Lightning density downscaling method and product based on multi-scale spatial attention
The lightning downscaling model using U-Net and a multi-scale spatial attention module solves the problem of traditional methods' difficulty in downscaling, enabling the generation of high-resolution lightning density prediction maps and accurately capturing local structure and extreme value features.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 内蒙古自治区气候中心(内蒙古自治区气候变化中心内蒙古自治区雷电防护中心)
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional downscaling methods are ineffective at downscaling spaceborne lightning observation images and cannot accurately capture their local structure and extreme value features.
We employ a lightning downscaling model based on U-Net and a multi-scale spatial attention module. By using parallel multi-scale convolution operations, we enhance the model's spatial attention capability and improve its expressive power at different spatial granularities.
It achieves high-resolution downscaling of spaceborne lightning observation maps, accurately captures local structure and extreme value features, and improves the spatial resolution and accuracy of lightning density prediction maps.
Smart Images
Figure CN122244574A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of meteorological data processing, and in particular to a lightning density downscaling method and product based on multi-scale spatial attention. Background Technology
[0002] Downscaling techniques, as an effective means of improving the spatial resolution of meteorological data, have been widely applied in the field of weather forecasting. They are primarily used to downscale low-resolution spaceborne meteorological data to obtain high-resolution meteorological data.
[0003] However, most current research focuses on downscaling continuous meteorological variables such as precipitation, temperature, and wind speed. For lightning, a strong convective indicator characterized by its instantaneous, localized, and extreme nature, a systematic approach to downscaling has yet to be developed. Lightning events are highly sparse in spatial distribution and exhibit dramatic intensity variations, making it difficult for traditional downscaling methods to accurately capture their local structure and extreme value characteristics. Therefore, traditional downscaling methods are ineffective for effectively downscaling spaceborne lightning observation images. Summary of the Invention
[0004] The purpose of this application is to provide a lightning density downscaling method and product based on multi-scale spatial attention, which can effectively downscale spaceborne lightning observation maps and solves the problem that traditional downscaling methods are difficult to effectively downscale spaceborne lightning observation maps.
[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a lightning density downscaling method based on multi-scale spatial attention, including: Obtain a target spaceborne lightning observation image to be scaled down; The target satellite-borne lightning observation map is downscaled by the trained lightning downscaling model to obtain a lightning density prediction map. The spatial resolution of the lightning density prediction map is higher than that of the target satellite-borne lightning observation map. The lightning downscaling model includes a U-Net and a multi-scale spatial attention module. The skip connections between the encoder and decoder in the U-Net are implemented through the multi-scale spatial attention module.
[0006] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the lightning density downscaling method based on multi-scale spatial attention as described above.
[0007] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the lightning density downscaling method based on multi-scale spatial attention as described above.
[0008] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the lightning density downscaling method based on multi-scale spatial attention as described above.
[0009] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a lightning density downscaling method and product based on multi-scale spatial attention. A trained lightning downscaling model is used to downscale the target spaceborne lightning observation map, thereby obtaining a downscaled, high-resolution lightning density prediction map. Since spaceborne lightning observation maps exhibit high local clustering and sparsity in their spatial distribution, traditional single-scale spatial attention modules (such as CBAM) have limitations in processing complex spatial distribution features. Therefore, a multi-scale spatial attention module (MSA) is introduced, which enhances the model's spatial attention capability through parallel multi-scale convolutional operations, improving its expressive power at different spatial granularities. Specifically, the skip connections between the encoder and decoder in the U-Net are implemented through the multi-scale spatial attention module. By adopting the above-described lightning downscaling model architecture, the target spaceborne lightning observation map can be effectively downscaled, solving the problem that traditional downscaling methods struggle to effectively downscale spaceborne lightning observation maps. Attached Figure Description
[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0011] Figure 1 This is a flowchart of a lightning density downscaling method based on multi-scale spatial attention in one embodiment of this application; Figure 2 This is an architectural diagram of a lightning downscaling model in one embodiment of this application; Figure 3 This is an architectural diagram of a multi-scale spatial attention module in one embodiment of this application; Figure 4 This is the input image and lightning density label image for the first region in an experiment of this application; Figure 5 This is a graph showing the lightning density predictions for the first region from various models in an experiment of this application. Figure 6 This is the input image and lightning density label image for the second region in an experiment of this application; Figure 7 This is a lightning density prediction map for the second region from various models in an experiment of this application. Detailed Implementation
[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0013] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0014] The lightning density downscaling method based on multi-scale spatial attention provided in this application can be applied to terminals or servers. The terminal can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, etc.; the server can be a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.
