You are here:
PREDICTING CARBON DIOXIDE AND METHANE EMISSION RATES FROM U.S. RESERVOIRS USING BOOSTED REGRESSION TREES
Citation:
Beaulieu, J., S. Waldo, W. Barnett, D. Balz, M. Platz, AND K. White. PREDICTING CARBON DIOXIDE AND METHANE EMISSION RATES FROM U.S. RESERVOIRS USING BOOSTED REGRESSION TREES. ASLO 2019 Aquatic Sciences Meeting, San Juan, Puerto Rico, February 23 - March 02, 2019.
Impact/Purpose:
Reservoirs (i.e. water impounded behind dams) may be the fourth largest anthropogenic source of methane, a potent greenhouse gas, in the United States. Emissions estimates are poorly constrained, however, due to a lack of measurements. In this research we developed a statistical model to predict reservoir emission rates based on reservoir and watershed characteristics. This model will allow for statistically-robust, national scale emission estimates. These estimates will be useful for reporting anthropogenic greenhouse gas emissions to the Intergovernmental Panel on Climate Change following our obligations under the United Nations Framework Convention on Climate Change treaty.
Description:
By one estimate, reservoirs are the fourth largest anthropogenic methane (CH4) source in the U.S., with emissions equivalent to about half of those from all U.S. landfills. Reservoir emission estimates are poorly constrained, however, due to large variation in emission rates among reservoirs and a lack of measurements from large portions of the country. In this study we use boosted regression trees (brt), a type of machine learning algorithm, to model CH4 and carbon dioxide (CO2) emission rates for U.S. reservoirs. The model was trained using data from a 2016 survey of CH4 and CO2 emission rates from 32 reservoirs in the mid-western U.S. and literature data from 11 reservoirs in the southeastern and northwestern U.S. Two sets of predictor variables were analyzed. The first set include variables available for all U.S. waterbodies represented in the National Hydrologic Dataset (NHD) and related data products, including information on watershed condition and reservoir morphology, and additional site-specific measurements of reservoir condition (i.e. nutrient concentrations). The second set of predictor variables included only those available at the national-scale. This approach allowed us to ask 1) how well can we predict emission rates using nationally available data, and 2) how much do predictions improve when additional site-specific data are available. Models predicted total CH4 emission rates better than diffusive or ebullitive emissions and models for CO2 emission rate performed worse than those for CH4. Important predictor variables for total CH4 emission rates included indices of morphology (i.e. mean depth) and reservoir-watershed coupling (i.e. relative drainage area). In the next stage of this research we will link the brt models and NHD to generate a national-scale estimate of CH4 and CO2 emissions for U.S. reservoirs.