PNNL

Abstract

The influence of coastal ecosystems on global greenhouse gas (GHG) budgets and how they will respond to increasing inundation and salinization is poorly constrained. Here we developed and applied an uncertainty quantification (UQ) and machine learning (ML) framework to identify and rank the most influential processes, properties, and conditions controlling methane behavior in a freshwater floodplain that is responding to recently restored seawater inundation. Our multiyear dataset includes tidal creek and floodplain porewater observations of water level; salinity; pH; temperature; the concentration of dissolved oxygen (DO), dissolved organic carbon (DOC), and total dissolved nitrogen (TDN); the partial pressure of carbon dioxide (pCO2), nitrous oxide (pN2O), and methane (pCH4); and the stable isotopic composition of methane (d13CH4). Topography, soil porosity, hydraulic conductivity, and water retention parameters were also used for UQ analysis of a previously developed 3D variably saturated flow and transport floodplain model. Principal component analysis indicated that porewater level and salinity are the most significant predictors of overall biogeochemical variability. Random forest ML identified d13CH4, DO, pCO2 and pN2O as the most important features with porewater CH4 as the target variable. Our ML analysis also emphasized the importance of water level and salinity, which were positively and negatively correlated with porewater pCH4, respectively. UQ analysis indicated that ~80% of the total variability in hourly water levels and ~60% of the total variability in hourly salinity can be explained by permeability, creek water level, and two van Genuchten water retention function parameters, alpha and m. These findings are a step forward for better understanding and representation of coastal ecosystem methane behavior in local to global scale Earth system models.