Delineation of Climte regions over Korean Peninsula Using Machine Learning Approaches

Abstract

Similarity of climate conditions is an useful index for representing the spatial environmental characteristic of a region. There are some approaches to regionalize climates such as rule- and clustering-based classification using temperature and precipitation data. Among them, Köppen-Geiger (K-G) climate classification has been widely used by researchers across a range of disciplines as a basis of the climatic characteristics of regions. However, because Köppen-Geiger (K-G) climate data have relatively course spatial resolution (0.05-, 0.17 and 0.08 degree, it is not often suitable for local to regional applications such as the Korean Peninsula, which has rugged and complex terrains. The purpose of this study is to classify climate regions at 1 km spatial resolution from satellite-based data using several state-of-the-art machine learning methods over the Korean Peninsula. The land surface temperature (T) provided by Moderate Resolution Imaging Spectroradiometer (MODIS) and the precipitation(P) provided by Tropical Rainfall Measuring Mission (TRMM) from 2001 to 2016 were used for calculating input variables such as annual mean of T and sum of P, max and min of T and P for each summer (from May to October) and winter (from November to April). Shuttle Radar Topography Mission (SRTM) digital elevation models (DEM) were also used to consider the terrains. The target variable was generated by rule of the K-G climate model at weather stations (a total of 90). Random forest (RF) and Artificial Neural Network (ANN) were applied to develop climate classification models. Results showed that overall accuracy of RF and ANN models was relatively high with some modeling errors caused by the small number of training samples for a few climate classes. The major difference between our models and the K-G climate classification model was found in mountainous areas with complex terrains. The results provided climate regions with much more detailed spatial variations at high resolution (i.e., at a sub-city scale) than the existing data. In addition, a temporal trend of climate regions over the Korean Peninsula was also examined.