Determination of Spatial and Temporal Drought Patterns Over CONUS Using Unsupervised Machine Learning Clustering Algorithms

The Standardized Precipitation Index (SPI) is computed over CONUS using daily precipitation estimates from the NOAA Daily U.S. Climate Gridded Dataset (NClimGrid-Daily). From the NClimGrid-SPI (1951-present; 0.05°x0.05°), we derive historical hydro-climatological conditions and drought information from the Drought Severity and Coverage Index (DSCI) which combines drought levels into a single areal value (from 0 to 500). One of our objective is to better understand drought dynamic and particularly how drought episodes evolve from short term rainfall deficit (i.e., less than three months) to persistent drought condition (i.e., beyond nine months). To investigate how those cascading effect work, we use a Machine Learning (ML) approach to identify spatio-temporal patterns of drought episodes over CONUS. Several unsupervised ML clustering algorithms are tested using an ensemble of features including drought duration, rainfall accumulation, drought severity (maximum DSCI, time of maximum DSCI), seasonality (drought beginning and end dates), location (latitude, longitude). Results show that the most severe drought events (i.e., DSCI > 350) are those that have the longest durations and for which drought relief is associated with higher rainfall accumulation regardless of the location considered. Furthermore, there is an apparent consistency across accumulation scales and the number of parameters selected with an optimum number of clusters around four. The Euclidian distance ML models tested seems to be able to define spatiotemporal areas of similar drought patterns. Differences between models are observed in terms of spatial definitions and predominance  of a cluster at a given location. The strongest prevalence of a given cluster has allowed to isolate areas of coherence such as the Pacific Northwest, the PNW, the Eastern Seaboard and the Southeast, and the Southwest area along the MX-US border. Domain delineations are weaker for areas such as the Rockies, the Midwest, and the Great Plains. While the SPI algorithm assumes a Gamma (McKee et al., 1993) or a Pearson III (Guttman, 1998) distribution for monthly rainfall accumulation periods, results show that this assumption might not be optimal depending on the domain considered and the accumulation period when applied to daily drought monitoring.