Related Experiment Video
Updated: Jul 5, 2026
Understanding Dissolved Organic Matter Biogeochemistry Through In Situ Nutrient Manipulations in Stream Ecosystems
Published on: October 29, 2016
Analyzing the deep learning approach-based modeling framework to understand the critical environmental factors of
Gourab Kumer Saha1, Farshid Rahmani2, Ashok Jacob3
1Department of Agricultural and Biological Engineering, The Pennsylvania State University, University Park, United States; Modeling Support Office, California Department of Water Resources, Sacramento, United States.
Abstract:
Continuous stream nitrate data is critical for understanding nutrient dynamics and informing management practices. Recent studies applied deep learning (DL) models on high-frequency nitrate sensor data to predict continuous stream nitrate concentrations during data gaps and at data-limited locations. However, limited knowledge exists about environmental factors that influence DL models' prediction performance. Understanding these factors is crucial for improving model performance, optimizing monitoring network design, and applying the framework to other regions. The study hypothesizes that similar stream nitrate concentration-discharge (c-Q) relationships and similar temporal variability of key environmental factors between target low-frequency and input high-frequency sites improve DL model performance. The c-Q relationship similarities and critical environmental factors are identified using a previously developed DL modeling framework that predicts daily stream nitrate concentrations at low-frequency nitrate monitoring locations in Iowa. Analysis of c-Q patterns at input high-frequency sites revealed that those contributing to accurate predictions consistently exhibited similar dominant c-Q patterns to the target sites. First-order sensitivity analysis of DL model inputs showed that day length (seasonality representation), high-frequency nitrate concentration, and stream discharges of both input and target sites are critical environmental factors for accurate daily nitrate predictions. Dynamic Time Warping (DTW)-based similarity analysis confirmed that similarity in critical environmental factors of input and target nitrate monitoring sites contributes to the model performance, while spatial distance between input and target sites had limited influence. The findings support the replicability and scalability of the developed deep learning framework, underscoring its potential as a valuable tool for stream nitrate monitoring.
