Lightning guidance attention coupling radar reflectivity inversion method and system

Abstract

The invention relates to the technical field of radar reflectivity inversion, and discloses a lightning guidance attention coupling radar reflectivity inversion method and system, and the method comprises the steps: collecting data, and carrying out the space-time alignment, and generating an input grid image; an attention coupling inversion model is trained, a lightning guide weight image is generated, soft enhancement is carried out on features in jump connection fusion and space attention calculation in the coding and decoding regression process, and hard gating is not carried out on a lightning-free area. And performing multiplicative enhancement processing on the features. And training an attention coupling inversion model by adopting the supervision loss weighted for the strong echo region. In the inversion stage, satellite multi-channel infrared observation raster data and lightning observation data are collected, an input raster image is generated and input into the attention coupling inversion model, and a radar combined reflectivity raster image is output as a radar reflectivity inversion result. According to the invention, the prediction precision of the strong echo region is improved, and the stability and accuracy of the inversion result in the strong convection region are enhanced.

Description

Technical Field

[0001] This invention relates to the field of radar reflectivity inversion technology, and more specifically, to a lightning-guided attention-coupled radar reflectivity inversion method and system. Background Technology

[0002] With the development of high-resolution satellite remote sensing technology, multi-channel infrared raster data has been widely used in meteorological monitoring and disaster early warning. Lightning observation, as a meteorological data source with high timeliness and high spatial resolution, can provide crucial information on thunderstorm activity and severe convective systems. Existing technologies typically rely on traditional radar reflectivity inversion methods, but due to the limitations of a single data source, it is difficult to fully utilize the complementary advantages of satellite remote sensing and lightning observation, thus affecting the accuracy and reliability of the inversion results. Therefore, using lightning-guided attention coupling technology to train deep learning models on satellite remote sensing data to optimize reflectivity inversion accuracy is a key technological direction for improving inversion results.

[0003] In existing technologies, radar reflectivity inversion methods typically rely on a single data source, either satellite remote sensing data or lightning observation data. While some existing technologies, such as the invention patent CN108445464A – a method for fusing satellite radar inversion and lightning observation data – attempt to integrate lightning observation and radar data, they still have several problems. For example, existing technologies often simply stitch together or weightedly fuse lightning observation data with satellite remote sensing data, failing to fully utilize the spatiotemporal distribution characteristics of lightning data and the crucial influence of strong echo areas. They also ignore situations where lightning data is missing and lack effective processing mechanisms. When lightning observation data is incomplete or time-series data is discontinuous, it can easily lead to errors in the inversion results.

[0004] Therefore, it is necessary to design a lightning-guided attention-coupled radar reflectivity inversion method and system to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a lightning-guided attention-coupled radar reflectivity inversion method and system, which aims to solve the problems of insufficient inversion accuracy caused by a single data source, errors caused by missing lightning data, and inaccurate prediction of strong echo areas in the current technology.

[0006] This invention proposes a lightning-guided attention-coupled radar reflectivity inversion method, comprising: Acquire satellite multi-channel infrared observation raster data and lightning observation data of the target area, and perform spatiotemporal alignment to generate an input raster image; The attention coupling inversion model is trained using training samples, which include the input raster image and the radar combined reflectivity raster image. During training, a lightning guidance weight image is generated from the lightning observation data, and the lightning guidance weight image is set to zero when all lightning observation data are missing. In the encoding, decoding and regression process, the lightning guidance weight image is used to perform soft enhancement on features in skip connection fusion and spatial attention calculation, and no hard gating is applied to areas without lightning. The features are multiplicatively enhanced based on the lightning-guided weight image at at least one level, and the multiplicative enhancement degenerates to not changing the features when the lightning-guided weight image is zero. The attention-coupled inversion model is trained using a supervised loss weighted for strong echo regions; During the inversion phase, satellite multi-channel infrared observation grid data and lightning observation data at the target time are acquired, an input grid image is generated and input into the trained attention-coupled inversion model, and the output radar combined reflectivity grid image is used as the radar reflectivity inversion result.

[0007] Furthermore, training the attention-coupled inversion model includes: The radar combined reflectivity raster image is resampled to a grid coordinate system consistent with the input raster image, and an effective pixel mask is generated based on the missing detection markers of the radar combined reflectivity raster image. During training iterations, a predicted radar combined reflectivity raster image is obtained based on the input raster image. The supervision error between the predicted radar combined reflectivity raster image and the radar combined reflectivity raster image at the pixel positions corresponding to the effective pixel mask is calculated, and the supervision error is converged to form a supervision loss. The model parameters of the attention coupling inversion model are updated according to the supervision loss.

[0008] Furthermore, training also includes: Based on the temporal continuity and validity indicators of the lightning observation data, it is determined whether the lightning observation data is missing. If a current time interval is missing but other valid data exists within the window, a lightning guidance weight image is generated by weighted fusion of lightning observation data from multiple time intervals within the time window. The weighted fusion assigns greater weight to higher quality and more recent lightning data based on the timeliness and quality indicators of the lightning observation data within the time window. If all lightning intervals within the window are missing, the lightning guidance weight image is set to zero. If the lightning observation data is not missing, a lightning guidance weight image is generated using the lightning observation data, and no time window fusion is performed on the data during the generation process.

[0009] Furthermore, when training the attention-coupled inversion model using a supervised loss weighted for strong echo regions, the following steps are included: Based on the radar combined reflectivity raster image, a strong echo pixel mask is generated according to a preset strong echo threshold, and the intersection of the strong echo pixel mask and the effective pixel mask is obtained to obtain a weighted pixel set; within the weighted pixel set, the supervision error between the predicted radar combined reflectivity raster image and the radar combined reflectivity raster image is calculated, and the strong echo pixels are assigned strong echo weights and then converged to obtain the supervision loss.

