The stakes are higher than they might first appear. Evapotranspiration, the combined flux of transpiration from vegetation and evaporation from soil, water, and other surfaces, is the dominant way water leaves a wetland. Separating it into its two components matters because transpiration is tightly linked to photosynthesis, the process by which ecosystems pull carbon dioxide out of the air. Wetlands, especially peatlands, are among the planet’s most important carbon sinks and play an outsized role in regulating regional hydrology and climate. If scientists cannot reliably say how much water moves through plants versus straight off the surface, they cannot accurately model how these ecosystems will respond to drought, warming, or changes in water management.

The difficulty in wetlands stems largely from the moss itself. Mosses photosynthesize and contribute to ecosystem carbon uptake, yet they contribute nothing to transpiration, which is by definition a vascular-plant process. That decoupling violates a core assumption of several popular partitioning methods, which presume that photosynthesis and transpiration share identical sources and sinks. Mosses also behave unlike vascular plants in other ways: they cannot actively regulate water transport, relying instead on passive capillary rise, and they can store water equivalent to 300 to 500 percent of their dry weight. When dry conditions limit capillary supply, moss cells rapidly equilibrate with the surrounding air, suppressing both evaporation and photosynthesis even while the atmosphere’s evaporative demand remains high. These traits can scramble the environmental relationships that partitioning algorithms depend on.

To find out how badly these violations matter, the team assembled data from four moss-covered wetlands: three sites in the Canadian Rocky Mountains, named Sibbald, Burstall, and Bonsai, measured during the 2021 growing season, and a forested peatland near Fort McMurray, Alberta, called Poplar, measured in 2013. The sites spanned a remarkable range of conditions. Canopy heights ranged from 0.8 meters at Bonsai to 4.2 meters at Poplar, leaf area indices varied more than threefold, and hydrology differed sharply: Sibbald and Bonsai stayed non-flooded throughout, Burstall transitioned to flooded conditions in mid-August 2021, and Poplar remained flooded for nearly the entire study period.

The ten methods evaluated fell into three methodological groups. The first exploits high-frequency eddy covariance data, using rapid fluctuations in carbon dioxide and water vapor to statistically separate ground-level and canopy-level fluxes. The second group relies on ecosystem carbon-water coupling, inferring transpiration from the relationship between photosynthesis and water use, often built on theories of optimal stomatal behavior. The third group applies machine learning, training models to predict evaporation from nighttime measurements, when transpiration is assumed negligible, and then subtracting that estimate from total evapotranspiration. Each group carries its own vulnerabilities in a mossy wetland, from the possibility that moss photosynthesis flips the surface carbon signal to the fact that nighttime transpiration, though often small, is not always zero.

Against what benchmark should the methods be judged? The researchers compiled independent, measurement-based estimates of the transpiration-to-evapotranspiration ratio using micro-lysimeters, leaf porometers, flux chambers, and sap flow sensors, reconstructed into continuous daily time series with a site-calibrated Shuttleworth-Wallace model that explicitly accounted for moss and litter effects on ground evaporation. The calibrated model reproduced measured latent heat flux and ground evaporation closely, with coefficients of determination of 0.83 and 0.85 respectively. Because even careful field reconstructions carry uncertainty, the team also evaluated every method against the ensemble mean of all ten approaches, a second reference that averages out method-specific errors.

The results dismantled any hope of a universal winner. No single method outperformed the others across all four sites, and no methodological group demonstrated a clear overall advantage. Most methods performed no better than simply predicting the observed mean, as reflected in predominantly negative or near-zero Nash-Sutcliffe efficiency values. Among the high-frequency methods, mean transpiration estimates exceeded the measurement-based values by 18.4 percent on average, with individual methods diverging by more than 26 percent. The carbon-water coupling ensemble came closest in magnitude, exceeding the measurement-based mean by just 0.32 percent, yet its individual members swung wildly, with one method underestimating the ratio by nearly 168 percent relative to measurements.

Flooding proved especially revealing. When Burstall transitioned to flooded conditions in August 2021, the measurement-based transpiration ratio dropped sharply, but most methods barely registered the change. The single exception was the TEA algorithm, which estimates a temporally varying water-use efficiency with a random forest model, allowing it to adapt as the relationship between carbon uptake and water flux shifted after inundation. The machine learning framework of Eichelmann and colleagues, which had performed well in California freshwater marshes, failed to capture the transition here, likely because its models were trained under non-flooded conditions and the environmental relationships they learned no longer applied once the site was inundated. Yet at Poplar, continuously flooded and densely forested, the machine learning estimates stayed close to the references, hinting that sparse, waterlogged vegetation poses particular difficulties.

Perhaps the most unsettling finding concerned the benchmarks themselves. The measurement-based time series and the all-method ensemble showed markedly different correlation patterns with the evaluated methods, and the two references were not significantly correlated with each other at any site. In other words, the verdict on any given method depends heavily on which reference you trust, exposing deep uncertainty in how partitioning performance is benchmarked. The authors also found that ecosystem heterogeneity mattered: at Sibbald, with its multi-layered shrub canopy, large gaps, and hummock-hollow microtopography, methods agreed least with one another, while at the structurally homogeneous Poplar forest they agreed most.

The practical takeaway is not that existing methods are useless, but that none should be trusted alone. The authors recommend applying multiple approaches from all three groups concurrently wherever possible, using disagreement among methods to diagnose uncertainty, and paying particular attention to how water-use efficiency is parameterized, which emerged as a major source of inter-method conflict. More fundamentally, the study makes a case for developing wetland-specific partitioning approaches that explicitly represent moss-mediated carbon uptake, dynamic carbon-water coupling, and the transitions between unsaturated and flooded states that define so many of these ecosystems. As climate change alters wetland hydrology worldwide, the humble moss carpet, it turns out, is a harder scientific problem than anyone’s equations anticipated.

Subject of Research: Evaluation of eddy covariance-based evapotranspiration partitioning methods in moss-covered wetland ecosystems

Article Title: Technical note: How well do evapotranspiration partitioning approaches perform in moss-covered wetlands?

Article References: Technical note: How well do evapotranspiration partitioning approaches perform in moss-covered wetlands?. (n.d.). https://doi.org/10.5194/bg-23-7067-2026

Image Credits: AI Generated

DOI: 10.5194/bg-23-7067-2026

Keywords: evapotranspiration, wetlands, peatlands, moss, eddy covariance, transpiration, evaporation, machine learning, carbon-water coupling, hydrology, biogeosciences, water-use efficiency