Loading...
Deep learning and aquatic sensing reveal nitrate dynamics in the Mississippi River Basin
Pandit, Aayush
Pandit, Aayush
Citations
Altmetric:
Abstract
High levels of nitrates can cause algal blooms, hypoxia, dead zones, and groundwater contamination, which are detrimental to human health and ecosystem function. Successful modeling of nitrate is challenging because nitrate is highly reactive and follow complex pathways in the environment. As a result, existing process-based models of nitrate have limited accuracy and transferability. Recent advances in machine learning, in conjunction with availability of high frequency sensor data, provide an opportunity to overcome existing limitations. In this study, we developed a continental-scale Long Short-Term Memory (LSTM) model for predicting daily nitrate using data from in situ nitrate sensors (n = 95), gridded meteorological forcings, and remotely sensed catchment attributes. The model showed robust performance with a median Kling Gupta Efficiency of 0.61, positioning the model as state-of-the-art for nitrate prediction. We utilized a domain adaptation technique, involving dimensionality reduction (t-SNE) and clustering (DBSCAN), to allow for transferring our model to novel sites in the Mississippi River Basin (MRB). The model was used to predict four decades (1980 to 2022) of daily flow, nitrate concentration, and nitrate yield from 526 HUC-8 basins (median size: 3500 km2) in the MRB. Model predicted average nitrate yield of 557 kg/km2/yr in MRB with hotspots in the Corn Belt having yields upwards of 5244 kg/km2/yr. Modeling results indicate that only 5% of the surplus nitrate, that is the sum of all nitrate inputs minus crop uptake, in the MRB is exported in a typical year. Wetter years are relatively more productive in the Midwest whereas in the Ohio River Valley, productivity is constant irrespective of annual wetness. We applied eXplainable AI (XAI) to identify dominant drivers of nitrate in MRB which were: sand and silt fraction, wetland extent, crop cover, fertilizer application, and road density. Targeted management for these variables can form the basis for the next generation of effective nitrate reduction strategies in MRB. Together, this study leverages advances in machine learning and XAI to improve our understanding of large scale hydrological and biogeochemical processes driving rising nitrates in rivers of United States.
Description
Date
2024-08-31
Journal Title
Journal ISSN
Volume Title
Publisher
University of Kansas
Collections
Archive Status
This item contains archived web content.
Files
Loading...
Pandit_ku_0099M_19659.pdf
Adobe PDF, 2.47 MB
Research Projects
Organizational Units
Journal Issue
Keywords
Environmental engineering, CONUS, explainable AI, LSTM, Mississippi, Nitrate, Water quality