Research On Tropical Cyclone Intensity And Scale Estimation Based On Multi-task Convolutional Neural Networ

Tropical cyclones are highly destructive cyclonic circulations that occur in the tropical oceans,often accompanied by severe weather events such as storms,showers,and storm surges.Accurate estimation of tropical cyclone intensity and size is the key to issuing early warning signals and taking preventive measures.However,traditional meteorological methods have a series of problems such as high subjective experience and complex preprocessing.Therefore,accurate and objective estimation of tropical cyclone intensity and size is a long-term challenge.In recent years,methods for estimating the intensity and size of tropical cyclone based on deep learning techniques have made positive progress.However,the current speed and accuracy of estimating the intensity and size of tropical cyclones are still insufficient to meet operational requirements.Therefore,based on satellite imageries of infrared,water vapor,and passive micro-wave rainfall,as well as environmental factors,and supported by convolutional neural network and multitask learning,this paper designs tropical cyclone intensity and size estimation models to achieve more accurate estimation of tropical cyclone intensity and size.The research work carried out in this article is as follows:(1)Aiming at the problem of single tropical cyclone information provided by infrared satellite imagery and two-dimensional convolutional overlapping multi-channel satellite imagery information,a multidimensional convolutional neural network based on multi-channel satellite data of infrared,water vapor,and passive micro-wave rainfall is designed to obtain three-dimensional atmospheric information related to tropical cyclone intensity from multi-channel data through the sliding of multidimensional convolution in the channel dimension.In addition,the convolutional attention module is employed to simulate visual attention to enhance the model’s attention to important features such as tropical cyclone eye,convective structure,and important channels.Compared to existing deep learning methods,the accuracy of tropical cyclone intensity estimation has been greatly improved.(2)Aiming at the problem of low accuracy in estimating the intensity and size of tropical cyclone,a multi-task learning model with adaptive loss balancing is designed to simultaneously estimate the intensity and size of tropical cyclone,considering the characteristics of multi-task learning that achieves common improvement of multiple tasks through sharing underlying parameters.An adaptive loss balancing strategy is used to dynamically select the optimal loss weight for multiple tasks,enabling multiple tasks to learn at similar speeds,avoiding multi task learning being dominated by one task.In addition,the dual attention module is used to strengthen the model’s attention to the global dependence and correlation of tropical cyclone.Experimental results show that the performance of the proposed method is superior to existing deep learning methods.(3)To address the challenge of overestimating and underestimating the intensity and size of tropical cyclone,a label distribution smoothing method is employed to convolution symmetric kernels with empirical label density distribution to obtain an effective label density distribution that reflects the true imbalance of the samples.Designing a loss function based on effective label density distribution to solve the problem of overestimating and underestimating of intensity and size of tropical cyclone has greatly improved the accuracy of tropical cyclone intensity and size estimation.In addition,deep learning visualization technology is employed to explore the interpretability of designed model to understand the process of model learning intensity and size features,and understand the contributions of various parts of satellite imageries to intensity and size estimation.