encoding unit is input into the bottleneck layer for feature extraction, resulting in the feature vector X. Bo The decoding units in the output decoder are then output.

[0012] The feature vector is input into each of the sequentially connected decoding units in the decoder, such that each decoding unit is based on the input feature vector X. En and the feature vector X Bo Cross-scale feature fusion is performed, and then two SwinTransformer modules are used to extract features from the fused feature map in sequence to obtain the second feature map;

[0013] The second feature map output by the last-stage decoding unit is downsampled using the downsampling unit to obtain a reconstructed image. Each pixel in the reconstructed image represents the radar reflectivity value at the corresponding radar site.

[0014] In one 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] The one-dimensional feature vector is input into the encoder, and each encoding unit connected in sequence in the encoder uses two Swing Transformer modules to extract features from its own input data to obtain the feature vector X. En and the feature vector X En The feature vector X is sent to at least one decoding unit in the upsampling module and the decoder, and further processed by the upsampling layer. En The steps for upsampling to obtain the first feature map include:

[0016] The one-dimensional feature vector is input into the first encoding unit, and the two SwinTransformer modules in the first encoding unit are used to extract features from the one-dimensional feature vector in sequence to obtain the feature vector. And respectively input to the upsampling module in the first encoding unit, the first decoding unit, the second decoding unit and the third decoding unit;

[0017] The feature vector is processed using the upsampling module in the first encoding unit. Upsampling is performed to obtain the first feature map F1, which is then input into the second encoding unit.

[0018] The first feature map F1 is extracted sequentially using two Swing Transformer modules in the second encoding unit to obtain feature vectors. And respectively input to the upsampling module in the second encoding unit, the first decoding unit and the second decoding unit;

[0019] The feature vector is processed using the upsampling module in the second encoding unit. Upsampling is performed to obtain the first feature map F2, which is then input into the third encoding unit.

[0020] The first feature map F2 is extracted sequentially using two Swing Transformer modules in the third encoding unit to obtain feature vectors. And input to the upsampling module in the third encoding unit and the first decoding unit;

[0021] The feature vector is processed using the upsampling module in the third encoding unit. Upsampling is performed to obtain the first feature map F3, which is then input into the bottleneck layer.

[0022] In one embodiment of the present invention, the bottleneck layer includes two sequentially connected Swing Transformer modules.

[0023] In one embodiment of the present invention, the feature vector is input into each decoding unit connected sequentially in the decoder, such that each decoding unit is based on the input feature vector X. En and the feature vector X Bo The steps for performing cross-scale feature fusion, and then using two Swin Transformer modules to sequentially extract features from the fused feature map to obtain the second feature map include:

[0024] The feature vector X Bo Input to 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, the two Swing Transformer modules in the first decoding unit are used to extract features sequentially to obtain the second feature map F4, which is then input into the second decoding unit and the third decoding unit.

[0025] The feature vector X is processed using the second cross-scale feature fusion module in the second decoding unit. Bo The second feature map F4, the feature vector and the feature vector After fusion, the two Swing Transformer modules in the second decoding unit are used to extract features sequentially to obtain the second feature map F5, which is then input into the third decoding unit.

[0026] The feature vector X is processed 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 are obtained. After fusion, the two Swing Transformer modules in the third decoding unit are used to extract features sequentially to obtain the second feature map F6.

[0027] In one embodiment of the present invention, the first cross-scale fusion module processes the feature vector X according to the following steps. Bo and the feature vector The feature vector and the feature vector To merge:

[0028] For the feature vector respectively Perform two upsampling operations on the feature vector Perform one upsampling operation on the feature vector X BoAfter performing one downsampling, and with the feature vector The features are concatenated to obtain the first sub-feature map, which is then fed into a fully connected layer to obtain the feature vector X. Bo The feature vector The feature vector and the feature vector The fused feature map.

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

[0030] For the feature vector respectively Perform a downsampling operation on the second feature map F4, perform an upsampling operation on the feature vector X. Bo After two upsampling operations, the feature vector is compared with the feature vector. The features are concatenated to obtain the second sub-feature map, which is then fed into a fully connected layer to obtain the feature vector X. Bo The second feature map F4, the feature vector and the feature vector The fused feature map.

