A method, system, equipment, and medium for identifying nearshore thunderstorms and strong winds based on DCSAM-TransUnet and transfer learning.

Abstract

This invention provides a method, system, device, and medium for identifying nearshore thunderstorms and strong winds based on DCSAM-TransUnet and transfer learning, belonging to the field of meteorological disaster monitoring technology. The method includes: acquiring multi-source meteorological data and creating thunderstorm and strong wind location labels; constructing satellite-retrieved radar data using a frequency-decoupled LightGBM radar echo inversion model to compensate for missing marine radar data; constructing a TransUnet model containing a two-layer, multi-scale channel spatial attention mechanism and pre-training it in the source domain; transferring the model to the target domain based on a dynamic freezing strategy using feature cosine similarity and an adversarial domain adaptation strategy; inputting marine data to obtain preliminary identification results, and introducing physical constraints, morphological filtering, and Kalman filtering for spatiotemporal correction, outputting the final result. This invention effectively solves the problems of scarce nearshore radar data and poor cross-domain adaptability of the model, improving the accuracy and spatiotemporal continuity of identification.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing and disaster early warning technology, and more particularly to a method, system, device and medium for identifying nearshore thunderstorms and strong winds based on DCSAM-TransUnet and transfer learning. Background Technology

[0002] Thunderstorms and strong winds are highly destructive convective weather events that pose a serious threat to near-shore shipping, maritime operations, and the safety of coastal cities. Radar data, capable of accurately detecting precipitation particles and airflow movement, is the core data for identifying thunderstorms and strong winds. In land areas, relying on a well-developed radar observation network, thunderstorm and strong wind identification technology is relatively mature.

[0003] However, nearshore areas face severe challenges: First, due to the curvature of the Earth and deployment costs, there is a lack of real-time radar data with high spatiotemporal resolution at sea, creating a "data gap"; second, the nearshore underlying surface differs significantly from that of land, resulting in different wind field evolution patterns, and directly transferring land-trained models to the sea will lead to a significant decrease in recognition accuracy due to domain shift; finally, existing methods mostly rely on single-moment pattern recognition, lack physical mechanism constraints and spatiotemporal continuity, and are prone to generating false alarms and spatial fragmentation noise.

[0004] While existing technologies attempt to use satellite data for inversion or direct identification, the physical quantities observed by satellite differ significantly from those in radar echoes, and it is difficult to capture microphysical processes. How to leverage abundant land-based data to overcome the lack of maritime data and regional differences, and achieve highly reliable identification of nearshore thunderstorms and strong winds, is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0005] This invention aims to provide a method, system, device, and medium for identifying nearshore thunderstorms and strong winds based on DCSAM-TransUnet and transfer learning. It fills data gaps through satellite inversion, solves the domain adaptation problem by utilizing an improved attention mechanism and transfer learning strategy, and improves the reliability of results by combining physical constraints.

[0006] The first aspect of this invention discloses a nearshore thunderstorm and strong wind identification method based on DCSAM-TransUnet and transfer learning, the method comprising: Step S1: Obtain historical data from the source and target domains, including real-time radar data, wind speed observation data, satellite observation data, and numerical model wind field data. Preprocess the data and create thunderstorm wind point labels. Step S2: Based on the preprocessed satellite feature data, a frequency decoupling mechanism is constructed using fast Fourier transform, the LightGBM radar echo inversion model is trained, and the simulated radar echo data retrieved from the satellite is output. Step S3: Based on the source domain real-time radar and numerical model wind field data, construct a DCSAM-TransUnet thunderstorm and strong wind identification model containing a two-layer multi-scale channel spatial attention mechanism DCSAM, pre-train the model, and obtain the optimal weights in the source domain. Step S4: Load the source domain optimal weights, and based on the target domain satellite inversion radar echo simulation data, source domain real radar data and numerical model wind field data, use dynamic freezing strategy and adversarial domain adaptation strategy to alternately retrain the model to obtain the target domain optimal weights. Step S5: Load the optimal weight of the target domain, input real-time marine satellite inversion radar echo and numerical model wind field data, and output the preliminary identification results of marine thunderstorms and strong winds; Step S6: Introduce a numerical model environment field, verify the preliminary identification results based on physical constraints, and perform spatiotemporal smoothing processing by combining morphological filtering and Kalman filtering to output the final identification results of marine thunderstorms and strong winds.

[0007] Furthermore, in step S1, the method for creating thunderstorm wind point tags includes: Set a wind speed threshold, and mark the observation point where the instantaneous wind speed reaches the threshold and its surrounding preset range as the potential area; Calculate the morphological indices of the radar situation within the potential area, including the maximum combined reflectivity and the convective structure intensity index; The convective structure intensity index is calculated based on a weighted average of gradient intensity characteristics, shape characteristics, and gradient variation characteristics. The calculation formula is as follows: ; Where w1, w2, w3 are weighting coefficients, F grad For gradient strength characteristics, F shape For shape characteristics, F var This is a gradient variation feature; When both the maximum combined reflectivity and the convective structure intensity index meet the preset conditions, the location of the thunderstorm and strong wind is confirmed.

[0008] Furthermore, in step S2, the specific implementation of the frequency decoupling mechanism includes: Fast Fourier transform is performed on real-time radar data and satellite-retrieved radar data to calculate the amplitude spectrum and phase spectrum, respectively; Construct a mask matrix to locally replace the amplitude spectrum of satellite-retrieved radar data while keeping the phase spectrum unchanged; The mask matrix is ​​used to retain the amplitude information of satellite inversion data in the low-frequency region at the center of the spectrum, and to introduce the amplitude information of real-time radar data in the high-frequency region; Enhanced satellite-inverted radar data is generated by using the mixed amplitude spectrum and the original phase spectrum through inverse Fourier transform. The formula for calculating the amplitude spectrum after mixing is: ; Where M is the mask matrix, For element-wise multiplication, A sat For the amplitude spectrum of satellite inversion data, A radar This is the amplitude spectrum of real-time radar data.

