Advancing Science to Better Characterize Drought and Groundwater-Driven Low-Flow Conditions in NOAA and USGS National-Scale Models
Research Team Members
Objective:
This project addresses the challenge of accurately forecasting streamflow during drought by developing open-source tools and machine learning models that improve low-flow predictions in groundwater-influenced streams, supporting more reliable water management and advancing CIROH operational capabilities.
Approach:
Our approach focuses on improving low-flow streamflow forecasts by developing scalable machine learning tools and hydrologic algorithms that quantify groundwater influence. We have released a new open-source Python package for digital baseflow separation, incorporating both standard and novel methods, supported by documentation and example notebooks. To identify baseflow-dominant (BFD) periods, we trained a random forest classifier with >90% accuracy using labeled streamflow data from over 200 U.S. gages. This classifier is now being applied CONUS-wide to identify regions where groundwater dominates flow, and to assess biases in NextGen model outputs. We are also translating the USGS-developed Baseflow Separation (BFS) state-space model into Python for full integration with the NWM. The BFS model will be tested and refined for operational low-flow prediction during BFD periods. Concurrently, we are building a national streamflow bias analysis framework, clustering long-term baseflow signatures and linking them to watershed features using interpretable ML. All tools are being prepared for integration into the HydroFabric/BMI framework, positioning them for use by NOAA, USGS, and broader hydrologic communities.
Impact:
Our project will improve the accuracy of low-flow forecasts by integrating open-source baseflow tools and machine learning models into NOAA’s NextGen system, reducing streamflow prediction bias in groundwater-influenced regions and supporting more effective drought management.Abstract:
This project improves the forecasting of streamflow during drought by developing new tools to identify and model the influence of groundwater on rivers and streams across the United States. It brings together machine learning, hydrology, and national-scale data to better predict how much water will be flowing when rainfall is low.
Most national models focus on predicting floods, but accurate low-flow forecasts are just as critical—for water supply, navigation, ecological protection, and infrastructure management—especially during droughts. Many streams are fed by groundwater, yet current models often miss or misrepresent this contribution, leading to significant prediction errors.
By improving low-flow forecasts and accounting for groundwater effects, this work supports smarter water management, reduces drought-related risk, and strengthens the nation’s water modeling capabilities.
Product and Improvements
Today:
• Groundwater influence on streamflow is not explicitly modeled in most national-scale systems.
• Manual or outdated methods are used for baseflow separation.
• Low-flow prediction bias is common in ungaged or groundwater-driven watersheds.
This project will:
• Deliver a Python-based tool for automated digital baseflow separation and classification.
• Translate and integrate a state-space baseflow model (BFS) into NOAA’s NextGen system.
• Use machine learning to identify baseflow-dominant periods with >90% accuracy.
• Provide national-scale bias detection and correction for groundwater-influenced stream segments.
Impact
Our tools will reduce low-flow prediction errors in the NextGen model, automate baseflow estimation at scale, and provide water managers and agencies with accurate, actionable forecasts during drought conditions.