Deriving Climate Zones of the Contiguous United States with Spatial Machine Learning

A spatially constrained unsupervised machine learning technique (SKATER) is implemented to objectively derive spatially contiguous climate zones for the United States. SKATER (Spatial ‘K’lustering by Tree Edge Removal) is a machine learning technique that grows and prunes a minimum spanning tree to create spatially contiguous, homogeneous clusters. Koppen and Trewartha, the mostly widely used climate systems, are classifications of climate based on some prior set of subjective rules. Machine learning techniques allow the climate data to ‘speak for itself’ and divide into a natural classification. Contemporary climate zones are presented based on recent (2017) monthly mean temperature and precipitation data from the very high resolution (approximately 32 km at the lowest latitudes) North American Regional Reanalysis (NARR) model. Potential future (2050) climate zones are presented based on the NCAR Community Climate System (CCSM) climate change scenarios. A time series of climate zone delineations (1979 through 2000) is constructed to explore the sensitivity of the clustering method to small to moderate perturbations.

Machine Learning (ML), a set of data-driven algorithms and techniques that automate the prediction, classification, and clustering of data, are increasing helping scientists understand past climates, better predict future climates and identify large-scale weather patterns. This presentation shows an approachable, practical application of machine learning for summarizing the vast and increasing volumes of data generated from numerical weather models.