A typhoon initial identification method, medium and program product based on a dual AI model architecture and typhoon asymmetry characteristics

Abstract

The present invention discloses a typhoon incipient identification method, medium and program product based on a dual AI model architecture and typhoon asymmetry characteristics, belonging to the technical fields of meteorological monitoring and artificial intelligence. The method first obtains multi-channel satellite image data and performs preprocessing; then locates the cloud cluster center through the first AI model; taking the cloud cluster center as a reference, extracts the symmetry characteristics and asymmetry characteristics of the cloud cluster; then analyzes the extracted characteristics through the second AI model to determine whether the target cloud cluster is in the typhoon incipient stage; finally outputs identification results including the typhoon center position, incipient time, development probability, etc. The present invention effectively improves the accuracy and reliability of typhoon incipient identification through phased identification by a dual AI model architecture and combines the unique asymmetry characteristics of the cloud cluster, and has high practical value.

Description

Technical Field

[0001] The present invention belongs to the technical fields of meteorological monitoring and artificial intelligence, and relates to technologies such as typhoon identification, artificial intelligence, image processing, and pattern recognition. Specifically, it is a typhoon initial identification method, medium, and program product based on a dual AI model architecture and typhoon asymmetry characteristics, which is used to accurately identify and judge the cloud clusters in the initial stage of typhoons, and improve the accuracy and reliability of early typhoon monitoring. Background Art

[0002] The formation and development process of typhoons is a complex dynamic and thermodynamic process, usually including four stages: typhoon initial stage, development stage, mature stage, and dissipation stage. Timely and accurately identifying the initial stage of typhoons is of great significance for improving typhoon warning capabilities and reducing disaster losses. The initial stage of typhoons usually refers to the initial formation stage of typhoon embryos over tropical oceans. The characteristics of this stage are that cyclonic disturbances have just begun to develop, and the rotational structure has not yet been fully formed, manifested as a loose and irregular cloud cluster distribution, and its intensity is usually much lower than that of severe tropical storms. Due to the lack of obvious vortex characteristics and weak convective activities in the cloud clusters of typhoons in the initial stage, and being significantly affected by sea surface temperature, vertical wind shear, etc., the cloud cluster structure is highly similar to that of general atmospheric disturbances (such as monsoon cloud clusters, mesoscale convective systems, etc.) that cannot develop into typhoons, which poses a great challenge to the identification of the initial stage of typhoons.

[0003] In recent years, with the rapid development of artificial intelligence (AI) technology, typhoon identification technology based on deep learning has been widely applied in the meteorological field. These technologies can effectively extract the spatial distribution characteristics and temporal evolution information of typhoons through the analysis of large-scale satellite remote sensing data. For example, convolutional neural network (CNN) is widely used to extract cloud cluster characteristics from satellite images, and object detection algorithms such as Single Shot Detector (SSD) are used to locate typhoon areas. In the development and mature stages of typhoons, especially for typhoon vortices with a strength of severe tropical storm level and above, the correct recognition rate of the AI model can reach over 90%. However, for typhoon vortices with a strength below severe tropical storm level, the recognition accuracy is only 40% - 80%. Especially in the initial stage of typhoons, due to the incomplete formation of the typhoon structure and the high similarity of its cloud cluster characteristics to those of ordinary atmospheric disturbance cloud clusters, the recognition accuracy of existing AI models further decreases, only being 10% - 30%. This low recognition rate seriously affects the timeliness and reliability of typhoon warnings.

[0004] In the application of typhoon incipient stage, the existing technology still has the following limitations: in terms of feature extraction, the existing AI models usually rely on the overall vortex structure, convection intensity and spatial distribution characteristics of cloud clusters for identification, while these characteristics are not obvious in the initial stage of typhoon, the cloud cluster structure is loose and the local characteristics are weak, resulting in a significant decrease in the model recognition ability; in terms of model architecture, the existing technology generally adopts a single model architecture for typhoon identification, and fails to fully consider the particularities of different tasks in the typhoon identification process. For example, although the positioning of the cloud center and the judgment of the initial typhoon are interrelated, there are significant differences in technical characteristics, and it is difficult to achieve the optimal effect at the same time using a single model; in terms of feature utilization, although the structure of the typhoon is not obvious in the initial stage, its asymmetric characteristics are more prominent. However, most of the existing AI models mainly focus on the improvement and improvement of AI algorithms, and fail to fully model and utilize the characteristics of the typhoon itself when it is generated, especially the asymmetric characteristics, thereby ignoring this important discriminant information.

[0005] In summary, the difficulty in identifying a typhoon in its infancy mainly comes from the obscurity of its structural features and its high similarity with other cloud clusters. Existing technologies face problems such as insufficient feature utilization, single model architecture, and low recognition accuracy in the identification of a typhoon in its infancy. Therefore, how to construct a multi-model collaborative recognition method that can fully utilize the asymmetric characteristics of typhoons and improve the recognition accuracy of typhoons in their infancy is a technical problem that needs to be urgently solved in the current meteorological field. Summary of the invention

[0006] 1. Purpose of the invention

[0007] In view of the defects and shortcomings of the prior art in the identification of the initial stage of typhoons, such as low recognition accuracy, poor reliability, insufficient modeling of typhoon asymmetric characteristics, and difficulty in meeting the requirements of different tasks with a single model architecture, in order to solve at least one of the above and other technical problems in the prior art, the present invention aims to provide a typhoon initial identification method, medium and program product based on a dual AI model architecture and typhoon asymmetric characteristics. By constructing a dual-stage AI model architecture, the typhoon identification task is decomposed into two subtasks: cloud center positioning and typhoon initial judgment. By introducing the Fourier decomposition technology of typhoon asymmetric characteristics and combining multi-channel satellite data such as infrared channels and water vapor channels, the present invention can effectively extract the key features of the initial stage of typhoons, significantly improve the recognition ability of newly born typhoons, improve the accuracy and reliability of early typhoon monitoring, and provide more reliable technical support for typhoon warning.

