Lightning target recognition model and method facing image input modality imbalance and high false alarm rate

Abstract

The application discloses a lightning target recognition model and method facing image input modal imbalance and high false alarm rate, and belongs to the technical field of image recognition and meteorological disaster monitoring. Firstly, Gaussian mixture density estimation and time-weighted fusion are performed on ground-based lightning location data to convert the image label with spatial probability distribution characteristics; then, Himawari-8 satellite multi-channel brightness temperature images and their derived feature maps and radar echo images are used as multi-source heterogeneous inputs. In view of the information imbalance between different observation modalities, an MDE-UNet deep learning model is constructed, a multi-scale feature fusion module, a MLP enhanced decoding unit based on a weighted sliding window and a radar image information enhancement module are innovatively designed, the key image features are enhanced and irrelevant noise is suppressed, and the feature imbalance problem in heterogeneous data fusion is solved. The model adopts an asymmetric weighted BCE-DICE loss function to strengthen the attention to lightning target pixels.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image recognition technology, and specifically relates to a multi-source fusion lightning target recognition model and method for image input modal imbalance and high false alarm rate. Background Technology

[0002] Image recognition and target detection technologies play a crucial role in meteorological disaster monitoring (such as lightning detection). Existing technologies include methods based on ground-based detection networks, satellite remote sensing images, and deep learning image segmentation models. Among these, methods based on deep learning and fusing multimodal heterogeneous image data are the closest to the present invention. Typical examples include: LightningNet, which for the first time systematically fused satellite, radar, and lightning location data for forecasting; StrikeNet, which uses an encoder-decoder CNN to achieve pixel-level lightning location prediction; and MLDYOLO, which uses the fusion of multiple radar features for lightning identification. However, the aforementioned closest existing technologies and current mainstream lightning detection and image reconstruction models (DeeplabV3, SegNet, BiSeNet, HRNet) generally have several significant shortcomings when processing such highly complex meteorological heterogeneous images, resulting in poor overall performance. Specifically, these shortcomings include:

[0003] The false alarm rate in pixel-level image classification is too high. Mainstream models such as StrikeNet and DeeplabV3 generally suffer from high false alarm rates, leading to a decrease in the reliability of warnings. The root cause lies in the model's failure to fully learn the mapping relationship between the low-level pixels of the image and the deep physical mechanisms, making it difficult to distinguish the target area that actually breeds lightning from non-lightning interference areas that only have similar macroscopic texture features at the image feature level.

[0004] The image spatial resolution is coarse, and pixel-level boundary delineation is insufficient. The feature maps output by the existing model have a relatively coarse resolution, making it impossible to achieve pixel-level accurate segmentation and localization of the lightning target region. The Critical Success Index (CSI) is low, resulting in blurry target localization results. This deficiency stems from the loss of spatial details and contextual information during image decoding, and the fact that the training loss function was not specifically optimized for the accuracy of image target boundaries.

[0005] Insufficient fusion of multi-scale features from heterogeneous image modalities leads to poor reconstruction quality. Lightning activity involves multi-scale physical processes, and existing models struggle to effectively align and fuse satellite multispectral channel image features with radar vertical structure feature maps, resulting in lagging image reconstruction metrics and an inability to accurately restore the original image structure of the target area.

[0006] Overall performance is unbalanced and reliability is insufficient. None of the mainstream image models have achieved synergistic optimization of target localization accuracy, pixel-level classification accuracy, and image reconstruction quality, resulting in insufficient recognition reliability under extremely imbalanced samples. Summary of the Invention

[0007] Purpose of the invention: To address the shortcomings of existing image segmentation and target recognition technologies in processing heterogeneous meteorological images, such as high false alarm rates, blurred boundaries, and imbalance of heterogeneous modal features, this invention aims to provide a lightning target recognition model and method oriented towards image input modal imbalance and high false alarm rates, in order to solve the core challenges faced by current computer vision technology in meteorological cross-applications.

[0008] The lightning recognition model of the present invention includes:

[0009] The encoder module is used to extract multi-scale image features from the input radar echo image data and satellite multi-channel brightness temperature image data, and output coded feature maps at different levels. The radar echo image data and satellite multi-channel brightness temperature image data are heterogeneous image data.

[0010] The decoder module, which is connected in skip connection to the encoder module, is used to upsample and restore the encoded feature map step by step, and fuse it with the encoded feature map of the corresponding level to output a multi-level decoded feature map;

[0011] A multi-scale feature fusion module is connected to the output of each level of the encoder module to realize cross-scale image feature alignment and global semantic fusion, and generate a global fusion feature map.

[0012] The radar information enhancement module is connected to the output of the multi-scale feature fusion module. It is used to perform gain correction on the fused image features using the spatial physical constraints of the radar image and output the final lightning target probability map.

[0013] Furthermore, the encoder module includes multiple cascaded encoding layers, each encoding layer containing a dual convolution module, a channel attention mechanism module, and a max pooling layer;

[0014] The dual convolution module includes two sets of two-dimensional convolutional layers. The dual convolution module is then connected to a channel attention mechanism module, which is used to perform channel-level recalibration on the feature map output by the convolution, and then output the encoded feature map through max pooling downsampling.

[0015] Furthermore, the satellite multi-channel brightness temperature image data includes Band9 channel data, TBB13 channel brightness temperature data, brightness temperature difference data between TBB15 and TBB13 channels, and composite index (Band11-TBB13)-(TBB13-TBB15).

[0016] Furthermore, the decoder module includes multiple cascaded decoding layers, each of which contains a multilayer perceptron-based enhanced decoding unit based on a weighted sliding window. The enhanced decoding unit is used to perform transposed convolutional upsampling on the decoded feature map of the previous layer, concatenate it with the encoded feature map of the corresponding layer, aggregate the context information of the neighboring image pixels through a weighted sliding window mechanism, and then generate the decoded feature map of the current layer through multilayer perceptron mapping.

[0017] Furthermore, the weighted sliding window-based multilayer perceptron enhanced decoding unit includes:

[0018] The upsampling subunit uses a transposed convolutional layer to double the spatial resolution of the input feature map;

[0019] The splicing subunit is used to splice the upsampled feature map with the corresponding level's encoded feature map along the channel dimension.

[0020] The sliding window aggregation subunit is used to perform reflection filling on the stitched feature map, extract the neighborhood image features of each pixel in a sliding window manner, and achieve local context aggregation through weighted summation;

[0021] The fully connected mapping subunit, consisting of multiple fully connected layers, ReLU activation functions, and batch normalization layers, is used to map the aggregated neighborhood features into an output feature map with the target number of channels.

[0022] Furthermore, the multi-scale feature fusion module includes:

[0023] Multiple feature transformation modules are connected one-to-one with the output terminals of each level of the encoder module. Each feature transformation module is used to unify the input encoded feature map to a preset target number of channels and target spatial size.

