Typhoon nascent identification method based on double-AI model architecture and typhoon asymmetry characteristics, medium and program product

Abstract

The invention discloses a typhoon nascent identification method based on a double-AI model architecture and typhoon asymmetry characteristics, a medium and a program product, and belongs to the technical field of meteorological monitoring and artificial intelligence, and the method comprises the steps: firstly, obtaining multi-channel satellite image data, and carrying out the preprocessing; then, the cloud cluster center is positioned through a first AI model; taking the center of the cloud cluster as a reference, and extracting symmetry features and asymmetry features of the cloud cluster; analyzing the extracted features through a second AI model, and judging whether the target cloud cluster is in a typhoon primary stage or not; and finally, outputting identification results including the typhoon center position, the birth time, the development probability and the like. According to the method, the accuracy and the reliability of typhoon nascent identification are effectively improved through staged identification of the double-AI model architecture in combination with the unique asymmetry characteristic of the cloud cluster, and the method has a relatively high practical value.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological monitoring and artificial intelligence technology, and involves 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 cloud clusters in the initial stage of a typhoon, thereby improving the accuracy and reliability of early typhoon monitoring. Background Art

[0002] The formation and development of a typhoon is a complex dynamic and thermodynamic process, which usually includes four stages: typhoon inception, development, maturity and extinction. Timely and accurate identification of the inception stage of a typhoon is of great significance for improving typhoon early warning capabilities and reducing disaster losses. The inception stage of a typhoon usually refers to the initial formation stage of a typhoon embryo over the tropical ocean. This stage is characterized by the fact that cyclonic disturbances have just begun to develop and the rotating structure has not yet fully formed, showing a loose and irregular distribution of clouds, and its intensity is usually far lower than that of a strong tropical storm. Since the typhoon cloud clusters in the inception stage do not have obvious vortex characteristics and weak convective activity, and are significantly affected by sea surface temperature, vertical wind shear, etc., their cloud structure is highly similar to other general atmospheric disturbance structures that cannot develop into typhoons (such as monsoon clouds, mesoscale convective systems, etc.), which makes the identification of typhoons in the inception stage extremely challenging.

[0003] In recent years, with the rapid development of artificial intelligence (AI) technology, typhoon recognition technology based on deep learning has been widely used in the field of meteorology. These technologies can effectively extract the spatial distribution characteristics and temporal evolution information of typhoons by analyzing large-scale satellite remote sensing data. For example, convolutional neural networks (CNN) are widely used to extract cloud features from satellite images, and target detection algorithms such as single shot detectors (SSD) are used to locate typhoon areas. In the development and maturity stages of typhoons, especially for typhoon vortices of strong tropical storms and above, the correct recognition rate of AI models can reach more than 90%. However, for typhoon vortices below the strong tropical storm level, the recognition accuracy is only 40% to 80%. Especially in the early stages of typhoons, since the typhoon structure has not yet been fully formed, its cloud characteristics are extremely similar to those of ordinary atmospheric disturbance clouds, and the recognition accuracy of existing AI models has further decreased to only 10% to 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 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.

[0007] (II) Technical solution In order to achieve the purpose of the invention and solve the technical problems, the present invention adopts the following technical solutions: The first invention object of the present invention is to provide a typhoon incipient identification method based on a dual AI model architecture and typhoon asymmetric characteristics, which is used to efficiently and accurately automatically identify the incipient stage of a typhoon by combining multi-channel satellite data, asymmetric feature extraction and a dual model collaborative architecture, thereby improving the timeliness and accuracy of typhoon warnings. The method comprises at least the following steps when implemented: SS1. Acquire multi-channel satellite image data in the target area, including at least infrared channel and water vapor channel brightness temperature data, and pre-process the acquired satellite data, including at least data quality control, standardization, unified adjustment of temporal and spatial resolutions, and conversion of brightness temperature data into a grayscale image format suitable for deep learning model input; SS2. A cloud center location AI model is constructed based on a convolutional neural network (CNN) as the first stage of typhoon initial identification. The model is used to analyze the preprocessed multi-channel satellite image data, extract the cloud features in the target area, and output the center position of the cloud 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 satellite images, the spatial attention mechanism layer is used to enhance the perception of local features of clouds, and the position regression layer predicts the coordinates of the cloud center based on the extracted features; SS3. Taking the cloud center identified and located in step SS2 as the reference, use Fourier decomposition technology to perform multi-order feature decomposition on the multi-channel satellite image data within a preset radius with the cloud center as the center, decompose the brightness temperature value in the cloud area into radially distributed multi-order fluctuation components, and extract the symmetry and asymmetry characteristics of the cloud. The symmetry characteristics represent the overall uniform distribution characteristics of the cloud, and the asymmetry characteristics are used to quantify the local disturbance characteristics and spiral morphology of the cloud in the initial stage. SS4. A typhoon inception identification AI model is constructed as the second stage of typhoon inception identification based on the single shot detector (SSD) target detection algorithm, and the symmetry and asymmetry features of the cloud cluster extracted in step SS3 are analyzed, wherein the typhoon inception identification AI model includes a feature input layer, a feature fusion layer, an SSD detection module and an output layer, wherein: 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 use the attention mechanism to enhance the perception of significant asymmetric features; the SSD detection module is used to perform target detection on the fused multi-scale features, identify and locate the potential typhoon inception area; the output layer determines whether the target cloud cluster belongs to the typhoon inception stage according to the classification results of the SSD detection module; SS5. Based on the typhoon incipient identification and judgment results, if the target cloud group belongs to the incipient stage of the typhoon, the typhoon incipient stage identification results are output, including the typhoon center position coordinates, incipient time, development probability and credibility, to provide support for subsequent typhoon path prediction, intensity development analysis and warning issuance.