[0008] (II) Technical solution

[0009] In order to achieve the purpose of the invention and solve the technical problems, the present invention adopts the following technical solutions:

[0010] The first object of the present invention is to provide a typhoon genesis identification method based on a dual AI model architecture and typhoon asymmetry characteristics, which is used to efficiently and accurately automatically identify the typhoon genesis stage by combining multi-channel satellite data, asymmetry feature extraction, and a dual-model collaborative architecture, and improve the timeliness and accuracy of typhoon warnings. When implementing this method, it at least includes the following steps:

[0011] SS1. Obtain multi-channel satellite image data within the target area, including at least brightness temperature data of the infrared channel and water vapor channel, and preprocess the obtained satellite data, including at least data quality control, standardization processing, unified adjustment of spatio-temporal resolution, and conversion of brightness temperature data into a grayscale image format suitable for input to the deep learning model;

[0012] SS2. Build a cloud cluster center localization AI model as the first stage of typhoon genesis identification based on a convolutional neural network (CNN). Use this model to analyze the preprocessed multi-channel satellite image data, extract cloud cluster features within the target area, and output the center position of the cloud cluster within the target area. And this model includes a feature extraction layer, a spatial attention mechanism layer, and a position regression layer: the feature extraction layer is used to extract the brightness temperature features and local convection intensity of the satellite image, the spatial attention mechanism layer is used to enhance the perception ability of local cloud cluster features, and the position regression layer predicts the coordinates of the cloud cluster center position based on the extracted features;

[0013] SS3. Taking the cloud cluster center identified and located in step SS2 as a reference, use the Fourier decomposition technique to perform multi-order feature decomposition on the multi-channel satellite image data within a preset radius range centered on the cloud cluster center, decompose the brightness temperature values within the cloud cluster area into multi-order fluctuation components distributed radially, and extract the symmetry features and asymmetry features of the cloud cluster. Among them, the symmetry feature represents the overall uniform distribution characteristic of the cloud cluster, and the asymmetry feature is used to quantify the local perturbation characteristic and spiral morphology of the cloud cluster in the genesis stage;

[0014] SS4. Build a typhoon genesis identification AI model as the second stage of typhoon genesis identification based on the single-shot detector (SSD) object detection algorithm, and analyze the cloud cluster symmetry features and asymmetry features extracted in step SS3. Among them, the typhoon genesis identification AI model includes a feature input layer, a feature fusion layer, an SSD detection module, and an output layer. Among them: the feature input layer is used to receive the cloud cluster symmetry and asymmetry features; the feature fusion layer is used to fuse the input symmetry and asymmetry features, and use the attention mechanism to enhance the perception ability of significantly asymmetric features; the SSD detection module is used to perform object detection on the fused multi-scale features, identify and locate potential typhoon genesis areas; the output layer judges whether the target cloud cluster belongs to the typhoon genesis stage according to the classification result of the SSD detection module;

[0015] SS5. Based on the typhoon genesis recognition and judgment results, if the target cloud cluster belongs to the typhoon genesis stage, the typhoon genesis stage recognition results are output, including the typhoon center position coordinates, genesis time, development probability, and credibility, providing support for subsequent typhoon path prediction, intensity development analysis, and early warning issuance.

[0016] (III) Technical Effects

[0017] Compared with the existing technology, the typhoon genesis recognition method, medium, and program product based on the dual AI model architecture and typhoon asymmetry characteristics of the present invention have the following beneficial and remarkable technical effects:

[0018] (1) The present invention integrates the brightness temperature data of the infrared channel (11-micron band) and the water vapor channel (6.7-micron band), and combines the Fourier decomposition technology to extract the symmetry characteristics (WN0) and asymmetry characteristics (WN1-WN4) of the cloud cluster, effectively capturing the global distribution characteristics and local perturbation characteristics of the cloud cluster in the typhoon genesis stage. Through the joint modeling of multi-channel data and Fourier features, the present invention realizes high-resolution and high-timeliness feature expression in the early cloud cluster dynamics analysis, significantly improving the perception ability of complex characteristics in the typhoon genesis stage.

[0019] (2) The present invention adopts a dual AI model architecture, including a feature extraction model and an object detection model. The feature extraction model extracts multi-scale features through multi-scale convolution operations and the FPN structure, and combines the attention mechanism to enhance the perception ability of key features; the object detection model is based on the SSD module, integrating a classification sub-network and a location regression sub-network to classify and accurately locate the cloud cluster target in the typhoon genesis stage. Through the synergistic effect of the dual models, the present invention can efficiently screen and identify typhoon genesis cloud clusters in a complex background, while taking into account the classification accuracy and detection precision. In addition, the object detection module further improves the accuracy and robustness of the candidate detection frame through the non-maximum suppression (NMS) and multi-scale comprehensive decision-making strategy, providing a strong technical guarantee for the accurate recognition of the typhoon genesis stage.

[0020] (3) Through the comprehensive analysis of multi-channel satellite data and the efficient application of the deep learning model, the present invention significantly improves the accuracy, reliability, and real-time performance of typhoon genesis stage recognition, and is not only applicable to typhoon early warning, but also can be extended to other fields involving symmetry and asymmetry feature analysis, such as tropical cyclone monitoring, atmospheric convection system identification, etc., with strong practicality and promotion value. Description of the Drawings

[0021] The accompanying drawings forming a part of this invention are used to provide a further understanding of the invention. The schematic embodiments and descriptions thereof of the invention are used to explain the invention and do not constitute an improper limitation of the invention. Hereinafter, the embodiments of the invention will be described in detail with reference to the drawings, where:

[0022] Figure 1 is the implementation flowchart of the typhoon genesis identification method based on the dual AI model architecture and typhoon asymmetry characteristics provided by the embodiment of the present invention;

[0023] Figure 2 is the initial embryo morphology of Typhoon No. 9 in 2021 and its Fourier decomposition structure diagram. The time is 12:00 on July 16, 2021. At this time, the typhoon intensity is 25 kt = 12.86 m / s. The red frame is within the range of 5° of the typhoon embryo. (a) is the infrared channel 11-micron data, and (b)-(f) are the Fourier 0-4 wave decomposition images within the range of 5° of the typhoon center. The abscissa in the figure is the longitude, and the ordinate is the dimension;