[0010] Furthermore, the strong echo weight is determined using a hierarchical weighting rule: pixels in the radar combined reflectivity grid image that are not less than a higher strong echo threshold are assigned a first weight, pixels that are not less than a lower strong echo threshold and are less than the higher strong echo threshold are assigned a second weight, and the remaining pixels are assigned a reference weight, wherein the first weight is greater than the second weight and the second weight is greater than the reference weight.

[0011] Furthermore, the multiplicative enhancement processing step includes: The features are weighted proportionally according to the lightning guiding weight image, wherein when the lightning guiding weight image is zero, the weighting ratio is set to 1.

[0012] Furthermore, the spatiotemporal alignment includes: aligning the satellite multi-channel infrared observation grid data and the lightning observation data to representative values ​​of the same time period, the time period being fifteen minutes in length; spatially registering and resampling the aligned data to a grid coordinate system with a spatial resolution of one kilometer; linearly normalizing the satellite multi-channel infrared observation grid data, the lightning observation data, and the radar combined reflectivity grid image to a range of zero to one according to preset upper and lower limits, and truncating values ​​exceeding the upper and lower limits.

[0013] Furthermore, constructing the training samples includes: cropping the input raster image and the radar combined reflectivity raster image into a 256×256 window, and sliding the slices with a step size of 128; pairing each sliced ​​input sub-image with the radar combined reflectivity sub-image as a training sample.

[0014] Furthermore, during training, when generating the lightning guidance weight image from the lightning observation data, the process includes: Within a preset time window, the lightning observation data is gridded and aggregated to obtain lightning density raster data. The lightning density raster data is input into the lightning guidance branch to extract lightning guidance features. The lightning guidance features are then scaled to the same spatial size as the output features of at least one coding layer to generate a lightning guidance weight image of the same scale. In the spatial attention calculation, a spatial attention weight map is generated based on the mean map, maximum map, and the lightning guidance weight image of the same scale of the output features of the coding layer. At the jump connection fusion point, the lightning guidance weight image of the same scale is used to perform soft enhancement on the fusion features.

[0015] Compared with existing technologies, the advantages of this invention are as follows: By combining satellite multi-channel infrared observation grid data and lightning observation data, an input grid image is generated using a spatiotemporal alignment method. Training and inversion are then performed using an attention coupling model from deep learning, effectively solving the problems of insufficient fusion of single data sources and improper handling of missing data in existing technologies. Lightning-guided weighted images are used to softly enhance features, and in the case of missing or incomplete lightning data, automatic zeroing or multi-time weighted fusion is employed to avoid errors caused by missing lightning data. Furthermore, the inversion model is optimized through supervised loss weighted for strong echo regions, improving the prediction accuracy in strong echo regions and enhancing the stability and accuracy of the inversion results in strong convection regions. This contributes to obtaining more accurate radar reflectivity inversion results in practical applications.

[0016] On the other hand, this application also provides a lightning-guided attention-coupled radar reflectivity inversion system for applying the above-mentioned lightning-guided attention-coupled radar reflectivity inversion method, including: The acquisition unit is configured to acquire satellite multi-channel infrared observation grid data and lightning observation data of the target area, and perform spatiotemporal alignment to generate an input grid image. The training unit is configured to train the attention coupling inversion model with training samples, including the input raster image and the radar combined reflectivity raster image. During training, a lightning guidance weight image is generated from the lightning observation data, and the lightning guidance weight image is set to zero when the lightning observation data is missing. In the encoding, decoding and regression process, the lightning guidance weight image is used to perform soft enhancement on features in skip connection fusion and spatial attention calculation, and no hard gating is performed on lightning-free areas. The first processing unit is configured to perform multiplicative enhancement processing on features based on the lightning-guided weight image at at least one level, and the multiplicative enhancement processing degenerates to not changing the features when the lightning-guided weight image is zero. The second processing unit is configured to train the attention-coupled inversion model using a supervised loss weighted for strong echo regions; The output unit is configured to acquire satellite multi-channel infrared observation grid data and lightning observation data at the target time during the inversion phase, generate an input grid image and input it into the trained attention-coupled inversion model, and output a radar combined reflectivity grid image as the radar reflectivity inversion result.

[0017] It is understandable that the above-mentioned lightning-guided attention-coupled radar reflectivity inversion method and system have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of the lightning-guided attention-coupled radar reflectivity inversion method provided in this embodiment of the invention; Figure 2 A functional block diagram of the lightning-guided attention-coupled radar reflectivity inversion system provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In some embodiments of this application, see Figure 1 As shown, this application proposes a lightning-guided attention-coupled radar reflectivity inversion method, including: S100: Acquire satellite multi-channel infrared observation raster data and lightning observation data of the target area, perform spatiotemporal alignment to generate input raster images.

[0021] S200: The attention-coupled inversion model is trained using training samples, including input raster images and radar combined reflectivity raster images. During training, a lightning-guided weight image is generated from lightning observation data, and the lightning-guided weight image is set to zero when all lightning observation data is missing. During the encoding, decoding, and regression processes, the lightning-guided weight image is used to softly enhance features in skip connection fusion and spatial attention calculation, without hard gating of lightning-free areas.

[0022] S300: Perform multiplicative enhancement on features based on the lightning-guided weight image at at least one level, and when the lightning-guided weight image is zero, the multiplicative enhancement degenerates into not changing the features.