[0008] (III) Technical Effect Compared with the prior art, the typhoon incipient identification method, medium and program product of the present invention based on the dual AI model architecture and typhoon asymmetry characteristics have the following beneficial and significant technical effects: (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 disturbance characteristics of the cloud cluster in the early stage of the typhoon. Through the joint modeling of multi-channel data and Fourier characteristics, the present invention achieves high-resolution and high-timeliness feature expression in the early cloud cluster dynamics analysis, significantly improving the perception of the complex characteristics of the typhoon in the early stage.

[0009] (2) The present invention adopts a dual AI model architecture, including a feature extraction model and a target detection model. The feature extraction model extracts multi-scale features through multi-scale convolution operations and FPN structures, and combines the attention mechanism to enhance the perception of key features; the target detection model is based on the SSD module, which integrates the classification sub-network and the position regression sub-network to classify and accurately locate the cloud targets in the early stages of the typhoon. Through the synergy of the dual models, the present invention can efficiently screen and identify the initial typhoon clouds in complex backgrounds, while taking into account the classification accuracy and detection accuracy. In addition, the target detection module further improves the accuracy and robustness of the candidate detection frame through non-maximum suppression (NMS) and multi-scale comprehensive decision-making strategies, providing a strong technical guarantee for the accurate identification of the initial stage of the typhoon.

[0010] (3) Through the comprehensive analysis of multi-channel satellite data and the efficient application of deep learning models, the present invention has significantly improved the accuracy, reliability and real-time performance of typhoon early warning. It is not only suitable for typhoon warning, but can also be extended to other fields involving symmetry and asymmetry feature analysis, such as tropical cyclone monitoring and atmospheric convective system identification. It has strong practicality and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein: Figure 1 It is an implementation flow chart of a typhoon incipient identification method based on a dual AI model architecture and typhoon asymmetry characteristics provided by an embodiment of the present invention; Figure 2 This is the embryonic 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 25kt=12.86m / s. The red box is within the range of 5° of the typhoon embryo. (a) is the 11-micron data of the infrared channel, (b)~(f) are the Fourier 0-4 wave decomposition images within 5° of the typhoon center. The horizontal axis in the figure is longitude and the vertical axis is latitude; Figure 3 The cloud cluster shape that cannot form a typhoon and its Fourier decomposition structure diagram, at the same time Figure 2 , the red box is within the range of 5° without the typhoon embryo, (a) is the 11-micron data of the infrared channel, (b)~(f) are the Fourier 0-4 wave decomposition images within the central 5° range, the horizontal axis is longitude, and the vertical axis is latitude; Figure 4 The typhoon's shape after formation and its Fourier decomposition structure diagram are shown in Figure 1. Figure 2 It is the same typhoon, but after the typhoon was formed, at 12:00 on July 20, 2021, the typhoon intensity was 80kt=41.1m / s. The red box is the box within 5° of the typhoon range. (a) is the 11-micron data of the infrared channel, (b)~(f) are the Fourier 0-4 wave decomposition images within 5° of the typhoon center, the horizontal axis is longitude, and the vertical axis is latitude. DETAILED DESCRIPTION

[0012] The following is a further explanation of the typhoon incipient identification method, medium and program product based on the dual AI model architecture and typhoon asymmetric characteristics of the present invention through specific implementation methods. By combining multi-channel satellite data, asymmetric feature extraction technology and dual model collaborative architecture, the cloud characteristics in the target area are comprehensively analyzed and modeled to significantly improve the recognition accuracy and reliability of the typhoon incipient stage. The specific implementation process of the present invention will be described in detail below in conjunction with the accompanying drawings, and the key technical solutions of the present invention and their implementation details will be explained, but the present invention is not limited to the specific contents described in the following embodiments.

