Predicting water quality across US watersheds using a time-series deep learning model

Abstract:

Stream water quality is driven by the physical and chemical processes within watersheds and is crucial to drinking-water security and ecosystem health. However, traditional process-based reactive transport simulations are challenging and remain computationally prohibitive at the watershed scale. Here we used Long-short term memory (LSTM), a widely-used time-series deep learning model, to predict water quality at basin outlets, with atmospheric forcings and geographical features of watersheds as inputs. Trained on over 1000 basins across the US, the framework can predict the temporal variations of concentration of multiple chemical species, including both rock-derived and biologically cycled elements, with promising accuracy. However, we found that the LSTM model could better predict rock-derived species (e.g., calcium and magnesium) compared to nutrient variables (e.g. nitrate and potassium), which suggests that the model is better able to capture deep weathering processes compared to biologically cycled and/or stream-moderated constituents. In addition, precipitation inputs can account for more than 50% of the variance of nutrient variables like sulfate and nitrate in the northeastern US. By examing the causal relationship between watershed characteristics and model performance, we can extract insights and hypotheses about biogeochemical and hydrological processes within watersheds. We argue that the data-driven models could advance our understanding of watershed functions and potentially improve the predictive capability of watershed models, both of which are critical for addressing the growing crisis of deteriorating water quality.