[0024] Figure 3 is the cloud mass morphology that cannot form a typhoon and its Fourier decomposition structure diagram. The time is the same as Figure 2 , the red frame is within the range of 5° of the non-typhoon embryo. (a) is the infrared channel 11-micron data, and (b)-(f) are the Fourier 0-4 wave decomposition images within the range of 5° of the center. The abscissa is the longitude, and the ordinate is the dimension;

[0025] Figure 4 is the morphology after the typhoon is formed and its Fourier decomposition structure diagram. It is the same typhoon as Figure 2 but after this typhoon is formed. The time is 12:00 on July 20, 2021. The typhoon intensity is 80 kt = 41.1 m / s. The square within the red frame is within the range of 5° of the typhoon. (a) is the infrared channel 11-micron data, and (b)-(f) are the Fourier 0-4 wave decomposition images within the range of 5° of the typhoon center. The abscissa is the longitude, and the ordinate is the dimension. Specific Embodiments

[0026] The following further illustrates the typhoon genesis identification method, medium and program product of the present invention based on the dual AI model architecture and typhoon asymmetry characteristics. By combining multi-channel satellite data, asymmetry feature extraction technology and a dual-model collaborative architecture, the cloud mass characteristics in the target area are comprehensively analyzed and modeled to significantly improve the identification accuracy and reliability in the typhoon genesis stage. The following will describe in detail the specific implementation process of the present invention with reference to the drawings, elaborate on the key technical solutions of the present invention and their implementation details, but the present invention is not limited to the specific content described in the following embodiments.

[0027] Example 1

[0028] AsFigure 1 As shown in Figure 1 , the typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry features provided by the embodiments of the present invention mainly includes the following steps when implemented:

[0029] SS1. Data acquisition and preprocessing

[0030] Obtain multi-channel satellite image data within the target area, including at least the brightness temperature data of the infrared channel and the water vapor channel. Preprocess the obtained satellite data, including at least data quality control, standardization processing, unified adjustment of spatio-temporal resolution, and conversion of the brightness temperature data into a grayscale image format suitable for input to the deep learning model.

[0031] Preferably, among the multi-channel satellite image data, the infrared channel data is the brightness temperature data in the 11-micron band, which is used to capture cloud top temperature and atmospheric convection information, and the water vapor channel data is the brightness temperature data in the 6.7-micron band, which is used to detect the water vapor content and distribution in the middle layer of the atmosphere. The horizontal resolution of the data is 4 - 10 km to balance the computational efficiency and the recognition ability of meso-scale and small-scale convective systems, and the time resolution is 1 - 6 hours to capture the rapid development and changes in the initial stage of the typhoon.

[0032] When preprocessing the obtained satellite data, data quality control includes removing strip noise, cloud detection, and outlier rejection from the satellite image data. Among them, wavelet transform method is used for strip noise removal, cloud detection is based on the brightness temperature difference threshold judgment between the infrared and water vapor channels, and outlier rejection adopts the 3 σ criterion; standardization processing includes min-max normalization or Z-score standardization of the brightness temperature data to eliminate the difference in the numerical range between different channel data. Among them, min-max normalization maps the data to the [0, 1] interval, and Z-score standardization converts the data into a distribution with a mean of 0 and a standard deviation of 1; the unified adjustment of spatio-temporal resolution includes using linear interpolation method to unify the time resolution of the multi-channel satellite image data to a 15-minute time interval, using bilinear interpolation method to unify the spatial resolution to a 1-km spatial resolution, and at the same time, according to the meteorological characteristics of the target area, performing regional cropping on the multi-channel satellite image data, only retaining the image data within the coverage range of the target area to reduce the interference of irrelevant data on the subsequent deep learning model processing.

[0033] SS2. Identification and positioning of the cloud cluster center position

[0034] Construct a cloud cluster center localization AI model based on a convolutional neural network (CNN) as the first stage of typhoon genesis recognition. Use this model to analyze the preprocessed multi-channel satellite image data, extract the cloud cluster features in the target area, and output the center position of the cloud cluster in the target area. The model includes a feature extraction layer, a spatial attention mechanism layer, and a position regression layer: The feature extraction layer is used to extract the brightness temperature features and local convection intensity of the satellite image. The spatial attention mechanism layer is used to enhance the perception ability of the local features of the cloud cluster. The position regression layer predicts the coordinates of the cloud cluster center position based on the extracted features.

[0035] Preferably, the feature extraction layer of the cloud cluster center localization AI model uses depthwise separable convolution to reduce the number of model parameters and computational complexity, while improving the extraction efficiency of the brightness temperature distribution features and local convection intensity, and normalizes the features output by each convolutional layer through batch normalization to accelerate model convergence and improve the generalization performance of the model; The spatial attention mechanism layer highlights the features of the target cloud cluster area by generating a dynamic weight matrix, where the dynamic weight matrix is generated based on the maximum pooling and average pooling results of the convolutional feature map, and is multiplied element-wise with the original feature map, thereby enhancing the perception ability of fuzzy or incomplete cloud cluster features.

[0036] The cloud cluster center localization AI model is trained based on a satellite image dataset with labeled cloud cluster center positions. The dataset is divided according to the ratio of 70% for training, 20% for validation, and 10% for testing. The training process is completed through supervised learning and data augmentation techniques are used to expand the training samples, including random rotation of ±30°, random scaling of ±10%, random translation of ±20 pixels, and horizontal flipping to improve the robustness of the model to cloud cluster morphological changes; and the localization error is used as the loss function for optimization. The loss function is defined as , where x i , y i is the predicted position of the i th sample, is the true labeled position of the i th sample, i = 1, 2, …, N , N is the total number of samples. The localization error is minimized by calculating the Euclidean distance between the predicted cloud cluster center position and the true labeled position; and the Adam optimizer with an adaptive learning rate is used to update the weights during model training. The initial learning rate is set to 0.001, and the learning rate is dynamically adjusted based on the loss value of the validation set to accelerate convergence.