[0013] Example 1 like Figure 1 As shown, the typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics provided by the embodiment of the present invention mainly includes the following steps when implemented: SS1. Data acquisition and preprocessing Acquire multi-channel satellite image data in the target area, including at least infrared channel and water vapor channel brightness temperature data, and preprocess the acquired satellite data, including at least data quality control, standardization processing, unified adjustment of temporal and spatial resolutions, and conversion of brightness temperature data into a grayscale image format suitable for deep learning model input.

[0014] Preferably, in 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 the cloud top temperature and atmospheric convection information; 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 to 10 kilometers, taking into account both computational efficiency and the ability to identify small and medium-scale convective systems; the time resolution is 1 to 6 hours, in order to capture the rapid development and changes of the typhoon in its infancy.

[0015] When preprocessing the acquired satellite data, data quality control includes strip noise removal, cloud detection and outlier removal of satellite image data. The strip noise removal adopts the wavelet transform method, the cloud detection is based on the brightness temperature difference threshold judgment of the infrared and water vapor channels, and the outlier removal adopts 3 σ Standardization processing includes minimum-maximum normalization or Z-score normalization of brightness temperature data to eliminate the difference in numerical range between data of different channels, where minimum-maximum normalization maps the data to the interval [0,1], and Z-score normalization converts the data into a distribution with a mean of 0 and a standard deviation of 1; unified adjustment of spatiotemporal resolution includes using linear interpolation to unify the temporal resolution of multi-channel satellite image data to a time interval of 15 minutes, and using bilinear interpolation to unify the spatial resolution to a spatial resolution of 1 km. At the same time, according to the meteorological characteristics of the target area, the multi-channel satellite image data is regionally cropped to retain only the image data within the coverage area of ​​the target area to reduce the interference of irrelevant data on the subsequent deep learning model processing.

[0016] SS2. Identification and location of cloud center Based on the convolutional neural network (CNN), an AI model for locating the cloud center is constructed as the first stage of typhoon initial identification. The model is used to analyze the preprocessed multi-channel satellite image data, extract the cloud features in the target area, and output the center position of the cloud 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 satellite images, the spatial attention mechanism layer is used to enhance the perception of local features of clouds, and the position regression layer predicts the coordinates of the center position of the cloud based on the extracted features.

[0017] Preferably, the feature extraction layer of the cloud center positioning AI model adopts deep separable convolution to reduce the number of model parameters and computational complexity, while improving the efficiency of extracting brightness temperature distribution characteristics and local convection intensity, and normalizes the features output by each convolution layer through batch normalization to accelerate model convergence and improve the generalization performance of the model; the spatial attention mechanism layer highlights the characteristics of the target cloud area by generating a dynamic weight matrix, wherein the dynamic weight matrix is ​​generated based on the maximum pooling and average pooling results of the convolution feature map, and is multiplied element-by-element with the original feature map, thereby enhancing the perception of blurred or incomplete cloud features.

[0018] The cloud center positioning AI model is trained based on a satellite image dataset with the cloud center location annotated. The dataset is divided into 70% for training, 20% for validation, and 10% for testing. The training process is completed through supervised learning and data enhancement technology is used to expand the training samples, including ±30° random rotation, ±10% random scaling, ±20 pixel random translation, and horizontal flipping to improve the model's robustness to cloud morphology changes; and the positioning error is used as the loss function for optimization. The loss function is defined as ,in, x i , y i For the i The predicted position of samples, For the i The true labeled position of each sample, i =1,2,…, N , N The Euclidean distance between the predicted cloud center position and the true annotation position is calculated to minimize the positioning error. The Adam optimizer with 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 according to the loss value of the validation set to accelerate convergence.

[0019] SS3. Cloud asymmetry feature extraction Taking the cloud center identified and located in step SS2 as the reference, the Fourier decomposition technology is used to perform multi-order feature decomposition on the multi-channel satellite image data within a preset radius with the cloud center as the center, and the brightness temperature value in the cloud area is decomposed into radially distributed multi-order fluctuation components, and the symmetry and asymmetry characteristics of the cloud are extracted, where the symmetry characteristics represent the overall uniform distribution characteristics of the cloud, and the asymmetry characteristics are used to quantify the local disturbance characteristics and spiral morphology of the cloud in the initial stage.

