Landing typhoon residual vortex detection method based on YOLOv5

Abstract

The invention discloses a landing typhoon residual vortex detection method based on YOLOv5, and the method comprises the steps: employing ERA5 data and landing typhoon optimal path data, screening a convolutional neural network YOLOv5 model, carrying out the screening of initial data based on a potential height field, taking the screened data as a model sample set, building a typhoon residual vortex detection model, and achieving the detection of weak vortexes. In the life stage after the landing typhoon is weakened into tropical low-voltage shutdown, the existence and development conditions of the landing typhoon residual vortex are accurately detected and monitored, accurate reference is provided for meteorological early warning and disaster management, and the accuracy rate of a test set is as high as 70.08%.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological technology, and in particular relates to a YOLOv5-based method for detecting residual vortices of a landing typhoon. Background Art

[0002] Since the typhoon best path data is no longer released after the typhoon is discontinued, there is currently no typhoon residual vortex database available for research. This creates two difficulties: first, due to the lack of a uniform, long-term typhoon residual vortex database, it is impossible to carry out climate statistical research on such weather systems; second, when developing AI-based typhoon residual vortex detection technology, there is no data set available for model training. At present, there is no automatic detection model specifically for such weak vortex systems. The following will give a detailed introduction to the existing work on the detection of landfalling typhoon residual vortices.

[0003] (1) Subjective judgment. Its disadvantages are: it is highly subjective, cannot be used for batch calculations, and cannot obtain uniform, long-time series, and large-sample data on the residual vortex of a landfalling typhoon.

[0004] (2) Objective algorithms, for example: Meng Zhaozhen et al. (Reference: Meng Zhaozhen, Han Shenyou, Lai Zhenquan, et al. Analysis of the characteristics of heavy rain in typhoon remnant vortex in Guangxi [J]. Meteorological Research and Application, 2017, 38(1): 20-25.) conducted a statistical analysis on the remnant vortex of typhoons that affected Guangxi and proposed an objective algorithm for detecting the remnant vortex of typhoons that landed. This method ignores the characteristic changes of other physical elements of the remnant vortex of typhoons that landed. At the same time, the algorithm provides a criterion for determining the starting time of the remnant vortex of typhoons that landed, but the accuracy and stability of the algorithm are poor when detecting after the remnant vortex intensity weakens, and when detecting fuzzy vortices or vortices under complex weather conditions.

[0005] (3) Artificial intelligence-based vortex detection algorithm. Currently, there is no artificial intelligence detection algorithm specifically for the remnant vortex of a landfalling typhoon. There is a similar one, such as an intelligent typhoon vortex detection technology based on deep image target detection proposed by Lv Xinyan et al. (Reference: Lv Xinyan, Qian Qifeng, Wang Dengke, et al. Intelligent typhoon vortex recognition technology based on deep image target detection [J]. Journal of Tropical Meteorology, 2022, 38(4): 492-501.). This algorithm is not specifically proposed for the remnant vortex of a landfalling typhoon, and the detection effect of weak vortices is low. However, due to the characteristics of weak intensity and asymmetry, the remnant vortex of a landfalling typhoon usually does not have significant statistical characteristics in satellite images. Therefore, this method is not suitable for the detection of the remnant vortex of a landfalling typhoon.

[0006] There is currently no detection model specifically proposed for the remnant vortex of a landing typhoon in the existing technology, and there are still great technical challenges in the accurate detection and positioning of the remnant vortex of a landing typhoon. Summary of the invention

[0007] Purpose of the invention: In view of a series of problems and limitations in the existing technology of the residual vortex detection algorithm of the landing typhoon, the present invention provides a residual vortex detection method of the landing typhoon based on YOLOv5.