[0015] Reference Figure 1 The lightning density downscaling method based on multi-scale spatial attention provided in this application embodiment includes steps S110 and S120.
[0016] Step S110: Obtain the target satellite-borne lightning observation image to be scaled down.
[0017] Step S120: The target satellite-borne lightning observation map is downscaled using the trained lightning downscaling model to obtain a lightning density prediction map. The spatial resolution of the lightning density prediction map is higher than that of the target satellite-borne lightning observation map.
[0018] The lightning downscaling model includes U-Net and a multi-scale spatial attention module. The skip connections between the encoder and decoder in U-Net are implemented through the multi-scale spatial attention module, which is used to enhance the lightning downscaling model's ability to capture multi-scale features of lightning activity.
[0019] In the above technical solution, a trained lightning downscaling model is used to downscale the target spaceborne lightning observation map, thereby obtaining a downscaled high-resolution lightning density prediction map. However, due to the high degree of local clustering and sparsity in the spatial distribution of spaceborne lightning observation maps, traditional single-scale spatial attention modules (such as CBAM) have limitations in processing complex spatial distribution features. Therefore, a multi-scale spatial attention module (MSA) is introduced. This module enhances the model's spatial attention capability through parallel multi-scale convolutional operations, improving its expressive power at different spatial granularities. Specifically, the skip connections between the encoder and decoder in U-Net are implemented through the multi-scale spatial attention module.
[0020] By adopting the lightning downscaling model with the above architecture, the target satellite lightning observation map can be effectively downscaled, solving the problem that traditional downscaling methods are difficult to effectively downscale satellite lightning observation maps.
[0021] In some embodiments, the U-Net encoder and decoder each include a three-level downsampling module and a three-level upsampling module, respectively, with skip connections between the downsampling and upsampling modules at the same level implemented through a separate multi-scale spatial attention module.
[0022] In the above embodiments, the encoder-decoder structure of U-Net employs a three-layer architecture for both the encoder and decoder. Correspondingly, there are also three multi-scale spatial attention modules, with skip connections between the downsampling and upsampling modules at the same level implemented through a separate multi-scale spatial attention module.
[0023] Furthermore, the feature processing actions of each downsampling module include two convolutions and max pooling operations in sequence, and the feature processing actions of each upsampling module include nearest neighbor interpolation upsampling, feature concatenation, and two convolutions in sequence. Feature concatenation is used to concatenate the sampled features of nearest neighbor interpolation upsampling and the output features of the same-level multi-scale spatial attention module.
[0024] The data processing flow of the lightning downscaling model is as follows: First, the input data is subjected to logarithmic transformation and normalization.
[0025] The data after logarithmic transformation and normalization then enters the encoding path. Each downsampling module sequentially performs two convolution and max pooling operations on its own input to gradually extract and compress features. The number of channels for each feature increases by a factor of two while the spatial size is halved.
[0026] The bottleneck layer expands the channels of the features extracted by the encoder through two convolutions, thereby capturing global features.
[0027] Subsequently, the global features enter the decoding path. Each upsampling module recovers the size of its input features through nearest neighbor interpolation upsampling, and concatenates the size-recovered features with the output features of the same-level multi-scale spatial attention module to obtain concatenated features. The concatenated features are then used as the output features of the upsampling module after two convolutions.
[0028] Finally, the channel number of the output features of the shallowest upsampling module is mapped to the target dimension through 1×1 convolution. After cropping and inverse transformation, the downscaling result is output, which is a high-resolution lightning density prediction map.
[0029] It should be noted that both convolution operations in the upsampling and downsampling modules can be implemented using 3×3 convolution.
