A Remote Sensing Technique to Upscale Methane Emission Flux in a Subtropical Peatland

Abstract

Quantification of methane (CH 4) gas emission from peat is critical to understand CH 4 budget from natural wetlands under a climate warming scenario. Previous studies have focused on prediction and mapping of CH 4 emission flux using process-based models, while application of statistical-empirical models for upscaling spatially sparse in situ measurements is scarce. In this study, we developed an empirical remote sensing upscaling approach to estimate CH 4 emission flux in the Everglades using limited in situ point-based CH 4 emission flux measurements and Landsat data during 2013-2018. We spatially and temporally linked in situ data with Landsat surface reflectance based on temporally composite data sets and developed an object-based machine learning framework to model and map CH 4 emission flux. An ensemble analysis of two machine learning models, k-Nearest Neighbor (k-NN) and Support Vector Machine (SVM), shows that the upscaling approach is promising for predicting CH 4 emission flux with a R 2 of 0.65 and 0.87 based on a fivefold cross-validation for a dry season and wet season estimation, respectively. We generated emission flux map products that successfully revealed the spatial and temporal heterogeneity of CH 4 emission within the dominant freshwater marsh ecosystem in the Everglades. We conclude that Landsat is promising for upscaling and monitoring CH 4 emission flux and reducing the uncertainty in emission estimates from wetlands.

WetCH4: A Machine Learning-based Upscaling of Methane Fluxes of Northern Wetlands during 2016–2022

“…The efficient computation cost makes it easier to apply the models over large regions at higher spatial resolutions. Among ML methods, decision-tree-based algorithms have been widely used in upscaling for the computation efficiency and prediction performance (Beaulieu et al, 2020;Jung et al, 2020;Virkkala et al, 2021;C. Zhang et al, 2020).…”

“…This approach involves using satellite products to quantify wetland characteristics and extent. For example, seasonal average surface reflectance of Landsat 8 images was used with pointbased gas trap measurements to estimate CH4 emissions in dry and wet seasons from Everglades' freshwater marshes (C. Zhang et al, 2020). Existing ML-based large-scale upscaling models used MODIS land surface temperature at night (LST) or enhanced vegetation index (EVI), vegetation canopy height and ancillary environmental variables from remote sensing products (McNicol et al, 2023;Ouyang et al, 2023;Peltola et al, 2019.…”

Using an object-based machine learning ensemble approach to upscale evapotranspiration measured from eddy covariance towers in a subtropical wetland

“…This matching can reduce the uncertainty caused by the inherent geolocation error of satellite data and better represent the flux tower footprint. Object-based modeling is valuable in EC CO 2 flux prediction (Zhang et al, 2021) and field measured CH 4 upscaling in wetlands using 30-m Landsat data (Zhang et al, 2020). Therefore, it is desirable for ET rate estimation given its ability to account for the spatial variation of plant community structure and composition.…”