AGU24 Preview: Predicting Streamflow Drought at Ungaged Locations Using Deep Learning

Big DataBig Data

Posted: December 4, 2024

AGU24 Preview: Predicting Streamflow Drought at Ungaged Locations Using Deep Learning

"Better drought prediction models for ungaged locations can support water management and planning efforts, helping communities adapt to changing environmental conditions."

At AGU24, the session titled "A Deep Learning Approach for Prediction of Streamflow Drought at Ungaged Locations" will discuss new research on drought prediction in the Colorado River Basin. The American Geophysical Union (AGU) hosts this annual conference to share advancements in earth and space sciences, serving as a key platform for scientists and policymakers.

The study applies long short-term memory (LSTM) networks, a type of deep learning model, to predict streamflow drought conditions. While LSTM models are common in hydrologic predictions, their use in streamflow drought forecasting is less explored. Improving models for ungaged locations could enhance early warning systems, which is crucial in the context of climate change.

Researchers trained LSTM models using 40 years of streamflow data from 411 gages in the Colorado River Basin. The models incorporated static watershed attributes and dynamic inputs from meteorological data, land surface models, and remote sensing datasets. Two approaches were used to predict drought occurrences: fixed drought thresholds calculated over all days and years, and variable drought thresholds that vary seasonally.

Validation at gage locations not included in training showed that models had better performance predicting fixed threshold drought events. Evaluation metrics included Kling-Gupta Efficiency for overall prediction accuracy and Cohen’s Kappa for accuracy under drought conditions. Event-based analysis assessed the models' ability to simulate drought onset, severity, and duration.

The study also examined the impact of incorporating remote sensing data and estimated water use on model predictions over a 2000-2020 period. Preliminary results indicated spatial patterns in model performance and highlighted the influence of water use and streamflow regulation on predictive skill. These findings suggest areas where model improvements could focus to enhance accuracy.

In the face of climate change, where temperature increases and altered precipitation patterns affect water resources, this research holds significance. Better drought prediction models for ungaged locations can support water management and planning efforts, helping communities adapt to changing environmental conditions.

Add to your AGU24 schedule.

Authors:

  • Scott D. Hamshaw
    U.S. Geological Survey
  • Jeremy Diaz
    U.S. Geological Survey, Water Mission Area
  • Phillip Goodling
    U.S. Geological Survey, Maryland-Delaware-D.C. Water Science Center
  • Roy R. Sando
    USGS Wyoming-Montana Water Science Center
  • Ryan McShane
    USGS Wyoming-Montana Water Science Center
  • William D. Watkins
    USGS Integrated Information Dissemination Division
  • Caelan Simeone
    U.S. Geological Survey
  • Elaheh White
    U.S. Geological Survey
  • Jacob A. Zwart
    USGS Integrated Information Dissemination Division, Data Science Branch
  • John Christopher Hammond
    U.S. Geological Survey