[0030] Specifically, the feature processing steps of the multi-scale spatial attention module include: performing channel-dimensional averaging and max pooling on its own input features to obtain average mapping and max pooling mapping respectively; extracting features from the fusion results of the average mapping and max pooling mapping through multiple convolutional branches of different scales; fusing the output features of each convolutional branch to obtain a spatial attention map; and applying the spatial attention map to its own input features through pixel-by-pixel multiplication to form its own output features.
[0031] Preferably, the convolution scales of the three convolutional branches are 1×1, 3×3, and 5×5, respectively. Each branch extracts spatial information at a different scale. Each convolutional branch learns a saliency response in the spatial dimension, and the final results are concatenated and fused to obtain spatial attention.
[0032] For example, refer to Figure 3 The multi-scale spatial attention module focuses on its own input features. F Perform average and max pooling along the channel dimension to generate two spatial mappings. and All shapes .Will and After fusion, the outputs are fed into multiple parallel convolutional branches with kernel sizes of 1×1, 3×3, and 5×5 to capture spatial context information from different receptive fields. The outputs of all convolutional branches are concatenated, followed by a 1×1 convolutional layer to fuse the channels, and a spatial attention map is generated using the sigmoid function. Finally, Applied to original input features F Above, we obtain the spatially enhanced self-output features. F This process can be represented by the following mathematical formula: ; ; ; .
[0033] In some embodiments, the training steps of the lightning downscaling model include: acquiring matching sample satellite-borne lightning observation maps and sample ground-based 3D lightning observation maps; performing temporal discretization on the sample ground-based 3D lightning observation maps; performing spatial discretization on the temporally discretized sample ground-based 3D lightning observation maps to obtain corresponding lightning density label maps, wherein the spatial resolution of the lightning density label maps is greater than that of the sample satellite-borne lightning observation maps; and using the sample satellite-borne lightning observation maps and lightning density label maps as sample pairs to train the lightning downscaling model.
[0034] Constructing the lightning density label map is a crucial step in model training, significantly impacting the downscaling performance of the lightning downscaling model. In the above embodiment, the sample ground-based 3D lightning observation map is matched with the sample spaceborne lightning observation map, both containing the same lightning information. The difference lies in their sources. The spaceborne lightning observation map, obtained from a spaceborne lightning imager, is captured from high altitudes and has relatively low spatial resolution. The ground-based 3D lightning observation map, on the other hand, is lightning data observed from ground-based equipment, offering relatively high spatial resolution and serving as the data basis for downscaling the sample spaceborne lightning observation map.
[0035] The three-dimensional lightning observation map of the sample foundation was sequentially discretized in time and spatially. Temporal discretization ensures that the lightning density label map only covers the lightning occurrence density within a certain time period, thus avoiding the loss of statistical significance due to large time spans, since lightning occurrence density changes over time, and this information will be lost if the time span is too large. Spatial discretization is used to further improve the spatial resolution of lightning density.
[0036] Specifically, the sample ground-based 3D lightning observation map includes several lightning location locations; the sample ground-based 3D lightning observation map, after time discretization, is spatially discretized to obtain the corresponding lightning density label map, which specifically includes: defining a two-dimensional spatial grid, the spatial resolution of which is greater than that of the sample spaceborne lightning observation map; and mapping each lightning location location to multiple grid points in the neighborhood of the two-dimensional spatial grid using latitude and longitude to obtain the lightning density label map.
[0037] By setting up a two-dimensional spatial grid with a larger spatial resolution and mapping each lightning location to multiple grid points in its neighborhood, the lightning location is mapped to these grid points. Crucially, the mapping between lightning locations and grid points is one-to-many; one lightning location maps to multiple grid points, further improving the resolution of the lightning density map. Correspondingly, a grid point may also be mapped by multiple lightning locations. Therefore, the lightning density at a given location can be determined by the number of times each grid point is mapped. For example, if a grid point is mapped multiple times, it indicates that multiple lightning strikes occurred near that grid point, meaning the lightning density at that location is high.