[0024] The splicing unit, connected to the output of all feature transformation modules, is used to splice the unified feature maps of each level in the channel dimension.

[0025] A dual convolution module, connected to the output of the stitching unit, is used to extract cross-scale convection organization features from the stitched multi-scale features.

[0026] The fusion output unit, connected to the output of the feature extraction unit, includes a channel attention mechanism module and a residual connection. It is used to perform channel weighted enhancement on the extracted features and then add them to the first-layer encoded feature map to output a global fusion feature map.

[0027] Furthermore, the radar image information enhancement module includes:

[0028] The radar image feature extraction unit is connected to the radar echo image data input terminal and is used to extract radar feature maps from the radar echo image data.

[0029] The confidence generation unit is connected to the output of the radar image feature extraction unit and is used to map the radar feature map into a radar confidence weight map of [0,1].

[0030] The residual compensation unit is connected to the output of the multi-scale feature fusion module and the output of the confidence generation unit, respectively. It is used to weight and fuse the global fused feature map with the radar confidence weight map according to the preset compensation weight to generate the corrected lightning probability prediction map.

[0031] Furthermore, the channel attention mechanism module includes:

[0032] The global pooling unit is used to perform global average pooling on the input feature map, compressing the spatial dimension to obtain channel-level global features;

[0033] The weight learning unit, comprising a dimension-reducing fully connected layer, a ReLU activation function, and a dimension-upgrading fully connected layer, is used to map the channel-level global features to channel attention weights between 0 and 1.

[0034] The feature recalibration unit is used to multiply the channel attention weights with the input feature map channel by channel to output the enhanced feature map.

[0035] Furthermore, the feature transformation module includes:

[0036] The channel adaptation subunit uses a 1×1 convolutional layer to convert the number of channels in the input feature map into a preset uniform number of channels;

[0037] Spatial alignment sub-units are used to interpolate the spatial dimensions of the feature map to a preset target size using bilinear interpolation.

[0038] The channel attention mechanism module is used to perform channel-level recalibration on spatially aligned feature maps.

[0039] This invention also proposes a lightning target identification method, comprising the following steps:

[0040] Step 1, Multimodal Image Data Preprocessing: Acquire and preprocess historical multi-source meteorological image data, including lightning location data, satellite multi-channel brightness and temperature image data, and radar echo image data. The lightning location data undergoes image-based preprocessing to generate high-risk lightning domain images with spatial probability distribution characteristics, serving as the true labels for image segmentation. The radar echo image sequence undergoes spatiotemporal alignment processing to ensure strict alignment between the radar images, satellite multi-channel brightness and temperature images, and target label images in both time and space dimensions.

[0041] (1) Generation of lightning target image labels

[0042] To address the difficulty of direct extrapolation caused by abrupt changes in the spatiotemporal distribution of discrete lightning location data, this invention proposes an image label generation method based on physical priors and probabilistic modeling, using ground-based lightning location data acquired at ten-minute intervals from a Very Low Frequency Lightning Location Network (VLF-LLN). This method constructs continuous images of areas of concentration (AOCs) that reflect the intensity of lightning activity through Gaussian mixture model density estimation and time-weighted fusion, serving as the ground-based labels. The specific steps are as follows:

[0043] 1) Two-dimensional Gaussian distribution modeling of discrete lightning location points

[0044] Based on lightning observations, the horizontal extension of lightning branches rarely exceeds 20 km. Therefore, a Gaussian model of lightning occurrence probability is constructed: the lightning location point ( , Transformed into an isotropic two-dimensional Gaussian distribution The standard deviation σ = 7.554 km, determined by the fitting, reduces the probability at x = 20 km to 3%, which meets the upper limit of extended statistics. The distribution peak is normalized to 1, indicating that the probability of occurrence at the location point is 100%, that is, 98% of the probability is concentrated within a radius of 3 km, which is consistent with the spatial distribution characteristics of actual lightning activity.

[0045]

[0046] 2) Weighted fusion of time-series image features

[0047] To address the timeliness differences of image sequences at different time steps, a time decay weighting strategy with ten-minute intervals is adopted to fuse multi-time Gaussian distributions. The weights are constructed and normalized based on the negative half-axis of the natural exponential function to ensure effective decay of long-term data: n=6 (using historical data from the past hour), the time steps are mapped to [1,10], and exp(−10)≈4× Achieving attenuation. The fused lightning probability density P follows a Gaussian distribution across time periods ( The weighted sum of these superimposed elements has the physical meaning of "the expected number of lightning strikes per 10 minutes," and the specific formula is as follows:

[0048]

[0049] 3) Data normalization and outlier handling

[0050] This invention addresses the problem that the spatiotemporal randomness of lightning data leads to extreme peak values ​​(exceeding 100) after weighted fusion, thus masking the characteristics of the low-value range (0-25). A peak truncation method is proposed. A maximum threshold of 25 is set, and the data is divided into: a core region of high-frequency lightning activity (…). >25), Second highest density area (1≤ ≤25) and areas with no significant activity ( <1). This method preserves the differences in lightning activity intensity while eliminating the interference of extreme outliers on the data distribution. Based on this, high-density and second-highest-density lightning areas are selected as high-risk regions, and binarization is performed with a threshold of 1 to generate labels. The formula is:

[0051]

[0052] (2) Construction of satellite multimodal image feature input

[0053] This invention uses Band 9, TBB13, TBB15 and their differences (TBB15-TBB13) and composite index ((Band11-TBB13)-(TBB13-TBB15)) from the Himawari-8 satellite as core inputs, with a time resolution of 10 minutes. The satellite and radar work in tandem; the former reveals cloud tops and microphysical structures through brightness temperature, while the latter provides the three-dimensional distribution of particles within the cloud, jointly pinpointing high-risk areas for lightning.

[0054] In visual feature extraction, images from different channels represent different underlying textures and high-level semantic structures. TBB13 (10.4 µm) below -52°C indicates a deep convective cloud top; Band 9 (6.9 µm) reflects water vapor transport. However, a single channel is insufficient to distinguish clouds with similar top heights but vastly different internal structures. The inter-channel brightness temperature difference (BTD) is key to unlocking cloud microphysics: TBB15-TBB13 is sensitive to the size of ice crystals at the cloud top; small ice crystals formed by strong convection result in a large positive value, a strong signal of impending lightning. The composite index (Band11-TBB13)-(TBB13-TBB15) can suppress water vapor interference and accurately reflect ice crystal concentration; its high-value area is highly consistent with the lightning outbreak area, serving as a precursor to lightning warnings.

