High-frequency Saltmarsh Dissolved Organic Carbon Estimates via Machine Learning

Oceanic dissolved organic carbon (DOC) is a major component of the global carbon cycle and the food source for microbes at the base of aquatic foodwebs. Saltmarshes are dynamic and productive ecosystems that may export DOC to the ocean, however the net DOC exchange between saltmarshes and adjacent coastal waters remains poorly constrained due in part to the lack of high-frequency, long-term studies of DOC dynamics. A field campaign was carried out in late summer 2013 through winter 2015 at Groves Creek, a tidal saltmarsh creek on the Georgia coast. Discrete water samples were collected over single tidal cycles every 2 weeks (n = 12-14), totaling nearly 450 samples. Samples were analyzed for dissolved black carbon (DBC), lignin, DOC, nutrients, salinity, temperature, and dissolved oxygen. An in situ spectrophotometer captured colored dissolved organic matter (C-DOM) attenuation spectra (220-730 nm) at high frequency (every 15 minutes) throughout this period. To begin addressing the need for high-frequency, long-term DOC data sets, numerical techniques to infill DOC, DBC, and lignin values between discrete samples at time intervals as short as 15 minutes have been investigated. Techniques such as multiple linear regression (MLR) and partial least squares (PLS) regression demonstrate some success in reproducing observations, but response variable relationships are often better modeled by nonlinear machine learning approaches. Artificial neural networks (ANN), support vector regression (SVR), and random forests (RF) have been tested in this capacity. Additional covariates (including tidal cycles, local meteorological data, and nearby river stage) have been incorporated in an effort to improve high frequency DOC estimates. Results demonstrate the potential for application of machine learning for the creation of high temporal resolution saltmarsh DOC datasets to facilitate greater understanding of DOC fluxes and transport mechanisms in these ecosystems.