The resolution of risk: covariance-informed spatiotemporal clustering improves the detection of hazardous weather events

Accurate delineation of the space time footprints of extreme events requires appropriate dataset resolutions matching the spatial and temporal scale of the natural hazard. More accurate extreme event databases subsequently improve risk estimates. Here, we describe Covariance-Informed Spatiotemporal Clustering (CIST), a novel methodology for quantifying the needed resolution of a given hazard before clustering extreme observations into discrete space time footprints representing individual extreme events. We first derive 24-hour precipitation and maximum daily heat index from ERA5 total precipitation, surface temperature, and dew point temperature. Then, we employ Bayesian Maximum Entropy (BME) to calculate space time metrics describing the rate of fluctuation in space and time, whose ratio describes whether spatial interpolation or coarsening is required. We determine that the spatial resolution of maximum daily heat index should be coarsened from 0.25° to 0.39° and that 24-hour precipitation should be interpolated from 0.25° to 0.05° for unbiased clustering. After resampling each dataset, extreme observations are quantified as exceedances of thresholds developed by NOAA Weather Forecast Offices and Atlas 14. Extreme observations are then clustered using a spatiotemporal extension of Density Based Spatial Clustering of Applications with Noise (DBSCAN). We validate heat waves and storms against the National Centers for Environmental Information (NCEI) NOAA Storm Events Database. We find that heat wave recall improves when using CIST over spatiotemporal clustering (0.94 vs 0.92) from 2019 to 2023. Furthermore, we observe that poor recall of flash floods, floods, heavy rain, and tropical storms reflects the need for higher resolution precipitation data.