[0055] (3) Spatiotemporal alignment and interpolation of high-resolution radar image frames

[0056] Radar echo images provide high-resolution maps of the internal spatial structure for lightning target identification. This invention utilizes six-minute interval radar echo images for spatiotemporal alignment, providing precise support for lightning identification. The radar reflectivity factor (dBZ) characterizes the three-dimensional distribution of particles within clouds; high reflectivity nuclei (>45 dBZ) extending above the freezing layer indicate strong updrafts colliding with ice phases, directly correlated with lightning activity. Combining satellite brightness-temperature differences allows for precise location of individual thunderstorm cells. Radar extrapolation is significant for strong convection early warning; high-intensity and wide-coverage echoes reflect strong convection, laying the physical basis for charge separation. Focusing on high-value samples enhances the model's ability to capture features such as precipitation system morphology, coverage, and intensity changes; these features are significantly correlated with lightning probability, intensity, and distribution.

[0057] To achieve spatiotemporal alignment with lightning and satellite data, the radar image was downsampled from 512×512 pixels to 256×256 pixels spatially. Temporally, optical flow combined with bilinear interpolation was used to generate 10-minute interval data. First, the Farneback algorithm was used to calculate the dense optical flow field of adjacent frames to obtain the pixel motion direction and displacement. The optical flow field was scaled according to the interpolation time ratio α, and the corresponding coordinates of the interpolated frame pixels in the previous frame were calculated in reverse. Then, bilinear interpolation was performed on the non-integer coordinates, and a new frame was generated by weighted summation of the four neighboring pixels. The resulting interpolated frame conforms to the physical motion law of radar echoes, and the pixel values ​​transition smoothly.

[0058] Step 2, Image Target Recognition Model Construction

[0059] A multi-scale U-Net network (MDE-UNet) based on an enhanced decoder is constructed as a lightning recognition model. This model employs an encoder-decoder backbone architecture to extract and reconstruct image features and includes three core computer vision processing modules:

[0060] (1) Channel Attention Mechanism Module (CABlock)

[0061] This invention addresses the problems of feature redundancy, background noise interference, and insufficient multimodal fusion accuracy in lightning target monitoring using multi-source satellite and radar image data. It introduces a channel attention mechanism module (CABlock) into a multi-scale U-Net network based on an enhanced decoder to achieve adaptive weighted fusion of multi-source heterogeneous image features. This module first extracts channel-level global features of the image through global average pooling. Then, it learns the attention weights of the channels through a bottleneck structure composed of fully connected layers, ReLU, and sigmoid activation functions, and multiplies these weights with the original input image features channel by channel. Through this mechanism, the model can adaptively enhance image feature channels representing key lightning targets such as deep convection and cloud top microphysics, while suppressing irrelevant background channels and pixel noise, thereby significantly improving the target pixel recognition accuracy. CABlock improves the model's feature extraction stability under various complex weather image scenarios while maintaining image processing computational efficiency, highly meeting the needs of operational real-time image monitoring systems.

[0062] (2) Multi-scale feature fusion module (Fusion Block)

[0063] Addressing the challenges of multi-scale feature dispersion, target feature signal fragmentation, and loss of subtle visual features commonly encountered in lightning target image segmentation, this invention innovatively integrates a multi-scale feature fusion module into the decoding path. This module spatially aligns and stitches together the multi-scale high-level semantic features output from each level of the decoder with shallow image detail features through cross-layer skip connections. The stitched feature map is then processed by a dual convolutional block (DC block) to effectively extract cross-scale local spatial organization features and reconstruct the vertical structural coherence of the image target. Simultaneously, the module embeds a channel attention mechanism (CA block) that weights features based on the importance of each channel in the physical image representation, strengthening deep convection target signals and suppressing background image interference. This module effectively breaks through the bottleneck of feature separation for target images of different morphologies, significantly reducing the false alarm rate of pixel-level image classification in operational scenarios.

[0064] (3) Enhanced decoding unit (WWMLP Block) of multilayer perceptron based on weighted sliding window

[0065] To address the decoding challenge of inversely mapping high-level semantic features to high-resolution physical fields when reconstructing lightning strike areas from radar and Sunflower satellite imagery data (Band9, TBB13, TBB15, etc.), this invention designs a weighted sliding window multilayer perceptron enhanced decoding unit (WWMLP Block) on the critical path of the decoder. This solves problems such as fragmented results, blurred boundaries, isolated false noise, and poor physical consistency that are common in traditional decoding. The WWMLP Block dynamically aggregates multi-source features from the neighborhood (such as satellite brightness temperature gradient and radar reflectivity) through a weighted sliding window, explicitly modeling spatial continuity and physical correlation, and suppressing isolated false signals. Its core MLP accurately learns the mapping from aggregated features to lightning probabilities, decoding the deep physical relationship between cloud top ice crystallization and lightning activity implicit in exponents such as TBB15-TBB13. Experiments show that this unit can generate smooth, clear-boundary, and physically reliable lightning strike areas, significantly reducing the false alarm rate and achieving accurate and robust decoding. The specific process is as follows: the input multi-channel feature map is transformed by a channel-by-channel weight matrix, then learned through a fully connected layer, ReLU, and batch normalization, and finally restored to the original output size.

[0066] (4) Radar Information Enhancement Module (REBlock)

[0067] This invention is based on the physical mechanism of lightning: in thunderstorm clouds, supercooled water droplets collide with ice crystals and soft hail in the mixed phase region of -10°C to -25°C, resulting in charge separation according to the non-inductive electrification theory; and ice phase particles are active when the radar reflectivity exceeds 31.29 dBZ at the -20°C layer altitude, showing a strong correlation with lightning. Addressing the problems of missed warnings in traditional satellite imagery warnings and the easy obscuring of weak radar image features by the background, this invention proposes a physically guided bidirectional collaborative neural network framework: satellite data provides large-scale physical priors as global constraints, while high-resolution radar features construct a complementary space; a "prediction-correction" mechanism is introduced to achieve functional decoupling—the satellite backbone network learns macroscopic precipitation patterns, and the physically guided radar enhancement module (PREM) generates attention weights based on ZR relationships, identifying high-confidence potential lightning areas (RSI), and using residual compensation to correct the initial inversion results. The output layer performs physical consistency verification, enabling the radar feature-guided model to focus on weak areas, restoring the fine structure lost by smoothing in the satellite imagery, and effectively recalling missed samples while maintaining a low false alarm rate. This framework enables efficient synergy between the global advantages of satellites and the local physical constraints of radar, thereby improving the precision of short-term early warning for severe convection.

[0068] (5) Overall Model Architecture

[0069] The MDE-UNet target recognition model of this invention is built on the U-Net image segmentation architecture as its basic framework. The specific structure is as follows: In the encoding stage, a channel attention mechanism module (CA Block) is introduced into the encoding layer. After adaptive weighting and labeling of the extracted features, they are downsampled by max pooling and passed to the next encoding layer. They are then transmitted to the corresponding level decoder through skip connections. In the decoding stage, a weighted sliding window MLP unit (WWMLPBlock) is set in the decoding layer. The features of the previous layer's decoding result after upsampling by transpose convolution (TC) are jointly reconstructed with the features of the encoder skip connections. The reconstruction results are respectively input into the next decoding layer and the multi-scale feature fusion module (FusionBlock). Finally, the radar information enhancement module (RE Block) performs residual compensation and fine correction on the output results of the U-Net backbone network.

