Forecasting Hydrological Drought Onset, Duration, and Intensity in the Colorado River Basin and Across the Conterminous United States Using Machine Learning Models

Drought events will likely be more impactful and difficult to predict incoming years given continued climate change. The U.S. Geological Survey Water Mission Area Drought Program is working to characterize and predict hydrological drought, defined as abnormally low streamflows and groundwater levels. Work at regional and national scales is focused on developing data-driven methods to advance early warning capacity for hydrological drought onset, duration, and severity. In this presentation we first focus on tree-based and long short-term memory neural network modeling approaches to forecast 0-90 day streamflow drought conditions in the Colorado River Basin using gridded meteorology, modeled snow and soil moisture storage, and watershed properties. Our modeling approaches predict drought for moderate (20%), severe (10%) and extreme (5%) intensity levels using both fixed (one threshold applied to all days and years) and seasonally varying drought thresholds (different drought thresholds for each day of the year). Quantile regression models are used to estimate streamflow percentiles and uncertainty ranges for 7 and 14 day forecasts, while drought likelihood and duration are modeled using classification approaches for 30 to 90 days. Preliminary regression models for 7 and 14 days had median KGE values of 0.79 and 0.66 for variable thresholds and 0.91 and 0.84 for fixed thresholds. Cohen’s Kappa for classifying moderate intensity drought periods ranged from0.45 for 30 day fixed threshold forecasts to 0.12 for 90 days, while Kappa values for predicting drought durations ranged from 0.80 for 7 day fixed threshold drought to 0 for 90 day fixed threshold droughts. Overall, models had weaker predictive ability for regulated basins, increasingly intense droughts, longer lead times, and variable thresholds, and for these reasons modified approaches are being explored to improve model performance. Following regional results, we present outcomes from models predicting streamflow and groundwater drought at sites across the conterminous United States. As we continue to develop prototype forecast tools for assessing and predicting hydrologic drought conditions, we are incorporating stakeholder input to design tools that complement existing drought and water supply prediction tools.