[0009] Furthermore, in step S3, the DCSAM-TransUnet thunderstorm and strong wind identification model includes an encoder, a DCSAM module, and a decoder; The DCSAM module includes a multi-scale channel attention mechanism MS-CAM and a multi-scale spatial attention mechanism MS-SAM. The MS-CAM extracts channel features using global and local branches, where the global branch includes pooling operations and the local branch includes convolution operations. The MS-SAM extracts spatial features using multi-scale convolutional branches and spatial saliency-guided branches, wherein the multi-scale convolutional branches contain convolutional layers with convolutional kernels of different sizes.

[0010] Furthermore, the calculation formulas for MS-CAM and MS-SAM are as follows: The feature fusion output of MS-CAM is: ; Where X represents the input features, σ is the sigmoid function, L(X) is the local branch output, G(X) is the global branch output, and ⊕ denotes convolutional fusion after feature concatenation. This represents element-wise multiplication; The feature fusion output of MS-SAM is: ; Where X′ is the output of MS-CAM, S scale For the output of multi-scale convolution branches, S spatial The spatial saliency guides the branch output.

[0011] Furthermore, in step S4, the dynamic freezing strategy includes: Calculate the cosine similarity between the feature vectors of satellite-inverted radar data and the feature vectors of real-world radar data; Based on the magnitude of cosine similarity, the different levels of the encoder and the Transformer module parameters of the DCSAM-TransUnet thunderstorm and strong wind recognition model are frozen in stages. The specific strategy is as follows: when the cosine similarity is greater than or equal to the first threshold, the number of frozen layers is the highest; when the cosine similarity is less than the second threshold, the number of frozen layers is the lowest; when the cosine similarity is between the first and second thresholds, the number of frozen layers is in the middle.

[0012] Furthermore, in step S4, the adversarial domain adaptation strategy includes: The domain discriminator module is connected after the encoder in the DCSAM-TransUnet thunderstorm and strong wind identification model. Construct a total loss function that includes the main model loss and the domain classifier loss; Use alternating training methods: Phase 1: Freeze the domain discriminator parameters and update the main model parameters to minimize the main model loss and maximize the domain classifier loss, making the extracted features difficult to distinguish their source. Phase 2: Freeze the main model parameters and update the domain discriminator parameters to minimize the domain classifier loss and improve the discriminator's ability to distinguish between source and target domain data.

[0013] Furthermore, in step S6, the verification based on physical constraints includes: Acquire the corresponding spatiotemporal numerical model environmental field data, including boundary layer wind speed, convective available potential energy (CAPE), and K exponent; Grid-by-grid evaluation: If any of the following conditions are met, the recognition result is retained; otherwise, it is discarded: The boundary layer wind speed is greater than or equal to the first preset threshold. CAPE is greater than or equal to the second preset threshold; The K-index is greater than or equal to the third preset threshold.

[0014] Furthermore, in step S6, the Kalman filtering process includes: Define the state vector of a thunderstorm cell as its center position coordinates and moving speed; A state prediction equation is constructed based on the assumption of uniform motion. The centroid of thunderstorm cells is extracted from the preliminary identification results as the observation value, and the nearest neighbor algorithm is used to associate the observation value with the predicted trajectory. The state vector is updated using the Kalman filter recursive formula to correct the identification position of thunderstorm cells and remove trajectories that have not been matched for multiple consecutive frames.

[0015] The second aspect of this invention discloses a nearshore thunderstorm and strong wind identification system based on DCSAM-TransUnet and transfer learning, the system comprising: The first processing module is configured to acquire historical data from the source and target domains, preprocess the data, and create thunderstorm and large-scale wind point labels. The second processing module is configured to use the fast Fourier transform to construct a frequency decoupling mechanism based on the preprocessed satellite feature data, train the LightGBM radar echo inversion model, and output the simulated radar echo data retrieved from the satellite. The third processing module is configured to construct a DCSAM-TransUnet thunderstorm and strong wind identification model based on source domain real-time radar and numerical model wind field data, which includes a two-layer multi-scale channel spatial attention mechanism DCSAM, and pre-train the model to obtain the optimal weights in the source domain. The fourth processing module is configured to load the source domain optimal weights, and based on the target domain satellite inversion radar echo simulation data, source domain real-time radar data and numerical model wind field data, alternately retrain the model using a dynamic freezing strategy and an adversarial domain adaptation strategy to obtain the target domain optimal weights. The fifth processing module is configured to load the optimal weight of the target domain, input real-time marine satellite inversion radar echo and numerical model wind field data, and output preliminary identification results of marine thunderstorms and strong winds. The sixth processing module is configured to introduce a numerical model environment field, verify the preliminary identification results based on physical constraints, and perform spatiotemporal smoothing processing by combining morphological filtering and Kalman filtering to output the final identification results of marine thunderstorms and strong winds.

[0016] A third aspect of this invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the nearshore thunderstorm and strong wind identification method based on DCSAM-TransUnet and transfer learning according to any one of the first aspects of this disclosure.

[0017] A fourth aspect of this invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a nearshore thunderstorm and strong wind identification method based on DCSAM-TransUnet and transfer learning, as described in any of the first aspects of this disclosure.

[0018] The beneficial effects of this invention are as follows: 1. Overcoming data bottlenecks: By using the frequency-decoupled LightGBM radar echo inversion model, high-quality simulated radar echoes are generated using satellite data. This preserves structural consistency and introduces the texture details of real radar, solving the problem of missing near-shore real-time radar data.

[0019] 2. Enhance cross-domain generalization ability: The DCSAM mechanism enhances the model's ability to extract multi-scale features of thunderstorms; the dynamic freezing and adversarial domain adaptation strategies effectively transfer knowledge learned on land to the nearshore area, significantly reducing the performance degradation caused by underlying surface differences.

