Pre-trained feature aggregated deep learning-based monitoring of overshooting tops using multi-spectral channels of GeoKompsat-2A advanced meteorological imagery. Issue 7 (3rd October 2021)

• Record Type: Journal Article
• Authors: Lee, Juhyun; Kim, Miae; Im, Jungho; Han, Hyangsun; Han, Deahyeon
• Is Part Of: GIScience & remote sensing. Volume 58:Issue 7(2021)
• Page Start: 1052
• Page End: 1071
• Publication Date: 2021-10-03
• Subjects: Feature-aggregated deep learning -- Geostationary satellite -- Geo-Kompsat-2A Advanced Meteorological Imager (GK2A AMI) -- Overshooting tops; Geodesy -- Periodicals; Cartography -- Periodicals; Aerial photogrammetry -- Periodicals; Remote sensing -- Periodicals; 526.05
• DOI: 10.1080/15481603.2021.1960075
• Languages: English
• ISSNs: 1548-1603

Abstract

ABSTRACT: Overshooting tops (OTs) play a crucial role in carrying tropospheric water vapor to the lower stratosphere. They are closely related to climate change as well as local severe weather conditions, such as lightning, hail, and air turbulence, which implies the importance of their detection and monitoring. While many studies have proposed threshold-based detection models using the spatial characteristics of OTs, they have shown varied performance depending on the seasonality and study areas. In this study, we propose a pre-trained feature-aggregated convolutional neural network approach for OT detection and monitoring. The proposed approach was evaluated using multi-channel data from Geo-Kompsat-2A Advanced Meteorological Imager (GK2A AMI) over East Asia. The fusion of a visible channel and multi-infrared channels enabled the proposed model to consider both physical and spatial characteristics of OTs. Six schemes were evaluated according to two types of data pre-processing methods and three types of deep learning model architectures. The best-performed scheme yielded a probability of detection (POD) of 92.1%, a false alarm ratio (FAR) of 21.5%, and a critical success index (CSI) of 0.7. The results were significantly improved when compared to those of the existing CNN-based OT detection model (POD increase by 4.8% and FAR decrease by 29.4%).