Response of stream ecosystems to climate change (III): Characterizing and predicting ecologically relevant flow regimes

Background/Question/Methods

Streamflow is expected to change with climate change, but the pattern of these changes will likely vary across the USA. We characterized streamflow regime based on several independent factors describing different ecologically relevant aspects of the streamflow regime.  We also examined trends in streamflow regime variables and how the streamflow regime classes we derived are projected to change with climate change.  We computed 16 streamflow variables that quantify ecologically relevant aspects of streamflow at 1512 reference quality stations from the USGS GAGES dataset. Variable normalization followed by principal component analysis (PCA) with varimax rotation identified the main components of spatial variation in streamflow regime variables. We next applied Ward’s hierarchical clustering method to the principal component axis factors to classify the reference streams into from three to eight flow regime classes. We then assembled a set of 52 watershed and climate attributes potentially associated with flow regime and developed Random Forest models to predict both class membership and 5 of the original flow metrics strongly associated with each PCA. We next used these models with climate projections from a downscaled climate model to predict changes in flow regime likely to occur by 2100. 

Results/Conclusions

Patterns in 16 stream flow variables were spatially consistent with historical long-term patterns in precipitation. The 5 PCA axes characterized differences among sites in low flow, magnitude, flashiness, timing and constancy aspects of the streamflow regime. Classification at the broadest level resulted in three classes characterized by geographic location, timing, low flow and constancy as (1) snow fed streams, (2) rain-fed streams that go dry and (3) rain-fed streams that do not go dry. More finely resolved classes resulted in maps of streamflow classes across the US that also make physical and ecological sense. The Random Forest models predicted classes with out of bag error ranging from 13% for three classes to 28% for eight classes. When future climate estimates were used in the Random Forest models, we observed that some, but not all sites, changed class membership. This observation implies that some sites will be more vulnerable to hydrologic alteration than others. By incorporating predictions of current and future hydrologic regime into species distribution models, we were able to identify those aspects of the flow regime most relevant to stream invertebrates and assess the likely consequences of flow alteration associated with climate change on stream biodiversity.