Estimation Of Tropical Cyclone Intensity And Classification Based On Improved Residual Network

Tropical cyclones are one of the most serious natural disasters,mainly including hurricanes in the Atlantic Ocean and the northeast Pacific Ocean,cyclones in the South Pacific and Indian Ocean,and tropical cyclones in the northwest Pacific Ocean.The strong winds and heavy rainfall brought by tropical cyclones when they land often have an important impact on people's property safety and social and economic development in coastal areas.How to accurately measure the intensity of tropical cyclones,accurately classify tropical cyclones and reduce the loss of people in coastal areas has become a hot research topic at present.The accuracy of the traditional measurement method,Dvorak,often depends on the subjectivity of the field experts,while the numerical analysis method needs to combine the prior knowledge of many related fields,so it is not efficient.In recent years,the wide application of deep learning in computer vision,natural language processing,speech recognition and other fields has truly realized the era of artificial intelligence.In particular,a Deep convolutional neural network(DCNN)can automatically extract the low-level and high-level features of the target data,and the efficiency of the model is far superior to that of the traditional model.Based on this,this paper constructs a deep residual network,improves the residual block,and combines the attention mechanism to assign weight to the training in terms of channel and time,thus classifying the grade of tropical cyclones.In addition,the network is applied to automatically measure the cyclone intensity through satellite remote sensing tropical cyclone images,aiming to use the residual network model algorithm to improve the automation degree of cyclone intensity measurement,reduce the measurement error and improve the accuracy of tropical cyclone classification.The specific research results are as follows:1)the traditional residual network under the environment of small data samples branch convolution waste layer characteristics,and the characteristics of tropical cyclone cloud complex,a close correlation between time and space,the characteristics of the use of Japan's National Institute of information science(National Institute of Informatics,NII)from multiple meteorological satellite in the northwestern Pacific Ocean cloud picture data of more than 8000 high resolution scene of tropical cyclones,build the training set and testing set.This chapter puts forward a double wide attention mechanism fusion residual piece of tropical cyclone classification algorithm,the algorithm to the original residual network residual block was improved,adapted to the characteristics of tropical cyclones is established and the identity map residual block structure of a data set that add convolution output to improve the utilization rate of branch channel and the introduction of attention mechanism make the network focused on tropical cyclone image sequence features,and suppress noise and redundant information.A prediction model DAW-RESNET,which is suitable for small sample tropical cyclone data set,is constructed.The experimental results show that the training accuracy of DAW algorithm on the self-built tropical cyclone data set reaches 99.90%,and the test accuracy reaches 86.19%,which is 5.94% higher than that of the traditional Resnet50 residual network.2)Due to the subjectivity of traditional tropical cyclone intensity estimation algorithm Dvorak technology,numerical simulation analysis requires a large number of prior physical quantities,resulting in low efficiency.Using Japan's National Institute of information science(National Institute of Informatics,NII)from "GMS-5" and other meteorological satellite in northwest Pacific Ocean cloud picture data of more than20000 high temporal resolution scene of tropical cyclones,establish residual neural network model of tropical cyclone intensity recognition Resnet-TC,applied in this paper,we adapt to the characteristics of tropical cyclone on identity map residual block structure,by increasing the convolution output to improve the utilization rate of branch channel and fully to extract characteristics of high and low level cyclone,improve the network performance.The experimental results show that RESNET-TC has good algorithm stability and high degree of automation,and can estimate the intensity of tropical cyclone effectively.The root mean square error of central maximum wind speed is 3.172m/s,and the mean absolute error is 2.146m/s,both of which are lower than the traditional statistical method and CNN model method.The correlation coefficient between the reference value and the predicted value is 0.9583.Moreover,through the visual analysis of the model,the features extracted by RESNET-TC are similar to those extracted by Dvorak TC mode,indicating the effectiveness of the proposed method in interpreting the intensity characteristics of tropical cyclones.