[0037] SS3. Cloud Cluster Asymmetry Feature Extraction

[0038] Taking the center of the cloud cluster identified and located in step SS2 as a reference, using Fourier decomposition technology to perform multi-order feature decomposition on multi-channel satellite image data within a preset radius range centered on the center of the cloud cluster, decomposing the brightness temperature values in the cloud cluster area into multi-order fluctuation components with radial distribution, and extracting the symmetry features and asymmetry features of the cloud cluster, where the symmetry features represent the overall uniform distribution characteristics of the cloud cluster, and the asymmetry features are used to quantify the local perturbation characteristics and spiral morphology of the cloud cluster in the initial stage.

[0039] Preferably, when using Fourier decomposition technology to perform multi-order feature decomposition on multi-channel satellite image data within a preset radius range centered on the center of the cloud cluster, it mainly includes the following sub-steps:

[0040] SS31. Target area selection: Taking the center of the cloud cluster identified and located in step SS2 as a reference, select a target area with a radius range of 1 - 5° centered on the center of the cloud cluster, perform grid division at a sampling interval of 0.1°, generate a brightness temperature distribution matrix of the target area, each grid point in the matrix corresponds to the satellite observation brightness temperature value, and perform interpolation processing on the data in the boundary area according to the cloud cluster distribution characteristics to supplement possible vacancies;

[0041] SS32. Radial data construction: Perform radial data processing on the brightness temperature values in the target area, convert the matrix data into radial distribution data based on the center of the cloud cluster, the sampling points of the radial data in the azimuthal direction are distributed within the range of 0 - 360° at a preset interval to ensure the continuity and integrity of the radial distribution data, and at the same time perform filtering processing on the strong noise part of the radial data to reduce the influence of external interference on the subsequent Fourier decomposition;

[0042] SS33. Fourier decomposition: Perform Fourier decomposition on the radial brightness temperature distribution data along the azimuthal direction, decompose it into several order fluctuation components, and the Fourier decomposition formula is as follows:

[0043]

[0044] Among them, T ( θ , r ) represents the radial distribution function of the brightness temperature, r is the radial distance, θ is the azimuth; i is the order of the Fourier decomposition and i = 1, 2, 3, …, n , n is the highest order of the Fourier decomposition and n ≥ 4; a 0( r ) is the amplitude of the 0th order component, ai ( r ) and b i ( r ) are the amplitudes of the cosine and sine components of order i respectively; WN 0( r ) represents the 0th order component of the Fourier decomposition, WNi ( r ) represents the i th order component;

[0045] SS34. Symmetry and asymmetry feature extraction: Extract the 0th order component WN0 in the Fourier decomposition result as the symmetry feature of the cloud cluster, which reflects the overall uniform distribution characteristic of the cloud cluster brightness temperature and characterizes the overall vortex shape of the cloud cluster in the initial stage of the typhoon. Extract the 1st - 4th order components WN1 - WN4 as the asymmetry features of the cloud cluster, where WN1 reflects the eccentric shape of the cloud cluster, and WN2 - WN4 capture the local perturbation characteristics and spiral shape features of the cloud cluster. Then reconstruct the extracted symmetry features and asymmetry features into a multi - channel feature map, with each channel corresponding to a component, as the input data for the subsequent typhoon initial recognition AI model.

[0046] SS4. Recognition and judgment of the initial stage of the typhoon

[0047] Construct a typhoon initial recognition AI model as the second stage of typhoon initial recognition based on the Single Shot Detector (SSD) object detection algorithm, and analyze the symmetry features and asymmetry features of the cloud cluster extracted in step SS3. The typhoon initial recognition AI model includes a feature input layer, a feature fusion layer, an SSD detection module, and an output layer, where: The feature input layer is used to receive the symmetry and asymmetry features of the cloud cluster; The feature fusion layer is used to fuse the input symmetry and asymmetry features and enhance the perception ability of significant asymmetry features using the attention mechanism; The SSD detection module is used to perform object detection on the fused multi - scale features to identify and locate potential typhoon initial regions; The output layer determines whether the target cloud cluster belongs to the initial stage of the typhoon according to the classification result of the SSD detection module.

[0048] Preferably, when the feature fusion layer of the typhoon initial recognition AI model fuses the input symmetry features and asymmetry features, it uses multi - scale convolution operations and combines a multi - scale feature pyramid (Feature Pyramid Network, FPN) structure to generate multiple feature maps of different scales to enhance the expression ability of the fused features for the multi - scale cloud cluster features in the initial stage of the typhoon, including:

[0049] When performing multi-scale convolution operations, the feature fusion layer separately performs multi-scale convolution extraction on the input symmetric features (WN0) and asymmetric features (WN1-WN4). The multi-scale convolution performs parallel convolution on the input features by setting convolution kernels of different sizes (such as 3×3, 5×5, 7×7) to extract local features and global features at different scales. After the convolution operation, the multi-scale features are fused into a unified high-dimensional feature representation through feature concatenation;

[0050] When optimizing the feature pyramid network (FPN) structure, a multi-scale feature pyramid network (FPN) structure is used to further process the fused features. The perception ability of typhoon initial cloud cluster targets at different scales is enhanced through bottom-up feature extraction and top-down feature fusion. The bottom-up feature extraction extracts features layer by layer through a deep convolutional network to generate feature maps of different resolutions. The lower-resolution feature maps are used to capture the global pattern information of the cloud clusters, and the higher-resolution feature maps are used to capture the local perturbation features of the cloud clusters. The top-down feature fusion fuses the low-resolution feature maps with the high-resolution feature maps layer by layer through upsampling operations, and reduces the dimension of the fused features through 1×1 convolution to achieve the full utilization of multi-scale information.

