Search

Methane flux (FCH4) measurements using the eddy covariance technique have increased over the past decade. FCH4 measurements commonly include data gaps, as is the case with CO2 and energy fluxes. However, gap‐filling FCH4 data are more challenging than other fluxes due to its unique characteristics including multidriver dependency, variabilities across multiple timescales, nonstationarity, spatial heterogeneity of flux footprints, and lagged influence of biophysical drivers. Some researchers have applied a marginal distribution sampling (MDS) algorithm, a standard gap‐filling method for other fluxes, to FCH4 datasets, and others have applied artificial neural networks (ANN) to resolve the challenging characteristics of FCH4. However, there is still no consensus regarding FCH4 gap‐filling methods due to limited comparative research. We are not aware of the applications of machine learning (ML) algorithms beyond ANN to FCH4 datasets. Here, we compare the performance of MDS and three ML algorithms (ANN, random forest [RF], and support vector machine [SVM]) using multiple combinations of ancillary variables. In addition, we applied principal component analysis (PCA) as an input to the algorithms to address multidriver dependency of FCH4 and reduce the internal complexity of the algorithmic structures. We applied this approach to five benchmark FCH4 datasets from both natural and managed systems located in temperate and tropical wetlands and rice paddies. Results indicate that PCA improved the performance of MDS compared to traditional inputs. ML algorithms performed better when using all available biophysical variables compared to using PCA‐derived inputs. Overall, RF was found to outperform other techniques for all sites. We found gap‐filling uncertainty is much larger than measurement uncertainty in accumulated CH4 budget. Therefore, the approach used for FCH4 gap filling can have important implications for characterizing annual ecosystem‐scale methane budgets, the accuracy of which is important for evaluating natural and managed systems and their interactions with global change processes. [ABSTRACT FROM AUTHOR]

In the Korea Flux Monitoring Network, Haenam Farmland has the longest record of carbon/water/energy flux measurements produced using the eddy covariance (EC) technique. Unfortunately, there are long gaps (i.e., gaps longer than 30 days), particularly in 2007 and 2014, which hinder attempts to analyze these decade-long time-series data. The open source and standardized gap-filling methods are impractical for such long gaps. The data-driven approach using machine learning and remote-sensing or reanalysis data (i.e., interpolating/extrapolating EC measurements via available networks temporally/spatially) for estimating terrestrial CO2/H2O fluxes at the regional/global scale is applicable after appropriate modifications. In this study, we evaluated the applicability of the data-driven approach for filling long gaps in flux data (i.e., gross primary production, ecosystem respiration, net ecosystem exchange, and evapotranspiration). We found that using a longer training dataset in the machine learning generally produced better model performance, although there was a greater possibility of missing interannual variations caused by ecosystem state changes (e.g., changes in crop variety). Based on the results, we proposed gap-filling strategies for long-period flux data gaps and used them to quantify the annual sums with uncertainties in 2007 and 2014. The results from this study have broad implications for long-period gap-filling at other sites, and for the estimation of regional/global CO2/H2O fluxes using a data-driven approach. [ABSTRACT FROM AUTHOR]

• A machine learning method (NN SIF) is developed to partition NEE using SIF. • NN SIF effectively reduces structural errors of standard partitioning methods. • NN SIF better partitions NEE under high temperature and water stress. • NN SIF unravels physiological controls on ecosystem–scale GPP–SIF relationships. Accurately partitioning net ecosystem exchange (NEE) into ecosystem respiration (ER) and gross primary productivity (GPP) is critical for understanding the terrestrial carbon cycle. The standard partitioning methods rely on simplified empirical models, which have inherent structural errors. These structural errors lead to biased GPP and ER estimation, especially during extreme events (e.g., drought) and human disturbances (e.g., crop harvest). Recently, solar-induced chlorophyll fluorescence (SIF) has been shown to be well correlated to GPP, thus offering a path to improve the NEE partitioning by constraining GPP. However, the ecosystem-scale relationship between GPP and SIF remains limited. Here, we show that neural networks informed by SIF observations (NN SIF) can be successfully used to partition NEE, while simultaneously learning the ecosystem-scale GPP-SIF relationship. NN SIF was compared against standard partitioning methods and NN without SIF constraint (NN noSIF), using field data from different ecosystems and synthetic data generated by a coupled fluorescence-photosynthesis model (SCOPE). NN SIF showed superior performance as: (1) it effectively improves the ER estimation, especially at high temperature, (2) it better captures the moisture limitation on ER, (3) it more accurately estimates LUE variations to stress, and (4) it uniquely captures the rapid GPP drop after land management (harvest). Furthermore, NN SIF can retrieve the GPP-SIF relationship at the ecosystem scale, and elucidate how this relationship responds to environmental conditions. Overall, our algorithm provides the first direct and non-empirical estimate of the ecosystem-scale GPP-SIF relationship, without relying on any prior empirical assumptions on the relationships between CO 2 fluxes, climatic drivers, and SIF. The new knowledge learned by NN SIF can help better estimate global-scale GPP using satellite SIF, especially during extreme events and in the presence of land management. [ABSTRACT FROM AUTHOR]

Within the past decade, flux towers which incorporate methane flux (FCH4) measurements have increased following the development of a robust and low-power CH4 gas analyzer. As is the case with other trace gases, CH4 flux data based on the eddy covariance technique commonly includes data gaps due to unsuitable atmospheric conditions and system failures, with FCH4 often having more gaps than CO2 and H2O fluxes. However, there is still no consensus on FCH4 gap-filling method. Some researchers have applied a marginal distribution sampling (MDS) algorithm for FCH4 gap-filling, which is a standard gap-filling method for CO2 and water fluxes. Others have tested and applied machine learning (ML) algorithms like artificial neural networks to gap-fill FCH4 data. One of the major challenges in FCH4 gap-filling using both MDS and ML algorithms concerns selection of input variables. Since MDS has been widely used in the Fluxnet community by virtue of a robust online algorithm and an R package, standardizing the selection of MDS input variables for gap-filling FCH4 is important. Nevertheless, discussions in the literature on MDS input parameters are rare. Likewise, use of ML algorithms for FCH4 gap-filling has increased with limited discussions on input selection. At present, we are also not aware of a comprehensive comparison between commonly used ML algorithms such as artificial neural networks (ANN), support vector machine (SVM), and random forest (RF) for FCH4 gap-filling. In this study we compare performance of MDS and the three ML algorithms (ANN, SVM, RF) using different combinations of ancillary parameters. In addition, we applied principal component analysis (PCA) as an input to the algorithms to resolve multi-parameter dependency of FCH4. We applied this approach to 5 benchmark FCH4 datasets from temperate and tropical wetlands and rice paddies. For performance evaluation, various artificial gap scenarios were added to FCH4 data and then gap-filled using MDS and ML algorithms, with and without PCA. Performance of the algorithms was measured using Taylor diagrams, and residual analysis was also conducted to understand limitations of each algorithm and input decision. Results indicate that PCA improves the performance of MDS compared to traditional MDS inputs (i.e. global radiation, air temperature, and vapour pressure deficit) and alternative MDS inputs (i.e., soil temperature, water table height, and net ecosystem exchange). On the other hands, all of the ML algorithms show better performance without PCA when all available biophysical variable are selected as inputs. As for the comparison among ML algorithms, RF was found to outperform other techniques for the 5 benchmark sites and all gap scenarios. In this presentation, we also discuss how annual CH4 budgets vary in relation to gap-filling methods based on biases introduced through the various gap-filing approaches. [ABSTRACT FROM AUTHOR]