[0070] Step 3: Train the lightning recognition model using the asymmetric weighted BCE-DICE loss function;

[0071] To address the challenges of extremely sparse positive samples and severe class imbalance in meteorological image monitoring data, as well as the high false alarm rate and model bias caused by negative samples dominating the gradient in traditional BCE loss, and its insensitivity to the operational characteristic that "the cost of false alarms is higher than that of false alarms," ​​this invention proposes an asymmetric weighted BCE-DICE composite loss function. Dice Loss effectively alleviates class imbalance by calculating the overlap between the predicted image region and the real image label, constraining the spatial continuity and localization accuracy of the region. The improved weighted BCE assigns higher gradient weights to sparse positive sample target pixels, guiding the model to focus on and learn challenging local target image features. The weighted fusion of these two loss functions enables the deep learning model to perform detailed probability calibration while ensuring overall consistency between the predicted target mask and the real landing area in the image spatial morphology, thus effectively avoiding target undersegmentation and meeting the needs of computer vision applications. This design solves the visual segmentation problems such as model learning failure and fragmented predicted image regions under sparse pixel samples, providing theoretical support for balancing high hit rate and low false alarm rate in image classification. The formula is as follows:

[0072] (7)

[0073] , The weighting coefficients for the weighted BCE loss and Dice loss are used to weigh the contribution ratio of the two losses.

[0074] The Dice loss is used to focus on the intersection of positive samples, correcting the gradient bias and guiding the model to focus on sparse target image pixels, as shown in the following formula:

[0075] (8)

[0076] Let be the predicted probability of the pixel at position (h,w) in the b-th image sample by the model; Let be the true label at position (h, w) in the b-th image sample.

[0077] Simultaneously, asymmetric weights for false positives and false negatives are introduced to accurately match business cost requirements. Combined with pixel-level loss preservation and region-level overlap measurement, this solves the problem of negative sample gradient dominance while also taking into account pixel accuracy and region positioning accuracy, thus ensuring effective learning of sparse lightning features. The asymmetric weights and weighted BCE loss are as follows:

[0078] (9)

[0079] (10)

[0080] Step 4, Pre-training and Transfer Learning:

[0081] The model was trained using the Adam optimizer, with an initial learning rate of 1× Furthermore, an adaptive decay strategy based on validation set performance is introduced: when the validation loss does not decrease for two consecutive rounds, the learning rate is multiplied by 0.5, with a lower bound of 1× Gradient clipping (L2 norm limited to 1.0) is applied to suppress gradient explosion. Training is performed for 30 epochs, with validation metrics monitored in real time and optimal model weights saved for testing to improve the model's generalization ability on unknown images.

[0082] To address the challenges of extracting visual features from sparse images and the low accuracy of pixel segmentation with few samples, a classic "pre-training-fine-tuning" transfer learning strategy in computer vision is employed.

[0083] (1) Pre-trained model construction: The initial model is trained using large-scale lightning labels after binarization, so that it learns the global spatial distribution, gray-scale differences and edge contour features of lightning images.

[0084] (2) Model parameter fine-tuning: Based on the pre-trained model, fine-tuning was performed using high-risk lightning region labels, with the learning rate set to 1×10⁻ 4 (One order of magnitude lower than pre-training) While preserving global features, it focuses on capturing subtle features such as local gray-scale abrupt changes and edge blurring, avoiding overfitting and insufficient feature extraction under small sample sizes.

[0085] This transfer learning strategy effectively improves the recognition accuracy of sparse lightning targets and the reconstruction quality of image masks, making it suitable for real-time lightning image monitoring applications.

[0086] Beneficial effects:

[0087] First, it significantly reduces the false alarm rate of images and improves prediction reliability. While maintaining a high hit rate, this invention learns the essential physical precursors of lightning from multimodal heterogeneous image observation data, and designs a targeted model structure and learning objectives to enhance the model's ability to distinguish between real lightning incubation areas and interference areas with similar macroscopic features. This significantly reduces false warnings while ensuring a high detection rate, thereby improving the practical value of the product.

[0088] Second, it achieves pixel-level lightning recognition with a fine scale of 2 kilometers. To overcome the shortcomings of existing visual segmentation methods, such as coarse spatial scale of output feature maps and blurred target boundaries, this invention optimizes the image processing model architecture, especially by introducing an efficient feature recovery and enhancement mechanism in the decoding path. This enables fine delineation of the potential lightning target activity area in terms of both spatial range and morphological boundaries, fully meeting the visual application requirements for high-precision target detection and regional risk avoidance.

[0089] Third, this invention enhances the deep feature fusion capability and network generalization robustness of multimodal heterogeneous image data. Addressing the issues of feature representation imbalance and over-reliance on specific modal sources in existing visual methods when fusing satellite multispectral images and radar echo images, this invention proposes a physically-guided image feature fusion strategy. This strategy adaptively coordinates multimodal visual information flow, effectively mining key high-frequency features such as the three-dimensional spatial mapping structure of radar images and deep semantic relationships between heterogeneous data, further strengthening the generalization capability and image feature extraction stability of deep learning models. Attached Figure Description

[0090] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0091] Figure 1 It is a visualization of the lightning processing flow;

[0092] Figure 2 It includes visualizations of lightning, radar, and satellite images;

[0093] Figure 3 This is a structural diagram of the CABlock channel attention mechanism module;

[0094] Figure 4 This is a structural diagram of the enhanced decoding unit module of a multilayer perceptron based on a weighted sliding window;

[0095] Figure 5This is a structural diagram of the multi-scale feature fusion module;

[0096] Figure 6 This is a schematic diagram of radar signal correction;

[0097] Figure 7 This is a diagram showing the overall structure of the lightning recognition model of this invention;

[0098] Figure 8 This is a comparison chart of the output results of different models;

[0099] Figure 9 These are residual plots of the output results from different models;

[0100] Figure 10 This is a comparison chart of the recognition results of the model of this invention with several mainstream image segmentation models;

[0101] Figure 11 This is a residual image showing the recognition results of the model of this invention compared with those of several mainstream image segmentation models;

[0102] Figure 12 It includes comprehensive performance indicators and a visual comparison chart. Detailed Implementation

[0103] 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 a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0104] The lightning identification method of the present invention includes the following steps:

[0105] Step 1: Multimodal image data processing, the complete processing flow is as follows Figure 1 ;

[0106] Step 1.1 Lightning target image diffusion, temporal weighting and binarization processing

[0107] Step 1.1.1: Model the discrete lightning location points using a two-dimensional Gaussian distribution with a standard deviation of σ=3.727 to generate a single-point Gaussian probability response image.

