Sensor Informed Predictive Model for Total Organic Carbon and Nutrients on the Upper Yampa River

Accurate, high-frequency monitoring of total organic
carbon (TOC)
and total nitrogen (TN) in rivers remains challenging but is critical
for water quality management. This study presents a machine-learning
framework that integrates in situ fluorescent dissolved organic matter
(FDOM) sensor data with high-resolution land use/land cover (LULC)
information to predict TOC and TN concentrations in the Upper Yampa
River, Colorado. Gradient boosting machine models were trained on
data collected from July 2023 to July 2024 (excluding winter months),
using FDOM measurements and LULC inputs derived from both 30 m and
10 cm resolution imagery. The optimal TOC model achieved a root-mean-square
error below 0.7 mg/L and prediction errors under 8% using spatial-temporal
cross-validation, with very high-resolution LULC data substantially
improving performance relative to standard-resolution inputs. A complementary
classification pipeline predicted TN concentration categorieslow
(<0.1 mg/L), medium (0.1–0.45 mg/L), and high (0.45 mg/L)with
an overall accuracy of approximately 62%. This integrated sensorML
approach enables nearreal-time nutrient estimation, supporting
adaptive river monitoring and regulatory compliance across diverse
watershed settings.