Variability and uncertainty in flux-site-scale net ecosystem exchange simulations based on machine learning and remote sensing: a systematic evaluation

Abstract

Abstract. Net ecosystem exchange (NEE) is an important indicator of carbon cycling in terrestrial ecosystems. Many previous studies have combined flux observations and meteorological, biophysical, and ancillary predictors using machine learning to simulate the site-scale NEE. However, systematic evaluation of the performance of such models is limited. Therefore, we performed a meta-analysis of these NEE simulations. A total of 40 such studies and 178 model records were included. The impacts of various features throughout the modeling process on the accuracy of the model were evaluated. Random forests and support vector machines performed better than other algorithms. Models with larger timescales have lower average R2 values, especially when the timescale exceeds the monthly scale. Half-hourly models (average R2 = 0.73) were significantly more accurate than daily models (average R2 = 0.5). There are significant differences in the predictors used and their impacts on model accuracy for different plant functional types (PFTs). Studies at continental and global scales (average R2 = 0.37) with multiple PFTs, more sites, and a large span of years correspond to lower R2 values than studies at local (average R2 = 0.69) and regional (average R2 = 0.7) scales. Also, the site-scale NEE predictions need more focus on the internal heterogeneity of the NEE dataset and the matching of the training set and validation set.

Evaluation of water flux predictive models developed using eddy-covariance observations and machine learning: a meta-analysis

“…In the above meta-analysis of the models, we found that water flux simulations based on EC observations can achieve high accuracy but also have high uncertainty through the modeling workflow. The R-squared of many water flux simulation models exceeds 0.8, possibly higher than some remotesensing-based and process-based models and possibly higher than carbon flux simulations such as the net ecosystem exchange (NEE) in a similar modeling framework (Shi et al, 2022). This may be because many data on important variables affecting carbon flux such as soil and biomass pools, disturbances, ecosystem age, management activities, and land use history are not yet effectively and continuously measured (Jung et al, 2011) with the global spatially and temporally explicit information.…”

“…In general, predictors related to meteorological, vegetation, and soil conditions were common to both ET and NEE simulations in a similar framework (Shi et al, 2022). However, in NEE predictions, explanatory variables such as soil organic content, photosynthetic photon flux density, and growing degree days (Shi et al, 2022) are not necessary for ET predictions. The selection of these variables requires our prior knowledge of the dominant drivers of ET and NEE anomalies of particular ecosystems and their differences.…”

“…The accuracy of NEE predictions (Shi et al, 2022) can be more limited by global variability across biomes and locations (Nemani et al, 2003) given the lack of locally measured data on soil and biomass pools, disturbances, ecosystem age, management activities, and land use history (Jung et al, 2011). It can result in a higher heterogeneity of the training data in large-scale modeling with multiple flux sites (Shi et al, 2022) and the weak ability to capture the NEE anomalies. In contrast, in ET predictions, meteorological variables and vegetation conditions appear to be already sufficient to capture a considerably large fraction of the ET variations in most conditions.…”

New data‐driven method for estimation of net ecosystem carbon exchange at meteorological stations effectively increases the global carbon flux data

“…When applied in regional or global extrapolation, it lacks a validation of the simulation results at regions without flux stations. More widely distributed meteorological stations have the potential to deliver more reliable NEE datasets to offset the limitation of the NEE validation in regions without flux stations (Shi et al, 2022).…”

Revisiting and attributing the global controls on terrestrial ecosystem functions of climate and plant traits at FLUXNET sites with causal networks

“…For example, eddy-covariance flux tower observations (Baldocchi, 2014) for carbon flux (i.e., net ecosystem exchange (NEE)) and water flux (i.e., evapotranspiration (ET)) have been widely used to investigate changes in ecosystem functions and their responses to climate change, vegetation condition changes, etc (Jung et al, 2020(Jung et al, , 2010Migliavacca et al, 2021;Peaucelle et al, 2019). With the increase in such observations, various statistical analysis methods such as emerging machine learning (Barnes et al, 2021;Migliavacca et al, 2021;Reichstein et al, 2019;Shi et al, 2022bShi et al, , a, 2020bTramontana et al, 2016) have been used to mine the hidden information on the effects of climate change and its induced changes in vegetation, etc. on ecosystem function variables such as carbon and water flux, which has not been understood in depth by process-based models (e.g., biogeochemistry models (Sakschewski et al, 2016)).…”