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

[0032] 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 upsampling operations, the feature vector is compared with the feature vector. The features are concatenated to obtain the third sub-feature map, which is then fed into a fully connected layer to obtain the feature vector X. Bo The second feature map F4, the second feature map F5, and the feature vector are obtained. The 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 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] in,

[0036]

[0037] In the formula, N represents the total number of pixels in the current reconstructed image obtained in each round of training, and y i w(y) represents the value of the position of the i-th pixel in the radar composite reflectivity image. i ) indicates that it depends on y i The weight function, This represents the value at the i-th pixel position in the currently reconstructed image, where SSIM stands for Structural Similarity Index, used to measure y. i and The similarity between them, where α represents the mean squared error loss L WMSE The weights, β, represent the structural similarity loss L. mySSIM The weight.

[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, 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 to invert radar reflectivity in areas where radar data is missing, 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 existing technology, but also provides more comprehensive and multi-dimensional input information for radar data reconstruction, which significantly improves the accuracy of radar reflectivity inversion.

[0041] (2) For the scattering model, this invention introduces the Swing 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, this invention enhances the ability of the scattering model to perceive features at different scales and levels, ensuring that the inversion model considers both local details and global context during the inversion process. This overcomes the limitation of existing technologies in losing global perception during spatial resolution improvement, and improves the accuracy and detail representation of the inversion results.

[0042] (3) For the training process of the inversion model, this invention designs a weighted loss function L WSSIM Different weights are assigned to different regions and weather phenomena, thereby optimizing the radar reflectivity inversion effect under complex weather conditions, especially in the monitoring of severe convective weather, typhoons and other extreme weather.

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0044] Figure 1 This is a flowchart of a radar reflectivity inversion method based on satellite infrared data and lightning observation data provided in this embodiment of the invention;

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

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

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

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

[0049] Figure 6 This is an example diagram of a meteorological event in the dataset provided in this embodiment of the invention. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0051] Figure 1 This is a flowchart of a radar reflectivity inversion method based on satellite infrared data and lightning observation data provided in this invention. Figure 2 This is a schematic diagram of the inversion model provided in an embodiment of the present invention. Please refer to [link / reference]. Figure 1-2 This invention provides a radar reflectivity inversion method based on satellite infrared data and lightning observation data, comprising:

[0052] S1. Acquire 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 coding unit connected in sequence in the encoder uses two Swing Transformer modules (ST modules) to extract features from its own input data in turn, and obtain the feature vector X. En and the feature vector X En The data is sent to at least one decoding unit in the upsampling module and the decoder, and the upsampling layer is used to further process the feature vector X. En Upsampling is performed to obtain the first feature map;

[0056] S203. Input the first feature map output by the last-level coding unit into the bottleneck layer for feature extraction to obtain the feature vector X. Bo The decoding units in the output decoder are then output.

[0057] S204. Input the feature vector into each of the sequentially connected decoding units in the decoder, so that each decoding unit is based on the input feature vector X. En and eigenvector X Bo Cross-scale feature fusion is performed, and then two Swin Transformer modules are used to extract features from the fused feature map in sequence to obtain the second feature map.

[0058] S205. The second feature map output by the last-stage decoding unit is downsampled using the downsampling unit to obtain the reconstructed image. Each pixel in the reconstructed image represents the radar reflectivity value at the corresponding radar station.

[0059] It should be noted that the two infrared images acquired in step S1 are from two different infrared channels: IR107 (10.7 μm) and IR069 (6.9 μm). These are two commonly used infrared channels in meteorological satellites, providing information at different altitudes and temperature levels. The IR107 channel primarily reflects cloud top temperature and can be used to identify convective clouds and cloud top cooling, while the IR069 channel primarily detects mid-level water vapor, aiding in the analysis of humidity distribution and convective activity in the mid-atmosphere.

[0060] These two infrared images are satellite observations from the same time and region. The IR107 channel focuses on cloud top features, while the IR069 channel reflects mid-level water vapor content. Combining them improves the ability to identify deep convection and precipitation systems, providing richer input features for radar reflectivity reconstruction. Therefore, this embodiment, by combining information from different channels, can more comprehensively characterize the atmospheric state.

