Research On CNN Algorithm And Application Of Typhoon Classification Based On Improved MCE Criterion

Typhoons often occur on the surface of the tropics,which seriously endangers the safety of human life and property.China is located in a multi-zone of typhoons.According to the statistics of the China Meteorological Administration,about 50% of the tropical cyclones in the western North Pacific will have an impact on China.In particular,the storm surge caused by strong typhoons can increase the sea level by 5 to 6 meters,causing flooding of seawater and great losses for marine vessels and marine engineering industries.The degree of typhoon affecting human life and property safety is closely related to its intensity,so the study of its intensity is of great significance.Dvorak analysis method and numerical simulation method are the traditional methods for typhoon intensity monitoring.Dvorak method uses the cloud structure characteristics and specific parameters in satellite cloud images to estimate the intensity through empirical rules and constraints.However,the Dvorak analysis method can't obtain the structural information of the typhoon kernel and requires professional knowledge to manually extract the typhoon features.The numerical simulation analysis method optimizes the mature typhoon prediction model by considering the prior physical quantities in the atmospheric environment such as different initial fields,boundary conditions and so on,but the numerical simulation analysis method needs to extract the explicit prior knowledge.In conclusion,both Dvorak analysis method and numerical simulation method have the shortcomings of extracting typhoon characteristics manually and explicitly.With the development of modern remote sensing technology,we can monitor the change of typhoon more accurately and stably by the weather satellite cloud image.Typhoons evolve from tropical cyclones,and typhoons with different intensities show spiral clouds with different characteristics in satellite cloud images,among which the size of dense cloud area is closely related to the intensity of typhoons.In the process of typhoon formation,the characteristics of cloud image are not obvious due to the complex atmospheric factors,so an algorithm is needed to automatically and quickly extract the implicit and complex features of cloud image.Convolutional neural network simulates the automatic learning of human brain and learns the implicit features of natural images through the approximation of complex functions,which overcomes the deficiency of traditional methods that require prior knowledge to extract features explicitly.In view of this,the loss function was optimized by introducing the improved MCE criterion,and a Typhoon-CNNs model suitable for Typhoon classification was built.Compared with the traditional model,the classification effect was better.The details are as follows:Firstly,the Minimum Classification Error(MCE)criterion was introduced to construct the loss function.The process of convolutional neural network simulating human brain is essentially a solving process by minimizing loss function,which is not only used to evaluate the classification effect of network structure,but also affects parameter gradient in gradient descent method.The traditional loss function based on MCE thought has the problem of reverse gradient when the samples are misclassified,which leads to the insufficiency of sample training due to the introduction of wrong signals in the model.Therefore,this paper takes the cross entropy as the basis element in the loss function space,defines the FMCE loss function with error correction term,and on this basis,proposes the loss function CE-FMCE with improved MCE criterion.The CE-FMCE loss function not only overcomes the gradient problem of the traditional MCE loss function,but also overcomes the deficiency that the cross entropy function does not distinguish the gradient of the non-label set.Secondly,the data set of self-built typhoon cloud chart is constructed for convolutional neural network.This paper is based on cloud image data obtained from the National Institute of Informatics(NII)of Japan,including infrared cloud images of nearly 1,000 typhoon processes taken by himawari1-8 satellite from 1978 to 2016.After the pre-processing of median filter denoising and nearest neighbor interpolation scaling,the data sets of 4000 training samples and 800 test samples were constructed,and the corresponding level labels of cloud images were made according to the typhoon level standards of Japan meteorological station.Finally,appropriate super parameters were selected to establish the TyphoonCNNs model for Typhoon classification.On the basis of the traditional convolutional neural network framework,the convolution unit is selected by the cyclic convolution strategy.Using 10-fold cross-validation to match Dropout zero setting rate.The fixed MCE loss function was introduced to optimize the gradient descent method,and the model was adjusted automatically,then the Typhoon-CNNs model suitable for Typhoon classification was established.In conclusion,by analyzing the typhoon cloud image data and loss function,this paper constructed the typhoon cloud image data set oriented to convolutional neural network,and proposed the CNN algorithm based on the improved MCE loss function.On this basis,the application of the improved algorithm in typhoon classification is realized.