USGS River DroughtCast Uses Century of Data to Predict Water Shortages With Machine Learning

The U.S. Geological Survey has released River DroughtCast, a machine learning-based forecasting tool designed to predict streamflow drought across more than 3,000 locations in the United States. The system fills a critical gap in water management — the weeks-long blind spot between short-range weather forecasts and seasonal water supply outlooks — giving water managers, farmers, and municipal planners a new window into emerging shortages.

How It Works

River DroughtCast draws on data from thousands of USGS streamgages, some with continuous records stretching back more than a century. The ML models were trained on this historical archive to learn the complex relationships between precipitation, snowpack, soil moisture, upstream conditions, and eventual streamflow at each location.

The system produces forecasts at 1-to-13-week intervals for every covered streamgage. It specifically targets streamflow drought — periods when rivers and streams drop below normal levels for extended durations. This distinction matters because streamflow drought directly impacts water availability even when rainfall returns to normal, since depleted groundwater and reservoirs take time to recover.

Accuracy and Reliability

Performance varies by forecast horizon, following a predictable gradient. For the first week, the system achieves roughly 75 percent accuracy in predicting severe and extreme drought conditions. Accuracy remains most reliable through the first four to six weeks, then gradually declines as forecasts extend further into the future. By week 13, accuracy drops to approximately 55 percent — still useful for planning, though less suitable for operational decisions.

The accuracy profile reflects a fundamental reality of hydrological forecasting: near-term streamflow is heavily constrained by existing conditions (current water levels, snowpack, soil saturation), while longer-term outcomes depend increasingly on future weather that is inherently less predictable.

Bridging the Forecast Gap

Water managers have long relied on two types of forecasts that leave a gap in the middle. Short-range weather forecasts from the National Weather Service provide detailed precipitation predictions for the next one to seven days. Seasonal water supply outlooks, issued by agencies like the Natural Resources Conservation Service, project total water availability months in advance but lack week-by-week granularity.

River DroughtCast occupies the space between these two tools, offering weekly resolution across the one-to-three-month horizon that is often most critical for operational decisions — when to restrict irrigation, when to implement conservation measures, or when to begin drought contingency planning.

Applications Across Sectors

Agriculture stands to benefit most directly, as farmers make planting and irrigation decisions weeks in advance based on expected water availability. Municipal water suppliers can use the forecasts to anticipate when treatment plants may face reduced intake volumes. Recreation managers can plan for river closures or boating restrictions. Ecologists can prepare for low-flow conditions that stress aquatic ecosystems and fish populations.

The tool is accessible through the USGS National Water Dashboard, which aggregates real-time and forecast data for waterways across the country. By publishing forecasts at individual streamgage locations rather than broad regional summaries, the system provides the spatial specificity that local decision-makers need.

A Data-Rich Foundation

The depth of the USGS streamgage network gives River DroughtCast an unusual advantage. With over 3,000 locations carrying 40 or more years of continuous data — and many stations recording for a century or longer — the training dataset captures a wide range of climatic conditions, including historical droughts that may serve as analogues for future events. This long historical baseline helps the models distinguish between normal seasonal variation and genuinely anomalous conditions that signal emerging drought.