This paper discusses different feature selection methods and CO2 flux data sets with a varying quality‐quantity balance for the application of a Random Forest model to predict daily CO2 fluxes at 250 m spatial resolution for the Rur catchment area in western Germany between 2010 and 2018. Measurements from eddy covariance stations of different ecosystem types, remotely sensed vegetation data from MODIS, and COSMO‐REA6 reanalysis data were used to train the model and predictions were validated by a spatial and temporal validation scheme. Results show the capabilities of a backwards feature elimination to remove irrelevant variables and an importance of high‐quality‐low‐quantity flux data set to improve predictions. However, results also show that spatial prediction is more difficult than temporal prediction by reflecting the mean value accurately though underestimating the variance of CO2 fluxes. Vegetated parts of the catchment acted as a CO2 sink during the investigation period, net capturing about 237 g C m−2 y−1. Croplands, coniferous forests, deciduous forests and grasslands were all sinks on average. The highest uptake was predicted to occur in late spring and early summer, while the catchment was a CO2 source in fall and winter. In conclusion, the Random Forest model predicted a narrower distribution of CO2 fluxes, though our methodological improvements look promising in order to achieve high‐resolution net ecosystem exchange data sets at the regional scale.

The gross primary productivity (GPP) of terrestrial ecosystems, of which forests are the dominant factor (Pan et al., 2011), is a key element of the global carbon cycle (Canadell et al., 2021). The resulting biomass further is important for human demands of food, energy, and construction materials (Taye et al., 2021). The assimilation of atmospheric CO 2 via photosynthesis is primarily driven by photosynthetically active radiation (PAR), though it is also sensitive to intertwined environmental and physiological variables, such as temperature, water, and nutrient availability, or chlorophyll content of the canopy (

No abstract

<p>Net Ecosystem Exchange (NEE) is an important factor regarding the impact of land use changes to the global carbon cycle and thus climate change. The Eddy Covariance technique is the most direct way of measuring CO<sub>2</sub> fluxes, however, it provides spatially discontinuous data from a sparse network of stations. Thus, generating high-resolution spatiotemporal products of carbon fluxes remains a major challenge. Machine Learning (ML) techniques are a promising approach to upscale this information to regional and global scales and can thereby help to produce better NEE datasets for earth-system modelling.</p><p>Our approach uses statistical relationships between NEE, vegetation indices and meteorological variables to train a Random Forest model with spatial feature selection to predict daily NEE values at 1 km spatial resolution for the Rur-catchment area (ca. 2400 km²) in western Germany. Data from twelve Eddy stations of different land use types of the TERENO Network Eifel/Lower Rhine Valley between 2010 and 2018 were used to train and test the ML model. Factors potentially affecting NEE such as vegetation indices (NDVI, EVI, LAI) extracted from MODIS products, incoming solar radiation from Heliosat (SARAH-2) and additional meteorological variables from COSMO REA6 reanalysis products served as independent variables, which were further evaluated in regard to their relative importance for NEE prediction.</p><p>A novel spatial cross-validation scheme has been applied and compared to a conventional random k-fold cross-validation. This is important for the assessment of the model performance regarding spatial predictions beyond the scope of training locations in contrast to mere data reproduction. Results indicate a lower model performance evaluated with spatial cross-validation and that conventional random cross-validation hence leads to an overoptimistic view of the prediction skills. Nonetheless, the ML approach displayed a feasible way to upscale carbon fluxes to a regional scale utilizing different datasets and produced high-resolution NEE-raster for an entire catchment area.</p>

Flux measurements over heterogeneous surfaces with growing vegetation and a limited fetch are a difficult task, as measurement heights that are too high or too low above the canopy adversely affect results. The aim of this study is to assess implications from measurement height in regard to low-pass filtering, footprint representativeness, and energy balance closure for a clear-cut site with regrowing vegetation of varying height. For this, measurements from two open-path eddy-covariance systems at different heights are compared over the course of one growing season. Particular attention is paid to low-pass-filtering corrections, for which five different methods are compared. Results indicate significant differences between fluxes from the upper and lower systems, which likely result from footprint differences and an insufficient spectral correction for the lower system. Different low-pass-filtering corrections add an uncertainty of 3.4% (7.0%) to CO2 fluxes and 1.4% (3.0%) to H2O fluxes for the upper (lower) system, also leading to considerable differences in cumulative fluxes. Despite limitations in the analysis, which include the difficulty of applying a footprint model at this study site and the likely influence of advection on the total exchange, the analysis confirms that information about the choice of spectral correction method and measurement-height changes are critical for interpreting data at complex sites.