[0108] Step 1.1.2: A time-decay weighted strategy is used to fuse the Gaussian distribution results of 6 historical time steps (n=6, matching the number of prediction frames). Specific steps are as follows:

[0109] 1) Define the mapping variable for the i-th time step. Mapping the time step range [1,6] to [1,10], the calculation formula is as follows:

[0110]

[0111] Where n=6, i=1,2,…,6;

[0112] 2) Construct a weighting function based on the negative half-axis of the natural exponential function, and calculate the weighting coefficient at the i-th time step. :

[0113]

[0114] 3) By fusing the Gaussian distribution results from multiple time periods, the final probability value of high-density lightning areas is calculated:

[0115]

[0116] in, The result of overlaying the Gaussian distribution at the i-th time step. The physical meaning is "the expected number of lightning strikes per 10 minutes".

[0117] Step 1.1.3: Truncate and binarize the fused image result. First, perform peak truncation, setting the maximum value threshold to 25, and truncate the fused result P using the following formula:

[0118]

[0119] in, >25 corresponds to the core area of ​​high-frequency lightning activity, 1≤ ≤25 indicates a secondary high-density lightning zone. A value <1 is considered an area with no significant lightning activity; subsequently, binarization is performed to obtain a high-risk region for lightning targets (including high-density and second-highest-density areas), and a threshold of 1 is used to classify... The specific formula for converting this into a high-risk lightning zone label is as follows:

[0120]

[0121] Step 1.2, Selection and preprocessing of satellite data

[0122] The following four types of channel images from the Himawari-8 satellite were selected as input:

[0123] 1) Band 9: The center wavelength of this band is 6.9 μm, representing the upper and middle layer water vapor channel.

[0124] 2) TBB 13: The center wavelength of the band is 10.45 μm, representing the infrared brightness temperature.

[0125] 3) TBB 15-TBB 13: Split window difference, characterizing cloud top height and temperature corrections, with the center wavelength of the TBB 15 band being 12.35 μm.

[0126] 4) Composite Index (Band11−TBB13)−(TBB13−TBB15): A three-channel linear combination used for cloud phase discrimination and strong convection detection, where the center wavelength of Band 11 is 8.6 μm.

[0127] Step 1.3 Radar echo image interpolation and spatiotemporal alignment processing

[0128] 1) Image spatial downsampling

[0129] The `cv2.resize()` function from the OpenCV-Python library is used to downsample a 512×512 pixel radar image to 256×256 pixels, specifying the interpolation mode as `cv2.INTER_LINEAR` (bilinear interpolation). This ensures that the spatial distribution features of the radar echo are not lost after resampling, achieving spatial window alignment with lightning and satellite data. Alternatively, the `skimage.transform.resize()` function from the scikit-image library can be used, specifying `anti_aliasing=True` in conjunction with bilinear interpolation to complete resampling, further optimizing the spatial feature preservation effect.

[0130] 2) Time interpolation processing

[0131] The dense optical flow field of adjacent radar echo frames is calculated using OpenCV's calcOpticalFlowFarneback() method. Based on the interpolation time ratio α, the optical flow is scaled using NumPy, and the non-integer coordinates of the interpolated frame pixels in the previous frame are inferred. Then, the smooth interpolated frames are generated by weighting using OpenCV bilinear interpolation (cv2.INTER_LINEAR) or manually using NumPy. All time windows are processed in a loop using NumPy array operations to ensure that the interpolation results conform to the laws of physical motion and that the pixel values ​​transition smoothly.

[0132] Step 1.4, Lightning Recognition Model Input and Label Construction

[0133] The radar image channel and the satellite 4-channel image data are stitched together in the channel dimension to obtain a 5x256x256 multimodal heterogeneous image tensor, which is used as the input of the lightning target recognition model. At the same time, the lightning high-risk domain data is used as the training label.

[0134] Step 2: Construct a lightning target image recognition model: the MDE-UNet model.

[0135] 1. Model Introduction

[0136] The MDE-UNet model in this embodiment includes an encoder module, a decoder module, a multi-scale feature fusion module, and a radar information enhancement module;

[0137] The encoder module is used to extract multi-scale features from the input multi-source meteorological data (radar and satellite 4-channel data) and output encoded feature maps at different levels.

[0138] The decoder module, which is connected in skip connection to the encoder module, is used to upsample the deepest encoded feature map step by step and fuse it with the encoded feature map of the corresponding level to output a multi-level decoded feature map.

[0139] The multi-scale feature fusion module is connected to the output of each level of the encoder module, and is used to perform cross-scale feature alignment and fusion on the encoded feature maps of different levels to generate a global fused feature map.

[0140] The radar information enhancement module is connected to the output of the multi-scale feature fusion module. Based on the radar echo features, it performs residual correction on the global fused feature map and outputs the final lightning occurrence probability.

[0141] The encoder module takes a multimodal heterogeneous image tensor as input (dimensions: B (batch size, specifically 10) × 5 (channels) × H (height, specifically 256) × W (width, 256)), such as... Figure 2 As shown, the first channel is the radar echo image, and the input to the first channel is the radar echo image. The following four channels are satellite data channels (Band 9, TBB 13, TBB 15-TBB 13, (Band11−TBB13)−(TBB13−TBB15)). The output of the radar information enhancement module is a probability map of high-risk lightning areas (dimension: B×1×H×W) in channel 1. In the encoder module, decoder module, multi-scale feature fusion module, and channel attention module, the hidden layer's basic dimension hidden_dim is set to 32. In the radar information enhancement module, the residual compensation weights are... Set it to 0.5.

[0142] 1) Encoder module

[0143] The encoder module comprises three cascaded encoding layers and one bottleneck layer. Each encoding layer consists of a dual convolutional module (DCBlock), a channel attention mechanism module (CABlock), and a max pooling layer. The DCBlock continuously stacks two sets of 2D convolutions (each set is a 3×3 convolution, padded with 1s, with no bias, followed by batch normalization and ReLU activation, inplace=True), converting the input channel count from in_channels to out_channels. The extracted features undergo adaptive channel recalibration via the CABlock, and the result is then connected to the corresponding decoding layer and downsampled via max pooling before being passed to the next encoding layer. The downsampled result from the third encoding layer is then passed to the bottleneck layer. The bottleneck layer, also composed of a dual convolutional module (DCBlock), sends its result to the decoder.