[0020] 3. Enhanced reliability of results: By introducing physical constraints such as dynamics, thermodynamics, and water vapor, as well as Kalman filter timing correction, non-physical false signals are eliminated, ensuring the spatiotemporal consistency of the recognition results and better meeting the needs of business early warning. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a nearshore thunderstorm and strong wind identification method based on DCSAM-TransUnet and transfer learning according to an embodiment of the present invention; Figures 2a-2b This is a basic model framework diagram (source domain) according to an embodiment of the present invention. Figure 3 This is a framework diagram of the migration model (target domain) according to an embodiment of the present invention. Figure 4 This is a structural diagram of a nearshore thunderstorm and strong wind identification system based on DCSAM-TransUnet and transfer learning according to an embodiment of the present invention. Figure 5 This is a structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0024] The first aspect of this invention discloses a method for identifying nearshore thunderstorms and strong winds based on DCSAM-TransUnet and transfer learning.

[0025] Example 1: Figure 1 This is a flowchart illustrating a nearshore thunderstorm and strong wind identification method based on DCSAM-TransUnet and transfer learning according to an embodiment of the present invention. Figure 1 As shown, the method includes: Step S1: Obtain historical data from the source and target domains, including real-time radar data, wind speed observation data, satellite observation data, and numerical model wind field data. Preprocess the data and create thunderstorm wind point labels. Specifically, historical data from land (source domain) and sea (target domain) are acquired, including real-time radar data, automatic weather station wind speeds, FY4B satellite data, and numerical model wind fields. The data is cleaned, resampled, and combined with wind speed thresholds and morphological indicators to create high-precision thunderstorm wind point labels.

[0026] In step S1, the method for creating thunderstorm large-scale point tags includes: Set a wind speed threshold, and mark the observation point where the instantaneous wind speed reaches the threshold and its surrounding preset range as the potential area; Calculate the morphological indices of the radar situation within the potential area, including the maximum combined reflectivity and the convective structure intensity index; The convective structure intensity index is calculated based on a weighted average of gradient intensity characteristics, shape characteristics, and gradient variation characteristics. The calculation formula is as follows: ; Where w1, w2, w3 are weighting coefficients, F grad For gradient strength characteristics, F shape For shape characteristics, F var This is a gradient variation feature; When both the maximum combined reflectivity and the convective structure intensity index meet the preset conditions, the location of the thunderstorm and strong wind is confirmed.

[0027] In some specific embodiments, data processing and label creation are as follows: (1) Satellite data: Six core channels (5, 9, 11, 13, 14, 15) of the FY4B satellite were selected, and three key channel differences were constructed (9-11 represent water vapor gradient, 13-14 represent cloud top cooling, and 13-15 represent cloud top texture). Outliers were removed using the 3σ criterion and aggregated to a 1-hour resolution.

[0028] (2) Radar data: The radar composite reflectivity mosaic was resampled to a resolution of 4km and 1h and aligned with the satellite data.

[0029] (3) Tag creation: A potential area with a radius of 5km is defined, centered on a station with a wind speed ≥17.2m / s. The morphological parameters of the radar echoes within the potential area are calculated: Maximum combined reflectivity, CR max ≥40dBz; Convection structure intensity index I conv =0.4×F grad +0.4×F shape+0.2×F var ≥0.5.

[0030] Among them, F grad The gradient strength, F, is calculated using the Sobel operator. shape The eccentricity F is calculated using the second-order central moment. var The gradient standard deviation is calculated. This ensures that the labels are not only based on wind speed, but also on typical thunderstorm convective structure characteristics.

[0031] Step S2: Based on the preprocessed satellite feature data, a frequency decoupling mechanism is constructed using fast Fourier transform, the LightGBM radar echo inversion model is trained, and the simulated radar echo data retrieved from the satellite is output. This step specifically involves radar echo inversion. Based on the multi-channel characteristics of satellites, a radar echo inversion model is established using LightGBM. To address the distribution discrepancy between the inverted data and the real data, an innovative frequency decoupling mechanism is introduced to mix the amplitude spectrum of the real radar in the frequency domain, generating enhanced satellite-inverted radar data.

[0032] In step S2, the specific implementation of the frequency decoupling mechanism includes: Fast Fourier transform is performed on real-time radar data and satellite-retrieved radar data to calculate the amplitude spectrum and phase spectrum, respectively; Construct a mask matrix to locally replace the amplitude spectrum of satellite-retrieved radar data while keeping the phase spectrum unchanged; The mask matrix is ​​used to retain the amplitude information of satellite inversion data in the low-frequency region at the center of the spectrum, and to introduce the amplitude information of real-time radar data in the high-frequency region; Enhanced satellite-inverted radar data is generated by using the mixed amplitude spectrum and the original phase spectrum through inverse Fourier transform. The formula for calculating the amplitude spectrum after mixing is: ; Where M is the mask matrix, For element-wise multiplication, A sat For the amplitude spectrum of satellite inversion data, A radar This is the amplitude spectrum of real-time radar data.

[0033] In some specific embodiments, the frequency-decoupled radar inversion is as follows: A radar echo retrieval model is established using LightGBM to map satellite data to radar echoes. To further improve the accuracy of the retrieval data, a frequency decoupling mechanism is employed. Perform an FFT transform on the real radar R and the inverted radar S to obtain the amplitude spectrum A. R A S and phase spectrum P R ,PS The phase spectrum determines the structure of the image (the location and shape of the thunderstorm), while the amplitude spectrum determines the texture (the details of the intensity distribution of the echoes).

[0034] A mask matrix M is constructed, preserving the amplitude information retrieved from satellites in the low-frequency region (center) and introducing the amplitude information from the real-time radar in the high-frequency region (periphery). The mixing formula is: ; Using A mix and the original phase spectrum P S An inverse transformation is performed to obtain enhanced data. This ensures that the retrieved data retains the accuracy of the thunderstorm locations observed by satellite while also possessing the texture characteristics of real radar, which is beneficial for subsequent transfer.

[0035] Step S3: Based on the source domain real-time radar and numerical model wind field data, construct a DCSAM-TransUnet thunderstorm and strong wind identification model containing a two-layer multi-scale channel spatial attention mechanism DCSAM, pre-train the model, and obtain the optimal weights in the source domain. This step specifically involves source domain pre-training to construct an improved DCSAM-TransUnet thunderstorm and strong wind recognition model. This model incorporates a two-layer, multi-scale channel spatial attention mechanism. Pre-training is performed using abundant real-time radar data from the source domain, enabling the model to learn the core features of thunderstorms and strong winds.

