Uprush cloud top detection method and system based on YOLOv8 algorithm

Abstract

The invention discloses an uprush cloud top detection method and system based on a YOLOv8 algorithm, and belongs to the technical field of target recognition. The method comprises the following steps: constructing a multi-channel satellite data set and generating a three-channel RGB image through a pseudo-color synthesis technology; performing truth value labeling by combining radar vertical profile data to form a high-quality data set; a convolution attention module is embedded in the YOLOv8 model, training is carried out, and small target detection performance is optimized; and preprocessing a real-time satellite image, inputting the preprocessed real-time satellite image into the model, outputting a detection result and linking an early warning system. The method gets rid of dependence on manual setting of a fixed physical threshold value, is high in generalization capability and robustness, is completely automatic in detection process, can reduce the burden of a predictor, improves the severe convection weather monitoring and early warning efficiency, and is especially suitable for precise recognition of small-scale and untypical-form uprush clouds.

Description

Technical Field

[0001] This invention relates to the field of target tracking and recognition technology, specifically to a method and system for detecting upward cloud tops based on the YOLOv8 algorithm. Background Technology

[0002] Overshooting Top (OT) is the peak of severe convective weather development, a violent convective phenomenon that penetrates the tropopause and enters the lower stratosphere. It is usually accompanied by severe weather such as strong winds, hail, and heavy rain. Accurate and rapid identification is crucial for severe weather warnings.

[0003] Currently, the detection of storm surges to the top of Genting Highlands mainly relies on meteorological satellite data, and the mainstream detection methods include: 1. Infrared Brightness Temperature Threshold Method: This method utilizes the brightness temperature difference (BTD) characteristics of the cloud top in the water vapor channel and the infrared window area. By setting a BTD threshold, the cloud top pixels are identified. This is one of the classic detection methods. 2. Texture feature analysis method: By calculating the brightness temperature gradient, variance and other texture parameters of the local area, combined with the average brightness temperature difference with the surrounding anvil clouds, the judgment is made based on the unique "V" or "U" shaped cold / warm zone texture structure of the cloud top. 3. Local Minimum Method: First, find the point with the lowest local brightness and temperature in the infrared window channel image as a potential candidate area, and then screen and confirm it by physical criteria such as the difference between the water vapor channel and the infrared window channel.

[0004] However, existing technologies have obvious drawbacks: on the one hand, traditional methods rely on fixed physical thresholds, which are highly regional and subjective. Differences in tropopause height and temperature and humidity profiles in different regions can lead to a large number of false alarms or missed alarms. Furthermore, the selection of thresholds depends on the forecaster's experience and lacks objective standards. On the other hand, the texture description of traditional methods is relatively simple and it is difficult to fully depict the complex spatial morphological features of upturned cloud tops. It also has insufficient ability to identify early upturned cloud tops that are atypical in shape, have blurred boundaries, or are small in size.

[0005] Therefore, there is an urgent need for a high-precision, robust, and highly automated top-mounted top detection technology to address the shortcomings of existing technologies. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a method and system for detecting upward cloud tops based on the YOLOv8 algorithm. The specific technical solution is as follows: A method for detecting upward cloud tops based on the YOLOv8 algorithm includes the following steps: S1. Construction and Fusion Preprocessing of Multi-channel Satellite Dataset: Acquire multi-channel data from high spatiotemporal resolution geostationary meteorological satellites for N consecutive years, and fuse the single-channel grayscale images into a three-channel pseudo-color RGB image using pseudo-color image synthesis technology.

[0007] S2. Truth Labeling and Dataset Augmentation: Using radar observation data, the cloud top areas confirmed by the observation data are accurately labeled with rectangular bounding boxes on multi-channel synthetic images to form a labeled dataset.

[0008] S3. Optimization and training of YOLOv8 model for small object detection: Based on the YOLOv8 model, a convolutional attention module is embedded in its backbone network or feature pyramid network structure to obtain an improved YOLOv8 model. The labeled dataset constructed in step S2 is input into the improved YOLOv8 model, and an adaptive momentum optimization algorithm is used to perform end-to-end training by minimizing the composite loss function that includes bounding box regression and classification confidence.

[0009] S4. Real-time monitoring and operational output: After processing the real-time geostationary meteorological satellite imagery through the preprocessing procedure in step S1, input it into the trained and improved YOLOv8 model, and output the bounding box position and category confidence of the cloud top.

[0010] As a further improvement of the present invention, the multi-channel data in step S1 includes data from the visible light channel, the water vapor channel, and the infrared window area channel.

[0011] As a further improvement of the present invention, the radar observation data in step S2 is the vertical profile data of CloudSat cloud profile radar and CALIPSO radar.

[0012] As a further improvement of the present invention, the convolutional attention module in step S3 is used to guide the model to adaptively learn the importance distribution of feature maps in the channel and spatial dimensions, enhance key cloud features and suppress background noise.

[0013] As a further improvement of the present invention, step S4 further includes visualizing the identification results and linking them with the business early warning system.

