Towards machine learning-based streamflow drought forecasts across the Colorado River Basin

Abstract

<p>Drought is among the most damaging environmental phenomena, affecting agricultural productivity, wildfire risks,  hydropower production, water quantity and quality, public health, ecosystem integrity, and recreation. Streamflow drought, where the streamflow declines below a threshold defining anomalously low flows, is one measure of hydrologic drought that can be interpreted as an integrative measure of the availability of water for specific uses. Early warning of streamflow drought onset, severity, spatial extent, and duration is needed to support improved water resource management. Streamflow drought forecasting is particularly important in the western United States where a changing climate threatens already-scarce water resources.</p><p>The U.S. Geological Survey is  applying a variety of machine learning and artificial intelligence modeling methods to predict streamflow drought in a 40-year retrospective analysis at 425 USGS stream gage locations within and surrounding the Colorado River basin. In this presentation, we briefly provide an overview of these approaches, then primarily focus on results from random forest binary classification models for streamflow drought onset and duration. For this study, streamflow drought is defined using seasonally variable streamflow exceedance thresholds developed from the Weibull distribution of observed flows or zero-flow durations from 1981-2020. We trained a large set of random forest models (n =72) , each of which predicts daily streamflow drought onset and duration probabilities at a particular forecast horizon and severity level. The models are trained using past observations of daily streamflow drought and a predictor dataset of daily hydrometeorological variables and static basin characteristics We combine the results of these models to provide holistic forecasts. In addition to streamflow drought prediction performance, we evaluate the opportunities for transitioning this modeling framework to operational forecasting and consider future directions for providing actionable forecasts to regional and national stakeholders.</p>