Abstract
Anthropogenic nutrient inputs have led to nutrient enrichment in many waterbodies worldwide, including Chesapeake Bay (USA). River water quality integrates the spatial and temporal changes of watersheds and forms the foundation for disentangling the effects of anthropogenic inputs. We demonstrate with the Chesapeake Bay Non-Tidal Monitoring Network that machine learning approaches - i.e., hierarchical clustering and random forest (RF) classification - can be combined to better understand the regional patterns and drivers of total nitrogen (TN) trends in large monitoring networks, resulting in information useful for watershed management. Cluster analysis revealed regional patterns of short-term TN trends (2007-2018) and categorized the stations into three distinct trend clusters, namely, V-shape (n = 23), monotonic decline (n = 35), and monotonic increase (n = 26). RF models identified regional drivers of TN trend clusters by quantifying the effects of watershed characteristics (land use, geology, physiography) and major N sources on the trend clusters. Results provide encouraging evidence that improved agricultural nutrient management has resulted in declines in agricultural nonpoint sources, which in turn contributed to water-quality improvement in our period of analysis. Moreover, water-quality improvements are more likely in watersheds underlain by carbonate rocks, reflecting the relatively quick groundwater transport of this terrain. By contrast, water-quality improvements are less likely in Coastal Plain watersheds, reflecting the effect of legacy N in groundwater. Notably, results show degrading trends in forested watersheds, suggesting new and/or remobilized sources that may compromise management efforts. Finally, the developed RF models were used to predict TN trend clusters for the entire Chesapeake Bay watershed at the fine scale of river segments (n = 979), providing fine spatial information that can facilitate targeted watershed management, including unmonitored areas. More broadly, this combined use of clustering and classification approaches can be applied to other regional monitoring networks to address similar water-quality questions.