Ecosystems at high latitudes are under increasing stress from climate
change. To understand changes in carbon fluxes, in situ measurements
from eddy covariance networks are needed. However, there are large
spatiotemporal gaps in the high-latitude eddy covariance network. Here
we used the relative extrapolation error index in machine learning-based
upscaled gross primary production as a measure of network
representativeness and as the basis for a network optimization. We show
that the relative extrapolation error index has steadily decreased from
2001 to 2020, suggesting diminishing upscaling errors. In experiments
where we limit site activity by either setting a maximum duration or by
ending measurements at a fixed time those errors increase significantly,
in some cases setting the network status back more than a decade. Our
experiments also show that with equal site activity across different
theoretical network setups, a more spread out design with shorter-term
measurements functions better in terms of larger-scale
representativeness than a network with fewer long-term towers. We
developed a method to select optimized site additions for a network
extension, which blends an objective modeling approach with expert
knowledge. Using a case study in the Canadian Arctic we show several
optimization scenarios and compare these to a random site selection
among reasonable choices. This method greatly outperforms an unguided
network extension and can compensate for suboptimal human choices.
Overall, it is important to keep sites active and where possible make
the extra investment to survey new strategic locations.