[0008] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution: A method for detecting residual vortex of a landfalling typhoon based on YOLOv5, comprising the following steps: S1, data acquisition: obtain the hourly ERA5 reanalysis data of the landfalling typhoon with a spatial resolution of 0.25°, and select the data under the potential height field of 850hPa in the lower atmosphere; S2, in the corresponding data under the potential height field of 850 hPa, the data of step S1 are screened according to the following two conditions that are met at the same time: (1) a closed vortex center appears in the 850 hPa flow field; (2) the 850 hPa potential height field is less than or equal to 1500 m, and the 850 hPa relative vorticity field is greater than or equal to ; S3, extracting the spatiotemporal potential height field values ​​of the data selected in step S2, and drawing a grayscale map of the potential height field using the ERA5 reanalysis data set as a model sample set; S4, based on YOLOv5 as the basic framework, builds a typhoon residual vortex detection model, and uses at least three convolutional layers and one pooling layer to extract typhoon residual vortex features. After the convolutional layer, a batch normalization layer and activation function are inserted to accelerate training and reduce overfitting; S5, training and optimizing the typhoon residual vortex detection model using the model sample set of step S3 to obtain an optimized typhoon residual vortex detection model TCRVNet.

[0009] Furthermore, in step S3, the ratio of training samples to test samples in the model sample set is 7:3.

[0010] Furthermore, the data in step S1 include the following physical quantities: time T, vortex center latitude LAT, vortex center longitude LON, vortex center potential height GHT, and vortex center vorticity VOR.

[0011] Furthermore, before the typhoon residual vortex detection model described in step S5 is trained and optimized, AchorBoxers clustering is performed on the model sample set through the K-means clustering algorithm to realize the use of anchor boxes of different sizes as the bounding boxes of the landing typhoon residual vortex, thereby completing the clustering of the landing typhoon residual vortex.

[0012] Furthermore, when the typhoon residual vortex detection model described in step S5 is trained, a binary cross entropy loss function is used to optimize the detection of the presence or absence of the landed typhoon residual vortex, and a smooth L1 loss function is used to optimize the residual vortex position and area information.

[0013] Furthermore, in step S5, the training process of the typhoon residual vortex detection model adjusts the weights and biases in the network through a directional propagation algorithm.

[0014] Furthermore, the training iteration number of the typhoon residual vortex detection model in step S5 is 500, the batch size is 16, and the learning rate is 1e -2 , the activation function uses LeakyReLU.

[0015] Beneficial effects: Compared with the prior art, the automatic detection system for residual vortices of a landfalling typhoon of the present invention solves the technical gap in the field of residual vortices of a landfalling typhoon by using deep learning technology and convolutional neural networks, and brings significant beneficial effects, especially in improving prediction accuracy, efficiency and automation level. It has outstanding advantages, including:

[0016] (1) The deep learning technology based on the YOLOv5 model in the present invention realizes the automatic detection of the residual vortex of the landing typhoon for the first time, significantly improving the detection accuracy and processing speed, providing a new technical means for the monitoring of the landing typhoon disaster, and can effectively fill the gap in the existing technology.

[0017] (2) Based on the understanding of the physical characteristics of the residual vortex of a landfalling typhoon, the present invention can enhance the multi-scale target detection capability through the YOLOv5 deep learning model, and realize the automatic extraction of the key features of the residual vortex of a landfalling typhoon with an accuracy rate of up to 70.08%. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The schematic diagram of the TCRVNet model architecture of the typhoon residual vortex detection model of the present invention includes four general modules: Input, Backbone, Neck and Head; Figure 2 It is a graph showing changes in three loss functions, namely, positioning loss (box_loss), confidence loss (obj_loss) and classification loss (cls_loss), with the number of iterations during the training process of the typhoon residual vortex detection model TCRVNet described in the present invention; Figure 3 1 is a comparison diagram of the true position of the residual vortex and the detection result of the residual vortex detection model TCRVNet of the present invention, wherein: (a) is a schematic diagram of the true position of the residual vortex; (b) is a schematic diagram of the detection result of the typhoon residual vortex detection model TCRVNet; Figure 4 This is an application flow chart of the typhoon residual vortex detection model of the present invention; Figure 5 This is a color image of the detection effect of an embodiment of the present invention, wherein the red frame is the detected residual vortex, and the white frame is the detected landfalling typhoon. DETAILED DESCRIPTION

[0019] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0020] The present invention aims to provide an artificial intelligence detection method specifically for a weak vortex system such as a landing typhoon remnant vortex, and uses the fifth-generation reanalysis data set ERA5 of the European Centre for Medium-Range Weather Forecasts to detect the landing typhoon remnant vortex, filling the current gap in the detection of weak vortexes of landing typhoons.