[0038] Specifically, each lightning location is mapped to multiple grid points in the neighborhood of a two-dimensional spatial grid using latitude and longitude. This includes: mapping each lightning location to a two-dimensional spatial grid using latitude and longitude; traversing each lightning location in the two-dimensional spatial grid, setting a search range with a preset radius centered on the lightning location, and incrementing the lightning count of all grid points within the search range by 1, with the lightning count of each grid point starting from 0.
[0039] By traversing each lightning location, a two-dimensional spatial grid containing the lightning counts of each grid point can be obtained. This two-dimensional spatial grid is a high-resolution lightning density label map.
[0040] The preset radius can be set empirically, and it must be at least greater than the diagonal length of a single grid cell in the two-dimensional space grid. For example, if the size of a single grid cell is 1km × 1km, the preset radius can be set to 3km.
[0041] For example, with a preset time resolution of 10 minutes, both the sample satellite-borne lightning observation image and the target satellite-borne lightning observation image can be lightning observation images acquired by FY-4A LMI (Lightning Imager carried by China's Fengyun-4A satellite). The spatial resolution of the lightning observation image acquired by FY-4A LMI is 0.08°×0.08° (latitude and longitude), and the spatial resolution of the two-dimensional spatial grid is 0.01°×0.01° (latitude and longitude).
[0042] The low-resolution sample satellite lightning observation map and the corresponding high-resolution lightning density label map constitute a sample pair.
[0043] Meanwhile, during the training process, data augmentation techniques can be used to expand the sample pairs by generating new sample pairs through random flipping at 90°, 180°, and 270°.
[0044] In some of these embodiments, the loss function used during training of the lightning downscaling model is... for: ; ; ; in, and These are the mean squared error loss and the structural similarity loss, respectively. and There are two weights, Lightning density labeling map Lightning density prediction map Similarity between them and Lightning density label chart Lightning density prediction map The mean, and Lightning density label chart Lightning density prediction map variance Lightning density labeling map Lightning density prediction map Covariance between and All are constants.
[0045] Specifically, in lightning density maps, class imbalance is extremely prominent: non-lightning pixels typically account for over 95%, while lightning pixels are sparsely distributed (<5%). If the standard cross-entropy loss function is used directly, the model will be dominated by the non-lightning class during gradient updates, leading to a severe decline in prediction performance for lightning regions. While mean squared error (MSE) can measure overall numerical differences, relying solely on pixel-level errors in image-related tasks can easily overlook structural information, especially in regions with low lightning intensity but clear distribution boundaries. Therefore, to more comprehensively improve the model's ability to model spatial structure and texture, this embodiment introduces Structural Similarity Index (SSIM) loss as an auxiliary metric.
[0046] The mean squared error loss and structural similarity loss are weighted together to form the complete loss. Preferably, the two weights can be set as follows: =0.8、 =0.2.
[0047] The following experimental data illustrates the advantages of the lightning downscaling model proposed in this application compared to other models.
[0048] 1. Experiment setup.
[0049] This experiment used the Linux Ubuntu 22.04.3LTS operating system, with 256GB of RAM and an Intel Xeon Silver 4210R CPU. It employed the Tensorflow deep learning framework and used a GeForce RTX 4090 graphics card for model training and validation. The ADAM algorithm was used for training optimization, with a learning rate set to 10. -3 An early stopping strategy was used during training. If the loss value did not improve within three periods, the learning rate was multiplied by a decay factor of 0.9. If the loss value did not improve within six periods, training was stopped and the model weights with the smallest loss value were saved.
[0050] 2. Evaluation indicators.
[0051] To comprehensively evaluate the model's performance in lightning intensity downscaling, this experiment selected several evaluation metrics to measure the model's effectiveness from four perspectives: regression accuracy, spatial structure similarity, lightning event detection capability, and sparse distribution consistency. RMSE measures the root mean square error between predicted and actual values; a smaller value indicates a smaller error. The calculation formula is: .
[0052] PSNR evaluates the peak signal-to-noise ratio of the predicted image and the real image; a higher value indicates higher image quality. It is expressed as: ; SSIM measures the similarity between two images in terms of brightness, contrast, and structure; the closer to 1, the more similar the structures.
