Tropical Cyclone Intensity Estimation Based On Convolutional Neural Networks And Relevance Vector Machine

Tropical Cyclone(TC)is a kind of cyclonic circulation,with a large volume and a small diameter of 300--400 km and a large diameter of 1,000-2,000 km.It often occurs on the tropical and subtropical sea surface and is severely destructive.In order to reduce the disaster impact,researchers are committed to the prediction of TC intensity and TC path.Based on FY-2 satellite scanning data and infrared brightness temperature data of tropical cyclones provided by Japan’s National Institute of Informatics(NII),a tropical cyclone intensity estimation method is proposed in this paper.The research mainly includes the following four steps:(1)A tropical cyclone detection model based on Faster-RCNN is proposed.The model can extract the region of interest containing tropical cyclone from infrared satellite images,and generate the bounding box and confidence score of tropical cyclone with any size of still infrared satellite images as input.The experimental results show that given any infrared satellite cloud image,the model can determine whether there is a tropical cyclone target instance,and if so,return to its spatial location and coverage.The higher the intensity level of tropical cyclone,the better the recognition accuracy of the model.(2)A tropical cyclone intensity classification model based on deep convolution neural network combined with tropical cyclone optimal path data is proposed.The model can classify the intensity of tropical cyclone detected in the first step into five categories: storm,strong storm,typhoon,strong typhoon and super typhoon.Convolution neural network can carry out autonomous learning,including four convolution layers,four pooling layers and two full linking layers.Using SOFTMAX layer as classifier,the accuracy rate of intensity classification reaches 96.2%.(3)Combining the static infrared satellite images of tropical cyclones with different intensity types classified in the second step and the best path data of tropical cyclones,a tropical cyclone classification and intensity estimation model based on depth convolution neural network is proposed.The model tests the data of five tropical cyclones with different intensity levels with an average absolute error between 1.80m/s and 2.97m/s.(4)The method of machine learning is used to build a tropical cyclone intensity estimation model.The center of the tropical cyclone is taken as the center of the circle,and within 450 km from the center of the tropical cyclone,starting from the center of the tropical cyclone,every 50 km in the radial direction is taken as an interval,and the captured cloud image is expanded outwards.Taking each point on the tropical cyclone cloud picture as a reference point,the deviation angle matrix is calculated and obtained,and the deviation angle-gradient co-occurrence matrix is constructed.Based on the statistical parameters in the co-occurrence matrix and the central wind speed of tropical cyclone,the characteristic factors closely related to the intensity of tropical cyclone are constructed,and the objective intensity estimation model of tropical cyclone is established by machine learning(relevance vector machine).The optimal intensity scale of tropical cyclone is obtained,and the prediction results are compared with the results of depth learning intensity model.The above three deep learning methods are combined to form a complete tropical cyclone intensity estimation system.The system inputs a static infrared satellite cloud picture,firstly detects whether there is a tropical cyclone target,if so,calibrates its coverage position,according to the identified position,the system cuts out a single tropical cyclone target,and then analyzes the intensity category of the target to determine its intensity category.Finally,the intensity of tropical cyclone with the intensity category determined is estimated to determine its maximum center wind speed near the ground.In this paper,the tropical cyclone intensity estimation method combined with three depth learning models has good effect and can effectively estimate the TC intensity.Compared with the traditional tropical cyclone intensity estimation method,the method in this paper improves the estimation accuracy of TC center wind speed and has better stability.