[0036] In step S3, the DCSAM-TransUnet thunderstorm and strong wind identification model includes an encoder, a DCSAM module, and a decoder; The DCSAM module includes a multi-scale channel attention mechanism MS-CAM and a multi-scale spatial attention mechanism MS-SAM. The MS-CAM extracts channel features using global and local branches, where the global branch includes pooling operations and the local branch includes convolution operations. The MS-SAM extracts spatial features using multi-scale convolutional branches and spatial saliency-guided branches, wherein the multi-scale convolutional branches contain convolutional layers with convolutional kernels of different sizes.

[0037] The calculation formulas for MS-CAM and MS-SAM are as follows: The feature fusion output of MS-CAM is: ; Where X represents the input features, σ is the sigmoid function, L(X) is the local branch output, G(X) is the global branch output, and ⊕ denotes convolutional fusion after feature concatenation. This represents element-wise multiplication; The feature fusion output of MS-SAM is: ; Where X′ is the output of MS-CAM, S scale For the output of multi-scale convolution branches, S spatial The spatial saliency guides the branch output.

[0038] In some specific embodiments, the model architecture is as follows: Figures 2a-2b As shown, the encoder uses ResNet50+Transformer, and the decoder uses cascaded upsampling.

[0039] The core innovation lies in the DCSAM module (dual-layer multi-scale channel spatial attention): MS-CAM (Channel Attention): Identifies key feature channels by fusing global branches (pooling) and local branches (convolution).

[0040] MS-SAM (Spatial Attention): Through multi-scale convolution (3x3, 5x5, 7x7) and spatial saliency branching, the model focuses on the core region of thunderstorm cells at different scales.

[0041] Step S4: Load the source domain optimal weights, and based on the target domain satellite inversion radar echo simulation data, source domain real radar data and numerical model wind field data, use dynamic freezing strategy and adversarial domain adaptation strategy to alternately retrain the model to obtain the target domain optimal weights. This step is transfer learning. By designing a dynamic freezing strategy (based on feature similarity) and an adversarial domain adaptation strategy (based on domain discriminator), the model is adapted to the target domain (sea) to solve the domain drift problem caused by land-sea differences.

[0042] In step S4, the dynamic freezing strategy includes: Calculate the cosine similarity between the feature vectors of satellite-inverted radar data and the feature vectors of real-world radar data; Based on the magnitude of cosine similarity, the different levels of the encoder and the Transformer module parameters of the DCSAM-TransUnet thunderstorm and strong wind recognition model are frozen in stages. The specific strategy is as follows: when the cosine similarity is greater than or equal to the first threshold, the number of frozen layers is the highest; when the cosine similarity is less than the second threshold, the number of frozen layers is the lowest; when the cosine similarity is between the first and second thresholds, the number of frozen layers is in the middle.

[0043] In step S4, the adversarial domain adaptation strategy includes: The domain discriminator module is connected after the encoder in the DCSAM-TransUnet thunderstorm and strong wind identification model. Construct a total loss function that includes the main model loss and the domain classifier loss; Use alternating training methods: Phase 1: Freeze the domain discriminator parameters and update the main model parameters to minimize the main model loss and maximize the domain classifier loss, making the extracted features difficult to distinguish their source. Phase 2: Freeze the main model parameters and update the domain discriminator parameters to minimize the domain classifier loss and improve the discriminator's ability to distinguish between source and target domain data.

[0044] In some specific embodiments, the transfer learning strategy is as follows: (1) Dynamic freezing: Calculate the cosine similarity Sim between the features of the target domain (satellite inversion) and the features of the source domain (real-time radar).

[0045] Sim≥0.75 (high similarity): Freeze the first two layers of the encoder and the Transformer to prevent overfitting.

[0046] $0.5\leSim $0.75: Freeze the first three layers and the Transformer.

[0047] Sim 0.5 (Large Difference): Only freezes the previous layer, allowing for more parameter fine-tuning to accommodate differences.

[0048] (2) Adversarial domain adaptation: Access domain discriminator, using alternating training: Phase 1: Update the main model to maximize the loss of the domain discriminator (deceive the discriminator so that it cannot distinguish the data source).

[0049] Phase 2: Update the discriminator and minimize the classification loss (improve discrimination ability).

[0050] Through game theory, the encoder is forced to extract thunderstorm features shared by both land and sea.

[0051] Step S5: Load the optimal weights of the target domain, input real-time marine satellite inversion radar echo and numerical model wind field data, and output the preliminary identification results of marine thunderstorms and strong winds; using the trained transfer model, input real-time marine satellite inversion data to obtain the preliminary identification results.

[0052] Step S6: Introduce the numerical model environmental field, verify the preliminary identification results based on physical constraints, and perform spatiotemporal smoothing processing using morphological filtering and Kalman filtering to output the final identification result of marine thunderstorms and strong winds. False alarms are eliminated by combining the numerical model environmental field (physical constraints), and Kalman filtering is used to track and smooth the trajectory of individual thunderstorm cells, outputting the final result.

[0053] In step S6, the verification based on physical constraints includes: Acquire the corresponding spatiotemporal numerical model environmental field data, including boundary layer wind speed, convective available potential energy (CAPE), and K exponent; Grid-by-grid evaluation: If any of the following conditions are met, the recognition result is retained; otherwise, it is discarded: The boundary layer wind speed is greater than or equal to the first preset threshold. CAPE is greater than or equal to the second preset threshold; The K-index is greater than or equal to the third preset threshold.

[0054] In step S6, the Kalman filtering process includes: Define the state vector of a thunderstorm cell as its center position coordinates and moving speed; A state prediction equation is constructed based on the assumption of uniform motion. The centroid of thunderstorm cells is extracted from the preliminary identification results as the observation value, and the nearest neighbor algorithm is used to associate the observation value with the predicted trajectory. The state vector is updated using the Kalman filter recursive formula to correct the identification position of thunderstorm cells and remove trajectories that have not been matched for multiple consecutive frames.

