search for




 

Development of deep-learning based present weather
Journal of the Korean Data & Information Science Society 2021;32:1007-21
Published online September 30, 2021;  https://doi.org/10.7465/jkdi.2021.32.5.1007
© 2021 Korean Data and Information Science Society.

Min Woo Kim1 · Reno Kyu-Young Choi2 · Yoon Sang Lee3 · Junho Choi4 · Do Youn Kim5 · Ki Hoon Kim6

1236Operational Systems Development Division, National Institute of Meteorological Sciences
4Department of Mechanical Engineering, Imperial College, UK
5ARA Consulting & Technology
Correspondence to: 1 Researcher, Operational Systems Development Department, National Institute of Meteorological Sciences, Seogwipo 63568, Korea.
2 Corresponding author: Principal Scientist, Operational Systems Development Department, National Institute of Meteorological Sciences, Seogwipo 63568, Korea. E-mail: renochoi@korea.kr
3 Researcher, Operational Systems Development Department, National Institute of Meteorological Sciences, Seogwipo 63568, Korea.
4 (SW7 2AZ) Department of Mechanical Engineering, Imperial College London, London, UK.
5 Director, ARA Consulting & Technology D-1510, Songdomirae-ro, Yeonsu-gu, Incheon, Korea.
6 Senior research scientist, Operational Systems Development Department, National Institute of Meteorological Sciences, Segwipo 63568, Korea.
This work was funded by the Korea Meteorological Administration Research and Development Program “Enhancement of Convergence Technology of Analysis and Forecast on Severe Weather” under Grant (KMA2018-00121).
Received August 2, 2021; Revised August 30, 2021; Accepted September 10, 2021.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Abstract
Present weather (PW) is a qualitative state of weather condition helping publics and industries. This intuitive information that has long been recorded by human-eyed observations needs be derived automatically. While number of conventional VPWs (visibility and present weather sensor) offer PW mainly from visibility, this study aims to determine it by deep learning with various quantitative meteorological measurements. Most probabilistic PW is determined by MLP model and softmax regression model with input, two hidden, and output layers. Additional korean abstract contents have to include. Model optimization yields best set of model parameters comparing with human-eyed PW observations. The model consistency is obtained by over 3 years of training data size, implying enough minimal training size of each type of PW. Performance of the model with VPW used statistical verification measures, such as probablity of detection (POD), false alarm ratio (FAR), and critical success index (CSI). Significant improvement is achieved in haze and mist in all indications, resulting that the deep learning model showed greater than around 70% of PODs over all PWs.
Keywords : ASOS, deep learning, MLP, present weather, softmax regression.