Research On Tropical Cyclone Intensity Estimation Based On Convolutional Neural Networks

Tropical Cyclone(TC)is a warm-hearted cyclonic vortex that occurs on the ocean surface in tropical and subtropical regions.It is a powerful and profound tropical disaster weather system.Tropical cyclone intensity is defined as the maximum continuous wind near the center of the tropical cyclone.Accurate estimation of tropical cyclone intensity is the key to tropical cyclone forecasting and disaster warning.Dvorak technology(Dvorak)is widely used to determine the intensity of tropical cyclones,and deep learning also shows a level equivalent to Dvorak in tropical cyclone intensity estimation.In addition to the Dvorak technology and its improved version,there are still a series of problems such as high subjective experience in tropical cyclone feature extraction,inconsistent error expression,complex preprocessing,and poor applicability in various ocean basins.Deep learning intensity estimation methods also have significant accuracy advantages.Insufficient utilization of multi-channel satellite observation space information and low interpretability.This article will jointly use multi-channel satellite observation data and deep learning technology to improve the problem of tropical cyclone intensity determination.To this end,this paper uses the tropical cyclone intensity estimation data set and designs a tropical cyclone intensity estimation model based on convolutional neural networks to further improve the accuracy of tropical cyclone intensity estimation.The main research contents are as follows:(1)Aiming at the problem of the uneven distribution of tropical cyclone sample intensity,a classification-regression-integrated multi-layer coupling intensity estimation model of tropical cyclone intensity was studied,and an integrated learning strategy was proposed.The model is composed of classification,regression and BP models,and is trained in sequence using ensemble learning strategies in the training phase.In the verification stage,the classification model divides tropical cyclone samples of different intensities into three categories: strong,medium,and weak,and uses the basis regression model to estimate the cyclone strength respectively;and uses the BP model to couple the regression and the output of the classification model to further reduce errors.The experimental results show that the use of the integrated model can effectively improve the accuracy of the model.The final accuracy is 8.91 kts,which is 1.52 kts higher than the Dvorak technology and 0.3 kts higher than the current deep learning method.(2)In order to further investigate the impact of satellite vertical stereo observation information on TC intensity estimation,the Res Net10 basic model is improved on the basis of infrared,water vapor and microwave channel data,and 3D convolution and convolution attention mechanisms are used to enhance the model's extraction of tropical cyclone stereo information.And the model's attention to the warm core.The model is further optimized by using domain knowledge such as tropical cyclone rotation invariance.The five-point smoothing accuracy reaches 7.2 kts,which is 3.23 kts higher than Dvorak technology and2.01 kts higher than the current deep learning method.Experiments and model visualization results show that the integrated learning strategy and multi-layer coupling strength estimation model proposed in this paper can effectively solve the regression problem under uneven data distribution,and can enhance the ability to extract satellite multi-channel stereo information and pay attention to warmth.Degree to replace the base model for higher accuracy.Compared with Dvorak technology and deep learning intensity estimation method,the method in this paper is significantly improved.