[0020] Preferably, when the Fourier decomposition technology is used to perform multi-order feature decomposition on the multi-channel satellite image data within a preset radius with the center of the cloud cluster as the center, the following sub-steps are mainly included: SS31. Target area selection: Based on the cloud center identified and located in step SS2, select the target area with a radius of 1-5° and the cloud center as the center, and divide it into grids according to the sampling interval of 0.1° to generate the brightness temperature distribution matrix of the target area. Each grid point in the matrix corresponds to the brightness temperature value observed by the satellite, and interpolate the data of the boundary area according to the cloud distribution characteristics to fill in possible gaps; SS32. Radial data construction: radial data processing is performed on the brightness temperature values ​​in the target area, and the matrix data is converted into radial distribution data with the cloud center as the base point. The radial data are distributed in the range of 0~360° at the preset intervals in the azimuth direction sampling points to ensure the continuity and integrity of the radial distribution data. At the same time, the strong noise part of the radial data is filtered to reduce the impact of external interference on the subsequent Fourier decomposition; SS33. Fourier decomposition: Perform Fourier decomposition on the radial brightness temperature distribution data along the azimuth direction and decompose it into several order wave components. The Fourier decomposition formula is as follows: in, T ( θ , r ) represents the radial distribution function of brightness temperature, r is the radial distance, θ is the azimuth; i is the order of Fourier decomposition and i =1,2,3,…, n , n is the highest order of 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 order i The amplitudes of the cosine and sine components of ; WN 0( r ) represents the 0th-order component of Fourier decomposition, Wl ( r ) indicates the i order component; SS34. Symmetry and asymmetry feature extraction: The 0th order component WN0 in the Fourier decomposition result is extracted as the symmetry feature of the cloud cluster, reflecting the overall uniform distribution characteristics of the brightness temperature of the cloud cluster, characterizing the overall vortex shape of the cloud cluster in the initial stage of the typhoon, and the 1st to 4th order components WN1~WN4 are extracted as the asymmetry features of the cloud cluster, where WN1 reflects the eccentric shape of the cloud cluster, and WN2~WN4 capture the local disturbance characteristics and spiral morphology characteristics of the cloud cluster. The extracted symmetry features and asymmetry features are reconstructed into a multi-channel feature map, with each channel corresponding to a component, as the input data of the subsequent typhoon initial recognition AI model.

[0021] SS4. Identification and judgment of typhoon in its early stages A typhoon inception identification AI model is constructed as the second stage of typhoon inception identification based on the single shot detector (SSD) target detection algorithm, and the symmetry features and asymmetry features of the cloud cluster extracted in step SS3 are analyzed. The typhoon inception identification AI model includes a feature input layer, a feature fusion layer, an SSD detection module and an output layer, wherein: 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 use the attention mechanism to enhance the perception of significant asymmetric features; the SSD detection module is used to perform target detection on the fused multi-scale features, identify and locate potential typhoon inception areas; the output layer determines whether the target cloud cluster belongs to the typhoon inception stage according to the classification results of the SSD detection module.

[0022] As a preferred method, the feature fusion layer of the typhoon incipient recognition AI model uses multi-scale convolution operations to fuse the input symmetric features and asymmetric features, and combines the multi-scale feature pyramid network (FPN) structure to generate multiple feature maps of different scales to enhance the ability of the fusion features to express the multi-scale cloud features in the incipient stage of the typhoon, including: When performing multi-scale convolution operations, the feature fusion layer performs multi-scale convolution extraction on the input symmetric features (WN0) and asymmetric features (WN1-WN4) respectively. The multi-scale convolution performs parallel convolution on the input features by setting convolution kernels of different sizes (for example, 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 splicing; When optimizing the feature pyramid structure (FPN), a multi-scale feature pyramid network (FPN) structure is used to further process the fused features. The perception ability of typhoon primary cloud targets of 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 map is used to capture the global pattern information of the cloud, and the higher resolution feature map is used to capture the local disturbance characteristics of the cloud. The top-down feature fusion fuses the low-resolution feature map with the high-resolution feature map layer by layer through upsampling operations, and reduces the dimension of the fused features through 1×1 convolution to fully utilize the multi-scale information.

[0023] Further preferably, the feature fusion layer of the typhoon incipient recognition AI model further combines the channel attention mechanism in the feature fusion process, and enhances the perception of asymmetric features (WN1-WN4) by learning the weights of different channel features, including: performing global average pooling on the fused multi-scale feature map to generate the global features of each channel; calculating the weight of each channel through the fully connected layer, and performing weighted processing on the original feature map. The calculation formula of the channel weight is: in, f pool is the feature after global pooling, W 1 and W 2 is the weight matrix of the fully connected layer, σ is the Sigmoid activation function, w c is the weight for each channel; Further preferably, the SSD detection module of the typhoon incipient recognition AI model mainly includes the following steps when performing target detection: SS41. Multi-scale feature map generation The multi-scale feature maps (generated by FPN) output by the feature fusion layer correspond to feature layers of different resolutions (for example, low-resolution feature maps are used to capture the global information of clouds in the initial stage of a typhoon, and high-resolution feature maps are used to extract local disturbance features in the target area). Each feature layer generates a fixed-size feature map through a convolution operation, and each scale of the feature map is used to detect cloud targets of a specific size, so as to fully cover possible typhoon initial areas. 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 fused features to extract spatial features; gradually reducing the spatial resolution of the feature map through a downsampling operation to generate feature maps of multiple scales (for example, feature maps of different spatial resolutions such as 8×8, 16×16 and 32×32).