[0055] In some specific embodiments, step S6 is described in detail, and the physical constraints and Kalman filtering post-processing are as follows: 1. Physical constraint verification Preliminary identification results may include strong echoes not caused by thunderstorms or strong winds (such as ordinary precipitation). A numerical model environmental field is introduced for multidimensional verification, retaining grid points that satisfy any of the following conditions: Dynamic conditions: Boundary layer wind speed ≥ 10 m / s (strong shear support).

[0056] Thermal conditions: Convective available potential energy (CAPE) ≥ 500 J / kg (energy support).

[0057] Water vapor conditions: K index ≥ 20 (stratification unstable).

[0058] 2. Morphological filtering The verified results are subjected to an "erosion-dilation" operation to remove isolated noise points with an area smaller than 3 grid points and smooth the edges of thunderstorms.

[0059] 3. Kalman Filter Spatiotemporal Smoothing Treating thunderstorm cells as moving targets, Kalman filtering is used to solve the problems of missed detections and inconsistent trajectories.

[0060] (1) Definition of state vector: The state of each thunderstorm cell at time k is defined as its center position and moving velocity. The thunderstorm cell is regarded as a moving target. The state vector is:

[0061] in and These are the longitude and latitude coordinates of the center point (centroid). and The speed of movement in the east-west and north-south directions, respectively.

[0062] (2) State prediction model: Assuming that the thunderstorm system approximately follows a uniform motion model in the short term, the state prediction equation is: ; in, It is a priori state estimation. This is a posterior state estimate, and the state transition matrix F is:

[0063] Δt is the time interval between adjacent recognition times.

[0064] (3) Observation model construction: The center position of the thunderstorm cell is extracted as the observation value from the preliminary identification results at the current moment. ; in To account for observation noise, the observation matrix H is: ; For scenarios involving multiple thunderstorm cells, a nearest neighbor data association algorithm is used to match trajectories with observations. Specifically, the Euclidean distance between the observation at time t and the centroid coordinates of all predicted trajectories at time t-1 is calculated, and the observation is assigned to the nearest trajectory. If no trajectory meets the conditions, it is determined to be a new thunderstorm cell and a new trajectory is initialized.

[0065] (4) Kalman filter recursive update Derivation of the current prior state and error covariance based on the historical best state:

[0066] in It is a priori state estimation. yes The posterior optimal state estimate at time t. yes The prior error covariance matrix at time t. yes The posterior error covariance matrix at time t, where Q is the process noise covariance, characterizing the uncertainty of thunderstorm movement, is statistically set based on historical thunderstorm movement data, and its value range is:

[0067] in These represent the process noise variances in the latitude and longitude directions, respectively. The variances of process noise are denoted as east-west and north-south velocity directions, respectively.

[0068] By correcting the prior state based on the current observations, the optimal posterior estimate is obtained:

[0069] in Here is the Kalman gain matrix. To observe the noise covariance matrix.

[0070] (5) Parameter initialization and trajectory management: the initial state x0 is determined by the first identification result, the initial velocity is set to zero, the initial error covariance P0 is set to a diagonal matrix, and the trajectory that has not matched the observation value for three consecutive iterations is automatically removed.

[0071] The Kalman filter is used to predict the position at the next time step, and then it is matched with the current observation (centroid) (nearest neighbor association).

[0072] By using observed values ​​to correct predicted values, the movement trajectory of thunderstorms can be smoothed, missed frames can be filled, and false targets with instantaneous jumps can be eliminated.

[0073] Example 2: The method in this embodiment includes the following steps: Step S1: Data Acquisition and Preprocessing Acquire historical radar data (source domain), wind speed observations, FY4B satellite observations (target domain / source domain), and numerical model wind field data.

[0074] 1. Satellite Data Processing: Six core channels of FY4B (5, 9, 11, 13, 14, 15) were selected as shown in Table 1, and outliers were removed using the 3σ criterion. The temporal resolution was aggregated to 1 hour (maximum value for channel 5, minimum value for the others). Key channel difference features were constructed: 9-11 (high and low level water vapor gradient), 13-14 (cloud top radiative cooling), and 13-15 (cloud top texture).

[0075] Table 1 2. Radar data processing: The radar data is resampled to a resolution of 4km and 1h. Maximum value synthesis is used temporally, and nearest neighbor interpolation is used spatially.

[0076] 3. Label making: The potential area is defined as a station with a wind speed of ≥17.2m / s and a radius of 5km.

[0077] Calculate the morphological parameters of radar echoes within the potential region: Maximum combined reflectance CRmax ≥ 40 dBz; The convective structure intensity index Iconv = 0.4 × Fgrad + 0.4 × Fshape + 0.2 × Fvar ≥ 0.5.

[0078] Fgrad is calculated as follows: ; ; ; ; ; in The horizontal gradient matrix, The vertical gradient matrix, The normalization constant is 15 dBz Fshape is calculated as follows: ; ; ; ; ; in For the second-order central torch of the binary region Ω with strong echo, , Let a and b be the centroid coordinates of the strong echo region, respectively; let a and b be the major and minor axes of the equivalent ellipse, respectively; and let E be the eccentricity. The shape F is obtained based on the ratio of the major and minor axes and the eccentricity. ; Fvar is calculated as follows:

[0079] in This represents the standard deviation of the gradient within the strong echo region. It is 0.1.

[0080] Among them, Fgrad represents the gradient intensity, Fshape represents the shape eccentricity, and Fvar represents the gradient variation characteristics. Computer vision algorithms are used to extract the edge gradients and geometric shapes of strong echo regions, effectively distinguishing disaster-causing structures such as gust fronts and bow echoes.

[0081] In some specific embodiments, Step S2: Construct a frequency-decoupled LightGBM radar echo inversion model 1. Basic Inversion: Input satellite 9-dimensional features (6 channels + 3 differences), output simulated radar echo ICR. Use the LightGBM model and employ joint loss MSE+MAE.

