A Development of High-Resolution Long-Term Gridded Meteorological Data for South Korea using Deep Learning

Climate change has a considerable impact on socioeconomic fields as well as on the natural environment. To effectively respond and adapt to climate change, we should analyze the long-term climate change trends and future impacts according to plausible climate scenarios. For this, the production of high-quality and high-resolution gridded meteorological data based on observation is essential, which is important for developing high-quality downscaled future projections. However, South Korea lacks the long term gridded meteorological data because a dense network from ASOS (Automated Synoptic Observing System) and AWS (Automated Weather System) Stations was available only after 2000. To address this problem, this study aims to produce high quality gridded meteorological data for a historical period (1973-1999), which could have been generated if a dense network existed. Specifically, we reconstruct spatial variations and features of meteorological variables for the historical (1973-1999) period by relating the gridded products for more recent period (2000-21) to that for preceding period (1973-1999) based on a deep learning algorithm. For this, MK-PRISM, an interpolation method for quantifying the effects of meteorological factors based on elevation in South Korea was applied to produce two different version of gridded products based on two different observation networks: a sparse network (ASOS) and a dense network (ASOS+AWS) over the recent 22 years (2000-2021). Then, we develop the Long-Short Term Memory (LSTM) for each grid cell using the gridded products by the sparse network as input and the denser network as output layer. Finally, we generate meteorological variables for the period of 1973-1999 using gridded product by a sparse network as input of the developed LSTM model for each grid cell. Our preliminary results showed that Nash-Sutcliffe Efficiency (NSE) was higher than 0.9 in most grid climate prediction models. Therefore, our development in this study has a potential to calculate high-quality and long-term meteorological data which can be used as important data to analyze the long-term climate trends and variability.

Acknowledgement:

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A4A3032838).