[0051] Further preferably, the feature fusion layer of the typhoon initial recognition AI model further combines the channel attention mechanism during the feature fusion process to enhance the perception ability of the asymmetric features (WN1-WN4) by learning the weights of different channel features, including: performing global average pooling on the fused multi-scale feature maps to generate the global features of each channel; calculating the weights of each channel through a fully connected layer and performing weighted processing on the original feature maps. The calculation formula for the channel weights is:

[0052]

[0053] Among them, f pool is the feature after global pooling, W 1 and W 2 are the weight matrices of the fully connected layer, σ is the Sigmoid activation function, w c is the weight of each channel;

[0054] Further preferably, when the SSD detection module of the typhoon initial recognition AI model performs target detection, it mainly includes the following steps:

[0055] SS41. Generation of multi-scale feature maps

[0056] The multi-scale feature maps output by the feature fusion layer (generated by FPN) respectively correspond to feature layers of different resolutions (for example, the low-resolution feature map is used to capture the global information of the cloud cluster in the initial stage of the typhoon, and the high-resolution feature map is used to extract the local perturbation features of the target area). Each feature layer generates a feature map of a fixed size through a convolution operation. The feature map of each scale is used to detect cloud cluster targets of a specific size to comprehensively cover the possible initial typhoon areas;

[0057] Further preferably, in sub-step SS41, each feature layer generates a feature map of a fixed size through a convolution operation, specifically including: performing multiple 3×3 convolution operations on the input fusion features to extract spatial features; gradually reducing the spatial resolution of the feature map through downsampling operations to generate feature maps of multiple scales (such as feature maps of different spatial resolutions such as 8×8, 16×16, and 32×32).

[0058] SS42. Candidate Region Generation

[0059] On the feature map of each scale, multiple candidate detection boxes are generated through a preset anchor box. The size and aspect ratio of the anchor box are adjusted according to the resolution of the feature map of different scales. The size range of the anchor box is from 16 pixels to 512 pixels, covering the detection requirements from small local perturbation cloud clusters to large convective systems. The aspect ratios of the anchor box include 1:1, 2:1, and 1:2 to adapt to the possible shapes and distribution characteristics of the initial typhoon cloud clusters. Each anchor box is associated with each pixel point on the feature map to generate candidate regions for subsequent classification and regression prediction;

[0060] SS43. Classification and Regression Prediction

[0061] The target category prediction and position regression correction of the candidate regions are respectively completed through a classification sub-network and a regression sub-network, where: the classification sub-network is used to classify the candidate regions in each feature map to determine whether they belong to the cloud clusters in the initial stage of the typhoon. The classification output calculates the probability that each candidate region belongs to the initial typhoon or non-initial typhoon category through the Softmax function, and the classification loss function adopts the cross-entropy loss; the regression sub-network is used to perform regression correction on the bounding boxes of the candidate regions, predict their precise positions and scales, and optimize the regression error using the smooth L1 loss;

[0062] SS44. Non-Maximum Suppression (NMS)

[0063] Post-process the results of classification and regression prediction. Use the Non-Maximum Suppression (NMS) method to remove redundant detection boxes, including: sorting the candidate detection boxes according to the confidence level output by the classification sub-network; for the detection boxes with confidence levels exceeding the preset threshold, calculate the intersection over union (IoU) with other detection boxes. If the IoU exceeds the set threshold (e.g., 0.5), then delete this detection box; only retain the detection result with the highest confidence level, including the position and scale of the detection box and the confidence score of the typhoon initial stage.

[0064] SS45. Multi-scale comprehensive decision-making

[0065] Combine the detection results of feature maps at different scales and generate the final typhoon initial identification result through weighted fusion. The weight allocation is dynamically adjusted according to the detection confidence levels of feature maps at different scales. If the detection confidence level of the low-resolution feature map is relatively high, then preferentially retain the detection result of the low-resolution feature map; if the detection confidence level of the high-resolution feature map is relatively high, then preferentially retain the detection result of the high-resolution feature map; the finally output result includes the central position of the typhoon initial cloud cluster, the scale of the detection box, and the confidence score of the typhoon initial stage.

[0066] Preferably, the typhoon initial identification AI model is trained based on a multi-channel feature image dataset containing labeled typhoon initial stage cloud clusters and non-typhoon initial cloud clusters and the corresponding asymmetry features. The dataset is divided according to the ratio of 70% for training, 20% for validation, and 10% for testing. During the training process, the classification cross-entropy loss function and the regression loss function are used to jointly optimize the object detection and classification performance. The Adam optimizer or stochastic gradient descent is used to optimize the model parameters, and new data (such as the latest observed typhoon initial stage cloud clusters) is regularly introduced to update the training set and continuously train the model.

[0067] SS5. Output of identification result

[0068] Based on the typhoon initial identification and judgment result in step SS4, if the target cloud cluster belongs to the typhoon initial stage, then output the typhoon initial stage identification result, including the typhoon center position coordinates, the initial time, the development probability, and the credibility, to provide support for subsequent typhoon path prediction, intensity development analysis, and early warning release.

[0069] Preferably, the output of the typhoon initial stage identification result includes at least: the central position coordinates of the target cloud cluster, the identification confidence level, the judgment result of the typhoon initial stage, and its credibility score, where the identification confidence level is calculated through the correlation between the classification output probability and the asymmetry feature weight, and the credibility score is calculated by comprehensively considering the consistency of the classification output result and the object detection result.

[0070] Embodiment 2

[0071] Based on the above-mentioned Embodiment 1, in this Embodiment 2, the 9th typhoon in 2021 will be taken as a specific case to further illustrate the specific application of the present invention. The main differences between this embodiment and Embodiment 1 lie in the selection of data sources and some parameters, as well as a more intuitive explanation of the asymmetry characteristics of cloud clusters.

[0072] Step 1: Data acquisition and preprocessing

[0073] Multi-channel satellite image data within the target area is acquired. In this embodiment, GridSat data released by the US NOAA is used. This data integrates various geostationary satellite data, including GOES data of the United States, Meteosat data of Europe, GMS and MTSAT data of Japan, and Fengyun satellite data of China, etc. The data includes brightness temperature data of the infrared channel at 11 microns and the water vapor channel at 6.7 microns. The horizontal resolution of the data is 0.07° latitude (about 8 km), and the time resolution is 180 minutes. The acquired satellite data is preprocessed, including steps such as data quality control and standardization processing, which is similar to Embodiment 1.