[0053] When the downscaling output value exceeds a set threshold (set to 0.1 in this experiment), it is considered that "a lightning event exists"; otherwise, it is considered that there is no lightning. Thus, a binary classification confusion matrix is constructed, and the following index is calculated: ; ; ; .
[0054] 3. Experimental results.
[0055] To comprehensively evaluate the performance of the proposed model in the lightning data downscaling task, this experiment compares it with two classical interpolation methods (Nearest Neighbor and Bilinear), a machine learning method (XGBoost), and deep learning baseline models (SRCNN and DeepSD). Evaluation metrics include root mean square error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), precision, recall, critical success index (CSI), and F1 score. Quantitative results for each method are summarized in Table 1. Ours represents the proposed lightning downscaling model.
[0056] Table 1 Comparison of quantitative results from different methods
[0057] As shown in Table 1, the lightning downscaling model proposed in this application outperforms the comparative methods in all evaluation metrics. While traditional interpolation methods are computationally efficient, they are weak in RMSE, PSNR, and structure preservation capabilities, especially in lightning images with complex spatial structure information, where structural similarity (SSIM) and critical success index (CSI) are low, demonstrating their limitations in sparse data reconstruction. XGBoost significantly outperforms traditional interpolation methods in several metrics, but it still falls slightly short in spatial structure preservation (SSIM = 0.9031), failing to fully capture the fine-grained spatial information of lightning data. Deep learning methods (SRCNN and DeepSD) show superior overall performance. SRCNN achieves a PSNR of 41.6357 and an SSIM of 0.9203, demonstrating the advantage of convolutional networks in image detail restoration. DeepSD, as an existing downscaling model, further improves performance. The lightning downscaling model proposed in this application has significant advantages over DeepSD (0.2093) and SRCNN (0.2183), with an RMSE of 0.1717, a PSNR of 44.6650, and an SSIM of 0.9768, which greatly surpasses DeepSD and SRCNN. In terms of lightning event recognition performance, the Precision is 0.8462, the Recall is 0.9863, the F1 score is 0.9106, and the CSI also reaches 0.8363, indicating that the lightning downscaling model proposed in this application can effectively maintain accuracy and integrity while reconstructing sparse lightning events.
[0058] These results demonstrate that the proposed lightning downscaling model can more accurately reconstruct spatial distribution patterns in sparse and high-resolution lightning scenes. The performance improvement is primarily attributed to the multi-scale spatial attention (MSA) module introduced into the model, which enhances the model's ability to perceive contextual information at different spatial scales. Simultaneously, the MSE and SSIM loss functions used jointly during training effectively improve structure preservation while maintaining pixel-level accuracy.
[0059] 4. Ablation experiment.
[0060] Table 2 compares the impact of different settings on lightning downscaling. The results show that the basic U-net model (using only MSE loss) can perform preliminary reconstruction of the lightning field structure, achieving an RMSE of 0.2196, a SSIM of 0.9144, and an F1 score of 0.8144. After introducing SSIM loss, both structural similarity and detection capabilities are improved; SSIM increases to 0.9501, and the F1 score improves to 0.8371, indicating that the structured loss helps improve the reconstruction of the lightning field's spatial structure and extreme features. With the addition of the MSA module, the model performs even better in capturing local structure and extreme region features, with an SSIM of 0.9615 and an F1 score of 0.8633, demonstrating the positive effect of cross-scale spatial feature modeling on lightning structure reconstruction. Finally, by introducing both the MSA module and SSIM loss, the model performance reaches its optimal level, with RMSE decreasing to 0.1717, SSIM increasing to 0.9768, and the F1 score reaching 0.9106. In summary, the ablation study results show that the proposed MSA module and SSIM hybrid loss can effectively improve the ability of lightning downscaling models to reconstruct structural and extreme value features, providing a reference for high-precision modeling of lightning fields.
[0061] Table 2 Ablation Experiment Results
[0062] 5. Case analysis.
