融合定位信息的热带气旋强度估计  

Location information-fused estimation method in relevance to tropical cyclone intensity

作　　者：刘英杰 张芮 刘青山 杭仁龙[1,2] Liu Yingjie;Zhang Rui;Liu Qingshan;Hang Renlong(School of Computer Science,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Computer Science,Jiangsu University of Science and Technology,Zhenjiang 212100,China)

机构地区：[1]南京信息工程大学计算机学院,南京210044 [2]南京信息工程大学自动化学院,南京210044 [3]江苏科技大学计算机学院,镇江212100

出　　处：《中国图象图形学报》2023年第8期2522-2535,共14页Journal of Image and Graphics

基　　金：科技创新2030—“新一代人工智能”重大项目(2021ZD0112200);国家自然科学基金项目(U21B2049,62206115);江苏省自然基金青年基金项目(BK20220646)。

摘　　要：目的精确估计热带气旋的强度有助于提升天气预报和预警的准确性。随着深度学习技术的不断发展,基于卷积神经网络(convolutional neural network,CNN)的方法已应用于强度估计任务中。然而,现有方法仍存在许多问题,例如无法充分利用不同波段的卫星图像信息、输入图像以热带气旋的定位为中心等限制,从而产生较大误差,影响实时估计的结果。针对以上问题,本文提出一种融合定位信息的强度估计网络IEFL(intensity estimation fusing location),提升强度估计的准确率。方法模型采用双分支结构,能有效融合不同波段的图像特征,同时可以同步优化两个任务,达到互相促进的效果。此外,模型对强度估计任务做了定位的特征融合,将得到的定位特征图与强度特征图进行拼接,共同输出最后的强度结果,通过利用定位信息达到提升强度估计精度的目的。结果本文在完成热带气旋强度估计的同时,可获取较好的热带气旋中心定位结果。收集了2015—2018年葵花-8卫星多通道图像用以训练模型,并在2019和2020年的数据上进行测试。结果表明,融合定位信息后模型的强度估计均方根误差为4.74 m/s,平均绝对误差为3.52 m/s。相比传统单一强度估计模型误差分别降低了7%和9%。结论IEFL模型在不依赖定位准确率的同时,能够有效提升强度估计的准确率。Objective A tropical cyclone can generate such severe weather condition like strong winds or heavy precipitation,as well as such secondary disasters derived of floods,landslides,and mudslides.Tropical cyclones may often threaten survival contexts in related to coastal community.The intensity of tropical cyclones(TC)can be estimated accurately and it is beneficial for weather forecasting and warning.Deep learning techniques-based convolutional neural networks(CNNs)methods have its optimal ability for estimation task of tropical cyclone intensity apparently.However,CNN-based methods are still challenging for such problem of insufficient use of multi-channel satellite images,and the input images are preferred to be centered on the location of tropical cyclones.To resolve large estimation errors and real-time estimation results.we develop a network called intensity-estimation-fusing-location(IEFL)to improve accuracy of intensity estimation further.Method The training data is captured from Himawari-8 satellite images from 2015 to 2018 in comparason with such data contexts from 2019 to 2020.The dataset contains 42028 training images and 5229 testing images.First,the data are preprocessed to remove non-TC cloud systems via clipping satellite images.Then,the implementation of data augmentation strategy is oriented to optimize the over-fitting problem and enhance the model robustness.Moreover,multiple channel images analysis is required to reveal varied features of TCs.Thus,a better combination for intensity estimation can be developed through fusing multi-channel images.The network is set up via a two-branch structure,which can be used to fuse different channel images effectively.Two sort of tasks can be optimized simultaneously and learnt mutually.In addition,the network can feed location task-extracted features into intensity estimation task.Specifically,their feature maps can be concatenated and intensity estimation results are generated as following.The experiment is segmented into two categories as mentioned below:f

关 键 词：强度估计 热带气旋(TC) 卷积神经网络(CNN) 中心定位 葵花-8 

分 类 号：P457.8[天文地球—大气科学及气象学]