A Continental-scale Deep Learning Model for Total Phosphorus Reveals Progress Toward Water Quality Goals over the Past 40 Years in the United States
Abstract
Phosphorus (P) is vital for human life and agricultural production, yet excess P in the environment has led to global water quality degradation. Understanding and quantifying P concentrations and fluxes in rivers rely on sparse observational P data. For example, U.S. Geological Survey monitoring stations have an average of 12 to 25 data points over 1 to 3 years with significant temporal gaps. Here we overcome such data limitations by training a continental-scale deep learning model for 513 basins across the U.S. and reconstructing their continuous daily total phosphorus (TP) records over the past 40 years. The trained model achieved a median Nash-Sutcliffe Efficiency of 0.83 and 0.94 for TP concentration and flux, respectively. The 40-year trend analysis shows a dominant decline in TP concentration for 62% of basins, while only 34% exhibited decreasing TP fluxes. On average, TP in urban basins has declined whereas TP in agricultural basins has increased with a noticeable deterioration of water quality in the Midwest. Change point analysis, which detects whether and when a change has taken place in time-series data, shows that TP from urban basins has the earliest change point (1994), followed by basins with mixed land uses (1997) and agricultural basins (2002). This reflects significant efforts to improve water quality in the U.S., such as policy aimed at reducing P point sources (e.g., from urban wastewater discharge) following the Clean Water Act of 1972 and management measures to control non-point P sources (e.g., agricultural runoff) in the 1990s. Further, considering the timing of delivery of P from source areas to rivers, agricultural P sources (e.g., runoff from land application of fertilizers and manure) often have longer lag times than urban P sources (which have a higher surface runoff proportion and direct pipeline discharges to rivers). Overall, our deep learning model reveals that policy and management strategies to decrease TP loadings to surface waters have had mixed success in these basins across the U.S. Further, results highlight the promise of deep learning techniques to improve predictability of riverine water quality.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H43E..06Z