[0144] 2) Decoder module

[0145] like Figure 4 As shown, the decoder module contains three cascaded decoding layers, each with a multilayer perceptron-based enhanced decoding unit based on a weighted sliding window. This unit first reduces the number of channels of the previous layer's decoded feature map to in_channels / 2 and upsamples it through a transposed convolutional layer (2×2 convolutional kernel, stride 2). Then, it is padded and aligned according to the size difference with the corresponding encoded features, and then concatenated along the channel dimension. The concatenated features are then expanded into a 3×3 sliding window (stride 1) after reflection padding (padding=1). Neighborhood features are extracted by weighted summation through channel-wise weight matrix transformation, and finally mapped to the current layer's decoded feature map through a fully connected layer (input dimension 4C, output dimension out_channels).

[0146] 3) Multi-scale feature fusion module (FusionBlock) such as Figure 5 As shown:

[0147] The multi-scale feature fusion module unifies the channels of the three-level features dec1-dec3 of the encoder into hidden_dim through FTBlock and upsamples and aligns them to the dec1 size, and then concatenates them along the channel dimension; then, it performs cross-scale convection feature extraction through DCBlock, and then reduces the channels to hidden_dim through 1×1 convolution, and sends it to CABlock for channel weighting enhancement to output a global fusion feature map.

[0148] 3.1) Feature Transform Block (FTBlock)

[0149] The input feature channel number is converted from in_channels to out_channels by 1×1 convolution and batch normalization is performed. Then, bilinear interpolation (align_corners=True) is used to align the feature space size to the target size H×W. Finally, the interpolated features are fed into the channel attention mechanism module CABlock for channel-level feature enhancement.

[0150] 3.2) Channel attention mechanism module (CABlock), such as Figure 3 As shown:

[0151] The input feature tensor (B×C×H×W) is subjected to global average pooling, compressing the spatial dimension to 1×1 to obtain B×C×1×1, which is then flattened into a one-dimensional vector of B×C. Subsequently, a fully connected layer with a bottleneck structure is passed: the first layer reduces the C dimension to C / 16 (no bias, ReLU activation), and the second layer restores it to C dimension (no bias, Sigmoid activation), outputting B×C channel attention weights (between 0 and 1). Finally, these weights are reconstructed into B×C×1×1 and multiplied with the original input channel by channel to achieve enhancement of high-contribution channels and suppression of low-contribution channels.

[0152] 4) Radar information enhancement module (REBlock), such as Figure 6 As shown;

[0153] For the radar echo image input to the first channel, features are first extracted using a 3×3 convolution (the number of channels is set to hidden_dim), and then the high-confidence region features are enhanced using the channel attention module CABlock. Subsequently, the number of channels is reduced to 1 using a 1×1 convolution, and a radar confidence weight map of 0-1 is generated by Sigmoid activation. This weight map and the backbone network output are then subjected to residual compensation by a×backbone output + a×confidence weight map (a=0.5). Finally, the output is restricted to the 0-1 interval by Sigmoid activation to obtain the probability of lightning occurrence.

[0154] 2. Introduction to the model forward propagation process, such as... Figure 7 As shown;

[0155] 1) Input preprocessing: Separate the radar echo image (channel 1) from the input tensor (B×5×256×256);

[0156] 2) Encoding stage:

[0157] The input tensor is processed sequentially through three encoding layers. Each encoding layer consists of a dual convolutional block (DCBlock), with output dimensions of B×32×256×256 (denoted as skip1), B×64×128×128 (denoted as skip2), and B×128×64×64 (denoted as skip3), respectively. Subsequently, the output of each layer is downsampled using 2×2 max pooling to obtain feature maps of size B×32×128×128 (denoted as enc1), B×64×64×64 (denoted as enc2), and B×128×32×32 (denoted as enc3), which serve as inputs to the next encoding layer. Meanwhile, skip1-skip3 are passed to the corresponding decoding layer via skip connections. Finally, enc3 is input to the bottleneck layer, which also consists of a dual convolutional block (DCBlock), and its output feature map size is B×128×32×32 (denoted as enc4). Finally, enc4 is sent to the first decoding layer in the decoder for further processing.

[0158] 3) Decoding stage:

[0159] The decoder consists of three decoding layers. The input feature maps of each layer (enc4, dec1, dec2 in sequence) are upsampled after being transposed by a 2×2 convolution and then concatenated and fused with the feature maps of the corresponding coding layers (skip3, skip2, skip1 in sequence). The fused feature maps are then fed into the next layer through the enhanced decoding unit WWMLPBlock of the multilayer perceptron based on a weighted sliding window (the output dimensions are B×128×64×64 (denoted as dec1), B×64×128×128 (denoted as dec2), and B×32×256×256 (denoted as dec3) in sequence).

[0160] 4) Multi-scale fusion: The feature maps output from the three decoding layers dec1-dec3 are fed into the multi-scale feature fusion module FusionBlock, and the fused features (B×32×256×256) are output.

[0161] 5) Backbone network output: 1×1 convolution converts the fused features into 1 channel, and Sigmoid activation obtains the initial probability map;

[0162] 6) Radar Enhancement: The initial probability map is sent to the radar information enhancement module REBlock module and fused with the radar echo characteristics to obtain the final lightning probability map (B×1×256×256).

[0163] Step 3, Loss Function Construction

[0164] This implementation method targets the training of a recognition model for lightning target images with extremely imbalanced pixels. It improves the accuracy of lightning region segmentation by differentially penalizing false positives and false negatives, combining weighted binary cross-entropy (BCE) loss with Dice loss, and employs vectorized computation to ensure efficiency. The specific steps are as follows:

[0165] 1) Initialize the loss function parameters: Set the weight of false positive (FP) to 2.0 and the weight of false negative (FN) to 1.5. These can be adjusted according to business tolerance (FP is recommended to be in the range of 1.0-5.0, and FN is recommended to be in the range of 1.0-3.0).

[0166] 2) Basic BCE loss calculation: Calculate the BCE loss between the model prediction and the true label pixel by pixel to obtain a tensor with shape [B,1,1,256,256].

[0167] 3) Error type mask and weight assignment: Without calculating gradients, generate two types of masks: false positives (prediction > 0.5 and true label = 0) and false negatives (prediction < 0.5 and true label = 1). Initialize a weight matrix of all 1s, replace the FP region with 2.0, replace the FN region with 1.5, and keep the rest at 1.0.

[0168] 4) Weighted BCE loss calculation: Multiply the basic BCE loss element by element with the weight matrix, and then take the average in the height and width dimensions to obtain the weighted BCE loss tensor of shape [B,1,1].

[0169] 5) Dice Loss Calculation: Flatten the predicted values ​​and labels to [B,1,1,65536] (65536=256×256), calculate the intersection sum (multiply each element and then sum). The Dice coefficient is (2×intersection sum+1) / (predicted value sum+label value sum+1), the Dice loss is 1-Dice coefficient, and the output shape is [B,1,1].