[0063] To visually demonstrate the performance of different downscaling methods in lightning data reconstruction, this experiment selects two representative regions, namely the first and second regions, within the study area. The reconstruction effects of different methods on the lightning density field at typical times are shown, and compared with the real labeled data. Figures 4 to 7The visualization results correspond to the first and second regions, respectively. Specifically, they include: the input map, the lightning density label map, lightning density prediction maps using two interpolation methods (Nearest and Bilinear), a traditional machine learning method (XGBoost), classic deep learning models (SRCNN and DeepSD), and the multi-scale spatial attention model (Ours) proposed in this application. Figures 4 to 7 In the diagram, the horizontal axis represents longitude, and the vertical axis represents latitude.
[0064] Figure 4 and Figure 5 This paper presents lightning density prediction maps for the first region using different models during the period from August 17, 2023 to September 17, 2023 (Beijing time). From... Figure 4 As can be seen, the lightning density label map reveals a locally high-density lightning center with distinct extreme regions and clear boundary structures. In contrast, the low-resolution input map only presents blurry bright spots, making it difficult to discern the true spatial structure. Figure 5 As can be seen, the traditional nearest neighbor interpolation method, by directly copying the original pixels, results in "blocky" artifacts and poor spatial continuity in its lightning density prediction map. While bilinear interpolation produces a relatively smooth edge transition in its lightning density prediction map, it cannot accurately restore the shape and intensity of the lightning extremum region. The XGBoost model has some perception capability of the main lightning region, but it still cannot completely reconstruct the true spatial morphology, and the extremum region is significantly underestimated. The SRCNN model shows some improvement in edge smoothness, but is limited by its shallow structure and struggles to recover high-frequency information. The DeepSD model outperforms the above methods, able to restore the spatial contour of the lightning region to some extent, but it still lacks in extremum representation and edge details. The multi-scale attention model (Ours) proposed in this application performs best; its reconstruction results not only accurately locate the position of high lightning density regions but also effectively restore the extremum intensity and boundary details in the real image, showing high consistency with the lightning density label map.
[0065] To further verify the generalization ability of the proposed model in different geographical regions, this experiment selected cumulative lightning data from the second study area between October 25, 2023 and November 25, 2023 (Beijing time) for case analysis to analyze the ability of different models to reconstruct the spatial distribution characteristics of lightning. Figure 6 and Figure 7 This shows the lightning density prediction maps generated by different downscaling methods in the second region during this severe convective event. (Refer to...) Figure 6The second region is similar to the first, with a concentrated and intense lightning core region in the lightning density label map. The low-resolution input map still fails to provide effective structural information, appearing as low-contrast, blurry areas. Figure 7 As can be seen, among the interpolation methods, the nearest neighbor method still suffers from significant block artifacts, while bilinear interpolation cannot effectively capture extreme values. The XGBoost model has made some progress in structural reconstruction, but it is still insufficient in the fine-grained restoration of high-value regions. The SRCNN model shows a slight improvement in boundary representation, but the intensity in the central region is lower than the actual value. The DeepSD model captures the shape of the main lightning region relatively accurately, but there is still some information loss. In contrast, the model proposed in this application once again demonstrates a significant advantage in this region. The lightning density map it generates not only has a clear structure and continuous boundaries, but also the high-value regions are basically consistent with the lightning density label map, reflecting the model's good adaptability to spatial heterogeneity and extreme value structures.
[0066] The case studies from both regions demonstrate that the downscaling model proposed in this application, which integrates a multi-scale spatial attention mechanism, significantly improves spatial structure reconstruction, extreme value restoration, and edge detail representation compared to traditional interpolation, machine learning, and classic deep networks. It also exhibits superior generalization ability under different geographical environments and lightning event types, verifying its application potential in refined lightning modeling and reconstruction.
[0067] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0068] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0069] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0070] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0071] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0072] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0073] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0074] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A lightning density downscaling method based on multi-scale spatial attention, characterized in that, include: Obtain a target spaceborne lightning observation image to be scaled down; The target satellite-borne lightning observation map is downscaled by the trained lightning downscaling model to obtain a lightning density prediction map. The spatial resolution of the lightning density prediction map is higher than that of the target satellite-borne lightning observation map. The lightning downscaling model includes a U-Net and a multi-scale spatial attention module. The skip connections between the encoder and decoder in the U-Net are implemented through the multi-scale spatial attention module.