[0021] The present invention uses ERA5 data and the best path data of the landing typhoon, and trains the convolutional neural network YOLOv5 model through deep learning technology, and finally realizes an estimation system (named TCRVNet) that automatically detects the position and intensity of the residual vortex of the landing typhoon based on the input meteorological data. The system can realize the detection of weak vortices, and accurately detect and monitor the existence and development of the residual vortex of the landing typhoon in the life stage after the landing typhoon weakens to a tropical depression and is discontinued, providing an accurate reference for meteorological warning and disaster management. In addition, based on the long-term landing typhoon path data and high-resolution ERA5 reanalysis meteorological data, the present invention has constructed a climatologically significant data set of residual vortices of landing typhoons since 1976, in order to support more in-depth meteorological research and prediction of extreme weather events in the future. Please refer to Figure 4 As shown, the specific steps are as follows: 1. Data Preprocessing

[0022] (1) Construction of typhoon remnant vortex dataset

[0023] Since there is currently no historical data set of typhoon residual vortex available for AI model training, the present invention first constructs a set of 48-year typhoon residual vortex data sets from 1976 to 2023 for pre-training of the YOLOv5 model. This data set is based on the understanding of the physical characteristics of the residual vortex of the landfalling typhoon. By analyzing the ERA5 reanalysis data with hourly spatial resolution of 0.25°, the potential height field of 850hPa in the lower atmosphere is combined with the characteristics of the vorticity field and flow field to determine the center position of the residual vortex, and further calculate or extract other physical quantities near the center of the residual vortex.

[0024] Since the center height of the residual vortex is generally located near 850 hPa, the preliminary criteria and basis for determining the center of the residual vortex in the present invention are as follows: (1) 850 hPa flow field: a closed vortex center must appear. If the closed center disappears, the life history of the residual vortex is determined to be over. On the premise of satisfying condition (1), (2) 850 hPa potential height field must also be satisfied: less than or equal to 1500 m; (3) 850 hPa relative vorticity field: Referring to the format of the typhoon optimal path data, the data set constructed by the present invention includes the following basic physical quantities, as shown in Table 1 below:

[0025] Table 1

[0026] Since this dataset is constructed based on ERA5 data, the types of physical quantities can be expanded according to research needs in the future.

[0027] (2) Input data processing

[0028] Based on the physical understanding of the environmental conditions for the formation and disappearance of landfalling typhoons and the background of climate change, the 850 hpa potential height field from 1998 to 2023 was selected from the landfalling typhoon remnant vortex dataset as input for model training. The main processing steps are as follows: Samples were screened based on the best path dataset of landfalling typhoons, and only the landfalling typhoon remnants that had landed in South China during their lifetime and were still south of 30°N when they finally disappeared were screened, and they were defined as the landfalling typhoon remnants that affected South my country. The 850pha potential height field at 100°E to 125°E east longitude and 10°N to 30°N north latitude (South China) at the time when the landfalling typhoon remnant existed was standardized and drawn into a grayscale map for detection processing.

[0029] (3) Label data processing

[0030] The present invention selects typhoon remnant vortex samples from 1998 to 2023 from the typhoon remnant vortex dataset mentioned above, draws a grayscale map of the potential height field based on the ERA5 reanalysis dataset, and processes it into a label file suitable for YOLOv5.

[0031] 2. Deep Learning Model Construction and Training

[0032] (1) Basic architecture of the model: The present invention uses YOLOv5 as the basic architecture to construct a landfall typhoon residual vortex detection model. YOLOv5 is an efficient convolutional neural network (CNN) architecture ( Figure 1 ), which improves the feature pyramid structure and enhances the detection capability of multi-scale objects.