[0061] This invention incorporates lightning observation data into the radar reflectivity inversion process. Particularly under severe convective weather and extreme climate conditions, the inclusion of lightning activity significantly enhances the model's predictive ability for high-threshold weather events. Lightning observation data provides crucial clues about severe convective activity in meteorological systems, especially in areas not covered by radar data (such as the open sea, plateaus, and deserts). Therefore, it can improve the accuracy of radar reflectivity inversion and contribute to enhancing the timeliness and accuracy of weather 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, so that each coding unit connected in sequence in the encoder uses two Swing Transformer modules to extract features from its own input data in turn, thereby obtaining the feature vector X. En and the feature vector X En The data is sent to at least one decoding unit in the upsampling module and the decoder, and the upsampling layer is used to further process the feature vector X. En The steps for upsampling to obtain the first feature map include:

[0064] A one-dimensional feature vector is input into the first encoding unit, and the two SwinTransformer modules in the first encoding unit are used to extract features from the one-dimensional feature vector in turn to obtain the feature vector. And respectively input to the upsampling module, the first decoding unit, the second decoding unit and the third decoding unit in the first encoding unit;

[0065] The feature vector is processed using the upsampling module in the first coding unit. Upsampling is performed to obtain the first feature map F1, which is then input into the second encoding unit.

[0066] The first feature map F1 is extracted sequentially using two Swing Transformer modules in the second coding unit to obtain the feature vector. And respectively input into the upsampling module, the first decoding unit and the second decoding unit in the second encoding unit;

[0067] The feature vector is processed using the upsampling module in the second coding unit. Upsampling is performed to obtain the first feature map F2, which is then input into the third encoding unit.

[0068] The first feature map F2 is extracted sequentially using two Swing Transformer modules in the third coding unit to obtain the feature vector. And input to the upsampling module in the third encoding unit and the first decoding unit;

[0069] The feature vector is processed using the upsampling module in the third coding unit. Upsampling is performed to obtain the first feature map F3, which is then input into the bottleneck layer.

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

[0071] In step S204, the feature vector is input into each decoding unit connected sequentially in the decoder, so that each decoding unit is based on the input feature vector X. En and eigenvector X Bo The steps for performing cross-scale feature fusion, and then using two SwinTransformer modules to sequentially extract features from the fused feature map to obtain the second feature map include:

[0072] The feature vector X Bo Input to the first decoding unit, and use the first cross-scale feature fusion module (MSFF module 1) in the first decoding unit to process the feature vector X. Bo and eigenvectors Feature vector and eigenvectors After fusion, the two Swing Transformer modules in the first decoding unit are used to extract features in sequence to obtain the second feature map F4, which is then input into the second and third decoding units.

[0073] The second cross-scale feature fusion module (MSFF module 2) in the second decoding unit is used to process the feature vector X. Bo Second feature map F4, feature vector and eigenvectors After fusion, the two SwinTransformer modules in the second decoding unit are used to extract features sequentially to obtain the second feature map F5, which is then input into the third decoding unit.

[0074] The third cross-scale feature fusion module (MSFF module 3) in the third decoding unit is used to process the feature vector X. Bo We obtain the second feature map F4, the second feature map F5, and the feature vectors. After fusion, the two Swing Transformer modules in the third decoding unit are used to extract features sequentially to obtain the second feature map F6.

[0075] Specifically, the first cross-scale fusion module processes the feature vector X according to the following steps. Bo and eigenvectors Feature vector and eigenvectors To merge:

[0076] For the feature vectors respectively Perform two upsampling operations on the feature vector Perform one upsampling operation on the feature vector X Bo After one downsampling, and with the feature vector The features are concatenated to obtain the first sub-feature map, which is then fed into a fully connected layer to obtain the feature vector X. Bo eigenvectors Feature vector and eigenvectors The fused feature map.

[0077] The second cross-scale fusion module processes the feature vector X according to the following steps. Bo Second feature map F4, feature vector and eigenvectors To merge:

[0078] For the feature vectors respectively Perform one downsampling, perform one upsampling on the second feature map F4, and perform one upsampling on the feature vector X. Bo After two upsampling operations, and with the feature vector The two sub-feature maps are concatenated to obtain the second sub-feature map, which is then fed into a fully connected layer to obtain the feature vector X. Bo Second feature map F4, feature vector and eigenvectors The fused feature map.

[0079] The third cross-scale feature fusion module processes the feature vector X according to the following steps. Bo We obtain the second feature map F4, the second feature map F5, and the feature vectors. To merge:

[0080] 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 upsampling operations, and with the feature vector The features are concatenated to obtain the third sub-feature map, which is then fed into a fully connected layer to obtain the feature vector X. Bo We obtain the second feature map F4, the second feature map F5, and the feature vectors. 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; each training sample includes sample lightning observation data and two sample infrared images from different infrared channels, labeled as radar composite reflectivity images.

