You are here:
Towards national mapping of aquatic condition (II): Predicting the probable biological condition of USA streams and rivers
Citation:
Hill, R., M. Weber, S. Leibowitz, AND Tony Olsen. Towards national mapping of aquatic condition (II): Predicting the probable biological condition of USA streams and rivers. 9th International Association of Landscape Ecology World Congress, Portland, OR, July 05 - 10, 2015.
Impact/Purpose:
A national map of stream resources and their prBC would provide a powerful tool for the USEPA Office of Water to prioritize sampling, monitoring, and restoration. We will present a model that predicts the probable (pr) biological condition (BC) of ~5.4 million km of streams within the conterminous USA. To model BC, we linked 1,883 streams that were previously determined by the EPA’s National Rivers and Stream Assessment to be in ‘good’ or ‘poor’ condition to a spatially extensive dataset of watershed-level landscape metrics: The Stream-Catchment (StreamCat) Dataset. The StreamCat Dataset will be described in a presentation by Marc Weber at the same conference (Towards national mapping of aquatic condition (I): The Stream-Catchment (StreamCat) Dataset). These paired presentations will introduce the StreamCat Dataset and our model of USA-wide BC to potential users.
Description:
The US EPA’s National River and Stream Assessment (NRSA) uses spatially balanced sampling to estimate the proportion of streams within the conterminous US (CONUS) that deviate from least-disturbed biological condition (BC). These assessments do not infer BC at un-sampled streams, nor the anthropogenic stressors that degrade BC. A national map of probable (pr) BC could provide an important tool for prioritizing monitoring and restoration of streams. We used random forest modeling and data from 1,883 NRSA sample sites that were previously determined to have ‘good’ or ‘poor’ BC to predict prBC for ~5.4 million km of stream. The NRSA sites were linked to the Stream-Catchment (StreamCat) Dataset, which contains >100 natural and anthropogenic landscape metrics for ~2.7 million watersheds across the CONUS. prBC was best predicted (70% correctly classified) by 3 natural (elevation, air temperature, and sand content of soils) and 4 human-altered (% riparian naturalness, population density, and % of watershed composed of forest cover or agriculture) landscape metrics. We applied the model to the StreamCat dataset to predict prBC nationally. The national map of prBC provided a unique assessment of model performance. Specifically, lower prBC was consistent with large-scale patterns of human-related land use. However, maps of local prBC were sometimes unrealistic, suggesting that predictions could be improved with regional, rather than national, models. Future work will seek to improve the national map of prBC for streams. In addition, our data and modeling framework will soon be extended to 356,044 lakes.