Classification with imbalanced cloud data using deep convolutional neural network

Daisuke Matsuoka; Masuo Nakano; Daisuke Sugiyama; Seiichi Uchida

Presentation information

[J] Oral

M (Multidisciplinary and Interdisciplinary) » M-AG Applied Geosciences

[M-AG39] Marine-Earth Informatics

Thu. May 30, 2019 1:45 PM - 3:15 PM A10 (TOKYO BAY MAKUHARI HALL)

convener:Seiji Tsuboi(JAMSTEC, Center for Earth Information Science and Technology), Keiko Takahashi(Japan Agency for Marine and Earth Science and Technology), Masaki Kanao(National Institute of Polar Research), Daisuke Matsuoka(Japan Agency for Marine-Earth Science and Technology), Chairperson:Daisuke Matsuoka, Seiji Tsuboi

2:45 PM - 3:00 PM

[MAG39-05] Classification with imbalanced cloud data using deep convolutional neural network

*Daisuke Matsuoka^1,2, Masuo Nakano¹, Daisuke Sugiyama¹, Seiichi Uchida³ (1.Japan Agency for Marine-Earth Science and Technology, 2.Japan Science and Technology Agency, 3.Kyushu University)

Keywords:Deep learning, Classification, Tropical cyclone

Image classification using deep convolutional neural network is effective technique to detect extreme phenomenafrom climate data. However, the number of extreme phenomena such as tropical cyclone (TC) is overwhelmingly small compared with others. It is knon that classification performance decline du to this imbalance between positive (TCs) and negative examples (non-TCs). In the present study, we developed a new negative data selection method for binary classification for cloud images. We analyze the relationship between the ease of classification and the characteristics of data, and optimize the training data in order to decrease the false alarm ratio (FAR). As the results, we succeeded in decrease FAR from 32.8–53.4% to 60.0–70.0% in the western North Pacific in the period from July to November.