[0023] S400: The attention-coupled inversion model is trained using a supervised loss weighted for strong echo regions.

[0024] S500: During the inversion phase, it acquires satellite multi-channel infrared observation grid data and lightning observation data at the target time, generates an input grid image and inputs it into the trained attention-coupled inversion model, and outputs a radar combined reflectivity grid image as the radar reflectivity inversion result.

[0025] Specifically, satellite multi-channel infrared observation raster data and lightning observation data for the target area are acquired, and temporal alignment and spatial registration are used to ensure that the two types of data fall under the same grid coordinate system, thus forming the input raster image. The input raster image includes at least raster data corresponding to multiple infrared channels and lightning density raster data obtained by aggregating lightning observation data. During the training phase, the input raster image and radar combined reflectivity raster image constitute the training samples to train the attention-coupled inversion model. During training, a lightning-guided weight image is generated from the lightning observation data. When all lightning observation data is missing within a preset time window, the lightning-guided weight image is set to zero. The model adopts an encoder-decoder regression structure, introducing the lightning-guided weight image to softly enhance features in skip connection fusion and spatial attention calculation, without hard gating of lightning-free areas. Simultaneously, multiplicative enhancement is performed on features based on the lightning-guided weight image at at least one level; when the lightning-guided weight image is zero, the multiplicative enhancement degenerates into not changing the features. To improve learning of strong echo regions, a supervised loss weighted for strong echo regions is used during training. In the inversion phase, the target time data undergoes preprocessing consistent with the training to generate an input raster image, which is then input into the trained attention-coupled inversion model, and the output radar combined reflectivity raster image is used as the radar reflectivity inversion result.

[0026] In one specific embodiment, satellite infrared channel raster data, lightning observation data, and radar combined reflectivity mosaic data of the target area are obtained from a data interface. The target area can be selected as the latitude and longitude range of Shandong Province, with a time resolution of 15 minutes and a spatial resolution of 1 kilometer. Using each 15-minute time interval as an alignment window, representative values ​​for that time interval are calculated for both satellite and lightning data. After spatial registration of each data point to the same grid, resampling is performed to obtain a consistent raster resolution. Specifically, the lightning observation data is first gridded and aggregated within the alignment window to obtain lightning density raster data. If there are no valid lightning observations or missing lightning observation data within the window, the lightning density raster data is set to zero, and a zero-valued lightning guidance weight image is generated accordingly. Minimum and maximum value normalization is performed on each channel to form the input raster image. The normalization upper and lower limits can be set as follows: infrared channel 1 range 200-405, infrared channel 2 range 110-305, infrared channel 3 range 110-340, lightning density range 0-6, and radar combined reflectivity range 0-75. To establish the training set, aligned raster samples were cropped to a window size of 256×256 and slid out in a step of 128. Each input sub-image was paired with a corresponding radar combined reflectivity sub-image, and the training, validation, and test sets were divided in a 7:2:1 ratio. In terms of model structure, the three infrared channels were fed into independent feature extraction branches to extract spectral features, and learnable channel gains were applied to each branch to enhance the expression weight of channels sensitive to strong convection, followed by feature fusion. Lightning density raster data was fed into a lightweight guiding branch to generate lightning guiding features, which were then scaled and projected to obtain lightning guiding weight images matched to different scale levels. The encoder-decoder regression backbone was downsampled and upsampled step-by-step, and lightning guiding weight images of the same scale were introduced at the skip connection fusion points to softly enhance the fused features. In spatial attention calculation, the feature statistics map and the lightning guiding weight image of the same scale were used together to generate the spatial attention weight map. Multiplicative enhancement is performed on the backbone features at at least one encoding layer, bottleneck layer, or decoding layer. This multiplicative enhancement uses the lightning-guided weight image as a pixel-by-pixel gain map for proportional modulation, and degenerates to an identity mapping when the lightning-guided weight image is zero. The training loss employs a combination of pixel-level supervision and structural similarity constraints, with weighting for strong echo regions: a strong echo pixel mask is generated based on the radar combined reflectivity raster image, assigning higher weights to pixels with target echoes of at least 35, even if the corresponding lightning density is zero, and including them in the weighted pixel set for supervision error convergence. The optimizer uses an adaptive moment estimation algorithm to iteratively update the model parameters and triggers early stopping based on the validation set to obtain a model with better generalization performance.

[0027] Understandably, by rigorously aligning satellite multi-channel infrared observation grid data with lightning observation data to form a unified input grid image, multi-source information can be compared and jointly modeled at the same grid scale. A lightning guidance mechanism using "soft enhancement rather than hard gating" continuously injects lightning priors into skip connection fusion and spatial attention computation. Combined with degenerate multiplicative enhancement, the model strengthens the representation of the core convection structure when lightning is present, and automatically degenerates to a robust inference path relying solely on spectral texture when there is no lightning or all lightning observation data is missing, avoiding the erroneous constraint of equating "no lightning" with "no strong echo." Furthermore, by using a supervised loss weighted towards strong echo regions to enhance the learning of strong echo samples, the model can still output more refined, continuous, and reliable radar combined reflectivity inversion results in radar blind zones, radar quality fluctuations, or far-sea monitoring scenarios.

[0028] In some embodiments of this application, training the attention-coupled inversion model includes: resampling the radar combined reflectivity raster image to a grid coordinate system consistent with the input raster image, and generating an effective pixel mask based on the missing detection markers of the radar combined reflectivity raster image. During training iterations, a predicted radar combined reflectivity raster image is obtained based on the input raster image. Supervision errors are calculated at the pixel positions corresponding to the effective pixel mask for the predicted radar combined reflectivity raster image and the radar combined reflectivity raster image, and these supervision errors are pooled to form a supervision loss. The model parameters of the attention-coupled inversion model are updated based on the supervision loss.