[0024] SS42. Candidate region generation On the feature map of each scale, multiple candidate detection boxes are generated by presetting anchor boxes. The size and aspect ratio of the anchor boxes are adjusted according to the resolution of the feature maps of different scales. The size of the anchor boxes ranges 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 typhoon's primary clouds. Each anchor box is associated with each pixel on the feature map to generate candidate regions for subsequent classification and regression prediction. SS43. Classification and Regression Prediction The target category prediction and position regression correction of the candidate area are completed by the classification subnetwork and the regression subnetwork respectively, wherein: the classification subnetwork is used to classify the candidate area in each feature map to determine whether it belongs to the cloud cluster in the initial stage of the typhoon. The classification output is calculated by the Softmax function to determine the probability that each candidate area belongs to the initial stage of the typhoon or the non-initial stage of the typhoon, and the classification loss function adopts the cross entropy loss; the regression subnetwork is used to perform regression correction on the bounding box of the candidate area, predict its precise position and scale, and use the smooth L1 loss to optimize the regression error; SS44. Non-maximum suppression (NMS) The results of classification and regression prediction are post-processed, and the non-maximum suppression (NMS) method is used to remove redundant detection frames, including: sorting candidate detection frames according to the confidence output by the classification subnetwork; for detection frames whose confidence exceeds the preset threshold, calculating the intersection over union (IoU) between them and other detection frames, and deleting the detection frame if the IoU exceeds the set threshold (for example, 0.5); only retaining the detection results with the highest confidence, including the position, scale and confidence score of the detection frame in the initial stage of the typhoon; SS45. Multi-scale integrated decision making Combining the detection results of feature maps of different scales, the final typhoon incipient recognition results are generated through weighted fusion. The weight distribution is dynamically adjusted according to the detection confidence of feature maps of different scales. If the detection confidence of the low-resolution feature map is higher, the detection results of the low-resolution feature map are retained first; if the detection confidence of the high-resolution feature map is higher, the detection results of the high-resolution feature map are retained first; the final output results include the central position of the typhoon's incipient cloud cluster, the scale of the detection frame, and the confidence score of the typhoon's incipient stage.

[0025] Preferably, the typhoon incipient recognition AI model is trained based on a multi-channel feature image dataset containing labeled typhoon incipient clouds and non-typhoon incipient clouds and corresponding asymmetry features. The dataset is divided into 70% for training, 20% for verification, 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 target 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 incipient clouds) are regularly introduced to update the training set to continuously train the model.

[0026] SS5. Recognition result output Based on the typhoon incipient identification judgment result of step SS4, if the target cloud group belongs to the incipient stage of the typhoon, the typhoon incipient stage identification result is output, including the typhoon center position coordinates, incipient time, development probability and credibility, to provide support for subsequent typhoon path prediction, intensity development analysis and warning issuance.

[0027] Preferably, the output of the recognition results in the initial stage of a typhoon includes at least: the center position coordinates of the target cloud, the recognition confidence, the judgment results in the initial stage of the typhoon and its credibility score, wherein the recognition confidence is calculated by 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 target detection result.

[0028] Example 2 Based on the above-mentioned Example 1, this Example 2 will use Typhoon No. 9 in 2021 as a specific case to further illustrate the specific application of the present invention. The main difference between this example and Example 1 lies in the selection of data sources and some parameters, as well as a more intuitive explanation of the asymmetric characteristics of cloud clusters.

[0029] Step 1: Data acquisition and preprocessing Acquire multi-channel satellite image data in the target area. This embodiment uses GridSat data released by NOAA of the United States, which integrates multiple 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. The data includes brightness temperature data of 11 microns in the infrared channel and 6.7 microns in the water vapor channel. The horizontal resolution of the data is 0.07° latitude (about 8km), and the time resolution is 180 minutes. The acquired satellite data is preprocessed, including steps such as data quality control and standardization, which is similar to Example 1.

[0030] Step 2: Identify and locate the center of the cloud This example focuses on analyzing the satellite image of the target area at 12:00 on July 16, 2021, which corresponds to the initial stage of Typhoon No. 9, when the typhoon intensity was 25kt (about 12.86m / s). This example uses a cloud center positioning AI model based on a convolutional neural network (CNN) similar to Example 1, and successfully identifies and locates the center position of the target cloud.

