Research Of Tropical Cyclone Intensity Estimation Based On Multi-View Feature Fusion

Tropical cyclones(TCs)are one of the most destructive weather phenomena,causing significant damage to human life and property.Previous studies have shown a close relationship between the losses caused by TCs and their intensity.Therefore,accurately estimating the intensity of TCs is crucial for weather forecasting and disaster warning.The current deep learning methods primarily rely on infrared satellite imagery and fail to fully exploit the overall structural information of TCs.In contrast,microwave channel data can provide vertical structure features of TCs,thereby offering more information for deep learning modeling.However,the current deep learning methods based on microwave channel data utilize brightness temperature data from the Advanced Technology Microwave Sounder(ATMS)to construct the tropical cyclone(TC)Warmcore Dataset for intensity estimation,which still lacks comprehensive research in terms of dataset quality and network models.Based on this,this thesis further studies and construct higher-quality datasets through data optimization processing,and proposes a Multi View Feature Fusion(MVFF)network suitable for TC Warm-core Dataset.The specific research content is as follows:(1)The revised label from the International Best Track Archive for Climate Stewardship(IBTr ACS)is used for data annotation to improve data quality.During the construction of the TC Warm-core Dataset,reasonable data selection is achieved by removing data affected by blind spots in ATMS scanning,further optimizing the data.(2)Considering the characteristics of warm core data having cross-sectional information in different directions(the along-track,cross-track and pressure-level),a multi-view learning architecture of one-view-one-network is used to extract TC structural information from different directions,that is,parallel backbone network branches are employed to extract features from different directions of the warm core data,and the output of the fully connected layers are concatenated and fused to improve the accuracy of the model.(3)The Single Shot Detector(SSD)is employed as the backbone network,and a Feature Effective Fusion(FEF)module is introduced to fuse features from different layers.The FEF module utilizes the Adaptive Spatial Feature Fusion(ASFF)module and Effective Channel Attention(ECA)module to fuse the three relatively large feature maps extracted from the convolutional layer,solving the problem of lack of connectivity between the upper and lower layers of the SSD backbone network,and further improving the accuracy of TC intensity estimation.In this study,499 TCs occurred in the North Atlantic and North Pacific from 2012 to 2019 are used for training and testing,with data from 2018 to 2019 are used for testing.The experimental results show that the MVFF network improves the estimation accuracy in terms of achieving TC intensity estimations with Mean Absolute Error(MAE)of 3.91m/s for the North Atlantic TCs and 3.94 m/s for the North Pacific TCs.