[0082] 2. Frequency decoupling data enhancement: To make the retrieved data more closely resemble real radar data in the frequency domain, a Fast Fourier Transform (FFT) is used to transform both the real and retrieved radar data into the frequency domain. In the frequency domain, the amplitude spectrum carries low-level statistical features such as radar echo texture details and intensity distribution, while the phase spectrum determines high-level structural information such as the overall morphology of thunderstorms and the arrangement of convective cells. Although satellite inversion data can simulate the approximate structure of thunderstorms through physical laws (the phase spectrum is effective), its low-level statistical features differ significantly from those of real radar data, which is a key factor in domain shift. Hybrid amplitude spectrum can preserve the structural integrity of satellite inversion data (phase spectrum unchanged) while imbuing it with the texture features of real radar data, effectively reducing the distribution difference between the source and target domains. This provides a high-quality data foundation for transfer learning and improves the model's cross-domain adaptability.

[0083] The size of the spectrum is The coordinates of the spectrum center are mask radius For any pixel in the spectrum Its Euclidean distance from the center ,when hour ,when hour .

[0084] Using the mask matrix M, the high-frequency amplitude spectrum information of the real-time radar is injected into the amplitude spectrum of the inverted radar. The formula is: Amix = M⊙Asat + (1 M)⊙Aradar.

[0085] By keeping the phase spectrum of the retrieved radar unchanged (the phase contains semantic structural information), an inverse phase transform (IFFT) is performed using Amix and the original phase spectrum to obtain enhanced retrieval data. This method preserves the structural consistency of satellite retrieval while introducing the textural details (high-frequency information) of the real-world radar, thus improving data quality.

[0086] Step S3: Construct and pre-train the DCSAM-TransUnet thunderstorm and strong wind recognition model 1. Model architecture: such as Figures 2a-2b As shown, the encoder uses ResNet50 to extract features, followed by linear projection and a 12-layer Transformer module to capture long-range dependencies; the decoder restores resolution through cascaded upsampling.

[0087] 2. DCSAM mechanism: Introduce a two-layer multi-scale channel spatial attention mechanism at the jump connection.

[0088] MS-CAM (Multi-Scale Channel Attention): It is divided into global branches (pooling) and local branches (convolution), which are then fused and activated by Sigmoid to make the model focus on important feature channels.

[0089] MS-SAM (Multi-Scale Spatial Attention): It consists of a multi-scale convolution branch (3x3, 5x5, 7x7 convolution) to capture different receptive field features, and a spatial saliency branch (pooling), which, when fused, allows the model to focus on the core region of thunderstorms.

[0090] 3. Pre-training: The model is trained using real-time radar and model wind field data from the source domain (land). The loss function is cross-entropy loss, and the optimizer is AdamW, to obtain the optimal weights from the source domain.

[0091] Step S4: Dynamic Freezing and Adversarial Domain Adaptation Migration like Figure 3 As shown, this addresses the domain differences between land and sea: 1. Dynamic Freeze: Calculate the cosine similarity Sim between satellite-inverted radar features and real-world radar features.

[0092] If Sim≥0.75 (high similarity), freeze the first two layers of the encoder and the Transformer; If 0.5 ≤ Sim 0.75, freeze the first three layers and the Transformer; If Sim 0.5 (significant difference), only freeze the previous layer.

[0093] This strategy avoids overfitting source domain features when there are large differences, or destroying learned features when there are similarities.

[0094] 2. Adversarial domain adaptation: A domain discriminator is added after the encoder. Alternating training is employed. Phase 1: Freeze the discriminator and train the main model so that the features it generates can "deceive" the discriminator (i.e., make the discriminator unable to distinguish whether the data is from the source domain or the target domain).

[0095] Phase Two: Freeze the main model and train the discriminator to accurately distinguish the data source.

[0096] By maximizing domain confusion, the encoder is forced to extract domain-invariant common features.

[0097] Step S5: Preliminary Identification Load the migrated model weights, input the satellite inversion radar data and model wind field at sea, and output a preliminary thunderstorm and gale probability map.

[0098] Step S6: Physical Constraints and Spatiotemporal Post-processing 1. Physical Verification: Introducing the environmental field from the numerical model. Only regions meeting one of the following conditions are retained: boundary layer wind speed ≥ 10 m / s (dynamic condition), CAPE ≥ 500 J / kg (thermal condition), K-index ≥ 20 (water vapor condition). This effectively eliminates false alarms of strong echoes not caused by thunderstorms.

[0099] 2. Morphological filtering: Erosion followed by dilation removes isolated noise with an area smaller than 3 grid points.

[0100] 3. Kalman filtering timing smoothing: In some specific embodiments, the details are as follows: (1) Definition of state vector: The state of each thunderstorm cell at time k is defined as its center position and moving velocity. The thunderstorm cell is regarded as a moving target. The state vector is: ; in and These are the longitude and latitude coordinates of the center point (centroid), respectively. and The speed of movement in the east-west and north-south directions, respectively.

[0101] (2) State prediction model: Assuming that the thunderstorm system approximately follows a uniform motion model in the short term, the state prediction equation is: ; The state transition matrix F is: ; Δt is the time interval between adjacent recognition times.

[0102] (3) Observation model construction: The center position of the thunderstorm cell is extracted as the observation value from the preliminary identification results at the current moment. ; The observation matrix H is: ; For scenarios involving multiple thunderstorm cells, a nearest neighbor data association algorithm is used to match trajectories with observations. Specifically, the Euclidean distance between the observation at time t and the centroid coordinates of all predicted trajectories at time t-1 is calculated, and the observation is assigned to the nearest trajectory. If no trajectory meets the conditions, it is determined to be a new thunderstorm cell and a new trajectory is initialized.

[0103] (4) Kalman filter recursive update Derivation of the current prior state and error covariance based on the historical best state: ; Where Q is the process noise covariance, characterizing the uncertainty of thunderstorm movement, and is set based on historical thunderstorm movement data statistics, with a value range of: ; By correcting the prior state based on the current observations, the optimal posterior estimate is obtained: ; (5) Parameter initialization and trajectory management: the initial state x0 is determined by the first identification result, the initial velocity is set to zero, the initial error covariance P0 is set to a diagonal matrix, and the trajectory that has not matched the observation value for three consecutive iterations is automatically removed. I is the identity matrix.

