Methane emissions from subtropical wetlands: An evaluation of the role of data filtering on annual methane budgets

Abstract

Widespread adoption of eddy covariance (EC) methods for methane (CH4) flux measurement has led to increased availability of continuous high-frequency CH4 data. However, unreliable data frequently occur during periods of atmospheric stability, rain or instrument malfunction, requiring filtering prior to subsequent analyses. While procedures for assessing CO2 have matured, processes to filter and gap-fill CH4 data are less studied, as their range and controls are not as well-understood. Moreover, publications often fail to describe procedures for data processing and filtering. Our primary objective was to study effects of common filtering thresholds and provide insight on how size and timing of gaps produced by filtering affect CH4 budgets. We utilized 4 years of data from two freshwater wetlands under the same climate regime but different hydroperiods. We applied friction velocity (U*) and signal strength filtering treatments to isolate site-specific effects and evaluate impacts of filtering on subsequent gap-filling via Random Forests (RF). We also tested sensitivity of results to predictor datasets with an “unrestricted predictors model” (using all possible predictors regardless of gaps), versus a “restricted predictors model” (using gap-filled predictors with no missing values). Depending on filtering treatment, 7 - 50% of CH4 data were removed over the study period. Using higher signal strength thresholds introduced more small gaps. U* filtering created small gaps (mostly nighttime), and corresponding annual budget estimates were generally different from those filtered solely on signal strength but with higher uncertainty, especially at the long-hydroperiod site. Regardless of filtering method, RF models using unrestricted predictors identified 2- to 32-day average CH4 flux as primary predictors, whereas heat and latent energy were most important when predictors were restricted. Although filtering may have less impact on CH4 budgets than selection and pre-processing of predictor variables, it can significantly impact uncertainty and should be considered in data curation protocols.