A Development of High-Resolution Long-Term Meteorological Grid Data Using Deep Learning over South Korea

It is unequivocal that climate change has a huge impact on natural environments as well as many different socio-economic sectors. To efficiently respond and adapt to such climate change, it is very important to analyze climate change trends in the long term and their future impacts according to climate scenarios. For this, it is essential to produce high-quality and high-resolution long-term grid-type climate data based on observations, which is a key piece for developing high-quality downscaled future projections. Also, It can be used as important information for big data-based modeling work. However, South Korea lacks long-term gridded climate data because a dense observation network from ASOS (Automated Synoptic Observing System) and AWS (Automated Weather System) has only been available since 2000. To address this issue, this study applies MK-PRISM, an interpolation scheme to quantify the influence of meteorological factors considering elevation according to the characteristics of South Korea. Also, we used a deep learning technique to reconstruct spatial variations over a historical period (1972-1999), which a dense network could produce if it existed. Specifically, we trained a deep learning model for each grid cell over the recent 20 years (2000-2021) using gridded products based on only ASOS station as input data and the one based on a dense network (ASOS+AWS) as an output. Finally, we produced daily maximum, minimum temperature, total precipitation, and average wind speed data with a spatial resolution of 5 km for a total of 50 years (1972-2021). This daily gridded meteorological data will be used as important data for long-term climate trend analysis in South Korea.