Zhu (CMS 2020): Prototyping a monitoring system of global wetland CH4 emissions with machine learning and satellite remote sensing
Qing Zhu, Lawrence Berkeley National Laboratory, qzhu@lbl.gov (Presenter)
Wetlands are highly dynamic terrestrial-aquatic interfaces widely distributed across tropical, temperate, and high latitude ecosystems. As the largest natural source to the atmosphere, wetlands are responsible for 30% of global methane (CH4) emissions, which accounts for about 25% of cumulative anthropogenic radiative forcing since the industrial revolution. Flux chambers and Eddy Covariance (EC) flux towers provide high-frequency measurements of CH4 exchanges between land and atmosphere. However, the complexity of wetland CH4 production and consumption processes makes it challenging to upscale and estimate the net emission flux at large scales and has resulted in large uncertainties in regional and global CH4 emission products. For example, non-linear and lagged relationships between air temperature and wetland CH4 emissions, particularly during late autumn, hinders process models to reproduce and upscale site-level eddy covariance measurements to regional and global scales. Advanced Machine Learning (ML) techniques, such as Artificial Intelligence frameworks, are skilled in modeling complex relationships between predictors (e.g., physical environment) and target variable (CH4 emission). Utilizing remote sensing data (e.g., land surface temperature, inundation area, vegetation growing conditions, and freeze/thaw state), ML models, globally distributed EC tower and flux chamber measurements, and ancillary datasets (e.g., soil properties), this proposal aims to develop a prototype monitoring system of global wetland CH4 emissions and to improve understanding of how the physical environment (i.e., temperature, water, and air pressure, etc.) and vegetation activity affect the spatial and temporal dynamics of CH4 emissions. Specifically, we will develop an open-source integrated software infrastructure that: (1) collects in situ and remote sensing data; (2) pre-processes data (e.g., gap-filling, feature engineering); (3) performs initial training of candidate machine learning models; (4) performs model selection and fine-tuning; and (5) predicts and evaluates CH4 emissions. This development will generate high-resolution monthly products of global wetland CH4 emissions and with uncertainty quantification. The resulting product will be used to diagnose the spatial and temporal evolution of the underlying drivers and evaluate existing wetland CH4 emission estimates in scientific publications and national and international reports.
The proposed project directly responds to the elements of ‘the use of satellite remote sensing as an alternative or a supplement to ground-based methods for quantifying net carbon emissions and/or storage’ and ‘Studies using remote sensing data that evaluate and enhance national reported carbon emissions inventories from bottom-up estimates from various sectors of emissions within the United States and have the potential to be applied to reported national inventories from other nations’.
The investigators request membership in the NASA CMS Science Team. The team will make the resulting software, code, data, and documentation publicly available via platforms such as GitHub and ORNL DAAC.
Associated Project(s):
Poster Location ID: 45
Presentation Type: Poster
Session: Poster Session 1
Session Date: Wednesday (9/27) 1:15 PM