[0029] Specifically, by acquiring satellite multi-channel infrared observation raster data and lightning observation data of the target area, spatiotemporal alignment is performed, and a unified input raster image is generated. Next, the radar combined reflectivity raster image is resampled to a grid coordinate system consistent with the input raster image to ensure consistency in spatial resolution and location. For the radar combined reflectivity raster image, the system generates a valid pixel mask based on its missing data identification information. This mask identifies which pixels in the radar image contain valid data to determine which locations participate in the calculation, avoiding calculation errors caused by missing data. During training iterations, the input raster image is passed to the trained model to predict the corresponding radar combined reflectivity raster image. Then, the system calculates the prediction error at the pixel location corresponding to the valid pixel mask by comparing it with the original radar combined reflectivity raster image. To ensure model accuracy, these errors are pooled to form the final supervised loss. The supervised loss is passed through the backpropagation algorithm to further adjust the model parameters to improve the model's prediction accuracy.

[0030] For example, for a specific target area, multi-channel infrared raster data from satellites is first acquired, and after spatiotemporal alignment, an input raster image is generated. Simultaneously, a radar composite reflectivity image of the area is used to generate an effective pixel mask based on its missing data markers. The missing data markers in the radar image are primarily based on missing data points in the image; pixels at these points are excluded from the calculation. When processing the input image, it is ensured that a uniform spatial resolution and temporal window are used for matching, thereby avoiding errors caused by spatiotemporal inconsistencies.

[0031] During training, the input raster image is used by the model to predict the corresponding radar combined reflectivity raster image, which is then compared with the actual radar reflectivity image. When calculating the error, only the effective pixels covered by the effective pixel mask are considered, thus avoiding calculations on invalid data and ensuring the accuracy of the results. After error calculation, gradient descent is used to optimize the model parameters, enabling the model to gradually converge to the optimal solution.

[0032] Understandably, by uniformly aligning the radar combined reflectivity raster image with the input raster image in both space and time, data compatibility and accuracy are ensured. Simultaneously, by generating an effective pixel mask, error calculation is performed only on valid data, avoiding the impact of invalid data on the training process. This method effectively improves the accuracy and robustness of the trained model, especially when radar data is missing. The effective pixel mask ensures the stability and accuracy of the calculation, avoiding errors caused by missing data in traditional methods. Furthermore, using supervised loss to weight the calculation of errors in strong echo regions further enhances the inversion model's learning ability in strong convection areas, improving the model's performance in complex meteorological scenarios. The efficiency of the training process is optimized, ensuring high accuracy and stability of the final inversion results.

[0033] In some embodiments of this application, training further includes: determining whether lightning observation data is missing based on the temporal continuity of the lightning observation data and the validity indicators of the observations. If a current time interval is missing but other valid data exists within the window, a lightning guidance weight image is generated by weighted fusion of lightning observation data from multiple time intervals within the time window. The weighted fusion assigns greater weight to higher quality and more recent lightning data based on the timeliness and quality indicators of the lightning observation data within the time window. When all lightning observations within the window are missing, the lightning guidance weight image is set to zero. When no lightning observation data is missing, a lightning guidance weight image is generated using the lightning observation data, and no time window fusion is performed on the data during the generation process.

[0034] In some embodiments of this application, when training the attention-coupled inversion model using a supervised loss weighted for strong echo regions, the process includes: generating a strong echo pixel mask based on a radar combined reflectivity raster image according to a preset strong echo threshold, and intersecting the strong echo pixel mask with the effective pixel mask to obtain a weighted pixel set. Within the weighted pixel set, the supervised error between the predicted radar combined reflectivity raster image and the radar combined reflectivity raster image is calculated, and the strong echo pixels are assigned strong echo weights before converging to obtain the supervised loss.

[0035] In some embodiments of this application, the strong echo weight is determined using a hierarchical weighting rule: pixels in the radar combined reflectivity grid image that are not less than a higher strong echo threshold are assigned a first weight, pixels that are not less than a lower strong echo threshold and are less than a higher strong echo threshold are assigned a second weight, and the remaining pixels are assigned a reference weight, wherein the first weight is greater than the second weight and the second weight is greater than the reference weight.

[0036] Specifically, during the training phase, the temporal continuity and validity of lightning observation data are first identified to determine if any data is missing in the current time frame. If lightning data for the current time frame is missing, but other valid lightning observation data exists within a preset time window, these observations are weighted and fused within the time window to generate a corresponding lightning guidance weight image. The weighting and fusion process assigns greater weight to more recent and higher-quality lightning observations based on their timeliness and quality within the time window, thereby improving the accuracy and reliability of the weight image. If all lightning observations are missing within the time window, the generated lightning guidance weight image is set to zero to avoid the negative impact of missing data on model training. If no lightning observation data is missing, the system directly uses this data to generate the lightning guidance weight image without additional time window fusion processing during the generation process.

[0037] During training, a supervised loss strategy based on strong echo regions was employed. First, a strong echo pixel mask was generated based on the radar combined reflectivity raster image, according to a preset strong echo threshold. This mask was used to filter out pixels in strong echo regions of the radar image. The intersection of the strong echo region pixels with the effective pixel mask yielded a weighted pixel set. Within this weighted pixel set, the supervised error between the predicted and actual radar combined reflectivity raster images was calculated, with strong echo pixels assigned higher weights and weighted during error calculation to enhance the model's learning of strong echo regions. Finally, the weighted errors were pooled to form the supervised loss, which was used to update the model's parameters.

