Machine learning analysis of streamflow recession patterns across climates in the contiguous United States

Abstract

Streamflow recession analysis has received substantial attention over the past decades. Such analysis strives to understand what controls low flow dynamics. Most studies use a power law relationship between streamflow and its time derivative and these variables are typically plotted on log-log scale to reveal the parameters of the power law. The exponent shows up in such plots as the slope of the central tendency of the point cloud. Recently, this procedure has been questioned, as individual recessions typically have slopes that are higher than the slope of the point cloud. To reconcile these two viewpoints, a machine learning method for catchment-scale recession analysis has been introduced that successfully integrates point clouds and individual event trajectories. The machine learning model demonstrated the existence of an attractor in phase space to which long individual recessions converge. In order to test the limits of this new approach and see how it performs under different circumstances, we applied the same methodology to a selection of catchments representing the variation in climate across the contiguous United States (CONUS). Several catchments were found to have potentially valid attractors in the modeled point cloud. Models for catchments with out-of-phase precipitation with significant snow cover struggled to capture the hysteresis in the observed point cloud, while still sometimes performing well on individual summer events. Arid, highly seasonal catchments performed poorly in the analysis. This work reveals the importance of out-of-phase precipitation and snow fraction on the success of the methodology, provides a reference point for the variety of behaviors that might be expected when conducting recession analysis on a range of climates, and further emphasizes the importance of careful recession selection criteria, especially in catchments outside of the standard humid, mild seasonality climates traditionally focused on in such recession analysis work.