[0074] Step 2: Identification and positioning of the cloud cluster center

[0075] In this embodiment, the satellite image of the target area at 12:00 on July 16, 2021 is mainly analyzed. This time point corresponds to the initial stage of the 9th typhoon, and the typhoon intensity is 25 kt (about 12.86 m / s) at this time. In this embodiment, a cloud cluster center positioning AI model based on a convolutional neural network (CNN) similar to that in Embodiment 1 is used to successfully identify and position the center position of the target cloud cluster.

[0076] Step 3: Extraction of cloud cluster asymmetry characteristics

[0077] Taking the cloud cluster center identified and positioned in Step 2 as a reference, multi-order feature decomposition is performed on the multi-channel satellite image data within a preset radius (set to 5° in this embodiment) centered on the cloud cluster center using Fourier decomposition technology to extract the symmetry feature WN0 and the asymmetry features WN1 - WN4 of the cloud cluster. The following Figures 2 to 4 more intuitively shows the feature differences of the initial stage of the typhoon, non-typhoon cloud clusters, and mature typhoons on the original satellite images and each component after Fourier decomposition.

[0078] Figure 2 shows the cloud cluster morphology and its decomposition structure of the 9th typhoon in 2021 at the initial stage (Time: 12:00 on July 16, 2021, Intensity: 25 kt), Figure 2 In Figure a in [], the red square represents the area with a radius of 5° centered on the cloud cluster center. The analysis shows that this stage has the following characteristics: WN0 at the initial stage of the typhoon ( Figure 2Figure (b) shows a relatively loose structure, indicating that the overall organization of the cloud cluster is not yet strong; the first-order component WN1 of the Fourier decomposition ( Figure 2 Figure (c) reflects the eccentric shape of the cloud cluster and has a significant spiral structure, which is a typical manifestation in the initial stage of a typhoon, indicating that the cloud cluster has strong asymmetry characteristics; the higher-order components WN2~WN4 of the Fourier decomposition ( Figure 2 Figures (d)~(f)) capture the local perturbations and spiral shape characteristics of the cloud cluster, where the WN2 component further refines the eccentric spiral structure, and the WN3 and WN4 components capture more subtle perturbations.

[0079] Figure 3 shows the morphology and Fourier decomposition characteristics of the cloud cluster that coexisted during the same period and failed to develop into a typhoon. The WN0 component ( Figure 2 Figure (b)) shows obvious symmetry characteristics, indicating that the brightness temperature distribution of non-typhoon cloud clusters is mainly uniform. The WN1 component ( Figure 3 Figure (c)) does not observe significant spiral structure and eccentric characteristics, indicating that non-typhoon cloud clusters lack the typical asymmetry characteristics of cloud clusters in the initial stage. The higher-order components WN2~WN4 ( Figure 3 Figures (d)~(f)) lack obvious perturbation characteristics, further indicating that the overall structure of non-typhoon cloud clusters is relatively simple. Figure 3

[0080] Figure 4 Figure 2 shows the same typhoon as Figure 2 , but its morphology and Fourier decomposition characteristics after it formed (time: 12:00 on July 20, 2021, intensity: 80 kt). It can be observed that as the typhoon develops, the WN0 component ( Figure 4 Figure (b)) becomes more compact, the range of the WN1 component ( Figure 4 Figure (c)) also relatively shrinks, and the higher-order components WN2~WN4 ( Figure 4 Figures (d)~(f)) basically no longer show significant perturbation characteristics, further indicating that the overall organization of the cloud cluster after the typhoon forms is enhanced.

[0081] By combining the symmetry and asymmetry characteristics extracted by Fourier decomposition and using them as the core input data of the typhoon initial identification AI model, the present invention can effectively identify the significant differences between cloud clusters in the initial stage of a typhoon and non-typhoon cloud clusters, and significantly improve the accuracy and reliability of typhoon initial identification.

[0082] Step 4: Identification and judgment of the initial stage of a typhoon

[0083] This embodiment uses a typhoon initial identification AI model similar to that in Embodiment 1, which is based on the Single Shot Detector (SSD) object detection algorithm. The input of the model is the cloud symmetry features and asymmetry features extracted in Step 3, and the output is the judgment result on whether the target cloud belongs to the typhoon initial stage. The typhoon initial identification AI model in this embodiment pays special attention to the following key points during the training process: focusing on the spiral structure features of the WN1 component; combining the structural compactness of WN0; analyzing the local perturbation features of WN2 - WN4; the model is trained using historical case data from 2000 to 2020, and verified using actual cases in 2021.

[0084] Step 5: Output of the identification result

[0085] Similar to Embodiment 1, the identification results of the typhoon initial stage are output, including information such as the coordinates of the typhoon center position, the initial time, the development probability, and the confidence level. The identification results for Typhoon No. 9 in 2021 show that: the initial signs of the typhoon were successfully identified 2.5 hours in advance; the identification confidence level reached 87%; the positioning error of the cloud center was within 20 kilometers.

[0086] Through the description of the above specific embodiments, it can be seen that the typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry features of the present invention realizes the efficient and accurate identification of the typhoon initial stage by introducing technical means such as multi-channel satellite data fusion, asymmetry feature extraction, and collaborative processing of the dual AI model architecture, significantly improving the timeliness and reliability of typhoon early warning, and providing reliable data support and technical guarantee for typhoon path prediction and disaster prevention and control. The above embodiments are only further elaborations of the present invention and do not constitute any limitation to the protection scope of the present invention. Those skilled in the art can make various forms of changes and improvements to the model structure, feature extraction method, and identification process, etc., without departing from the essence of the present invention, and these changes and improvements should all fall within the protection scope of the present invention.

