CHARACTERIZING AND PREDICTING HYDROLOGICAL DROUGHT FOR IMPROVED EARLY WARNING OF DROUGHT ONSET, DURATION, AND SEVERITY

Drought events will be more impactful and more difficult to predict in coming years given continued climate change. The USGS Drought Program is working to characterize and predict hydrological drought (a lack of water in the hydrological system) reflected by abnormally low streamflow, groundwater, and surface water reservoir storage. Ongoing work at regional and national scales focuses on (i) better understanding historical dynamics of hydrological drought, (ii) developing data-driven methods to predict drought onset, duration, and severity (iii) examining the nexus of drought and other hazards including floods and wildfires. In this presentation we first focus on streamflow drought variability and trends across the U.S. Long-term streamflow drought variability was greatest in areas that experienced the most severe streamflow droughts and, for much of the western and southern U.S., drought duration and severity increased from 1951-2020. We then present random forest and long short-term memory neural network modeling approaches to predict streamflow drought in the Colorado River Basin using gridded meteorology, modeled snow and soil moisture storage, satellite observations, and watershed properties. Our modeling approaches account for non-stationary drought occurrence and predict drought using both fixed and seasonally varying drought thresholds of varying severities. Initial modeling efforts focused on the Colorado River Basin for 1981-2020 and demonstrated a stronger ability to predict streamflow droughts defined using fixed drought thresholds, and in predicting less severe droughts. With a fixed drought threshold, performance decayed more slowly moving from 0-60 days lead time (median NSE 0 day: 0.90, median NSE 60 day: 0.65) compared to the same predictions made with variable drought thresholds (median NSE 0 day: 0.80, median NSE 60 day: 0.20). Models had weaker predictive capacity in highly regulated locations, and alternate approaches are being explored to improve model performance for mainstem locations and below reservoirs. Ultimately, our goal is to develop operational tools for hydrologic drought early warning forecasts tailored to stakeholder needs at catchment scale and weekly timesteps, with forecasts available several months in advance.