Abstract
Time series of wetland methane fluxes measured by eddy covariance require gap-filling to estimate daily, seasonal, and annual emissions. Gap-filling methane fluxes is challenging because of high variability and complex responses to multiple drivers. To date, there is no widely established gap-filling standard for wetland methane fluxes, with regards both to the best model algorithms and predictors. This study synthesizes results of different gap-filling methods systematically applied at 17 wetland sites spanning boreal to tropical regions and including all major wetland classes and two rice paddies. Procedures are proposed for: 1) creating realistic artificial gap scenarios, 2) training and evaluating gap-filling models without overstating performance, and 3) predicting half-hourly methane fluxes and annual emissions with realistic uncertainty estimates. Performance is compared between a conventional method (marginal distribution sampling) and four machine learning algorithms. The conventional method achieved similar median performance as the machine learning models but was worse than the best machine learning models and relatively insensitive to predictor choices. Of the machine learning models, decision tree algorithms performed the best in cross-validation experiments, even with a baseline predictor set, and artificial neural networks showed comparable performance when using all predictors. Soil temperature was frequently the most important predictor whilst water table depth was important at sites with substantial water table fluctuations, highlighting the value of data on wetland soil conditions. Raw gap-filling uncertainties from the machine learning models were underestimated and we propose a method to calibrate uncertainties to observations. The python code for model development, evaluation, and uncertainty estimation is publicly available. This study outlines a modular and robust machine learning workflow and makes recommendations for, and evaluates an improved baseline of, methane gap-filling models that can be implemented in multi-site syntheses or standardized products from regional and global flux networks (e.g., FLUXNET).
Original language | English (US) |
---|---|
Article number | 108528 |
Journal | Agricultural and Forest Meteorology |
Volume | 308-309 |
DOIs | |
State | Published - Oct 15 2021 |
Keywords
- Machine learning
- flux
- gap-filling
- imputation
- methane
- time series
- wetlands
ASJC Scopus subject areas
- Forestry
- Global and Planetary Change
- Agronomy and Crop Science
- Atmospheric Science
Access to Document
Other files and links
Fingerprint
Dive into the research topics of 'Gap-filling eddy covariance methane fluxes: Comparison of machine learning model predictions and uncertainties at FLUXNET-CH4 wetlands'. Together they form a unique fingerprint.Cite this
- APA
- Standard
- Harvard
- Vancouver
- Author
- BIBTEX
- RIS
In: Agricultural and Forest Meteorology, Vol. 308-309, 108528, 15.10.2021.
Research output: Contribution to journal › Article › peer-review
}
TY - JOUR
T1 - Gap-filling eddy covariance methane fluxes
T2 - Comparison of machine learning model predictions and uncertainties at FLUXNET-CH4 wetlands
AU - Irvin, Jeremy
AU - Zhou, Sharon
AU - McNicol, Gavin
AU - Lu, Fred
AU - Liu, Vincent
AU - Fluet-Chouinard, Etienne
AU - Ouyang, Zutao
AU - Knox, Sara Helen
AU - Lucas-Moffat, Antje
AU - Trotta, Carlo
AU - Papale, Dario
AU - Vitale, Domenico
AU - Mammarella, Ivan
AU - Alekseychik, Pavel
AU - Aurela, Mika
AU - Avati, Anand
AU - Baldocchi, Dennis
AU - Bansal, Sheel
AU - Bohrer, Gil
AU - Campbell, David I.
AU - Chen, Jiquan
AU - Chu, Housen
AU - Dalmagro, Higo J.
AU - Delwiche, Kyle B.
AU - Desai, Ankur R.
AU - Euskirchen, Eugenie
AU - Feron, Sarah
AU - Goeckede, Mathias
AU - Heimann, Martin
AU - Helbig, Manuel
AU - Helfter, Carole
AU - Hemes, Kyle S.
