Research On Large-scale Deep Learning Algorithms For Tropical Cyclones Recognition

A tropical cyclone is a severe weather phenomenon that usually originates on the ocean surface in tropical or subtropical areas.It is one of the most important natural disasters in China.To study the occurrence mechanism,life cycle and even the development trend of tropical cyclones,we should recognize them effectively.The observation data obtained with the help of satellites are often not accurate enough,while tropical cyclones themselves are very destructive,making them difficult to observe at close range.Therefore,climate data analysts usually use climate simulation software on supercomputers to perform simulations,and rely entirely on multi-variate threshold conditions for prescribing these extreme weather patterns.Criteria for different climate data analysts are often inconsistent,which is not conducive to the objectification of results Thus far,there are some precedents in the field of climate research for analyzing and processing scientific data with the help of neural networksIn this subject,we use deep learning methods to achieve large-scale,pixel-level segmentation,and try to learn a common pattern from the complex climate science data to avoid the effects of subjective thresholds.We utilize the Community Atmosphere Model 5.0(CAM5)module in Community Earth System Model(CESM)to generate a five-year climate simulation dataset with a resolution of 0.5-degree.Then,we propose a complete scheme for transforming traditional climate simulation data into neural network training data.After using the Toolkit for Extreme Climate Analysis(TECA)to mark the cyclone center,pre-processing of data labels,we finally generate the labels with a series of label generation algorithms.Then,we carry out some data post-processing and regularization strategies,which can effectively and scientifically convert scientific simulation data into training data files.Finally,due to the scientific dataset on which this subject is based,it has the characteristics of larger overall size,larger single file size,and extremely small segmentation targets.We analyze the existing pipeline requirements for data input.Based on the network architecture of DeepLabV3+,a deep neural network model is designed to solve the problem of tropical cyclone recognition.We use a weighted cross-entropy loss function to solve the class imbalance problem caused by extremely small segmentation targets.These special data put forward new requirements for network construction.In this paper,a series of optimization strategies such as data-preprocessing,data pipeline,and data shuffle are used for the construction and optimization of this neural network,and have achieved certain resultsIn the end,the network achieves effective convergence on both the training set and the test set,achieving a mIOU performance of 51.43%,which is also pretty good.The analysis of tropical cyclone recognition effect also shows that the model has a very significant effect.