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

[0083] in,

[0084]

[0085] In the formula, y i w(y) represents the value of the position of the i-th pixel in the radar composite reflectivity image. i ) indicates that it depends on y i The weight function, This represents the value at the i-th pixel position in the currently reconstructed image, where SSIM stands for Structural Similarity Index, used to measure y. i and The similarity between them, where α represents the mean squared error loss L WMSE The weights, β, represent the structural similarity loss L. mySSIM The weights are N, where N represents the total number of pixels in the current reconstructed image obtained in each round of training. It should be understood that the total number of pixels in the two sample infrared images and the radar composite reflectivity image is also N.

[0086] The training process of the inversion model will be described below.

[0087] Step 1: Select representative storm event images from the Storm EVent ImagRy (SEVIR) dataset. Two infrared images and lightning observation data are selected as input sample images for each round of training of the evolution model. Infrared images contain infrared radiation information during the storm's occurrence, while lightning observation data provides spatiotemporal information related to radar observation data, effectively supplementing the deficiencies of radar data. Furthermore, radar composite reflectivity images are selected as labels for supervised learning in the evolution model training. Composite reflectivity images provide reflectivity information observed by radar, serving as high-precision ground observation data for training the deep learning model. During data processing, the input infrared images and lightning observation data are matched one-to-one with the radar composite reflectivity images to ensure accurate input-label pairing. Simultaneously, the dataset undergoes preprocessing such as normalization and denoising to address different meteorological conditions and geographical locations, ensuring the model can learn effective features.

[0088] Step 1 includes the following steps:

[0089] (1) Save the SEVIR data as integers in the HDF file. Decode the SEVIR data using linear scaling or exponential transformation. The processing steps are as follows: Infrared images from the IR107 and IR069 channels are converted to floating-point format using linear scaling; lightning observation data (lght) is stored in matrix form and converted to standard format after preprocessing; radar composite reflectivity images are used as labels and stored in integer form with a value range of 0-254, and 255 is used to represent areas with missing data. Figure 6 This is an example diagram of a meteorological event in the dataset provided in this embodiment of the invention.

[0090] (2) Select samples from the dataset that are consistent in time and region, including two infrared images from the IR069 and IR107 channels and lightning observation data (lght). Use radar composite reflectivity images as labels for model training. Ensure temporal and spatial consistency of the samples and ensure data matching.

[0091] (3) Flatten the two infrared images, lightning observation data lght and radar composite reflectivity image into a one-dimensional vector and perform uniform formatting.

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

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

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

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

[0096] By uniformly formatting and spatiotemporally aligning the composite reflectance image, two infrared images, and lightning observation data (lght), the consistency of multi-source data during training is ensured, improving the model's generalization ability and training efficiency, and solving the problems caused by heterogeneity of data sources.

[0097] Step 2: Construct an inversion model based on satellite infrared data and lightning observation data. The inversion model architecture consists of an ST module and a multi-scale feature fusion module. The multi-scale feature fusion module is designed to fuse low-level and high-level feature information during the decoding process, thereby improving the model's ability to learn detailed and global information. The ST module is designed to extract multi-scale contextual information from the input satellite images, enhancing the inversion model's ability to perceive features at different scales.

[0098] For radar reflectivity inversion tasks, this invention also designs a weighted loss function L WSSIM =αL WMSE +βL mySSIM Among them, the weighted mean square error loss function L WMSE By weighting the errors in different regions, it is ensured that the scattering model pays more attention to the inversion accuracy of important regions during training; structural similarity loss L mySSIM This is used to optimize the structural similarity of the inverted image, ensuring that the inverted result is structurally consistent with the real radar reflectivity image.

[0099] Step 2 includes the following steps:

[0100] (1) Construct a deep learning model based on the ST module and a multi-scale feature fusion module. The inversion model consists of an encoder, a bottleneck layer, and a decoder. The encoder is responsible for extracting features from the input image, the bottleneck layer is used to enhance the global information interaction of features, and the decoder restores the spatial resolution of the image. The inversion model introduces the ST module, which is a deep learning architecture that effectively captures both local and global features. The inversion model achieves feature recovery from low resolution to high resolution through progressive downsampling and upsampling. Between the decoder and encoder, a multi-scale feature fusion module is used to preserve semantic information at different scales, ensuring the effective fusion of multi-layer feature information.