[0170] 6) Combination loss and batch averaging: The combination loss is 0.7 × weighted BCE loss + 0.3 × Dice loss, with shape [B,1,1]; then the scalar loss is obtained by averaging over all dimensions and used for backpropagation.

[0171] Step 4, Model Training

[0172] The hardware used was an NVIDIA GeForce RTX 3090, and the software environment was Python 3.12 and PyTorch 1.10+. The dataset was divided into training and testing sets in an 8:2 ratio. The training set was further divided into training and validation subsets in a 9:1 ratio. The final proportions of the training, validation, and testing sets were 72%, 8%, and 20%, respectively, and each subset was randomly shuffled.

[0173] Training hyperparameter configuration: Adam optimizer is selected, with an initial learning rate of 1e-3; adaptive learning rate scheduling based on validation set performance is adopted. If the validation set loss does not decrease for two consecutive rounds, the learning rate is multiplied by a decay factor of 0.5, and the lower limit is set to 1e-6; gradient clipping restricts the L2 norm to 1.0; the total number of training rounds is 30, the validation set performance is monitored in real time, and the optimal model is saved for testing.

[0174] A training strategy combining pre-training and fine-tuning is employed: First, Gaussian smoothed labels are binarized with a threshold of 0 to obtain a wide range of lightning labels. Based on this, the pre-trained model learns to identify salient targets. Then, based on the pre-trained model, high-risk region labels are used, and a learning rate of 1e-4 is set for fine-tuning to optimize the recognition and reconstruction results. The specific performance of each module in the lightning recognition model of this invention and its advantages compared to other mainstream models are as follows:

[0175] (1) Effects of each module

[0176] To verify the improvement of target region spatial localization and pixel-level mask reconstruction performance by each image feature optimization module, this embodiment adopts a progressive approach, embedding the modules sequentially into the basic image segmentation network (UNet), and conducting ablation experiments to test the core visual metrics of the model at each stage. Combining quantitative evaluation metrics and visualization of the predicted residual image, the system analyzes the localization accuracy and reconstruction quality of the baseline image model and the final MDE-UNet model under different improvement strategies (metrics are shown in Table 1, recognition results are shown in...). Figure 8 See residual heatmap Figure 9 The configurations of each model module are as follows: the basic UNet contains no additional modules; CA-UNet adds CA·Block; Fusion-UNet further adds Fusion·Block; WWMLP-UNet adds WWMLP-Block; and MDE-UNet contains all modules.

[0177] Table 1. Indicators for different models

[0178]

[0179] This invention addresses the task of localizing and reconstructing lightning target images. Through joint analysis of quantitative evaluation metrics and residual visualization results, it fully verifies the effectiveness of various improvement strategies and the superiority of technical solutions. The core metric CSI improves stepwise from 0.4594 in the basic UNet to 0.5530 in the MDE-UNet of this invention, an increase of 20.37%, confirming the scientific validity of the phased embedding optimization module strategy. Specifically, after introducing the channel attention mechanism in CA-UNet, CSI increases to 0.4661, FAR decreases by 3.04 percentage points to 32.49%, and SSIM, PSNR, and accuracy all show slight improvements. The number of false positive regions in the residual map is reduced, indicating that this module effectively enhances feature collaboration and suppresses redundant interference. Fusion-UNet, through a multi-scale fusion module, further improves CSI to 0.4710, reduces FAR to 30.91%, slightly optimizes reconstruction metrics, and the residual map shows improved spatial matching of lightning regions and reduced missed details, proving its enhanced contour and detail capture capabilities. WWMLP-UNet, employing a weighted sliding window perceptron, achieved a performance breakthrough: FAR was significantly reduced to 17.12%, CSI jumped to 0.5306, and SSIM, PSNR, and accuracy were significantly optimized. The false alarm area in the residual map was greatly reduced, verifying its strong suppression of the false alarm rate. Finally, MDE-UNet, incorporating a radar enhancement module, saw POD rise to the highest value among all models at 0.6259, FAR only slightly increased to 17.40%, and CSI reach a peak of 0.5530. Reconstruction metrics were simultaneously optimized, and the residual results were consistent with the quantized data, highlighting the synergistic improvement of radar physical prior information on positioning accuracy, reconstruction quality, and detection reliability.

[0180] The various visual feature optimization modules complement each other strongly: channel attention lays the foundation for noise reduction, multi-scale fusion enhances spatial consistency, weighted sliding window significantly suppresses false alarms, and radar enhancement specifically improves missed detections. The quantitative indicators and residual visualization results are in high agreement, fully demonstrating that this invention achieves optimal comprehensive computer vision performance in target image recognition and reconstruction tasks. The final MDE-UNet model achieves optimal comprehensive performance in the task of locating and reconstructing lightning target images, possessing outstanding practical value.

[0181] (2) Advantages of comparing with mainstream models

[0182] To verify the segmentation performance of MDE-UNet, mainstream models such as Deeplabv3, SegNet, BiSeNet, and HRNet were selected as benchmarks for comparative experiments. (See Table 2 for details.) Figure 10-11As can be seen, MDE-UNet significantly outperforms the comparison models in all core metrics, demonstrating outstanding comprehensive technical advantages. The results show that the functional modules and overall network architecture designed in this invention are reasonable and practical, effectively overcoming the performance bottlenecks of existing models, verifying the innovation and application value of the technical solution, and providing a better implementation path for image segmentation tasks.

[0183] Table 2. Indicators for Each Model

[0184]

[0185] In the task of localizing lightning target images, the MDE-UNet model proposed in this invention leads in all core metrics: the Critical Success Index (CSI) reaches 0.5430, ranking first, which is 26.47%, 13.02%, 8.58%, and 8.56% higher than DeeplabV3, SegNet, BiSeNet, and HRNet, respectively; the Point of Detection (POD) is 0.6210, which is better than the comparison models except HRNet; and the False Alarm Rate (FAR) is as low as 0.1878, which is the lowest among all models, achieving "high hit rate and low false alarm rate" for target localization.

[0186] In terms of image reconstruction quality, MDE-UNet also performs exceptionally well: its structural similarity (SSIM) reaches 0.9367, a 30.67% improvement over BiSeNet; its peak signal-to-noise ratio (PSNR) is 24.32 dB, a 1.91 dB improvement over SegNet; and its accuracy reaches 0.9904, ranking first among all models. These metrics demonstrate that the model achieves synergistic optimization in localization accuracy and reconstruction quality, effectively overcoming the shortcomings of existing models in areas such as false alarm control, target acquisition integrity, and reconstruction distortion.

[0187] Combined with visualization results ( Figure 12 Further verification shows that MDE-UNet has outstanding technical advantages and practical value in lightning target image detection and image reconstruction tasks. The network architecture is reasonably designed, the optimization modules are effective, and the overall performance is at the leading level in the industry.

