H056-0020
A Partial Least Squares Modelling Approach for Absorbance-Based Continuous Monitoring of Stream Dissolved Organic Carbon

Wednesday, 9 December 2020
Poster
Christian Gaviria, Memorial University of Newfoundland, Earth Science, St John's, NL, Canada, Allison Myers-Pigg, Pacific Northwest National Laboratory, Marine Sciences Laboratory, Sequim, WA, United States, Alan Roebuck, Memorial University of Newfoundland, Earth Sciences, St John's, NL, Canada, Andrea Skinner, Canadian Forest Service, Narural Resources Canada, Corner Brook, NF, Canada and Susan E Ziegler, Memorial University, Earth Sciences, St John's, NF, Canada
Abstract:
Accurate quantification of dissolved organic carbon (DOC) fluxes from terrestrial to aquatic systems is imperative for understanding landscape and downstream aquatic responses to climate change. High-frequency DOC analysis improves accuracy of reported fluxes, yet is difficult to achieve in remote sampling locations, including boreal forest headwater streams. In-situ spectrophotometers can be used to obtain high-frequency estimates by modelling DOC from chromophoric dissolved organic matter (CDOM) absorbance spectra.

Accuracy of DOC flux measurements can be improved through better representation of changes in the relationship between CDOM and DOC with stream DOC composition as a result of hydrologic connectivity, DOC sources, and processing. Therefore, we use absorbance from a wide range of wavelengths and optical-derived indicators of DOC composition from a discrete dataset capturing baseflow and stormflow conditions from a boreal headwater stream to train partial least squares (PLS) models, integrating these qualities for bulk DOC predictions. PLS was employed here as it can use information from the entire absorbance spectra, which are multicollinear variables, yet collectively contain information about CDOM composition, enhancing the ability to predict DOC accurately.

The performance of each model was tested for turbidity corrections using an independent validation dataset and then applied to high-frequency in-situ absorbance data. Models from turbidity-corrected spectra were found to make adequate DOC predictions during discharge events using high-frequency in-situ data, where stream turbidity varies greatly. The best performing model had a validation root mean square error of 0.71 mg C L-1 and a mean absolute error of 0.59 mg C L-1 over a range of 4-16 mg C L-1, suggesting a low margin of error and no extreme outliers among predictions.

Further model development will involve using high-resolution stream pH and conductivity data that vary during hydrologic events and likely relate to stream DOC or its composition. This will enable further refinement of the models with anticipated improvement in accuracy and precision of DOC predictions.

<< Previous Abstract | Next Abstract >>