You are here:
A CONUS-scale long short-term memory model for prediction of nitrate dynamics
Citation:
Pandit, A., C. Lane, J. Christensen, H. Golden, AND A. Husic. A CONUS-scale long short-term memory model for prediction of nitrate dynamics. AGU 2023 Fall Meeting, San Francisco, CA, December 10 - 15, 2023.
Impact/Purpose:
Modeling nitrate is a challenging task due to the reactive nature of nitrate in the environment, the variable sourcing and residence times in the landscape, and the limitations of numerical models to resolve physical processes. Recent developments in interpretable machine learning have opened the possibility of overcoming these challenges to better understand the fate of nitrate in rivers. In this study, we train a long short-term memory (LSTM) model for daily prediction of nitrate on a combination of data from grab samples and high-frequency aquatic sensors, from 1980 to 2022, at 148 sites spanning climatic, geomorphologic, and anthropogenic settings.
Description:
Rising nitrate concentrations have polluted waterways and led to the development of harmful algae blooms and other water and human health hazards. Treating nutrient-rich waters is costly and amounts to approximately $4 billion per year for the United States, necessitating an understanding of what causes water quality degradation. Modeling nitrate is a challenging task due to the reactive nature of nitrate in the environment, the variable sourcing and residence times in the landscape, and the limitations of numerical models to resolve physical processes. Recent developments in interpretable machine learning have opened the possibility of overcoming these challenges to better understand the fate of nitrate in rivers. In this study, we train a long short-term memory (LSTM) model for daily prediction of nitrate on a combination of data from grab samples and high-frequency aquatic sensors, from 1980 to 2022, at 148 sites spanning climatic, geomorphologic, and anthropogenic settings. Potential catchment-scale drivers of nitrate, such as land use, precipitation, and soils, are retrieved from CAMELS and GAGES-II datasets. The LSTM model does an exceptional job at simulating daily evolution of nitrate with median statistics of approximately 0.5 R2 and KGE. The highest performing sites are located around the agricultural Midwest (R2 and KGE > 0.7), whereas the poorest performers are in data-sparse regions of the intercontinental West (R2 and KGE < 0.3) where sensor data is not available. We present results from the first continental-scale LSTM model of nitrate and highlight the necessity of sensor data for accurate simulations. Future work will involve applying interpretable metrics to the LSTM outputs to identify human and natural drivers of stream nitrate.