Assessing Environmental Flow Violations in US Rivers: Exploring the Impact of Human Activities and Climate Change Using Machine Learning

Abstract

Human activities and climate change have significantly altered natural river flow regimes, adversely affecting ecosystems globally. This study uses the GAGES-II dataset (1981–2016) to examine relationships between environmental flow violations (EFVs) at stream gauges in the United States and characteristics of their upstream basins over two periods (1981–1998 and 1999–2016). The Variable Minimum Flow (VMF) approach was used to estimate environmental flows based on natural flow conditions for each hydrometric station. Basin characteristics were categorized into climate variability (precipitation and temperature changes), water withdrawals and geographical attributes. The variables representing these basin characteristics were then used as predictors or inputs to the random forest (RF) machine learning algorithm to analyse and predict temporal and spatial variations of EFVs based on the observed variations of predictors. The results of this study showed that approximately 55% of the 1625 stream gauges analysed exhibited EFV percentages exceeding 80% in both periods. Mapping these stations highlighted critical areas requiring intervention. Temporal EFV changes were assessed by comparing the two periods, and K-means clustering grouped stations into two clusters with distinct geographical and climatic characteristics. The RF models trained for prediction of average EFV differences between the two periods showed acceptable accuracy, with Kling–Gupta efficiency (KGE) values ranging from 0.5 to 0.7, although accuracy was higher in the stations in Cluster 2 covering more arid areas in the southwest. The feature importance analysis revealed that the dam storage-to-streamflow ratio (DSSR) and precipitation were key factors in humid areas (Cluster 1), while water withdrawal and temperature were more significant in arid areas (Cluster 2). A noticeable temporal shift was also observed as the relative importance of DSSR (water withdrawal) diminished (intensified) overtime. Given the large dataset and the diversity of factors considered, this methodology can be applied to the rest of the streamflow gauges in the United States, providing valuable insights for water resource management and environmental policy making.