[0033] This is because the scale of the typhoon itself varies greatly. During its landing attenuation process, it will undergo dynamic evolution due to the influence of terrain and environmental fields, and its scale will also change greatly, with multi-scale situations. Therefore, the YOLOv5 used in the present invention is more suitable for target detection tasks with large scale differences such as the residual vortex of a landing typhoon. The model architecture of the present invention includes 3 convolutional layers, 1 batch normalization layer and an activation layer, which aims to extract deep features from the input image. It uses standard convolution operations and AnchorBoxes mechanisms to achieve target detection, and can provide fast inference speed while maintaining a high accuracy rate.

[0034] (2) Typhoon residual vortex feature extraction: The feature extraction of typhoon residual vortex that the present invention focuses on is mainly carried out through three convolution layers and pooling layers. In view of the large difference in the spatial scale of typhoon residual vortex, the present invention has been tested and selected a 3x3 convolution kernel to gradually extract the detailed features of the typhoon residual vortex in the image, which helps to accurately detect the shape, position and intensity of the vortex. After the convolution layer, a batch normalization layer and an activation function (such as LeakyReLU) are connected to accelerate training and reduce overfitting, while improving the nonlinear representation ability of the model.

[0035] (3) Application of attention mechanism: Since the shape, intensity and position of the residual vortex of the landing typhoon are constantly evolving, the TCRVNet model of the present invention adopts the AnchorBoxes attention mechanism to have more advantages in identifying such targets, and before model training, the size of AnchorBoxes is automatically calculated by the K-means clustering algorithm. It uses anchor boxes of different sizes to predict the bounding box of the residual vortex of the landing typhoon, and can adaptively adjust the size of the anchor box according to the actual size of the residual vortex of the typhoon. The prediction result of each anchor box includes the target category (whether there is a residual vortex of the landing typhoon) and the coordinates of the bounding box. In the detection of the residual vortex of the landing typhoon, by analyzing the shape and position changes of the residual vortex, the model can quickly and accurately output the vortex center and its surrounding area. The final output of the model includes the existence probability, position coordinates and confidence interval of the residual vortex of the landing typhoon.

[0036] (4) Model training: After debugging the control variables, the best hyperparameter combination was obtained on the test set. The specific settings are shown in Table 2. The change of loss function with the number of training iterations is shown in Figure 2As shown. In the model training process of the Landing Typhoon Remnant Vortex Detection Model (TCRVNet), firstly, the binary cross entropy loss function is used to optimize the detection of the presence or absence of the landing typhoon remnant vortex; secondly, the smooth L1 loss function is used to optimize the location and regional information of the remnant vortex (including the central longitude and latitude of the landing typhoon remnant vortex, as well as the north-south and east-west widths), which can prevent some outliers from having a significant impact on the training results. The training process adjusts the weights and biases in the network through the back-propagation algorithm, so that the output prediction is as close to the true label as possible, thereby achieving accurate estimation of the location and morphology of the remnant vortex.

[0037] Table 2

[0038] Through this method, the present invention can automatically process a large amount of reanalysis data and realize fast and accurate detection of residual vortices of landing typhoons. After the training is completed, the model can receive new landed typhoon geopotential height maps or other meteorological data as input, automatically detect the position and evolution trend of the residual vortices of the landing typhoon, and output the presence or absence, position coordinates and confidence of the residual vortices of the landing typhoon. The trained model can accurately detect the existence and development of the residual vortices of the landing typhoon by learning a large amount of landed typhoon data, and can effectively distinguish between the landed typhoon and the residual vortex when they exist at the same time, providing data support for the meteorological early warning system.

