Upscaling Net Ecosystem Exchange Over Heterogeneous Landscapes With Machine Learning

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