[0031] Step 3: Extract cloud asymmetry features Taking the cloud center identified and located in step 2 as the reference, the Fourier decomposition technology is used to perform multi-order feature decomposition on the multi-channel satellite image data within a preset radius (set to 5° in this embodiment) with the cloud center as the center, and extract the symmetry feature WN0 and asymmetry features WN1-WN4 of the cloud. Figures 2 to 4 , which more intuitively displays the characteristic differences of typhoon incipient stages, non-typhoon clouds and mature typhoons in original satellite images and in each component after Fourier decomposition.

[0032] Figure 2 The cloud morphology and decomposition structure of Typhoon No. 9 in 2021 at its infancy (time: 12:00 on July 16, 2021, intensity: 25kt) are shown. Figure 2 The red box in Figure a represents the area with a radius of 5° centered on the center of the cloud cluster. Analysis shows that this stage has the following characteristics: WN0 ( Figure 2 The structure of the cloud cluster is relatively loose, indicating that the overall organization of the cloud cluster is not strong; the first-order component WN1 ( Figure 2 Figure c in the middle reflects the eccentric shape of the cloud cluster and has a significant spiral structure. This feature is a typical manifestation of the initial stage of a typhoon, indicating that the cloud cluster has a strong asymmetric characteristic; the high-order components WN2~WN4 ( Figure 2 Figures d~f in the middle capture the local disturbance and spiral morphology characteristics of the cloud, among which the WN2 component further refines the eccentric spiral structure, while the WN3 and WN4 components capture more subtle disturbances.

[0033] Figure 3 Shown with Figure 2 The morphology and Fourier decomposition characteristics of clouds that did not develop into typhoons during the same period. WN0 component ( Figure 3 Figure b in the middle shows obvious symmetry, indicating that the brightness temperature distribution of non-typhoon clouds is mainly uniform. Figure 3 No significant spiral structure and eccentricity were observed in Figure (c), indicating that non-typhoon clouds lack the typical asymmetric characteristics of clouds in their infancy. Figure 3 The lack of obvious disturbance characteristics in Figures d~f in the middle further indicates that the overall structure of non-typhoon clouds is relatively simple.

[0034] Figure 4 Shown with Figure 2 The same typhoon, but in its shape and Fourier decomposition characteristics after formation (time: 12:00 on July 20, 2021, intensity: 80kt). It can be observed that as the typhoon develops, the WN0 component ( Figure 4 Figure b) is more compact, and the WN1 component ( Figure 4 The range of the high-order components WN2~WN4 ( Figure 4 In Figures d~f in the middle, there is basically no longer any significant disturbance characteristics, which further illustrates that the overall organization of the cloud cluster is enhanced after the typhoon takes shape.

[0035] By combining the symmetry and asymmetry features extracted by Fourier decomposition and using them as the core input data of the typhoon incipient identification AI model, the present invention can effectively identify the significant differences between typhoon incipient clouds and non-typhoon clouds, thereby significantly improving the accuracy and reliability of typhoon incipient identification.

[0036] Step 4: Identification and judgment of the initial stage of a typhoon This embodiment adopts a typhoon incipient identification AI model based on a single-shot detector (SSD) target detection algorithm similar to that of Embodiment 1. The input of the model is the symmetry characteristics and asymmetry characteristics of the cloud cluster extracted in step 3, and the output is the judgment result of whether the target cloud cluster belongs to the incipient stage of the typhoon. The typhoon incipient identification AI model of this embodiment pays special attention to the following key points during the training process: focus on the spiral structure characteristics of the WN1 component; combine the structural compactness of WN0; analyze the local disturbance characteristics of WN2-WN4; the model training uses historical case data from 2000 to 2020, and the verification uses actual cases in 2021.

[0037] Step 5: Identification result output Similar to Example 1, the identification results of the initial stage of the typhoon are output, including the coordinates of the typhoon center, the initial time, the probability of development, the credibility and other information. The identification results of Typhoon No. 9 in 2021 show that the initial signs of the typhoon were successfully identified 2.5 hours in advance; the identification confidence reached 87%; the positioning error of the cloud center was within 20 kilometers.

[0038] Through the description of the above specific implementation methods, it can be seen that the typhoon initial identification method based on the dual AI model architecture and typhoon asymmetric characteristics of the present invention, by introducing multi-channel satellite data fusion, asymmetric feature extraction, dual AI model architecture collaborative processing and other technical means, has achieved efficient and accurate identification of the initial stage of the typhoon, significantly improved the timeliness and reliability of typhoon warnings, and provided reliable data support and technical guarantees 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 scope of protection of the present invention. Those skilled in the art may make various forms of changes and improvements to the model structure, feature extraction method and identification process based on actual needs without departing from the essence of the present invention, and these changes and improvements should all fall within the scope of protection of the present invention.

