Exploring the Potential of Unsupervised Machine Learning and Satellite Measurements to Derive Climate Zones in the USA

Identifying climate zones is important for assessing population and ecosystem vulnerabilities to climate changes. We explore the potential of using three different unsupervised machine learning techniques, k-means clustering, self-organizing maps, and hierarchical clustering, with satellite data to identify climate zones within the United States (US) over the time period 2000-2020 using the following datasets: terrestrial water storage (TWS) from the GRACE satellite, surface air temperature from the AIRS satellite, precipitation (P) from the TRMM satellite, vapor pressure deficit VPD from the AIRS satellite, surface solar radiation from the CERES satellite, and elevation data from ASTER. While climate zone maps typically rely solely on temperature and precipitation, we also included TWS, solar radiation, and VPD as they are commonly used to inform on drought and fire occurrence, both of which are increasing issues within the United States and could drive changes in climate zones. We found that the k-means, SOM, and hierarchical clustering all found similar climate zones, and best performed compared to a Köppen-Geiger climate classification map with 9 clusters. Examination of silhouette plots for each machine learning method showed that k-means clustering performed the best. Our method could be useful in monitoring how climate zones change in the future, especially in areas of the world where in-situ measurements are sparse.