Sunchai PHUNGERN, Yuji GOTO, Liya DING, Iain MCTAGGART, Kosuke NOBORIO
{"title":"Gap-filling greenhouse gas fluxes for the closed chamber method at paddy fields using machine learning techniques","authors":"Sunchai PHUNGERN, Yuji GOTO, Liya DING, Iain MCTAGGART, Kosuke NOBORIO","doi":"10.2480/agrmet.d-22-00024","DOIUrl":null,"url":null,"abstract":"Measurements of greenhouse gas (GHG) emissions from paddy fields can often include flux measurement errors due to either instrument errors or unfavorable weather. Therefore, data post-processing, including the gap-filling process, is required to improve data quality and quantify the GHG flux budget. This study applied machine learning (ML) techniques with polynomial and multivariate polynomial regression models for gap-filling methane (CH4) and carbon dioxide (CO2) fluxes from closed chamber (CC) method measurements and compared results with mean diurnal variation (MDV) and look-up table (LUT) techniques. The most influential factors affecting methane emissions in the paddy field were used for input variables in the models: air temperature, soil temperature, soil redox potential, soil water content, solar radiation, and days after transplanting. The models' performances were compared using mean absolute error (MAE) and root mean square error (RMSE). The results showed that MAE and RMSE for gap-filling CH4 fluxes were 1.299-2.984 and 2.499-4.981 mg CH4 m-2 h-1, respectively. Also, the multivariate polynomial regression models performed better for gap-filling CH4 fluxes (RMSE = 2.499 mg CH4 m-2 h-1) than the polynomial regression models, MDV (RMSE = 3.210 mg CH4 m-2 h-1), and LUT (RMSE = 3.339 mg CH4 m-2 h-1) techniques. The MAE and RMSE for gap-filling CO2 fluxes were 0.282-0.949 and 0.435-1.078 g CO2 m-2 h-1, respectively. The ML techniques with polynomial regression using solar radiation (RMSE = 0.435 g CO2 m-2 h-1) and multivariate models (RMSE = 0.445 g CO2 m-2 h-1) perform better on gap-filling CO2 fluxes than MDV (RMSE = 0.544 g CO2 m-2 h-1), and LUT (RMSE = 0.553 g CO2 m-2 h-1) techniques. The gap-filling using the multivariate polynomial regression models used in this study improved the reliability of the diurnal variation in GHG fluxes. Therefore, ML techniques could be a proper alternative for gap-filling GHG fluxes.","PeriodicalId":56074,"journal":{"name":"Journal of Agricultural Meteorology","volume":null,"pages":null},"PeriodicalIF":1.4000,"publicationDate":"2023-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Agricultural Meteorology","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.2480/agrmet.d-22-00024","RegionNum":4,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"AGRICULTURE, MULTIDISCIPLINARY","Score":null,"Total":0}
引用次数: 0
Abstract
Measurements of greenhouse gas (GHG) emissions from paddy fields can often include flux measurement errors due to either instrument errors or unfavorable weather. Therefore, data post-processing, including the gap-filling process, is required to improve data quality and quantify the GHG flux budget. This study applied machine learning (ML) techniques with polynomial and multivariate polynomial regression models for gap-filling methane (CH4) and carbon dioxide (CO2) fluxes from closed chamber (CC) method measurements and compared results with mean diurnal variation (MDV) and look-up table (LUT) techniques. The most influential factors affecting methane emissions in the paddy field were used for input variables in the models: air temperature, soil temperature, soil redox potential, soil water content, solar radiation, and days after transplanting. The models' performances were compared using mean absolute error (MAE) and root mean square error (RMSE). The results showed that MAE and RMSE for gap-filling CH4 fluxes were 1.299-2.984 and 2.499-4.981 mg CH4 m-2 h-1, respectively. Also, the multivariate polynomial regression models performed better for gap-filling CH4 fluxes (RMSE = 2.499 mg CH4 m-2 h-1) than the polynomial regression models, MDV (RMSE = 3.210 mg CH4 m-2 h-1), and LUT (RMSE = 3.339 mg CH4 m-2 h-1) techniques. The MAE and RMSE for gap-filling CO2 fluxes were 0.282-0.949 and 0.435-1.078 g CO2 m-2 h-1, respectively. The ML techniques with polynomial regression using solar radiation (RMSE = 0.435 g CO2 m-2 h-1) and multivariate models (RMSE = 0.445 g CO2 m-2 h-1) perform better on gap-filling CO2 fluxes than MDV (RMSE = 0.544 g CO2 m-2 h-1), and LUT (RMSE = 0.553 g CO2 m-2 h-1) techniques. The gap-filling using the multivariate polynomial regression models used in this study improved the reliability of the diurnal variation in GHG fluxes. Therefore, ML techniques could be a proper alternative for gap-filling GHG fluxes.
期刊介绍:
For over 70 years, the Journal of Agricultural Meteorology has published original papers and review articles on the science of physical and biological processes in natural and managed ecosystems. Published topics include, but are not limited to, weather disasters, local climate, micrometeorology, climate change, soil environment, plant phenology, plant response to environmental change, crop growth and yield prediction, instrumentation, and environmental control across a wide range of managed ecosystems, from open fields to greenhouses and plant factories.