[0038] Furthermore, a hierarchical weighting rule was employed when determining the weights of pixels with strong echoes. Specifically, pixels in strong echo regions were divided into three levels: First, pixels in the radar combined reflectivity raster image with reflectivity values ​​greater than or equal to a higher strong echo threshold were assigned a first weight. Second, pixels with reflectivity between a lower and higher strong echo threshold were assigned a second weight. Finally, the remaining pixels with reflectivity below the lower strong echo threshold were assigned a baseline weight. The first weight is greater than the second weight, and the second weight is greater than the baseline weight. This hierarchical weighting rule ensures more accurate learning of strong echo regions while avoiding overtraining in weak echo regions.

[0039] In practical implementation, taking a specific satellite observation area as an example, the target area is a city and its surrounding area in eastern China, and the data for this area comes from the FY-4B satellite. Assume that the satellite multi-channel infrared raster data and lightning observation data for this area are observed hourly, with a spatial resolution of 1 kilometer and a time window of 15 minutes. The satellite infrared observation data and lightning observation data are first spatiotemporally aligned to generate the input raster image. At this time, the radar combined reflectivity raster image is also resampled to a grid coordinate system consistent with the input raster image.

[0040] Next, the lightning observation data is processed according to the missing data flag. If lightning data is missing for the current time period, the missing data is filled by calculating a weighted average of valid data from the preceding and following time periods. If the missing data is severe within a time window, the lightning steering weight image is set to zero. If all data is valid, the lightning steering weight image is directly generated using the lightning observation data and fed into the deep learning model along with the input raster image as training samples for training.

[0041] During the weighting process for strong echo regions, the values ​​of the radar combined reflectivity raster image are used to generate strong echo pixel masks. For example, a higher strong echo threshold of 50 and a lower strong echo threshold of 30 are set. These thresholds are used to determine which regions in the radar image belong to strong echo regions and which belong to weak echo regions. Pixels in strong echo regions are assigned higher weights, while those in weak echo regions are assigned lower weights, ensuring that the features of strong echo regions receive more attention during training.

[0042] Understandably, by introducing a lightning-guided weighted image generation mechanism and weighted fusion of lightning data within a time window, the problem of missing lightning observation data can be effectively addressed, enhancing the model's robustness to missing lightning data. Furthermore, the supervised loss strategy using strong echo regions weighting allows the model to focus more on the characteristics of strong echo regions during training, effectively improving the accuracy of reflectivity inversion. The hierarchical weighting rule for strong echo regions ensures refined learning of these regions while avoiding overtraining on weak echo regions. This improves the accuracy of radar reflectivity inversion, especially in situations with missing or incomplete data, still providing high-quality inversion results.

[0043] In some embodiments of this application, the multiplicative enhancement processing step includes: proportionally weighting the features according to the lightning guiding weight image, wherein when the lightning guiding weight image is zero, the weighting ratio is set to 1.

[0044] In some embodiments of this application, spatiotemporal alignment includes: aligning satellite multi-channel infrared observation grid data and lightning observation data to representative values ​​for the same time period, the time period being fifteen minutes in length. Spatially registering and resampling the aligned data to a grid coordinate system with a spatial resolution of one kilometer. Linearly normalizing the satellite multi-channel infrared observation grid data, lightning observation data, and radar combined reflectivity grid image to a range of zero to one according to preset upper and lower limits, and truncating values ​​exceeding these limits.

[0045] Specifically, this embodiment describes how to implement multiplicative enhancement processing and spatiotemporal alignment in the radar reflectivity inversion method. Multiplicative enhancement processing optimizes the reflectivity inversion effect of the model by weighting features using lightning-guided weight images. In the spatiotemporal alignment part, satellite multi-channel infrared observation grid data and lightning observation data are spatiotemporally synchronized to generate a unified input data format, providing high-quality data input for subsequent model training.

[0046] Specifically, the multiplicative enhancement step includes: during training, proportionally weighting the model features based on the lightning guidance weight image, where each pixel in the lightning guidance weight image represents the intensity of lightning activity in that region. The weighting coefficient for each pixel in the feature image is provided by the lightning guidance weight image. When the value of the lightning guidance weight image is zero, it means there is no lightning data in that region, so the weighting ratio for that region is set to 1, meaning no change is made to the feature value. This method ensures that the model's learning process is not affected in regions lacking lightning data, while it can better enhance the features in regions with denser lightning data.

[0047] The spatiotemporal alignment process first aligns satellite multi-channel infrared raster data with lightning observation data to the same time period, with a time period length of fifteen minutes. During this process, the timestamps of the synchronized satellite and lightning data are used to ensure they correspond to the same time window. Next, the aligned data is spatially registered and resampled to a uniform spatial resolution (e.g., a 1-kilometer grid coordinate system). This ensures spatial consistency among different data types and provides standardized input data for subsequent model training.

[0048] In the aligned data, satellite infrared raster data, lightning observation data, and radar combined reflectivity raster images undergo linear normalization, normalizing the data values ​​to the range of zero to one. This process ensures dimensional consistency across different data sources and improves model training performance. Simultaneously, values ​​exceeding the upper and lower limits are truncated according to preset limits to ensure data quality and stability.