2. The multi-scale spatial attention based lightning density downscaling method according to claim 1, wherein, The encoder and decoder of the U-Net include a three-level downsampling module and a three-level upsampling module, respectively. The skip connection between the downsampling module and the upsampling module at the same level is implemented through an independent multi-scale spatial attention module.
3. The multi-scale spatial attention based lightning density downscaling method of claim 2, wherein, The feature processing actions of the downsampling module at each level include two convolution and max pooling operations in sequence. The feature processing actions of the upsampling module at each level include nearest neighbor interpolation upsampling, feature concatenation and two convolutions in sequence. The feature concatenation is used to concatenate the sampled features of the nearest neighbor interpolation upsampling and the output features of the multi-scale spatial attention module at the same level. The bottleneck layer between the encoder and decoder of the U-Net expands the channels through two convolutions.
4. The multi-scale spatial attention based lightning density downscaling method of claim 2, wherein, The feature processing steps of the multi-scale spatial attention module include: The input features are subjected to channel-dimensional averaging and max pooling respectively to obtain the average mapping and max pooling mapping. Feature extraction is performed on the fusion result of the average mapping and the max pooling mapping through multiple convolutional branches of different scales; The output features of each convolutional branch are fused to obtain a spatial attention map; The spatial attention map is applied to its own input features through pixel-by-pixel multiplication to form its own output features.
5. The lightning density downscaling method based on multi-scale spatial attention according to claim 1, characterized in that, The loss function employed by the lightning downscaling model at training time is: ； ； ； in, and These are the mean squared error loss and the structural similarity loss, respectively. and There are two weights, Lightning density labeling map Lightning density prediction map Similarity between them and Lightning density label chart Lightning density prediction map The mean, and Lightning density label chart Lightning density prediction map variance Lightning density labeling map Lightning density prediction map Covariance between and All are constants.
6. The lightning density downscaling method based on multi-scale spatial attention according to claim 1, characterized in that, The training steps for the lightning downscaling model include: Obtain matching sample spaceborne lightning observation maps and sample ground-based three-dimensional lightning observation maps; The three-dimensional lightning observation map of the sample foundation is discretized in time; The sample ground-based three-dimensional lightning observation map, after time discretization, is spatially discretized to obtain the corresponding lightning density label map. The spatial resolution of the lightning density label map is greater than that of the sample spaceborne lightning observation map. The lightning downscaling model is trained using the sample satellite-borne lightning observation map and the lightning density label map as sample pairs.
7. The lightning density downscaling method based on multi-scale spatial attention according to claim 1, characterized in that, The sample ground-based three-dimensional lightning observation map includes several lightning location points; Spatial discretization is performed on the time-discretized sample ground-based three-dimensional lightning observation map to obtain the corresponding lightning density label map, specifically including: A two-dimensional spatial grid is defined, wherein the spatial resolution of the two-dimensional spatial grid is greater than the spatial resolution of the sample satellite lightning observation map; The lightning density label map is obtained by mapping each lightning location to multiple grid points in the neighborhood of the two-dimensional spatial grid using latitude and longitude.
8. The lightning density downscaling method based on multi-scale spatial attention according to claim 7, characterized in that, The preset time resolution is 10 minutes; Both the sample satellite-borne lightning observation image and the target satellite-borne lightning observation image are lightning observation images acquired by FY-4A LMI. The spatial resolution of the lightning observation image acquired by FY-4A LMI is 0.08°×0.08°, and the spatial resolution of the two-dimensional spatial grid is 0.01°×0.01°.
9. The lightning density downscaling method based on multi-scale spatial attention according to claim 7, characterized in that, Mapping each lightning location to multiple grid points in the neighborhood of the two-dimensional spatial grid using latitude and longitude specifically includes: The locations of each lightning strike are mapped to the two-dimensional spatial grid using latitude and longitude. In the two-dimensional spatial grid, each lightning location is traversed, and a search range with a preset radius is set with the lightning location as the center. The lightning count of all grid points within the search range is increased by 1, and the lightning count of each grid point starts from 0.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the lightning density downscaling method based on multi-scale spatial attention as described in any one of claims 1-9.