Claims

1. A lightning target recognition model, characterized in that, include: The encoder module is used to extract multi-scale image features from the input radar echo image data and satellite multi-channel brightness temperature image data, and output coded feature maps at different levels. The decoder module, which is connected in skip connection to the encoder module, is used to upsample and restore the encoded feature map step by step, and fuse it with the encoded feature map of the corresponding level to output a multi-level decoded feature map; A multi-scale feature fusion module is connected to the output of each level of the encoder module to realize cross-scale image feature alignment and global semantic fusion, and generate a global fusion feature map. The radar information enhancement module is connected to the output of the multi-scale feature fusion module. It is used to perform gain correction on the fused image features using the spatial physical constraints of the radar image and output the final lightning target probability map.

2. The lightning target recognition model according to claim 1, characterized in that, The encoder module includes multiple cascaded encoding layers, each encoding layer containing a dual convolution module, a channel attention mechanism module, and a max pooling layer; The dual convolution module includes two sets of two-dimensional convolutional layers. The dual convolution module is then connected to a channel attention mechanism module, which is used to perform channel-level recalibration on the feature map output by the convolution, and then output the encoded feature map through max pooling downsampling.

3. The lightning target recognition model according to claim 1, characterized in that, The decoder module includes multiple cascaded decoding layers, each containing a multilayer perceptron-based enhanced decoding unit based on a weighted sliding window. The enhanced decoding unit includes: The upsampling subunit uses a transposed convolutional layer to double the spatial resolution of the input feature map; The splicing subunit is used to splice the upsampled feature map with the corresponding level's encoded feature map along the channel dimension. The sliding window aggregation subunit is used to perform reflection filling on the stitched feature map, extract the neighborhood image features of each pixel in a sliding window manner, and achieve local context aggregation through weighted summation; The fully connected mapping subunit, consisting of multiple fully connected layers, ReLU activation functions, and batch normalization layers, is used to map the aggregated neighborhood features to an output feature map with the target number of channels.

4. The lightning target recognition model according to claim 1, characterized in that, The multi-scale feature fusion module includes: Multiple feature transformation modules are connected one-to-one with the output terminals of each level of the encoder module; The splicing unit is used to splice feature maps from different levels along the channel dimension; A dual convolutional module is used to extract cross-scale convection organization features; and a fusion output unit includes a channel attention mechanism module and residual connections; The feature transformation module includes: a channel adaptation subunit, which uses a 1×1 convolutional layer to transform the number of channels; a spatial alignment subunit, which uses bilinear interpolation to interpolate the spatial size to a preset target size; and a channel attention mechanism module for channel-level recalibration.

5. The lightning recognition model according to claim 1, characterized in that, The radar image information enhancement module includes: The radar image feature extraction unit is connected to the radar echo image data input terminal and is used to extract radar feature maps from the radar echo image data. The confidence generation unit is connected to the output of the radar image feature extraction unit and is used to map the radar feature map into a radar confidence weight map of [0,1]. The residual compensation unit is connected to the output of the multi-scale feature fusion module and the output of the confidence generation unit, respectively. It is used to weight and fuse the global fused feature map with the radar confidence weight map according to the preset compensation weight to generate the corrected lightning probability prediction map.

6. The lightning target recognition model according to claim 1, characterized in that, The channel attention mechanism module includes: The global pooling unit is used to perform global average pooling on the input feature map, compressing the spatial dimension to obtain channel-level global features; The weight learning unit, comprising a dimension-reducing fully connected layer, a ReLU activation function, and a dimension-upgrading fully connected layer, is used to map the channel-level global features to channel attention weights between 0 and 1. The feature recalibration unit is used to multiply the channel attention weights with the input feature map channel by channel to output the enhanced feature map.

7. A lightning target recognition method addressing image input modal imbalance and high false alarm rate, characterized in that, Includes the following steps: Historical multi-source meteorological image data is acquired and preprocessed. The multi-source meteorological image data includes lightning location data, satellite multi-channel brightness temperature image data, and radar echo image data. The lightning location data is preprocessed into images to generate high-risk lightning domain images with spatial probability distribution characteristics as lightning target labels. The radar echo image data is spatiotemporally aligned to align the radar echo image sequence with the satellite multi-channel brightness temperature image data and the lightning target image labels in both time and space. A lightning target recognition model is constructed using an encoder-decoder backbone architecture; The lightning recognition model is trained using a loss function to optimize its lightning target recognition capability, resulting in a well-trained lightning target recognition model. After real-time acquisition of satellite multi-channel brightness temperature image data and radar echo image data, spatiotemporal alignment processing is performed, and the data is used as a multi-source heterogeneous image input to a trained lightning target recognition model. The lightning target recognition model outputs a probability map of lightning target occurrence in the future time period.

8. The multi-source fusion lightning target identification method according to claim 7, characterized in that, The satellite multi-channel brightness temperature image data includes Band9 channel data, TBB13 channel brightness temperature data, brightness temperature difference data between TBB15 and TBB13 channels, and composite index (Band11-TBB13)-(TBB13-TBB15) from geostationary meteorological satellites.

9. The multi-source fusion lightning identification method according to claim 7, characterized in that, The lightning location data is preprocessed to generate high-risk lightning regions as lightning labels. The specific steps are as follows: An isotropic two-dimensional Gaussian distribution is constructed for the location of a single lightning strike to characterize the probability of lightning occurring in the surrounding area. , defined as the following formula, where σ is the standard deviation of the Gaussian distribution, ( , ( ) represents the coordinates of the lightning location point: A time decay weighting strategy is used to fuse Gaussian distribution results from multiple historical time steps to construct a probability value P for high-density lightning regions. Peak truncation and binarization are then performed on the probability value P to generate labels for high-risk lightning regions.

10. The lightning target identification method according to claim 7, characterized in that, The spatiotemporal alignment processing of the radar echo data includes: downsampling the original radar image from a first resolution to a second resolution to match the spatial resolution of the satellite image data and the lightning tag image; and interpolating the radar echo data using optical flow combined with bilinear interpolation to generate an interpolated frame radar echo image that matches the lightning tag time window; and training the lightning target recognition model using an asymmetric weighted BCE-DICE composite loss function to optimize the model's lightning recognition performance under conditions of extreme imbalance between positive and negative image samples. The BCE-DICE composite loss function is expressed as follows: in, , These are the weighting coefficients for the weighted BCE loss and the Dice loss, respectively. in, Let be the predicted probability of the pixel at position (h,w) in the b-th sample by the model; Let H be the true label at position (h,w) in the b-th sample, B be the total number of image samples, W be the number of pixels in the horizontal direction of the image, and H be the number of pixels in the vertical direction of the image. FP indicates a false positive, and FN indicates a false negative. This is the false positive weight, i.e., the misreporting weight. This is the false negative weight, i.e., the underreporting weight. Time weighting.

Images