[0049] Specifically, in this embodiment, a target area located in eastern China is selected. The satellite multi-channel infrared observation data and lightning observation data for this area are obtained from the FY-4B satellite. The satellite has a temporal resolution of 15 minutes and a spatial resolution of 1 kilometer. The lightning observation data has the same temporal and spatial resolution. During data processing, the satellite multi-channel infrared observation grid data and the lightning observation data are first spatiotemporally aligned to ensure that their timestamps and spatial coordinate systems are consistent.

[0050] Next, the aligned data is resampled to ensure they have the same spatial resolution (1 km). For example, if some data has a low spatial resolution, the system will upsample it using an interpolation algorithm to match the resolution of the input raster image.

[0051] The system then performs linear normalization on the aligned data. For infrared observation grid data, the data range may vary between different channels, so normalization is performed separately for each channel. Assuming the original data range for channel one is 200 to 405, it is normalized to a value between 0 and 1, a normalized value ranging from zero to one. For lightning observation data, it is normalized to a value between zero and six, and its values ​​are compressed to this range.

[0052] During the multiplicative enhancement process, the lightning-guided weight image adjusts the weights of the feature images based on the distribution of lightning observation data. Specifically, in areas with lightning activity, the weight values ​​are increased accordingly, thereby amplifying the impact of features from that region on model learning. If some areas lack lightning data, the weight image values ​​are set to zero, and the feature weighting ratio is set to 1, keeping the features unchanged and ensuring that areas without lightning data do not affect the inversion results.

[0053] Understandably, precise spatiotemporal alignment and linear normalization effectively ensure that data from different sources can be trained with a uniform spatial resolution and temporal window. Multiplicative enhancement, combined with lightning-guided weighted images, enhances features in relevant regions when data is complete, but degenerates to not changing features when lightning data is missing, ensuring model stability in the absence of lightning data. In particular, the use of supervised loss weighted for strong echo regions allows the model to focus more on these areas, improving the accuracy of reflectivity inversion. This not only optimizes the training process but also improves the accuracy of the inversion model, especially in providing high-quality inversion results even when lightning data is missing or incomplete.

[0054] In some embodiments of this application, constructing training samples includes: cropping the input raster image and the radar combined reflectivity raster image into a 256×256 window, and sliding the image through a step of 128. Each sub-image obtained from the image is paired with the radar combined reflectivity sub-image to form a training sample.

[0055] In some embodiments of this application, the generation of a lightning guidance weight image from lightning observation data during training includes: gridding and pooling the lightning observation data within a preset time window to obtain lightning density raster data; inputting the lightning density raster data into the lightning guidance branch to extract lightning guidance features; and scaling the lightning guidance features to the same spatial size as the output features of at least one coding layer to generate a lightning guidance weight image at the same scale. In spatial attention calculation, a spatial attention weight map is generated based on the mean map, maximum map, and lightning guidance weight image at the same scale of the coding layer output features, and soft enhancement of the fused features is performed using the lightning guidance weight image at the skip connection fusion point.

[0056] Specifically, the construction and processing of training data samples involves multiple steps. First, window cropping and sliding image extraction are performed on the input raster image and the radar combined reflectivity raster image. In this step, the input raster image and the radar combined reflectivity raster image are cropped into 256×256 pixel windows, with a step size of 128 pixels between each cropped window to ensure the diversity of the training data and the comprehensiveness of the image information. For each cropped sub-image, it is paired with the corresponding radar combined reflectivity sub-image to form a training sample. In this way, the training dataset can fully cover different parts of the input image, ensuring that the network can learn the features of different regions during training.

[0057] During training, a lightning guidance weight image is generated using lightning observation data. Specifically, lightning observation data is gridded and aggregated within a preset time window to obtain lightning density raster data. This data reflects the density of lightning activity within the target area over different time periods. Subsequently, the lightning density raster data is input into the lightning guidance branch to extract lightning guidance features. Through this branch, the system can obtain the spatial distribution features of lightning activity within the area and scale them to the same spatial scale as the output features of at least one coding layer. Ultimately, the lightning guidance weight image generated by the system has the same spatial scale as the output features of the coding layer, ensuring that this lightning guidance information can be effectively fused with other feature information in subsequent processing.

[0058] In the spatial attention computation process, the mean and maximum maps output by the encoding layer are combined with a lightning-guided weight map of the same scale to generate a spatial attention weight map. This weight map is used to weight features from different spatial regions, enhancing the feature representation of regions closely related to lightning activity. Then, during the skip connection fusion process, this spatial attention weight map is used to perform soft enhancement on the fused features, i.e., strengthening features in important regions while keeping other regions relatively unchanged. This process ensures that the model can focus on lightning-active regions during the learning process, thereby effectively improving the accuracy of reflectivity inversion.

[0059] Specifically, taking a specific satellite observation area as an example, the target area is a city in eastern China and its surrounding region. The satellite multi-channel infrared raster data and lightning observation data for this area come from the FY-4B satellite. The satellite data has a temporal resolution of 15 minutes and a spatial resolution of 1 kilometer, while the lightning observation data also has spatial and temporal resolutions of 1 kilometer and 15 minutes, respectively. During the data processing stage, the satellite multi-channel infrared raster data and lightning observation data are first spatiotemporally aligned to ensure synchronization between the two data sources in time and space. The aligned data is then resampled to a uniform spatial resolution.

[0060] During the construction of training data samples, the input raster image and the radar combined reflectivity raster image are cropped into 256×256 pixel windows, with a stride of 128 pixels between each cropped window to ensure the diversity and comprehensiveness of image information. For each cropped sub-image, it is paired with the corresponding radar combined reflectivity sub-image to generate a training sample. This ensures that the network covers information from different spatial locations during training, improving the model's learning ability.

