Nocturnal Evapotranspiration in Eddy-Covariance Records from Three Co-Located Ecosystems in the Southeastern U.S.: Implications for Annual Fluxes

Abstract

Nocturnal evapotranspiration (ETN) is often assumed to be negligible in terrestrial ecosystems, reflecting the common assumption that plant stomata close at night to prevent water loss from transpiration. However, recent evidence across a wide range of species and climate conditions suggests that significant transpiration occurs at night, frustrating efforts to estimate total annual evapotranspiration (ET) from conventional methods such as the eddy-covariance technique. Here, the magnitude and variability of ETN is explored in multiple years of eddy-covariance measurements from three adjacent ecosystems in the Southeastern U.S.: an old grass field, a planted pine forest, and a late-successional hardwood forest. After removing unreliable data points collected during periods of insufficient turbulence, observed ETN averaged 8–9% of mean daytime evapotranspiration (ETD). ETN was driven primarily by wind speed and vapor pressure deficit and, in the two forested ecosystems, a qualitative analysis suggests a significant contribution from nocturnal transpiration. To gapfill missing data, we investigated several methodologies, including process-based multiple non-linear regression, relationships between daytime and nighttime ET fluxes, marginal distribution sampling, and multiple imputation. The utility of the gapfilling procedures was assessed by comparing simulated fluxes to reliably measured fluxes using randomly generated gaps in the data records, and by examining annual sums of ET from the different gapfilling techniques. The choice of gapfilling methodology had a significant impact on estimates of annual ecosystem water use and, in the most extreme cases, altered the annual estimate of ET by over 100 mm year−1, or ca. 15%. While no single gapfiling methodology appeared superior for treating data from all three sites, marginal distribution sampling generally performed well, producing flux estimates with a site average bias error of <10%, and a mean absolute error close to the random measurement error of the dataset (12.2 and 9.8 W m−2, respectively).