[0101] (2) The input image is segmented into fixed-size, non-overlapping image patches (e.g., 16×16 or 32×32 pixels), which are then fed independently into the encoder for processing. This segmentation method helps reduce memory consumption and allows for parallel processing of each image patch, improving computational efficiency. Furthermore, the segmented image patches reduce spatial redundancy in the image, allowing the model to focus more on extracting detailed features.

[0102] (3) The encoder uses three sets of ST modules, each with a different window size, to extract contextual features at different scales of image patches, which are then fed into the multi-scale feature fusion module for processing. The window size and stride of each module 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 contextual information at a larger scale. This allows the model to flexibly cope with weather systems of different sizes and complex meteorological phenomena;

[0103] (4) Save the feature maps of each layer of the encoder. Each layer of the encoder's feature maps will be saved for later use in the decoder. During multi-scale feature fusion, the low-level feature maps of the encoder can provide detailed information, while the high-level feature maps contain more global information. Saving the feature maps of each layer helps to combine this information in the decoder stage, ensuring the consistency of the inverted image in terms of spatial resolution and semantic content;

[0104] (5) At the bottleneck layer, a set of ST modules with larger window sizes is used. The function of the bottleneck layer is to enhance the global information interaction between features. At this layer, a set of ST modules with larger window sizes are used to improve the ability to capture global contextual information by expanding the receptive field of the window. Larger windows can span a wider area and capture the correlation information of the global weather system, thereby improving the model's ability to handle large-scale meteorological events;

[0105] (6) The decoder upsamples the image using three sets of symmetrical ST modules to restore the spatial resolution of the image. Corresponding to the encoder's downsampling process, the decoder's upsampling process gradually restores the resolution of the feature map to the size of the input image. Each ST module recovers higher resolution details by refining the feature map while reducing information loss.

[0106] (7) To ensure that information at different scales is fully integrated in the decoder, two methods are used to adjust the size of the output at different levels: merging adjacent image patches and patch expansion layers. The purpose of merging adjacent image patches is to combine adjacent image patches of low-scale output, while the patch expansion layer is used to expand the patches of high-scale output to ensure that the input and output sizes of the decoder are consistent. 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, low-scale and high-scale outputs are merged to ensure effective fusion of information at different scales. First, the outputs of merging adjacent image patches and patch extension layers are adjusted to the same size as the decoder, and then merged with the encoder output at 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 resizing, the decoder merges the encoder's feature map with the decoder's output map to form the final skip connection. Through skip connections, low-scale detail features and high-scale global information are effectively preserved, thereby improving the image reconstruction accuracy. This step ensures the comprehensive fusion of multi-scale features, making full use of semantic information at different levels;

[0109] (10) During training, we calculate the probability distribution based on the values ​​of the tensor elements in the dataset and assign different weights to different regions. The weights are set according to the reciprocal of the region proportion, meaning that elements in larger regions are assigned smaller weights, while elements in smaller regions are assigned larger weights. This strategy helps the model focus more on key regions when calculating the loss, improves the adaptability to different regions of data during training, and ultimately improves the model's prediction accuracy. The calculated weights of the data in each region are shown in Table 1 below:

[0110] Table 1

[0111] Threshold range 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 the model on an unordered SEVIR storm event dataset to optimize its parameters. During training, batch processing is used to process the input data, ensuring that each update effectively learns the features from the data. Hyperparameters such as learning rate, batch size, and number of training epochs are adjusted during training to optimize model performance. A validation set can be used to monitor model performance and avoid overfitting. The trained model generates high-resolution radar reflectivity inversion products. These inversion results will be used as model outputs, providing higher spatiotemporal resolution than traditional radar observations, and are particularly valuable in radar-covered areas (such as the open sea, deserts, and plateaus). The inversion results are evaluated using real radar reflectivity images, comparing the errors between the model's inversion results and the real data 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 each training session, thereby improving computational efficiency and preventing overfitting.

[0116] (3) The learning rate lr is set to 0.0001 to control the magnitude of each gradient update and avoid instability caused by excessive parameter updates during model training.

