Machine Learning Analysis of Streamflow Recession Patterns Across Climates in the Contiguous United States

Streamflow recession analysis aims to understand the controls on low‐flow dynamics in catchments. Traditionally, this involves analyzing baseflow Q and its time derivative dQ⁄dt on log‐log plots, often revealing power‐law relationships. The slope of these power laws has been interpreted through hydraulic groundwater theory. However, recent studies challenge this interpretation, noting discrepancies between individual recession slopes and the aggregate power‐law slope. To address this, Kim et al. (2023), https://doi.org/10.1029/2022wr032690 introduced a machine learning (ML) method for recession analysis, which explains the spread of point clouds and predicts individual event trajectories. This method reveals an attractor in phase space, suggesting individual recessions converge to a common trajectory, independent of past flows. Unlike traditional power‐law fits, the ML approach offers a more objective framework for analyzing recession dynamics. We applied this method to catchments across the contiguous United States (CONUS), chosen to reflect climate variability. Results show that in some catchments, attractors align with power‐law fits, while in others, they deviate significantly, suggesting unique low‐flow dynamics unexplained by hydraulic theory. In certain cases, attractors exhibit non‐linear patterns in log‐log space, highlighting hysteresis and climate‐driven variability. Our findings provide insights into the diversity of recession behaviors across climates, moving beyond the conventional focus on humid, mild‐seasonality catchments. The ML method establishes a foundation for interpreting complex low‐flow dynamics, offering a broader perspective on how climate influences catchment storage and release processes.