[0039] like Figure 3 As shown, the present invention is based on the understanding of the physical characteristics of the remnant vortex of a landing typhoon, and automatically extracts the key features of the remnant vortex of a landing typhoon from the reanalysis data through deep learning technology. Compared with the traditional manual analysis method and satellite cloud image detection method, especially in complex weather environments, complex shapes and dynamic changes of the remnant vortex of a landing typhoon, the use of the algorithm of the present invention can more accurately detect the existence and characteristics of the remnant vortex of a landing typhoon, reduce the probability of false detection and missed detection, and maintain a high detection accuracy. The accuracy of the present invention on the test set is as high as 70.08%. Figure 3 Part (b) shows that the recognition position is accurate and the reliability can reach 0.9. The detailed comparison results are shown in Table 3.

[0040] Table 3

[0041] like Figure 5 As shown in the figure, the residual vortex of the 2023 typhoon "Sura" and the 2023 typhoon "Hai Kui" are selected for detection and application at the same time. In order to better show the detection effect, the grayscale image is drawn as a color image. The red frame in the figure is the detected residual vortex, and the white frame is the detected typhoon. It can be seen that the system of the present invention can successfully detect the residual vortex of the typhoon, and can also distinguish the typhoon and the residual vortex as two types.

[0042] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for detecting residual vortex of a landfalling typhoon based on YOLOv5, characterized in that: The following steps are involved: S1, data acquisition: obtain the hourly ERA5 reanalysis data of the landfalling typhoon with a spatial resolution of 0.25°, and select the data under the potential height field of 850hPa in the lower atmosphere; S2, in the corresponding data under the potential height field of 850 hPa, the data of step S1 are screened according to the following two conditions that are met at the same time: (1) a closed vortex center appears in the 850 hPa flow field; (2) the 850 hPa potential height field is less than or equal to 1500 m, and the 850 hPa relative vorticity field is greater than or equal to ; S3, extracting the spatiotemporal potential height field values ​​of the data selected in step S2, and drawing a grayscale map of the potential height field using the ERA5 reanalysis data set as a model sample set; S4, based on YOLOv5 as the basic framework, builds a typhoon residual vortex detection model, and uses at least three convolutional layers and one pooling layer to extract typhoon residual vortex features. After the convolutional layer, a batch normalization layer and activation function are inserted to accelerate training and reduce overfitting; S5, training and optimizing the typhoon residual vortex detection model using the model sample set of step S3 to obtain an optimized typhoon residual vortex detection model TCRVNet.

2. According to the YOLOv5-based landing typhoon residual vortex detection method of claim 1, it is characterized in that: In step S3, the ratio of training samples to test samples in the model sample set is 7:

3.

3. The method for detecting residual vortex of a landfalling typhoon based on YOLOv5 according to claim 1, characterized in that: The data in step S1 include the following physical quantities: time T, vortex center latitude LAT, vortex center longitude LON, vortex center potential height GHT, vortex center vorticity VOR.

4. The method for detecting residual vortex of a landfalling typhoon based on YOLOv5 according to claim 1, characterized in that: Before the typhoon residual vortex detection model described in step S5 is trained and optimized, AchorBoxers clustering is performed on the model sample set through the K-means clustering algorithm to realize the use of anchor boxes of different sizes as the bounding boxes of the landing typhoon residual vortex, thereby completing the clustering of the landing typhoon residual vortex.

5. The method for detecting residual vortex of a landfalling typhoon based on YOLOv5 according to claim 4, characterized in that: When the typhoon residual vortex detection model described in step S5 is trained, a binary cross entropy loss function is used to optimize the detection of the presence or absence of the landed typhoon residual vortex, and a smooth L1 loss function is used to optimize the residual vortex position and area information.

6. According to the YOLOv5-based landing typhoon residual vortex detection method of claim 4, it is characterized in that: The training process of the typhoon residual vortex detection model described in step S5 adjusts the weights and biases in the network through the directional propagation algorithm.

7. The method for detecting residual vortex of a landfalling typhoon based on YOLOv5 according to claim 4, characterized in that: The training iteration number of the typhoon residual vortex detection model in step S5 is 500, the batch size is 16, and the learning rate is 1e -2 , the activation function uses LeakyReLU.

Images