AU - Hirano, Takashi
AU - Iwata, Hiroki
AU - Jurasinski, Gerald
AU - Kalhori, Aram
AU - Kondrich, Andrew
AU - Lai, Derrick YF
AU - Lohila, Annalea
AU - Malhotra, Avni
AU - Merbold, Lutz
AU - Mitra, Bhaskar
AU - Ng, Andrew
AU - Nilsson, Mats B.
AU - Noormets, Asko
AU - Peichl, Matthias
AU - Rey-Sanchez, A. Camilo
AU - Richardson, Andrew D.
AU - Runkle, Benjamin RK
AU - Schäfer, Karina VR
AU - Sonnentag, Oliver
AU - Stuart-Haëntjens, Ellen
AU - Sturtevant, Cove
AU - Ueyama, Masahito
AU - Valach, Alex C.
AU - Vargas, Rodrigo
AU - Vourlitis, George L.
AU - Ward, Eric J.
AU - Wong, Guan Xhuan
AU - Zona, Donatella
AU - Alberto, Ma Carmelita R.
AU - Billesbach, David P.
AU - Celis, Gerardo
AU - Dolman, Han
AU - Friborg, Thomas
AU - Fuchs, Kathrin
AU - Gogo, Sébastien
AU - Gondwe, Mangaliso J.
AU - Goodrich, Jordan P.
AU - Gottschalk, Pia
AU - Hörtnagl, Lukas
AU - Jacotot, Adrien
AU - Koebsch, Franziska
AU - Kasak, Kuno
AU - Maier, Regine
AU - Morin, Timothy H.
AU - Nemitz, Eiko
AU - Oechel, Walter C.
AU - Oikawa, Patricia Y.
AU - Ono, Keisuke
AU - Sachs, Torsten
AU - Sakabe, Ayaka
AU - Schuur, Edward A.
AU - Shortt, Robert
AU - Sullivan, Ryan C.
AU - Szutu, Daphne J.
AU - Tuittila, Eeva Stiina
AU - Varlagin, Andrej
AU - Verfaillie, Joeseph G.
AU - Wille, Christian
AU - Windham-Myers, Lisamarie
AU - Poulter, Benjamin
AU - Jackson, Robert B.
N1 - Funding Information: BRKR was supported by NSF Award 1752083. OS was supported through the Canada Research Chairs and Natural Sciences and Engineering Research Council Discovery Grants programs. TS and CW were supported by the Helmholtz Association of German Research Centres (Grant No. VH-NG-821). GB was supported by a US Department of Energy (Grant DE-SC0021067) and NOAA Davidson Fellowship Award administered by ODNR OWC-NERR (Subaward N18B 315-11). Funding from the SNF projects DiRad and InnoFarm (146373 and 407340_172433), from the ETH Board and from ETH Zurich is greatly acknowledged. DP, CT and DV were supported by the Department of Excellence 2018 Program MIUR Project “Landscape 4.0 - food, wellbeing and environment" and the ICOS Ecosystem Thematic Centre. CT was supported by the E-SHAPE (GA820852) H2020 European project. DB, JV, DS, CS, SK, EE, KSH, KK, AV, CRS, RS (or sites US-MYB, US-TW1, US-TW3, US-TW4, US-TW5, US-TWT, US-SND, US-SNE) were supported by the California Department of Water Resources through a contract from the California Department of Fish and Wildlife and the United States Department of Agriculture (NIFA grant #2011-67003-30371). Funding for the AmeriFlux core sites was provided by the U.S. Department of Energy's Office of Science (AmeriFlux contract #7079856). KK was supported by the Estonian Research Council grants No. PSG631 and PRG352. KSH was supported by the California Sea Grant Delta Science Fellowship (programs R/SF-70, grant no. 2271). The contents of this material do not necessarily reflect the views and policies of the Delta Stewardship Council or California Sea Grant, nor does mention of trade names or commercial products constitute endorsement or recommendation for use. MU was supported by the Arctic Challenge for Sustainability II (JPMXD1420318865) and the JSPS KAKENHI (20K21849). Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. IM thanks H2020 RINGO project (Grant Agreement 730944), the Academy of Finland Flagship funding (grant no. 337549) and ICOS-Finland by University of Helsinki funding. Funding Information: This study was supported by the Gordon and Betty Moore Foundation through Grant GBMF5439 “Advancing Understanding of the Global Methane Cycle” to Stanford University supporting the Methane Budget activity for the Global Carbon Project (globalcarbonproject.org). Funding Information: This study was supported by the Gordon and Betty Moore Foundation through Grant GBMF5439 ?Advancing Understanding of the Global Methane Cycle? to Stanford University supporting the Methane Budget activity for the Global Carbon Project (globalcarbonproject.org). BRKR was supported by NSF Award 1752083. OS was supported through the Canada Research Chairs and Natural Sciences and Engineering Research Council Discovery Grants programs. TS and CW were supported by the Helmholtz Association of German Research Centres (Grant No. VH-NG-821). GB was supported by a US Department of Energy (Grant DE-SC0021067) and NOAA Davidson Fellowship Award administered by ODNR OWC-NERR (Subaward N18B 315-11). Funding from the SNF projects DiRad and InnoFarm (146373 and 407340_172433), from the ETH Board and from ETH Zurich is greatly acknowledged. DP, CT and DV were supported by the Department of Excellence 2018 Program MIUR Project ?Landscape 4.0 - food, wellbeing and environment" and the ICOS Ecosystem Thematic Centre. CT was supported by the E-SHAPE (GA820852) H2020 European project. DB, JV, DS, CS, SK, EE, KSH, KK, AV, CRS, RS (or sites US-MYB, US-TW1, US-TW3, US-TW4, US-TW5, US-TWT, US-SND, US-SNE) were supported by the California Department of Water Resources through a contract from the California Department of Fish and Wildlife and the United States Department of Agriculture (NIFA grant #2011-67003-30371). Funding for the AmeriFlux core sites was provided by the U.S. Department of Energy's Office of Science (AmeriFlux contract #7079856). KK was supported by the Estonian Research Council grants No. PSG631 and PRG352. KSH was supported by the California Sea Grant Delta Science Fellowship (programs R/SF-70, grant no. 2271). The contents of this material do not necessarily reflect the views and policies of the Delta Stewardship Council or California Sea Grant, nor does mention of trade names or commercial products constitute endorsement or recommendation for use. MU was supported by the Arctic Challenge for Sustainability II (JPMXD1420318865) and the JSPS KAKENHI (20K21849). Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. IM thanks H2020 RINGO project (Grant Agreement 730944), the Academy of Finland Flagship funding (grant no. 337549) and ICOS-Finland by University of Helsinki funding. AN, BP, RBJ, SHK, and LWM acquired funding for and conceived of the project. EFC, FL, GM, JI, SZ, VL, and ZO conceived of, and AA, ALM, CT, DP, DV, and IM contributed to, the design and execution of the machine learning analysis. ALM was consulted with on machine learning model and artificial gap evaluation. CT and DP contributed the marginal distribution sampling analysis. DV was consulted with on multi-imputation methods. EFC, FL, GM, JI, SZ VL, and ZO wrote the initial draft of the manuscript and ALM, ACRS, ACV, ADR, AK, AL, AM, AN, ARD, BM, BRKR, CH, CS, CT, DDB, DIC, DP, DV, DYFL, DZ, EE, EJW, ESH, GB, GJ, GLV, GXW, HC, HI, HJD, IM, JC, KBD, KSH, KVRS, LM, LWM, MA, MBN, MG, MHelbig, MHeimann, MP, MU, OS, PA, RBJ, RV, SB, SF, SHK, and TH contributed edits to subsequent drafts. ADR, ARD, CH, DDB, DIC, DYFL, GB, GLV, HI, HJD, IM, JC, KVRS, MA, MBN, MH, MU, OS, TF, TH, and TS were affiliated as principal investigators for the 17 core analysis sites. All other coauthors contributed data as principal investigators or were named as affiliated team members at other FLUXNET-CH4 sites. Publisher Copyright: © 2021 Elsevier B.V.
