A deep learning-based novel approach to generate continuous daily stream nitrate concentration for nitrate...

Area of Science:

• Environmental Science
• Hydrology
• Data Science

Background:

• Traditional nitrate monitoring relies on infrequent lab analysis of water samples.
• High-frequency nitrate data collection using optical sensors is increasingly common.
• Data-driven models can potentially fill temporal and spatial data gaps in nitrate monitoring.

Purpose of the Study:

• To develop and evaluate a data-driven model for estimating continuous daily stream nitrate concentrations.
• To assess the model's performance in capturing nitrate variability and trends.
• To provide a framework for generating nitrate data in data-limited regions.

Main Methods:

• A Long Short-Term Memory (LSTM) model framework was developed.
• The model was trained and tested using high-frequency sensor-based nitrate data from 42 sites in Iowa, USA.
• Model performance was evaluated using Nash-Sutcliffe efficiency (NSE) and Root Mean Square Error (RMSE), and seasonal performance was assessed.

Main Results:

• The LSTM model achieved a median NSE of 0.75 and RMSE of 1.53 mg/L, demonstrating unprecedented performance.
• 50% of sites had NSE > 0.75, and 76% had NSE > 0.50.
• Neighboring site nitrate concentration was a key factor, and the model performed well in summer and fall. Four dominant nitrate transport patterns were identified.

Conclusions:

• The developed LSTM model framework successfully generates continuous daily stream nitrate data, addressing limitations of traditional monitoring.
• This approach provides valuable, high-resolution nitrate data for data-limited areas, aiding watershed management and conservation planning.
• The model's ability to capture temporal variability and identify transport patterns offers significant insights for ecosystem sustainability.