[0104] The Kalman filter is used to predict the position at the next time step, and then it is matched with the current observation (centroid) (nearest neighbor association).

[0105] By using observed values ​​to correct predicted values, the movement trajectory of thunderstorms can be smoothed, missed frames can be filled, and false targets with instantaneous jumps can be eliminated.

[0106] In summary, the proposed solution effectively overcomes the technical bottleneck of scarce near-shore radar data. By generating radar data through satellite inversion and employing cross-domain transfer learning strategies, it significantly reduces the model domain shift problem caused by differences in land and sea underlying surfaces. Simultaneously, by combining physical constraints and spatiotemporal filtering corrections, it greatly improves the accuracy, reliability, and spatiotemporal continuity of near-shore thunderstorm and gale identification, providing precise and efficient meteorological disaster early warning support for near-shore shipping, offshore operations, and the safety of coastal cities.

[0107] The second aspect of this invention discloses a nearshore thunderstorm and strong wind identification system based on DCSAM-TransUnet and transfer learning. Figure 4 This is a structural diagram of a nearshore thunderstorm and strong wind identification system based on DCSAM-TransUnet and transfer learning according to an embodiment of the present invention; as shown. Figure 4 As shown, the system 100 includes: The first processing module 101 is configured to acquire historical data from the source domain and the target domain, preprocess the data, and create thunderstorm large-scale point labels. The second processing module 102 is configured to construct a frequency decoupling mechanism based on the preprocessed satellite feature data using fast Fourier transform, train the LightGBM radar echo inversion model, and output simulated radar echo data retrieved from the satellite. The third processing module 103 is configured to construct a DCSAM-TransUnet thunderstorm and strong wind identification model based on source domain real-time radar and numerical model wind field data, including a two-layer multi-scale channel spatial attention mechanism DCSAM, and pre-train the model to obtain the optimal weights in the source domain. The fourth processing module 104 is configured to load the source domain optimal weights, and based on the target domain satellite inversion radar echo simulation data, source domain real-time radar data and numerical model wind field data, alternately retrain the model using a dynamic freezing strategy and an adversarial domain adaptation strategy to obtain the target domain optimal weights. The fifth processing module 105 is configured to load the optimal weight of the target domain, input real-time marine satellite inversion radar echo and numerical model wind field data, and output preliminary identification results of marine thunderstorms and strong winds. The sixth processing module 106 is configured to introduce a numerical model environment field, verify the preliminary identification results based on physical constraints, and perform spatiotemporal smoothing processing by combining morphological filtering and Kalman filtering to output the final identification result of marine thunderstorms and strong winds.

[0108] A third aspect of this invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the nearshore thunderstorm and strong wind identification method based on DCSAM-TransUnet and transfer learning according to any one of the first aspects of this invention.

[0109] Figure 5 This is a structural diagram of an electronic device according to an embodiment of the present invention, such as... Figure 5 As shown, the electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, Near Field Communication (NFC), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0110] Those skilled in the art will understand that Figure 5 The structure shown is merely a structural diagram of the part related to the technical solution of this disclosure and does not constitute a limitation on the electronic device to which the solution of this application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0111] A fourth aspect of this invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a nearshore thunderstorm and strong wind identification method based on DCSAM-TransUnet and transfer learning, as described in any of the first aspects of this invention.

[0112] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0113] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying nearshore thunderstorms and strong winds based on DCSAM-TransUnet and transfer learning, characterized in that, Includes the following steps: Step S1: Obtain historical data from the source and target domains. The historical data includes real-time radar data, wind speed observation data, satellite observation data, and numerical model wind field data. Preprocess the historical data, including extracting satellite feature data and creating thunderstorm wind point labels after preprocessing the satellite observation data. Step S2: Based on the satellite feature data, a frequency decoupling mechanism is constructed using Fast Fourier Transform, the LightGBM radar echo inversion model is trained, and the simulated radar echo data retrieved from the satellite is output. Step S3: Based on the source domain real-time radar and numerical model wind field data, construct a DCSAM-TransUnet thunderstorm and strong wind identification model containing a two-layer multi-scale channel spatial attention mechanism DCSAM, and pre-train the thunderstorm and strong wind identification model to obtain the optimal weights in the source domain. Step S4: Load the source domain optimal weights, and based on the target domain satellite inversion radar echo simulation data, source domain real radar data and numerical model wind field data, use dynamic freezing strategy and adversarial domain adaptation strategy to alternately retrain the model to obtain the target domain optimal weights. Step S5: Load the optimal weight of the target domain, input real-time marine satellite inversion radar echo and numerical model wind field data, and output the preliminary identification results of marine thunderstorms and strong winds; Step S6: Introduce a numerical model environment field, verify the preliminary identification results based on physical constraints, and perform spatiotemporal smoothing processing by combining morphological filtering and Kalman filtering to output the final identification results of marine thunderstorms and strong winds.

2. The method according to claim 1, characterized in that, In step S1, the method for creating thunderstorm large-scale point tags includes: Set a wind speed threshold, and mark the observation point where the instantaneous wind speed reaches the threshold and its surrounding preset range as the potential area; Calculate the morphological indices of the radar situation within the potential area, including the maximum combined reflectivity and the convective structure intensity index; The convective structure intensity index is calculated based on a weighted average of gradient intensity characteristics, shape characteristics, and gradient variation characteristics. The calculation formula is as follows: ； Where w1, w2, w3 are weighting coefficients, F grad For gradient strength characteristics, F shape For shape characteristics, F var This is a gradient variation feature; When both the maximum combined reflectivity and the convective structure intensity index meet the preset conditions, the location of the thunderstorm and strong wind is confirmed.