[0117] (4) During training, two loss functions are used to optimize the model: WMSE (Weighted Mean Squared Error Loss) and WSSIM (Weighted Structural Similarity Loss). WMSE can handle the distribution differences between regions by weighting the errors of different regions, while WSSIM helps to improve the structural similarity of the inverted image. The training process will run 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 trained in the previous step. Loading the model parameters means restoring the model's weights and biases from the saved file, ensuring that the inverted model can continue to use the knowledge from the training process in subsequent stages. 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 will not update the model parameters, but instead use the validation set to test the performance of the current model, ensuring that the model can effectively invert radar data. The results of the validation set will help evaluate the model's performance under different weather scenarios.

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

[0121] (8) After completing the hyperparameter tuning, we load the optimal model parameters obtained earlier. This is to ensure that the optimized and effective model is used during the testing phase, 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 is usually not involved in the model training or validation process. By performing inference on the test set, we can evaluate the model's performance in real-world scenarios, ensuring that it can provide stable and accurate inversion results under different data and environments. The performance on the test set ultimately determines whether the model can be applied to a real-world meteorological disaster early 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 to invert radar reflectivity in areas where radar data is missing, 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 existing technology, but also provides more comprehensive and multi-dimensional input information for radar data reconstruction, which significantly improves the accuracy of radar reflectivity inversion.

[0125] (2) For the scattering model, this invention introduces the Swing 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, this invention enhances the ability of the scattering model to perceive features at different scales and levels, ensuring that the inversion model considers both local details and global context during the inversion process. This overcomes the limitation of existing technologies in losing global perception during spatial resolution improvement, and improves the accuracy and detail representation of the inversion results.

[0126] (3) For the training process of the inversion model, this invention designs a weighted loss function L WSSIM Different weights are assigned to different regions and weather phenomena, thereby optimizing the radar reflectivity inversion effect under complex weather conditions, especially in the monitoring of severe convective weather, typhoons and other extreme weather.

[0127] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0128] The use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" indicates that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0129] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A radar reflectivity retrieval method based on satellite infrared data and lightning observation data, characterized in that, The method comprises: obtaining lightning observation data and two infrared images from different infrared channels to form a three-channel input image; inputting the input image into an inversion model, so that the inversion model performs the following steps: performing feature extraction on the input image, and mapping the extracted features to a one-dimensional feature vector by using a mapping unit; The one-dimensional feature vector is input into an encoder, so that each encoding unit connected in sequence in the encoder uses two Swin Transformer modules to sequentially extract features of input data of the encoding unit, to obtain a feature vector X En The feature vector X En is sent to at least one decoding unit in a decoder and an upsampling module, and the feature vector X En is further subjected to upsampling processing by the upsampling module, to obtain a first feature map. The first feature map output by the last coding unit is input into the bottleneck layer for feature extraction to obtain a feature vector X Bo Each decoding unit in the post-output decoder The feature vectors are input into each decoding unit connected in sequence in the decoder respectively, so that each decoding unit performs feature extraction on the input feature vectors X En and the feature vectors X Bo Cross-scale feature fusion is performed, and then two SwinTransformer modules are used to sequentially perform feature extraction on the fused feature maps to obtain a second feature map. performing down-sampling processing on the second feature map output by the last decoding unit by using a down-sampling unit to obtain a reconstructed image, and each pixel point in the reconstructed image represents a radar reflectivity value at a corresponding radar station.

2. The method of radar reflectivity retrieval based on satellite infrared data and lightning observation data according to claim 1, 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 an encoder, so that each encoding unit connected in sequence in the encoder uses two Swin Transformer modules to sequentially extract features of the input data of itself to obtain a feature vector X En The feature vector X En is sent to at least one decoding unit in an upsampling module and a decoder, and the feature vector X En is further subjected to upsampling processing by the upsampling module to obtain a first feature map, comprising: The one-dimensional feature vector is input into a first encoding unit, two SwinTransformer modules in the first encoding unit are used to sequentially perform feature extraction on the one-dimensional feature vector, and a feature vector is obtained and is respectively input into an up-sampling module in the first encoding unit, the first decoding unit, the second decoding unit, and the third decoding unit. performing up-sampling on the feature vector in the first encoding unit by using the up-sampling module performing up-sampling to obtain a first feature map F1 and inputting the first feature map F1 into the second encoding unit The two Swin Transformer modules in the second encoding unit are used to sequentially perform feature extraction on the first feature map F1 to obtain a feature vector and are respectively input into an up-sampling module in the second encoding unit, the first decoding unit and the second decoding unit. performing up-sampling on the feature vector by using an up-sampling module in the second encoding unit performing up-sampling to obtain a first feature map F2 and inputting the first feature map F2 into the third encoding unit The two Swin Transformer modules in the third encoding unit are used to sequentially perform feature extraction on the first feature map F2 to obtain a feature vector and input into an up-sampling module in the third encoding unit and the first decoding unit. using an up-sampling module in the third encoding unit on the feature vector performing up-sampling processing to obtain a first feature map F3 and inputting the first feature map F3 into the bottleneck layer.