Claims

1. A typhoon incipient identification method based on a dual AI model architecture and typhoon asymmetry characteristics, characterized in that: include: SS1. Acquire multi-channel satellite image data in the target area, including at least infrared channel and water vapor channel brightness temperature data, and pre-process the acquired multi-channel satellite image data; SS2. Based on CNN, an AI model for locating the cloud center in the first stage of typhoon initial identification is constructed. The pre-processed multi-channel satellite image data is analyzed to extract the cloud features in the target area and output the cloud center position. The model includes: a feature extraction layer, which is used to extract the brightness temperature features and local convection intensity of satellite images; a spatial attention mechanism layer, which is used to enhance the perception of local features of clouds; and a position regression layer, which is used to predict the coordinates of the cloud center position based on the extracted features. SS3. Based on the identified and located cloud center, multi-order feature decomposition is performed on the multi-channel satellite image data within a preset radius with the cloud center as the center, and the brightness temperature value in the cloud area is decomposed into radially distributed multi-order fluctuation components to extract the symmetry and asymmetry characteristics of the cloud. SS4. Based on the SSD target detection algorithm, the typhoon inception identification AI model of the second stage of typhoon inception identification is constructed to analyze the extracted cloud symmetry and asymmetry features, including: feature input layer, used to receive cloud symmetry and asymmetry features; feature fusion layer, used to fuse the input symmetry and asymmetry features, and use the attention mechanism to enhance the perception of asymmetric significant features; SSD detection module, used to perform target detection on the fused multi-scale features, identify and locate potential typhoon inception areas; output layer, used to judge whether the target cloud belongs to the typhoon inception stage according to the classification results of the SSD detection module; SS5. Based on the typhoon incipient identification judgment result, if the target cloud cluster belongs to the typhoon incipient stage, the typhoon incipient stage identification result is output.

2. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1 is characterized in that: In step SS1, the infrared channel data is the brightness temperature data in the 11-micron band, which is used to capture the cloud top temperature and atmospheric convection information; 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 to 10 kilometers, and the time resolution is 1 to 6 hours.

3. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1 is characterized in that: In step SS1, data preprocessing includes data quality control, standardization, unified adjustment of spatiotemporal resolution, and conversion of brightness temperature data into a grayscale image format suitable for deep learning model input, wherein: the data quality control includes strip noise removal, cloud detection and outlier removal of satellite image data, wherein strip noise removal adopts wavelet transform method, cloud detection is based on brightness temperature difference threshold judgment of infrared and water vapor channels, and outlier removal adopts 3 σ The standardization process includes performing minimum-maximum normalization or Z-score normalization on the brightness temperature data; the unified adjustment of the spatiotemporal resolution includes unifying the temporal resolution of the multi-channel satellite image data to a 15-minute time interval using a linear interpolation method, and unifying the spatial resolution to a 1-kilometer spatial resolution using a bilinear interpolation method, and at the same time, performing regional cropping on the multi-channel satellite image data according to the meteorological characteristics of the target area, and retaining only the image data within the coverage area of ​​the target area.

4. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1 is characterized in that: In step SS2, the feature extraction layer of the cloud center positioning AI model uses depthwise separable convolution to reduce the number of model parameters and computational complexity, and normalizes the features output by each convolution layer through batch normalization; The spatial attention mechanism layer highlights the characteristics of the target cloud 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-by-element with the original feature map to enhance the perception of blurred or incomplete cloud features.

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

6. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1 is characterized in that: In step SS3, a multi-order eigendecomposition is performed using Fourier decomposition technology, including the following sub-steps: SS31. Based on the cloud center located in step SS2, select a target area with a radius of 1-5° and a cloud center as the center, and divide it into grids at a sampling interval of 0.1° to generate a brightness temperature distribution matrix for the target area. Each grid point in the matrix corresponds to the satellite observed brightness temperature value, and interpolate the data in the boundary area according to the cloud distribution characteristics to fill in possible gaps; 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 center as the base point, distribute the sampling points in the azimuth direction within the range of 0~360° according to the preset interval, and filter the strong noise part of the radial data; SS33. Perform Fourier decomposition on the radial brightness temperature distribution data along the azimuth direction and decompose it into several order wave components. The Fourier decomposition formula is as follows: in, T ( θ , r ) represents the radial distribution function of brightness temperature, r is the radial distance, θ is the azimuth; i is the order of Fourier decomposition and i =1,2,3,…, n , n is the highest order of 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 order i The amplitudes of the cosine and sine components of ; WN 0( r ) represents the 0th-order component of Fourier decomposition, Wl ( r ) indicates the i order component; SS34. The 0th-order component WN0 in the Fourier decomposition result is extracted as the symmetry feature of the cloud, and the 1st to 4th-order components WN1~WN4 are extracted as the asymmetry features, where WN1 reflects the eccentric shape of the cloud, and WN2~WN4 capture the local disturbance characteristics and spiral morphology characteristics of the cloud. The extracted symmetry and asymmetry features are reconstructed into a multi-channel feature map, with each channel corresponding to one component.

7. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1 is characterized in that: In step SS4, the feature fusion layer of the typhoon incipient recognition AI model uses multi-scale convolution operations to fuse the input symmetry and asymmetry features, and combines the FPN structure to generate multiple feature maps of different scales, including: When performing multi-scale convolution extraction on the symmetric and asymmetric features of the input, convolution kernels of different sizes are set to perform parallel convolution to extract local 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. A multi-scale FPN structure is used to further process the fused features. The perception ability of typhoon primary cloud targets of 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 map is used to capture the global pattern information of the cloud, and the higher resolution feature map is used to capture the local disturbance characteristics of the cloud. The top-down feature fusion fuses the low-resolution feature map with the high-resolution feature map layer by layer through upsampling operations, and reduces the dimension of the fused features through 1×1 convolution to make full use of multi-scale information.

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

9. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 7 is characterized in that: The SSD detection module of the typhoon incipient identification AI model includes the following steps when performing target detection: SS41. The multi-scale feature maps output by the feature fusion layer correspond to feature layers of different resolutions. Each feature layer generates a fixed-size feature map through convolution operation. The feature map of each scale is used to detect cloud targets of a specific size to fully cover the possible typhoon initiation area. SS42. On the feature map of each scale, multiple candidate detection boxes are generated by presetting anchor boxes. The size and aspect ratio of the anchor boxes are adjusted according to the resolution of the feature maps of different scales. The size of the anchor boxes ranges 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 typhoon's primary clouds. Each anchor box is associated with each pixel 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 area are completed by the classification subnetwork and the regression subnetwork respectively, wherein: the classification subnetwork is used to classify the candidate area in each feature map to determine whether it belongs to the cloud cluster in the initial stage of the typhoon. The classification output is calculated by the Softmax function to determine the probability of each candidate area belonging to the initial stage of the typhoon or the non-initial stage of the typhoon, and the classification loss function adopts the cross entropy loss; the regression subnetwork is used to perform regression correction on the bounding box of the candidate area, predict its precise position and scale, and use the smooth L1 loss to optimize the regression error; SS44. Post-process the results of classification and regression prediction, and use the non-maximum suppression method to remove redundant detection frames, including: sorting candidate detection frames according to the confidence output by the classification subnetwork; for detection frames whose confidence exceeds the preset threshold, calculate the intersection over union (IoU) ratio between them and other detection frames, and delete the detection frame if the IoU exceeds the set threshold; only retain the detection results with the highest confidence, including the detection frame's position, scale, and confidence score at the initial stage of the typhoon; SS45. The final typhoon incipient recognition result is generated by combining the detection results of feature maps of different scales through weighted fusion. The weight distribution is dynamically adjusted according to the detection confidence of feature maps of different scales. If the detection confidence of the low-resolution feature map is higher, the detection result of the low-resolution feature map is retained first; if the detection confidence of the high-resolution feature map is higher, the detection result of the high-resolution feature map is retained first; the final output results include the central position of the typhoon incipient cloud cluster, the scale of the detection frame, and the confidence score of the typhoon incipient stage.

10. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 9 is 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 fusion features to extract spatial features; gradually reducing the spatial resolution of the feature map through a downsampling operation to generate feature maps of multiple scales.

11. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1 is characterized in that: In step SS4, the typhoon incipient recognition AI model is trained based on a multi-channel feature image dataset containing labeled typhoon incipient clouds and non-typhoon incipient clouds and corresponding asymmetry features. The dataset is divided into 70% for training, 20% for verification, 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 target 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 to continuously train the model.

12. The typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics according to claim 1 is characterized in that: In step SS5, the output of the typhoon initial stage recognition result includes at least the following information: the center position coordinates of the target cloud group, recognition confidence, the judgment result of the typhoon initial stage and its credibility score, wherein the recognition confidence is calculated by 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 target detection result.

13. A computer program product comprising computer instructions, characterized in that: The computer instructions are used to execute the typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics as described in any one of claims 1 to 12.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the typhoon incipient identification method based on the dual AI model architecture and typhoon asymmetry characteristics as described in any one of claims 1 to 12.

Images