[0061] In generating the lightning guidance weight image, lightning observation data is first gridded and aggregated to obtain lightning density raster data. Assuming a 15-minute time window for lightning data, a lightning density raster image is generated based on the lightning observation data within that time window, and lightning guidance features are further extracted through a lightning guidance branch. These features are scaled to the same spatial size as the output features of the coding layer, thus generating a lightning guidance weight image with a scale consistent with other features. In this way, the model can fuse features from lightning observation data and infrared data during the learning process, thereby improving the model's reflectance inversion accuracy.

[0062] Understandably, this embodiment achieves efficient matching and integration of input raster images from different data sources through precise spatiotemporal alignment and data normalization. Generating training samples via a sliding patch method ensures the network can learn features from different regions, avoiding overfitting. Introducing a lightning-guided weight image generation mechanism, combined with the spatial distribution information of lightning-guided features, allows the model to focus on regions closely related to lightning activity during training, further improving the model's reflectivity inversion accuracy. Using a spatial attention mechanism for soft feature enhancement, particularly combined with weighting of the lightning-guided weight map, enhances the focus on strong convection regions, improving the model's learning ability in strong echo regions. Through multi-source data fusion, feature weighting, and spatial enhancement techniques, especially in handling data gaps and strong echo regions, the accuracy of radar reflectivity inversion is improved.

[0063] Understandably, in this embodiment, training with the input raster image and the combined radar reflectivity raster image, and introducing lightning observation data to generate a lightning-guided weight map, enhances the feature representation of key areas and improves the accuracy and robustness of the inversion results. Especially in areas with strong echoes and frequent lightning, it avoids the errors and instabilities of traditional methods, thereby effectively improving the accuracy and reliability of radar reflectivity inversion.

[0064] In summary, this application combines satellite multi-channel infrared observation grid data with lightning observation data, generates input grid images using a spatiotemporal alignment method, and utilizes an attention coupling model from deep learning for training and inversion. This effectively addresses the problems of insufficient fusion of single data sources and improper handling of missing data in existing technologies. Lightning-guided weighted images are used to softly enhance features, and automatic zeroing or multi-time weighted fusion is employed in cases of missing or incomplete lightning data to avoid errors caused by missing lightning data. Furthermore, the inversion model is optimized through supervised loss weighted for strong echo regions, improving prediction accuracy in strong echo areas and enhancing the stability and accuracy of the inversion results in areas of strong convection. This improved inversion accuracy and enhanced robustness to missing data contribute to obtaining more accurate radar reflectivity inversion results in practical applications.

[0065] Based on another preferred embodiment described above, see [link to preferred embodiment]. Figure 2 As shown, this embodiment provides a lightning-guided attention-coupled radar reflectivity inversion system for applying the above-described lightning-guided attention-coupled radar reflectivity inversion method, including: The acquisition unit is configured to acquire satellite multi-channel infrared observation raster data and lightning observation data of the target area, and perform spatiotemporal alignment to generate an input raster image.

[0066] The training unit is configured to train the attention-coupled inversion model using training samples, including input raster images and radar combined reflectivity raster images. During training, a lightning-guided weight image is generated from lightning observation data, and the lightning-guided weight image is set to zero when lightning observation data is missing. In the encoding-decoding-regression process, the lightning-guided weight image is used to softly enhance features in skip connection fusion and spatial attention calculation, without hard gating of lightning-free areas.

[0067] The first processing unit is configured to perform multiplicative enhancement processing on features based on a lightning-guided weight image at at least one level, and the multiplicative enhancement processing degenerates to not changing the features when the lightning-guided weight image is zero.

[0068] The second processing unit is configured to train the attention-coupled inversion model using a supervised loss weighted for strong echo regions.

[0069] The output unit is configured to acquire satellite multi-channel infrared observation grid data and lightning observation data at the target time during the inversion phase, generate an input grid image and input it into the trained attention-coupled inversion model, and output a radar combined reflectivity grid image as the radar reflectivity inversion result.

[0070] Understandably, by combining satellite multi-channel infrared raster data with lightning observation data, and generating input raster images using a spatiotemporal alignment approach, and employing an attention coupling model from deep learning for training and inversion, the problems of insufficient fusion from a single data source and improper handling of missing data in existing technologies are effectively addressed. Lightning-guided weighted images are used to softly enhance features, and in cases of missing or incomplete lightning data, automatic zeroing or multi-time weighted fusion is employed to avoid errors caused by missing lightning data. Furthermore, the inversion model is optimized through supervised loss weighted for strong echo regions, improving prediction accuracy in these areas and enhancing the stability and accuracy of the inversion results in areas of strong convection. This improved inversion accuracy and enhanced robustness to missing data contribute to obtaining more accurate radar reflectivity inversion results in practical applications.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A lightning-guided attention-coupled radar reflectivity inversion method, characterized in that, include: Acquire satellite multi-channel infrared observation raster data and lightning observation data of the target area, and perform spatiotemporal alignment to generate an input raster image; The attention coupling inversion model is trained using training samples, which include the input raster image and the radar combined reflectivity raster image. During training, a lightning guidance weight image is generated from the lightning observation data, and the lightning guidance weight image is set to zero when all lightning observation data are missing. In the encoding, decoding and regression process, the lightning guidance weight image is used to perform soft enhancement on features in skip connection fusion and spatial attention calculation, and no hard gating is applied to areas without lightning. The features are multiplicatively enhanced based on the lightning-guided weight image at at least one level, and the multiplicative enhancement degenerates to not changing the features when the lightning-guided weight image is zero. The attention-coupled inversion model is trained using a supervised loss weighted for strong echo regions; During the inversion phase, satellite multi-channel infrared observation grid data and lightning observation data at the target time are acquired, an input grid image is generated and input into the trained attention-coupled inversion model, and the output radar combined reflectivity grid image is used as the radar reflectivity inversion result.

2. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 1, characterized in that, Training the attention-coupled inversion model includes: The radar combined reflectivity raster image is resampled to a grid coordinate system consistent with the input raster image, and an effective pixel mask is generated based on the missing detection markers of the radar combined reflectivity raster image. During training iterations, a predicted radar combined reflectivity raster image is obtained based on the input raster image. The supervision error between the predicted radar combined reflectivity raster image and the radar combined reflectivity raster image at the pixel positions corresponding to the effective pixel mask is calculated, and the supervision error is converged to form a supervision loss. The model parameters of the attention coupling inversion model are updated according to the supervision loss.

3. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 2, characterized in that, Training also includes: Based on the temporal continuity and validity indicators of the lightning observation data, it is determined whether the lightning observation data is missing. If a current time interval is missing but other valid data exists within the window, a lightning guidance weight image is generated by weighted fusion of lightning observation data from multiple time intervals within the time window. The weighted fusion assigns greater weight to higher quality and more recent lightning data based on the timeliness and quality indicators of the lightning observation data within the time window. If all lightning intervals within the window are missing, the lightning guidance weight image is set to zero. If the lightning observation data is not missing, a lightning guidance weight image is generated using the lightning observation data, and no time window fusion is performed on the data during the generation process.

4. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 2, characterized in that, When training the attention-coupled inversion model using a supervised loss weighted for strong echo regions, the following steps are included: Based on the radar combined reflectivity raster image, a strong echo pixel mask is generated according to a preset strong echo threshold, and the intersection of the strong echo pixel mask and the effective pixel mask is obtained to obtain a weighted pixel set; within the weighted pixel set, the supervision error between the predicted radar combined reflectivity raster image and the radar combined reflectivity raster image is calculated, and the strong echo pixels are assigned strong echo weights and then converged to obtain the supervision loss.

5. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 4, characterized in that, The strong echo weight is determined using a hierarchical weighting rule: pixels in the radar combined reflectivity grid image that are not less than a higher strong echo threshold are assigned a first weight, pixels that are not less than a lower strong echo threshold and are less than the higher strong echo threshold are assigned a second weight, and the remaining pixels are assigned a reference weight, wherein the first weight is greater than the second weight and the second weight is greater than the reference weight.

6. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 1, characterized in that, The multiplicative enhancement processing steps include: The features are weighted proportionally according to the lightning guiding weight image, wherein when the lightning guiding weight image is zero, the weighting ratio is set to 1.

7. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 1, characterized in that, The spatiotemporal alignment includes: aligning the satellite multi-channel infrared observation grid data and the lightning observation data to representative values ​​of the same time period, the time period being fifteen minutes in length; spatially registering and resampling the aligned data to a grid coordinate system with a spatial resolution of one kilometer; linearly normalizing the satellite multi-channel infrared observation grid data, the lightning observation data, and the radar combined reflectivity grid image to a range of zero to one according to preset upper and lower limits, and truncating values ​​exceeding the upper and lower limits.

8. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 1, characterized in that, Constructing the training samples includes: cropping the input raster image and the radar combined reflectivity raster image into a 256×256 window, and sliding the slices with a step size of 128; pairing each sliced ​​input sub-image with the radar combined reflectivity sub-image to form a training sample.

9. The lightning-guided attention-coupled radar reflectivity inversion method according to claim 8, characterized in that, During training, when generating the lightning guidance weight image from the lightning observation data, the process includes: Within a preset time window, the lightning observation data is gridded and aggregated to obtain lightning density raster data. The lightning density raster data is input into the lightning guidance branch to extract lightning guidance features. The lightning guidance features are then scaled to the same spatial size as the output features of at least one coding layer to generate a lightning guidance weight image of the same scale. In the spatial attention calculation, a spatial attention weight map is generated based on the mean map, maximum map, and the lightning guidance weight image of the same scale of the output features of the coding layer. At the jump connection fusion point, the lightning guidance weight image of the same scale is used to perform soft enhancement on the fusion features.

10. A lightning-guided attention-coupled radar reflectivity inversion system, used to apply the lightning-guided attention-coupled radar reflectivity inversion method as described in any one of claims 1-9, characterized in that, include: The acquisition unit is configured to acquire satellite multi-channel infrared observation grid data and lightning observation data of the target area, and perform spatiotemporal alignment to generate an input grid image. The training unit is configured to train the attention coupling inversion model with training samples, including the input raster image and the radar combined reflectivity raster image. During training, a lightning guidance weight image is generated from the lightning observation data, and the lightning guidance weight image is set to zero when the lightning observation data is missing. In the encoding, decoding and regression process, the lightning guidance weight image is used to perform soft enhancement on features in skip connection fusion and spatial attention calculation, and no hard gating is performed on lightning-free areas. The first processing unit is configured to perform multiplicative enhancement processing on features based on the lightning-guided weight image at at least one level, and the multiplicative enhancement processing degenerates to not changing the features when the lightning-guided weight image is zero. The second processing unit is configured to train the attention-coupled inversion model using a supervised loss weighted for strong echo regions; The output unit is configured to acquire satellite multi-channel infrared observation grid data and lightning observation data at the target time during the inversion phase, generate an input grid image and input it into the trained attention-coupled inversion model, and output a radar combined reflectivity grid image as the radar reflectivity inversion result.

Images