Remarkably fast enhancement of machine learning for estimating spatiotemporal dataset suggest new ways to remote sensing and atmospheric fields in using data-driven modeling. Especially, many kinds of optimizable model structure for spatio-temporal database such as LSTM (Long Short Term Memory) or token-mixing based methodology have been suggested and replaced ‘state of the art’ models in representing numerical phenomenon over the earth. In addition to this, vast amount of observation database has been reanalyzed and assimilated as large amount of atmospheric variables such as EAR5 or MERRA-2 which can cover densely along spatial and temporal aspect. In this research, we compare LSTM based model (PredRNN) with token-mixing based model(AFNO) in forecasting precipitation as quantitative guideline for application of newly suggested machine learning model. 8 different atmospheric variables (wind components, temperature, relative humidity, water vapor and geopotential, total precipitation) over 4 pressure level (500, 750, 850, 1000) from ERA5 and RAR(Radar-AWS Rainrates) database from KMA (Korea Meteorological Administration) is adopted as input and output data to train and evaluate models.
|