PY - 2021/10/15
Y1 - 2021/10/15
N2 - Time series of wetland methane fluxes measured by eddy covariance require gap-filling to estimate daily, seasonal, and annual emissions. Gap-filling methane fluxes is challenging because of high variability and complex responses to multiple drivers. To date, there is no widely established gap-filling standard for wetland methane fluxes, with regards both to the best model algorithms and predictors. This study synthesizes results of different gap-filling methods systematically applied at 17 wetland sites spanning boreal to tropical regions and including all major wetland classes and two rice paddies. Procedures are proposed for: 1) creating realistic artificial gap scenarios, 2) training and evaluating gap-filling models without overstating performance, and 3) predicting half-hourly methane fluxes and annual emissions with realistic uncertainty estimates. Performance is compared between a conventional method (marginal distribution sampling) and four machine learning algorithms. The conventional method achieved similar median performance as the machine learning models but was worse than the best machine learning models and relatively insensitive to predictor choices. Of the machine learning models, decision tree algorithms performed the best in cross-validation experiments, even with a baseline predictor set, and artificial neural networks showed comparable performance when using all predictors. Soil temperature was frequently the most important predictor whilst water table depth was important at sites with substantial water table fluctuations, highlighting the value of data on wetland soil conditions. Raw gap-filling uncertainties from the machine learning models were underestimated and we propose a method to calibrate uncertainties to observations. The python code for model development, evaluation, and uncertainty estimation is publicly available. This study outlines a modular and robust machine learning workflow and makes recommendations for, and evaluates an improved baseline of, methane gap-filling models that can be implemented in multi-site syntheses or standardized products from regional and global flux networks (e.g., FLUXNET).
AB - Time series of wetland methane fluxes measured by eddy covariance require gap-filling to estimate daily, seasonal, and annual emissions. Gap-filling methane fluxes is challenging because of high variability and complex responses to multiple drivers. To date, there is no widely established gap-filling standard for wetland methane fluxes, with regards both to the best model algorithms and predictors. This study synthesizes results of different gap-filling methods systematically applied at 17 wetland sites spanning boreal to tropical regions and including all major wetland classes and two rice paddies. Procedures are proposed for: 1) creating realistic artificial gap scenarios, 2) training and evaluating gap-filling models without overstating performance, and 3) predicting half-hourly methane fluxes and annual emissions with realistic uncertainty estimates. Performance is compared between a conventional method (marginal distribution sampling) and four machine learning algorithms. The conventional method achieved similar median performance as the machine learning models but was worse than the best machine learning models and relatively insensitive to predictor choices. Of the machine learning models, decision tree algorithms performed the best in cross-validation experiments, even with a baseline predictor set, and artificial neural networks showed comparable performance when using all predictors. Soil temperature was frequently the most important predictor whilst water table depth was important at sites with substantial water table fluctuations, highlighting the value of data on wetland soil conditions. Raw gap-filling uncertainties from the machine learning models were underestimated and we propose a method to calibrate uncertainties to observations. The python code for model development, evaluation, and uncertainty estimation is publicly available. This study outlines a modular and robust machine learning workflow and makes recommendations for, and evaluates an improved baseline of, methane gap-filling models that can be implemented in multi-site syntheses or standardized products from regional and global flux networks (e.g., FLUXNET).
KW - Machine learning
KW - flux
KW - gap-filling
KW - imputation
KW - methane
KW - time series
KW - wetlands
UR - http://www.scopus.com/inward/record.url?scp=85109612362&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85109612362&partnerID=8YFLogxK
U2 - 10.1016/j.agrformet.2021.108528
DO - 10.1016/j.agrformet.2021.108528
M3 - Article
AN - SCOPUS:85109612362
SN - 0168-1923
VL - 308-309
JO - Agricultural and Forest Meteorology
JF - Agricultural and Forest Meteorology
M1 - 108528
ER -