Hurricane Wind Speed Estimation Based On Satellite Microwave Data And Deep Learning

Hurricane is one of the most serious natural disasters for human.Accurate estimation of hurricane wind speed(or hurricane intensity)can help people to forecast and manage disasters more effectively and reduce a lot of losses caused by it.The traditional hurri-cane wind speed estimation method requires the use of anemometer equipment(buoys,aircrafts,ships,etc.)for on-site measurement,which is difficult to observe the vast ocean area for a long time and the cost is too high.Satellite remote sensing instruments can con-tinuously observe hurricanes without being affected by the climate.However,the existing methods can not effectively process the satellite remote sensing data,resulting in a low accuracy of hurricane wind speed estimation.Therefore,it is very important and chal-lenging to estimate hurricane wind speed accurately.In recent years,Deep Learning has shown excellent performance in tasks dealing with large amounts of highly complex data.In this thesis,a Multi-Scale Feature Distribution Learning Network(MFDLNet)suitable for hurricane wind speed estimation is proposed by using satellite microwave brightness temperature data and Deep Learning method.The specific research contents are as fol-lows:1.A hurricane three-dimensional(3D)structural feature extraction method on satel-lite microwave data is proposed.This article uses the brightness temperature data of ATMS on the US Suomi National Polar-orbiting Partnership Satellite(SNPP).Using the ATMS Temperature Retrieval Algorithm,the temperature anomaly data of the hurricane warm-core was synthesized at different pressure.Then the warm-core temperature anomaly data is sampled from three directions of along-track,the cross-track,and the pressure level.Finally,the cross-section samples in these three directions are concatencated to form a two-dimensional(2D)data stream,which constructs the Hurricane Warm-core Dataset.2.In the construction of Hurricane Warm-core Dataset,a Background Iterative Ex-pansion algorithm and a Non-warm-core Suppression algorithm are proposed.Background Iterative Expansion solves the problem that the effective background field is lacked in the range of the hurricane center±10°,which disables the generation of the warm-core tem-perature anomaly data.The Non-Warm-core Suppression algorithm solves the problem that the non-warm-core high temperature anomaly data at the edge of the hurricane inter-feres with the actual warm-core structure.These two algorithms improve the confidence of Hurricane Warm-core Dataset and make the wind speed estimation more accurate.3.In order to further improve the accuracy of hurricane wind speed estimation,inno-vatively proposes a Multi-scale Feature Distribution Learning network,MFDLNet.The proposed network consists of two main parts:The first part introduces a multi-scale fea-ture extraction mechanism.This mechanism can effectively identify hurricane warm-core features of different scales,and its structure includes a series of convolutional layers and a combined framework of six different scale sampling features.The second part intro-duces a distribution learning mechanism,using the probability distribution of estimated wind speed values.Its structure includes two fully connected(FC)layers and the joint loss function proposed.The MFDLNet makes full use of the respective advantages of direct regression and label distribution learning.This network is trained by the Hurricane Warm-core Dataset,and makes full use of the advantages of multi-scale feature structure and label distribution learning to achieve high accuracy hurricane wind speed estimation.By using North Atlantic(NA)hurricane data from 2018 to 2019,MFDLnet achieved a Mean Absolute Error(MAE)of 4.37m/s.The MAE is reduced by 37%and 14%,re-spectively,compared with the traditional algorithm and the latest estimation algorithm.The experiment results verify the effectiveness of hurricane wind speed estimation based on satellite microwave data and Deep Learning.