[0014] On the other hand, the present invention also provides an upward surge detection system based on the YOLOv8 algorithm, comprising: Data preprocessing module: Used for multi-channel data from high spatiotemporal resolution geostationary meteorological satellites over N consecutive years, the module fuses the single-channel grayscale images into a three-channel pseudo-color RGB image using pseudo-color image synthesis technology; Dataset construction module: Used to accurately annotate the cloud top areas confirmed by the radar observation data with rectangular bounding boxes on multi-channel synthetic images, forming an annotated dataset; Model optimization and training module: This module is used to obtain an improved YOLOv8 model by embedding a convolutional attention module into the backbone network or feature pyramid network structure of the YOLOv8 model as the basic architecture. The labeled dataset constructed by the dataset construction module is input into the improved YOLOv8 model, and the adaptive momentum optimization algorithm is used to perform end-to-end training by minimizing the composite loss function that includes bounding box regression and classification confidence. Real-time monitoring and operational output module: This module processes real-time geostationary meteorological satellite images through the preprocessing process of the data preprocessing module, inputs them into the trained and improved YOLOv8 model, and outputs the bounding box position and category confidence of the cloud top.

[0015] The beneficial effects of this invention are as follows: This invention utilizes an artificial intelligence model to automatically learn the deep, multi-dimensional abstract features of cloud tops from data, eliminating the reliance on manually set, geographically specific physical thresholds. The model has stronger robustness and generalization ability, and can adapt to the differences in meteorological environments in different regions. Once the model is trained, the entire detection process is fully automated without human intervention. It can effectively utilize massive amounts of satellite data for timely severe weather warnings, greatly reducing the burden of manual interpretation for forecasters and improving the efficiency of large-scale meteorological monitoring. By embedding a convolutional attention module into the YOLOv8 model, the feature capture capability for small-scale, atypical cloud tops is enhanced, improving detection accuracy and positioning accuracy, and solving the problem of insufficient early cloud top identification by traditional methods. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, the upward cloud top detection method based on the YOLOv8 algorithm of the present invention includes the following steps: S1. Construction and fusion preprocessing of multi-channel satellite datasets; S2, Truth Labeling and Dataset Augmentation; S3. Optimization and training of YOLOv8 model for small object detection; S4. Real-time monitoring and business-oriented output.

[0019] I. Construction and Fusion Preprocessing of Multi-channel Satellite Datasets To acquire multi-channel data from geostationary meteorological satellites (such as the Chinese Fengyun series, the US GOEs series, the European Meteosat series, and the Japanese Himawari 8 / 9 satellites) with high spatiotemporal resolution for N consecutive years (e.g., 2015-2024, where N is a positive integer greater than or equal to 5), and to comprehensively utilize the morphology, texture, and physical characteristics of cloud tops, this study uses the Himawari 8 / 9 satellite data as an example, selecting data from the visible light channel (channel 3), the water vapor channel (channel 8), and the infrared window channel (channel 13). Using pseudo-color image synthesis technology, the aforementioned single-channel grayscale images with different physical meanings are fused into a three-channel pseudo-color RGB image, thereby encoding key physical information (such as cloud top height, temperature, and texture) into visual features that can be directly processed by deep learning models.

[0020] II. Truth Labeling and Dataset Augmentation To construct a highly reliable labeled dataset, vertical profile data from CloudSat cloud profiling radar and CALIPSO radar were used. This data can accurately confirm the actual location and extent of the cloud tops. On multi-channel synthetic imagery, the cloud top regions confirmed by the vertical observation data were precisely labeled with rectangular bounding boxes, forming a high-quality labeled dataset that provides reliable supervision information for model training.

[0021] III. Optimization and Training of YOLOv8 Model for Small Object Detection The YOLOv8 model was chosen as the basic architecture due to its advantages of fast detection speed and high accuracy. Targeted optimizations were made to address the characteristic that uplifting cloud tops typically appear as small-scale targets within the vast convective cloud systems they belong to. Specifically, a convolutional attention module was embedded into the model's backbone network or feature pyramid network structure. This attention mechanism guides the model to adaptively learn the importance distribution of feature maps in both channel and spatial dimensions, thereby enhancing the focus on key cloud top features and effectively suppressing broad background noise. Ultimately, this significantly improves the model's detection performance and localization accuracy for small-scale uplifting cloud tops.

[0022] The constructed labeled dataset is input into the improved YOLOv8 model for end-to-end training. During training, an adaptive momentum optimization algorithm is used, which can adaptively adjust the learning rate and improve training stability. By minimizing the composite loss function that includes bounding box regression and classification confidence, the model parameters are driven to converge quickly to the optimal state.

[0023] IV. Real-time monitoring and operational output The latest real-time geostationary meteorological satellite imagery, after undergoing the multi-channel fusion preprocessing described above, is input into the trained improved YOLOv8 model. The model automatically outputs the bounding box location and category confidence score of the cloud tops, achieving accurate identification and positioning of cloud tops. The identification results can be overlaid on geographic base maps or satellite imagery in real time for visualization, facilitating intuitive viewing for forecasters. Simultaneously, it can be linked to operational early warning systems to provide timely decision support for the monitoring and forecasting of severe convective weather.

