Streamflow recession analysis, relating discharge Q and its time rate of change −dQ/dt, has been widely used to understand catchment-scale flow dynamics. However, data points in the recession plot, the plot of −dQ/dt versus Q, typically form a wide point cloud due to noise and hysteresis in the storage-discharge relationship, and it is still unclear what information we can extract from the data points and how to understand the information. There seem to be two contrasting approaches to interpret the flow recession plot. One emphasizes the importance of analyzing the ensembles of many recessions (i.e., the lower envelope or a measure of central tendency), and the other highlights the importance of the event-scale analysis and questions the meaning of the lower envelope and the measure of central tendency. In this study, we examine if those approaches can be combined. We utilize a machine learning tool to capture the point cloud using the past trajectory of discharge. Our results show that most of the data points can be captured using 5 days of past discharge. While analyzing the machine learning model structure and the trained parameters is a daunting task, we show that we can learn the catchment-scale flow recession dynamics from what the machine learned. We analyze patterns learned by the machine and explain and hypothesize why the machine learned those characteristics. The hysteresis in the plot mainly occurs during the early time dynamics, and the flow recession dynamics eventually converge to an attractor in the plot, which represents the master recession curve. We also illustrate that a hysteretic storage-discharge relationship can be estimated based on the attractor.
