Clustering Watersheds Based On Performance of Deep Learning Streamflow Drought Models In Colorado River Basin

Hydrological drought is a very complex phenomenon and defies a clear consensus on its definition. In earth system and environmental sciences, multiple types of droughts have been identified based on different attributes such as duration, thresholds, and impacts/stressors. Streamflow is an important and direct indicator of hydrological drought. Therefore, predicting streamflow drought conditions accurately is of significant interest for the hydrological forecasting community. As part of a regional drought early warning system project, the U.S. Geological Survey is developing models to predict streamflow drought conditions at lead times of 0, 7, 14, 30, and 60 days using a long short-term memory (LSTM) deep learning model. Models are being developed using USGS streamflow data from 425 stream gages throughout the Colorado River Basin and the surrounding area. Model performance is being characterized using a variety of performance measures such as Nash Sutcliffe efficiency (NSE) and Kling-Gupta efficiency (KGE) on log streamflow, low flow bias. Model performance ranges from reliable to no predictive skill across watersheds depending on selection of streamflow drought definitions (using fixed or daily variable thresholds over a year). To support this early warning system project, we use a data set comprised of preliminary model performance measures for predictions at the 425 gages within the Colorado river basin along with compiled watershed attributes associated with each gage. We apply machine learning clustering algorithms to categorize watersheds based on model performance metrics. We then investigate which hydrological signatures and watershed attributes are associated with different groups of model performance. Results from this study will assist modelers in determining watersheds likely to be associated with categories of expected model performance. Additionally, insights into the factors that influence the predictive ability of the LSTM model will be used to identify strategies for further forecasting model improvements.