3. The method according to claim 1, characterized in that, In step S2, the specific implementation of the frequency decoupling mechanism includes: Fast Fourier transform is performed on real-time radar data and simulated radar echo data retrieved from satellites to calculate the amplitude spectrum and phase spectrum, respectively. Construct a mask matrix to locally replace the amplitude spectrum of satellite-retrieved radar data while keeping the phase spectrum unchanged; The mask matrix is ​​used to retain the amplitude information of satellite inversion data in the low-frequency region at the center of the spectrum, and to introduce the amplitude information of real-time radar data in the high-frequency region; Enhanced satellite-inverted radar data is generated by using the mixed amplitude spectrum and the original phase spectrum through inverse Fourier transform. The formula for calculating the amplitude spectrum after mixing is: Where M is the mask matrix, For element-wise multiplication, A sat For the amplitude spectrum of satellite inversion data, A radar This is the amplitude spectrum of real-time radar data.

4. The method according to claim 1, characterized in that, In step S3, the DCSAM-TransUnet thunderstorm and strong wind identification model includes an encoder, a DCSAM module, and a decoder; The DCSAM module includes a multi-scale channel attention mechanism MS-CAM and a multi-scale spatial attention mechanism MS-SAM. The MS-CAM extracts channel features using global and local branches, where the global branch includes pooling operations and the local branch includes convolution operations. The MS-SAM extracts spatial features using multi-scale convolutional branches and spatial saliency-guided branches, wherein the multi-scale convolutional branches contain convolutional layers with convolutional kernels of different sizes; The calculation formulas for MS-CAM and MS-SAM are as follows: The feature fusion output of MS-CAM is: ； Where X represents the input features, σ is the sigmoid function, L(X) is the local branch output, G(X) is the global branch output, and ⊕ denotes convolutional fusion after feature concatenation. This represents element-wise multiplication; The feature fusion output of MS-SAM is: ； Where X′ is the output of MS-CAM, S scale For the output of multi-scale convolution branches, S spatial The spatial saliency guides the branch output.

5. The method according to claim 1, characterized in that, In step S4, the dynamic freezing strategy includes: Calculate the cosine similarity between the feature vectors of satellite-inverted radar data and the feature vectors of real-world radar data; Based on the magnitude of cosine similarity, the different levels of the encoder and the Transformer module parameters of the DCSAM-TransUnet thunderstorm and strong wind recognition model are frozen in stages. The specific strategy is as follows: when the cosine similarity is greater than or equal to the first threshold, the number of frozen layers is the largest; when the cosine similarity is less than the second threshold, the number of frozen layers is the smallest; when the cosine similarity is between the first and second thresholds, the number of frozen layers is in the middle. The adversarial domain adaptation strategy includes: The domain discriminator module is connected after the encoder in the DCSAM-TransUnet thunderstorm and strong wind identification model. Construct a total loss function that includes the main model loss and the domain classifier loss; Use alternating training methods: Phase 1: Freeze the domain discriminator parameters and update the main model parameters to minimize the main model loss and maximize the domain classifier loss, making the extracted features difficult to distinguish their source. Phase 2: Freeze the main model parameters and update the domain discriminator parameters to minimize the domain classifier loss and improve the discriminator's ability to distinguish between source and target domain data.

6. The method according to claim 1, characterized in that, In step S6, the verification of the preliminary identification result based on physical constraints includes: Acquire the corresponding spatiotemporal numerical model environmental field data, including boundary layer wind speed, convective available potential energy (CAPE), and K exponent; Grid-by-grid evaluation: If any of the following conditions are met, the recognition result is retained; otherwise, it is discarded: The boundary layer wind speed is greater than or equal to the first preset threshold. CAPE is greater than or equal to the second preset threshold; The K-index is greater than or equal to the third preset threshold.

7. The method according to claim 1, characterized in that, In step S6, the Kalman filter performs spatiotemporal smoothing processing, including: Define the state vector of a thunderstorm cell as its center position coordinates and moving speed; A state prediction equation is constructed based on the assumption of uniform motion. The centroid of thunderstorm cells is extracted from the preliminary identification results as the observation value, and the nearest neighbor algorithm is used to associate the observation value with the predicted trajectory. The state vector is updated using the Kalman filter recursive formula to correct the identification position of thunderstorm cells and remove trajectories that have not been matched for multiple consecutive frames.

8. A nearshore thunderstorm and strong wind identification system based on DCSAM-TransUnet and transfer learning, wherein the system employs the method described in any one of claims 1-7, characterized in that... The system includes: The first processing module is configured to acquire historical data from the source domain and the target domain, and preprocess the historical data, including extracting satellite feature data and creating thunderstorm wind point labels after preprocessing the satellite observation data. The second processing module is configured to construct a frequency decoupling mechanism based on the satellite feature data using fast Fourier transform, train the LightGBM radar echo inversion model, and output simulated radar echo data retrieved from the satellite. The third processing module is configured to construct a DCSAM-TransUnet thunderstorm and strong wind identification model based on source domain real-time radar and numerical model wind field data, which includes a two-layer multi-scale channel spatial attention mechanism DCSAM, and pre-train the thunderstorm and strong wind identification model to obtain the optimal weights in the source domain. The fourth processing module is configured to load the source domain optimal weights, and based on the target domain satellite inversion radar echo simulation data, source domain real-time radar data and numerical model wind field data, alternately retrain the model using a dynamic freezing strategy and an adversarial domain adaptation strategy to obtain the target domain optimal weights. The fifth processing module is configured to load the optimal weight of the target domain, input real-time marine satellite inversion radar echo and numerical model wind field data, and output preliminary identification results of marine thunderstorms and strong winds. The sixth processing module is configured to introduce a numerical model environment field, verify the preliminary identification results based on physical constraints, and perform spatiotemporal smoothing processing by combining morphological filtering and Kalman filtering to output the final identification results of marine thunderstorms and strong winds.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the steps in the nearshore thunderstorm and gale identification method based on DCSAM-TransUnet and transfer learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the nearshore thunderstorm and gale identification method based on DCSAM-TransUnet and transfer learning as described in any one of claims 1 to 7.

Images