[0024] Embodiments of the present invention also provide an upswing cloud top detection system based on the YOLOv8 algorithm, comprising: The system includes a data preprocessing module, a dataset construction module, a model optimization and training module, and a real-time monitoring and business-oriented output module. The functions implemented by these modules correspond to the method steps of this invention.

[0025] In addition, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method of the present invention.

[0026] Embodiments of the present invention also provide an electronic device, comprising: One or more processors; and A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the method of the present invention.

[0027] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention. For those skilled in the art, several simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered within the protection scope of the present invention.

Claims

1. A method for detecting upward cloud tops based on the YOLOv8 algorithm, characterized in that, The method includes the following steps: S1. Construction and Fusion Preprocessing of Multi-channel Satellite Dataset: Acquire multi-channel data from high spatiotemporal resolution geostationary meteorological satellites for N consecutive years, and fuse the single-channel grayscale images into a three-channel pseudo-color RGB image using pseudo-color image synthesis technology. S2. Truth Labeling and Dataset Augmentation: Using radar observation data, the cloud top areas confirmed by the observation data are accurately labeled with rectangular bounding boxes on multi-channel synthetic images to form a labeled dataset. S3. Optimization and training of YOLOv8 model for small object detection: Based on the YOLOv8 model, a convolutional attention module is embedded in its backbone network or feature pyramid network structure to obtain an improved YOLOv8 model. The labeled dataset constructed in step S2 is input into the improved YOLOv8 model. The adaptive momentum optimization algorithm is used to perform end-to-end training by minimizing the composite loss function that includes bounding box regression and classification confidence. S4. Real-time monitoring and operational output: After processing the real-time geostationary meteorological satellite imagery through the preprocessing procedure in step S1, input it into the trained and improved YOLOv8 model, and output the bounding box position and category confidence of the cloud top.

2. The method for detecting upward cloud tops based on the YOLOv8 algorithm according to claim 1, characterized in that, The multi-channel data in step S1 includes data from the visible light channel, the water vapor channel, and the infrared window area channel.

3. The method for detecting upward cloud tops based on the YOLOv8 algorithm according to claim 1, characterized in that, The radar observation data in step S2 are vertical profile data from CloudSat cloud profile radar and CALIPSO radar.

4. The method for detecting upward cloud tops based on the YOLOv8 algorithm according to claim 1, characterized in that, The convolutional attention module in step S3 is used to guide the model to adaptively learn the importance distribution of feature maps in the channel and spatial dimensions, enhance key cloud features, and suppress background noise.

5. The method for detecting upward cloud tops based on the YOLOv8 algorithm according to claim 1, characterized in that, Step S4 also includes visualizing the identification results and linking them to the business early warning system.

6. A system for detecting upward surges to cloud tops based on the YOLOv8 algorithm, characterized in that, The system includes: Data preprocessing module: Used for multi-channel data from high spatiotemporal resolution geostationary meteorological satellites over N consecutive years, the module fuses the single-channel grayscale images into a three-channel pseudo-color RGB image using pseudo-color image synthesis technology; Dataset construction module: Used to accurately label the Shangchong cloud top area confirmed by vertical measurement data with rectangular bounding boxes on multi-channel synthetic images using radar observation data, forming a labeled dataset; Model optimization and training module: This module is used to obtain an improved YOLOv8 model by embedding a convolutional attention module into the backbone network or feature pyramid network structure of the YOLOv8 model as the basic architecture. The labeled dataset constructed by the dataset construction module is input into the improved YOLOv8 model, and the adaptive momentum optimization algorithm is used to perform end-to-end training by minimizing the composite loss function that includes bounding box regression and classification confidence. Real-time monitoring and operational output module: This module processes real-time geostationary meteorological satellite images through the preprocessing process of the data preprocessing module, inputs them into the trained and improved YOLOv8 model, and outputs the bounding box position and category confidence of the cloud top.

7. The cloud top detection system based on the YOLOv8 algorithm according to claim 6, characterized in that, The multi-channel data in the data preprocessing module includes data from the visible light channel, the water vapor channel, and the infrared window area channel.

8. The cloud top detection system based on the YOLOv8 algorithm according to claim 6, characterized in that, The radar observation data are vertical profile data from CloudSat cloud profiling radar and CALIPSO radar.

9. The cloud top detection system based on the YOLOv8 algorithm according to claim 6, characterized in that, The convolutional attention module is used to guide the model to adaptively learn the importance distribution of feature maps in the channel and spatial dimensions, enhance key cloud features, and suppress background noise.

10. The cloud top detection system based on the YOLOv8 algorithm according to claim 6, characterized in that, The real-time monitoring and business output module is also used to visualize the identification results and link them with the business early warning system.

Images