期刊
GLOBAL BIOGEOCHEMICAL CYCLES
卷 34, 期 12, 页码 -出版社
AMER GEOPHYSICAL UNION
DOI: 10.1029/2020GB006703
关键词
carbon dioxide; climate change; machine learning; stream network; remote sensing
资金
- COWIfonden
Stream networks transport and emit substantial volumes of carbon dioxide (CO2) into the atmosphere. We gathered open monitoring data from streams in three Scandinavian countries and estimated CO2 partial pressure (pCO(2)) at 2,298 sites. Most of the sites (87%) were supersaturated when averaged across the year with an overall mean pCO(2) of 1,464 mu atm (range: 17-15,646). Using remote sensing data, we modeled a realistic stream network including streams above similar to 2.5 m wide and calculated catchment averages of multiple variables associated with geomorphometry, stream network proximity, and land cover. We compared the ability of eight machine learning models to predict pCO(2) and found that the Random Forest model achieved the highest accuracy, with a root-mean-square error of 0.22 (log(10)(pCO(2))) and R-2 of 0.66. Mean catchment elevation, slope, and permanent water cover were the most important predictor variables. We used the predictive model to create a high-resolution (25-m resolution) map with predicted stream pCO(2) throughout the 268.807 km stream network in Denmark, Sweden, and Finland. Predicted pCO(2) averaged 1,134 mu atm (range: 154-8,174). We used surface runoff, air temperature, and stream channel slope to estimate gas transfer velocity and CO2 flux throughout the network. Mean stream CO2 fluxes ranged from 1.0 and 1.2 in Sweden and Finland, respectively, 3 to 3.2 g C m(-2) day(-1) in Denmark. Better-performing models improve our ability to predict pCO(2) in stream networks and reduce the uncertainty of upscaling estimates of carbon emissions from inland waters to countries and continents.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据