Claims

1. A typhoon initial identification method based on a dual AI model architecture and typhoon asymmetry characteristics, characterized in that, Including: SS1. Obtain multi-channel satellite image data within the target area, including at least brightness temperature data of the infrared channel and the water vapor channel, and preprocess the obtained multi-channel satellite image data; SS2. Based on CNN, construct a cloud cluster center localization AI model for the first stage of typhoon genesis identification, analyze the preprocessed multi-channel satellite image data, extract cloud cluster features within the target area and output the cloud cluster center position, including: a feature extraction layer for extracting brightness temperature features and local convection intensity of the satellite image; a spatial attention mechanism layer for enhancing the perception ability of local cloud cluster features; a position regression layer for predicting the coordinates of the cloud cluster center position according to the extracted features; SS3. Taking the identified and located cloud cluster center as a reference, perform multi-order feature decomposition on the multi-channel satellite image data within a preset radius range centered on the cloud cluster center, decompose the brightness temperature values within the cloud cluster area into multi-order fluctuation components with radial distribution, and extract the symmetry and asymmetry features of the cloud cluster; SS4. Based on the SSD object detection algorithm, construct a typhoon genesis identification AI model for the second stage of typhoon genesis identification, analyze the extracted cloud cluster symmetry and asymmetry features, including: a feature input layer for receiving cloud cluster symmetry and asymmetry features; a feature fusion layer for fusing the input symmetry and asymmetry features and using the attention mechanism to enhance the perception ability of significantly asymmetric features; an SSD detection module for performing object detection on the fused multi-scale features, identifying and locating potential typhoon genesis areas; an output layer for judging whether the target cloud cluster belongs to the typhoon genesis stage according to the classification result of the SSD detection module; SS5. Based on the typhoon genesis identification judgment result, if the target cloud cluster belongs to the typhoon genesis stage, output the typhoon genesis stage identification result.

2. The typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, characterized in that In step SS1, the infrared channel data is brightness temperature data in the 11-micron band, used to capture cloud top temperature and atmospheric convection information; the water vapor channel data is brightness temperature data in the 6.7-micron band, used to detect the water vapor content and distribution in the middle layer of the atmosphere; the horizontal resolution of the data is 4 - 10 km, and the time resolution is 1 - 6 hours.

3. The typhoon genesis identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, characterized in that, In step SS1, data preprocessing includes data quality control, standardization, unified adjustment of spatio-temporal resolution, and conversion of brightness temperature data into a grayscale image format suitable for input to a deep learning model, where: the data quality control includes removing strip noise, cloud detection, and outlier rejection from satellite image data. The strip noise removal uses the wavelet transform method, the cloud detection is based on the brightness temperature difference threshold of the infrared and water vapor channels, and the outlier rejection uses the 3 σ criterion; the standardization includes performing min-max normalization or Z-score standardization on the brightness temperature data; the unified adjustment of spatio-temporal resolution includes uniformly changing the time resolution of multi-channel satellite image data to a 15-minute time interval using linear interpolation, and changing the spatial resolution to a 1-kilometer spatial resolution using bilinear interpolation. At the same time, according to the meteorological characteristics of the target area, the multi-channel satellite image data is regionally cropped, and only the image data within the coverage area of the target area is retained.

4. The typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, wherein, In step SS2, the feature extraction layer of the cloud cluster center localization AI model uses depthwise separable convolution to reduce the model parameter quantity and computational complexity, and performs normalization processing on the features output by each convolution layer through batch normalization; The spatial attention mechanism layer highlights the features of the target cloud cluster area by generating a dynamic weight matrix, where the dynamic weight matrix is generated based on the maximum pooling and average pooling results of the convolutional feature map and multiplied element-wise with the original feature map to enhance the perception ability of fuzzy or incomplete cloud cluster features.

5. The typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, characterized in that, In step SS2, the cloud cluster center positioning AI model is trained based on the satellite image dataset with labeled cloud cluster center positions. The dataset is divided according to the ratio of 70% for training, 20% for validation, and 10% for testing. The training process is completed through supervised learning and data augmentation techniques are used to expand the training samples, including random rotation of ±30°, random scaling of ±10%, random translation of ±20 pixels, and horizontal flipping. The positioning error is used as the loss function for optimization. The loss function L loc is defined as , where x i , y i is the predicted position of the i th sample, is the true labeled position of the i th sample, i = 1, 2, …, N , N is the total number of samples; and during model training, the Adam optimizer with an adaptive learning rate is used to update the weights, and the learning rate is dynamically adjusted through the loss value of the validation set to accelerate convergence.

6. The typhoon genesis identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, wherein In step SS3, use the Fourier decomposition technique for multi-order feature decomposition, including the following sub-steps: SS31. Taking the center of the cloud cluster located in step SS2 as a reference, a target area with a radius ranging from 1 to 5° centered on the cloud cluster center is selected, and it is divided into grids at a sampling interval of 0.1°, generating a brightness temperature distribution matrix of the target area. Each grid point in the matrix corresponds to the satellite observed brightness temperature value, and interpolation processing is performed on the data in the boundary area according to the cloud cluster distribution characteristics to supplement possible vacancies; SS32. Perform radial data processing on the brightness temperature values in the target area, convert the matrix data into radial distribution data with the cloud cluster center as the base point, and the sampling points in the azimuth direction are distributed within the range of 0 to 360° at a preset interval. At the same time, filter the strong noise part of the radial data; SS33. Perform Fourier decomposition on the radial brightness temperature distribution data along the azimuth direction, decompose it into several order fluctuation components, and the Fourier decomposition formula is as follows: Among them, T ( θ , r ) represents the radial distribution function of the brightness temperature, r is the radial distance, θ is the azimuth angle; i is the order of the Fourier decomposition and i = 1, 2, 3, …, n , n is the highest order of the Fourier decomposition and n ≥ 4; a 0( r ) is the amplitude of the 0th order component, a i ( r ) and b i ( r ) are the cosine and sine component amplitudes of the order i respectively; WN 0( r ) represents the 0th order component of the Fourier decomposition, WNi ( r ) represents the i th order component; SS34. Extract the 0th order component WN0 in the Fourier decomposition result as the symmetry feature of the cloud cluster, and the 1st to 4th order components WN1~WN4 as the asymmetry features. Among them, WN1 reflects the eccentric shape of the cloud cluster, and WN2~WN4 capture the local perturbation characteristics and spiral shape characteristics of the cloud cluster, and reconstruct the extracted symmetry and asymmetry features into a multi-channel feature map, with each channel corresponding to one component.

