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

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 R-2 values, especially when the timescale exceeds the monthly scale. Half-hourly models (average R-2 = 0.73) were significantly more accurate than daily models (average R-2 = 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 R-2 = 0.37) with multiple PFTs, more sites, and a large span of years correspond to lower R 2 values than studies at local (average R-2 = 0.69) and regional (average R-2 = 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.