3. The method of radar reflectivity retrieval based on satellite infrared data and lightning observation data according to claim 2, characterized in that, The bottleneck layer comprises two Swin Transformer modules connected in sequence.

4. The method of radar reflectivity retrieval based on satellite infrared data and lightning observation data according to claim 3, characterized in that, The feature vectors are input into each decoding unit connected in sequence in the decoder respectively, so that each decoding unit performs feature extraction on the input feature vectors X En and the feature vectors X Bo The step of performing cross-scale feature fusion and then using two Swin Transformer modules to sequentially perform feature extraction on the fused feature maps to obtain a second feature map comprises: The feature vector X Bo The feature vector X Bo The feature vector X The feature vector X The feature vector X After fusion, two Swin Transformer modules in the first decoding unit are used for feature extraction in sequence to obtain a second feature map F4 and input the second decoding unit and the third decoding unit. The second cross-scale feature fusion module in the second decoding unit is used for fusing the feature vector X Bo , the second feature map F4, the feature vector , and the feature vector After fusion, two Swin Transformer modules in the second decoding unit are used for feature extraction in sequence to obtain a second feature map F5 and input the third decoding unit. The third cross-scale feature fusion module in the third decoding unit is used to fuse the feature vector X Bo , the obtained second feature map F4, the second feature map F5, and the feature vector After fusion, two Swin Transformer modules in the third decoding unit are used to sequentially perform feature extraction to obtain a second feature map F6.

5. The method of claim 4, wherein the method is based on satellite infrared data and lightning observation data. The first cross-scale fusion module fuses the feature vectors X Bo and the feature vectors The feature vectors and the feature vectors are fused as follows: respectively, twice up-sampling is performed on the feature vectors respectively, once up-sampling is performed on the feature vectors respectively, once down-sampling is performed on the feature vectors Bo respectively, and then splicing is performed on the feature vectors respectively to obtain a first sub-feature map and send it to a full connection layer to obtain the feature vectors Bo , the feature vectors , the feature vectors and the feature vectors after fusion.

6. The method of radar reflectivity retrieval based on satellite infrared data and lightning observation data according to claim 4, characterized in that, The second cross-scale fusion module fuses the feature vector X Bo , the second feature map F4, the feature vector and the feature vector according to the following steps: respectively, one down-sampling is performed on the feature vectors , two up-samplings are performed on the feature vectors X Bo , and then the feature vectors X are spliced to obtain a second sub-feature map and sent to a full connection layer to obtain a feature vector X Bo , the second feature map F4, the feature vectors X , and the feature vectors X after fusion.

7. The method of claim 4, wherein the method is based on satellite infrared data and lightning observation data. The third cross-scale feature fusion module fuses the feature vector X Bo , the obtained second feature map F4, the second feature map F5 and the feature vector according to the following steps: two times of down-sampling on the second feature map F4, one time of down-sampling on the second feature map F5, and three times of up-sampling on the feature vector X Bo After the three times of up-sampling, the feature vector X is spliced to obtain a third sub-feature map and is sent into a full connection layer to obtain the feature vector X Bo , the obtained second feature map F4, the second feature map F5 and the feature vector After the fusion, the feature map.

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

9. The method of radar reflectivity retrieval 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 = aL WMSE + bL mySSIM ; wherein, where N denotes the total number of pixels of the current reconstructed image obtained in each round of training, y i denotes the value of the i-th pixel position in the radar composite reflectivity image, w(y i ) denotes a weight function dependent on y i , denotes the value of the i-th pixel position in the current reconstructed image, SSIM denotes the structural similarity index, which is used to measure the similarity between y i and , α denotes the weight of the mean square error loss L WMSE , and β denotes the weight of the structural similarity loss L mySSIM .

Images