7. The typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, characterized in that In step SS4, when the feature fusion layer of the typhoon initial identification AI model fuses the input symmetry and asymmetry features, it uses multi-scale convolution operations and combines the FPN structure to generate multiple feature maps of different scales, including: When performing multi-scale convolution extraction on the input symmetry and asymmetry features respectively, parallel convolution is performed by setting convolution kernels of different sizes to extract local features and global features at different scales. After the convolution operation, the multi-scale features are fused into a unified high-dimensional feature representation through feature splicing; Use the multi-scale FPN structure to further process the fused features. Through bottom-up feature extraction and top-down feature fusion, the perception ability of typhoon initial cloud cluster targets at different scales is enhanced. The bottom-up feature extraction extracts features layer by layer through a deep convolutional network to generate feature maps of different resolutions. The lower-resolution feature maps are used to capture the global mode information of the cloud cluster, and the higher-resolution feature maps are used to capture the local perturbation features of the cloud cluster. The top-down feature fusion fuses the low-resolution feature maps with the high-resolution feature maps layer by layer through upsampling operations, and reduces the dimension of the fused features through 1×1 convolution to achieve the full utilization of multi-scale information.

8. The typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 7, characterized in that In the typhoon initial identification AI model, the channel attention mechanism is further combined during the feature fusion process to enhance the perception ability of the asymmetry features by learning the weights of different channel features, including: performing global average pooling on the fused multi-scale feature maps to generate the global features of each channel; calculating the weights of each channel through a fully connected layer and performing weighted processing on the original feature maps: Among them, f pool is the feature after global pooling, W 1 and W 2 are the weight matrices of the fully connected layers, σ is the Sigmoid activation function, w c is the weight for each channel.

9. The typhoon genesis recognition method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 7, wherein, When the SSD detection module of the typhoon initial identification AI model performs target detection, it includes the following steps: SS41. The multi-scale feature maps output by the feature fusion layer respectively correspond to feature layers of different resolutions. Each feature layer generates a feature map of a fixed size through a convolution operation. The feature maps of each scale are used to detect cloud cluster targets of specific sizes to comprehensively cover the possible typhoon initial regions. SS42. On the feature maps of each scale, multiple candidate detection boxes are generated through preset anchor boxes. The size and aspect ratio of the anchor boxes are adjusted according to the resolutions of the feature maps of different scales. The size range of the anchor boxes is from 16 pixels to 512 pixels, covering the detection requirements from small local disturbance clouds to large convective systems. The aspect ratios of the anchor boxes include 1:1, 2:1, and 1:2 to adapt to the possible shapes and distribution characteristics of the initial typhoon clouds. Each anchor box is associated with each pixel point on the feature map to generate candidate regions for subsequent classification and regression prediction. SS43. The target category prediction and position regression correction of the candidate regions are respectively completed through the classification sub-network and the regression sub-network, where: the classification sub-network is used to classify the candidate regions in each feature map to determine whether they belong to the clouds in the initial typhoon stage. The classification output calculates the probability that each candidate region belongs to the initial typhoon or non-initial typhoon category through the Softmax function, and the classification loss function uses the cross-entropy loss; the regression sub-network is used to perform regression correction on the bounding boxes of the candidate regions, predict their accurate positions and scales, and uses the smooth L1 loss to optimize the regression error. SS44. Post-process the results of the classification and regression predictions, and use the non-maximum suppression method to remove redundant detection boxes, including: sorting the candidate detection boxes according to the confidence levels output by the classification sub-network; for the detection boxes with confidence levels exceeding the preset threshold, calculate their intersection over union (IoU) with other detection boxes. If the IoU exceeds the set threshold, then delete this detection box; only retain the detection result with the highest confidence level, including the position and scale of the detection box and the confidence score of the initial typhoon stage. SS45. Combine the detection results of the feature maps of different scales, and generate the final typhoon initial recognition result through weighted fusion. The weight assignment is dynamically adjusted according to the detection confidence levels of the feature maps of different scales. If the detection confidence level of the low-resolution feature map is higher, then preferentially retain the detection result of the low-resolution feature map; if the detection confidence level of the high-resolution feature map is higher, then preferentially retain the detection result of the high-resolution feature map; the final output result includes the central position of the initial typhoon clouds, the scale of the detection box, and the confidence score of the initial typhoon stage.

10. The typhoon genesis identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 9, characterized in that In sub-step SS41, each feature layer generates a feature map of a fixed size through a convolution operation, including: performing multiple 3×3 convolution operations on the input fused features to extract spatial features; gradually reducing the spatial resolution of the feature map through downsampling operations to generate feature maps of multiple scales.

11. The typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, wherein In step SS4, the typhoon genesis recognition AI model is trained based on a multi-channel feature image dataset containing labeled typhoon genesis stage cloud clusters and non-typhoon genesis cloud clusters, as well as the corresponding asymmetry features. The dataset is divided according to the ratio of 70% for training, 20% for validation, and 10% for testing. During the training process, a classification cross-entropy loss function and a regression loss function are used to jointly optimize the object detection and classification performance. The Adam optimizer or stochastic gradient descent is used to optimize the model parameters, and new data is regularly introduced to update the training set for continuous training of the model.

12. The typhoon initial identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1, wherein In step SS5, the output of the typhoon genesis stage recognition result includes at least the following information: the central position coordinates of the target cloud cluster, the recognition confidence, the judgment result of the typhoon genesis stage and its credibility score. The recognition confidence is calculated through the correlation between the classification output probability and the asymmetry feature weight, and the credibility score is calculated by comprehensively considering the consistency between the classification output result and the object detection result.

13. A computer program product comprising computer instructions, characterized in that, The computer instructions are used to execute the typhoon genesis recognition method based on the dual AI model architecture and typhoon asymmetry features according to any one of claims 1 to 12.

14. A computer-readable storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by a processor, it implements the typhoon genesis recognition method based on the dual AI model architecture and typhoon asymmetry features according to any one of claims 1 to 12.

Images