Technical note: Uncertainties in eddy covariance CO<sub>2</sub> fluxes in a semiarid sagebrush ecosystem caused by gap-filling approaches

Abstract

Abstract. Gap-filling eddy covariance CO2 fluxes is challenging at dryland sites due to small CO2 fluxes. Here, four machine learning (ML) algorithms including artificial neural network (ANN), k-nearest neighbors (KNNs), random forest (RF), and support vector machine (SVM) are employed and evaluated for gap-filling CO2 fluxes over a semiarid sagebrush ecosystem with different lengths of artificial gaps. The ANN and RF algorithms outperform the KNN and SVM in filling gaps ranging from hours to days, with the RF being more time efficient than the ANN. Performances of the ANN and RF are largely degraded for extremely long gaps of 2 months. In addition, our results suggest that there is no need to fill the daytime and nighttime net ecosystem exchange (NEE) gaps separately when using the ANN and RF. With the ANN and RF, the gap-filling-induced uncertainties in the annual NEE at this site are estimated to be within 16 g C m−2, whereas the uncertainties by the KNN and SVM can be as large as 27 g C m−2. To better fill extremely long gaps of a few months, we test a two-layer gap-filling framework based on the RF. With this framework, the model performance is improved significantly, especially for the nighttime data. Therefore, this approach provides an alternative in filling extremely long gaps to characterize annual carbon budgets and interannual variability in dryland ecosystems.

Quantifying the Spatial Representativeness of Carbon Flux Footprints of a Grassland Ecosystem in the Semi‐Arid Region

“…Yao et al. (2021) constructed a two‐layer gap‐filling framework based on RF for site‐level carbon fluxes and comprehensively selected 12 carbon flux driving factors such as solar radiation, temperature, relative humidity, and NDVI as input variables for the ML algorithm through importance evaluation. When analyzing the relative importance of predictor variables on the area of carbon flux footprints, we used the percentage increases in the mean square error (Increase in MSE(%), %IncMSE) of variables to quantify the importance of these predictors, with higher %IncMSE values implying more important predictors (Breiman, 2001; Jiao et al., 2018).…”

“…Huang et al (2021) used the variable importance measurement method of RF to construct the most suitable set of candidate predictors for predicting net ecosystem exchange, including 13 variables such as temperature, precipitation, EVI, and leaf area index, and realized the upscaling of net ecosystem exchange at global FLUXNET sites. Yao et al (2021) constructed a two-layer gap-filling framework based on RF for site-level carbon fluxes and comprehensively selected 12 carbon flux driving factors such as solar radiation, temperature, relative humidity, and NDVI as input variables for the ML algorithm through importance evaluation. When analyzing the relative importance of predictor variables on the area of carbon flux footprints, we used the percentage increases in the mean square error (Increase in MSE(%), %IncMSE) of variables to quantify the importance of these predictors, with higher %IncMSE values implying more important predictors (Breiman, 2001;Jiao et al, 2018).…”

Gap-filling carbon dioxide, water, energy, and methane fluxes in challenging ecosystems: Comparing between methods, drivers, and gap-lengths

“…It is also a standard gap-filling approach particularly for gaps longer than one month (Delwiche et al, 2021;Mahabbati et al, 2021). The SVR was also an established gap-filling algorithm, it converts non-linear regressions into higher-dimensional linear regression by a predefined kernel function (Khan et al, 2021;Yao et al, 2021b).…”

“…Second, most early eddy covariance towers were installed in productive and/or less-disturbed natural ecosystems, and this sampling distribution poses challenges to the development of gap-filling (Moffat et al, 2007;Irvin et al, 2021;Zhu et al, 2022). Flux gap-filling for other types of ecosystems can be challenging, these include managed ecosystems and ecosystems with flux rates close to zero (Lucas- McKenzie et al, 2021;Yao et al, 2021a). In managed ecosystems, for example, agricultural activities, can substantially alter flux temporal dynamics (McCalmont et al, 2021;Cardenas et al, 2022), quantifying the frequency and intensity of management activities can be challenging for training a machine learning model to gap-fill these timeseries.…”

The turning of ecological change in the Yellow River Basin

“…3.4 | GPP changes with respect to the TP Different with the NDVI, which is used to describe vegetation coverage, GPP represents the total influx of carbon into the ecosystem from the atmosphere and is typically used to describe regional carbon balance (Cheng et al, 2017;Lu et al, 2022;Yao, Gao, et al, 2021;Yao, Mao, et al, 2021;Yu et al, 2013). Figure 8c shows that the GPP displayed an increase in 74.1% of the YRB, with 39.8% exhibiting a significant increase, distributing in the source region and the western part of the middle reaches.…”