From data to insights: Upscaling riverine GHG fluxes in Germany with machine learning.

Global fluvial ecosystems are important sources of greenhouse gases (CO, CH and NO) to the atmosphere, but their estimates are plagued by uncertainties due to unaccounted spatio-temporal variabilities in the fluxes. In this study, we tested the potential of modeling these variabilities using several machine learning models (ML) and three different input datasets (remotely sensed vegetation indices, in-situ water quality, and a combination of both) from 20 headwater catchments in Germany that differ in catchment land use and stream size. We also upscaled fluvial GHG fluxes for Germany using the best ML model and explored the role of catchment land use on the GHG spatial-temporal trends. Model performance depended on the choice of ML model, input data and GHG type. Complex decision-tree-based models better predicted GHG concentrations and fluxes than other ML model types (r = 0.33 to 0.72). Our upscaled fluxes from catchment scale remotely sensed vegetation indices showed that total annual riverine CO equivalent fluxes from 2934 catchments in Germany ranged from 1.7 to 96.4 kg m yr (mean ± SE: 23.2 ± 0.001). The highest fluxes came from urban and intensively cropped catchments, while lower fluxes came from extensively cropped, forestry, and pasture-dominated catchments. Our study demonstrates that spatially and temporally resolved catchment vegetation indices from remotely sensed data in conjunction with machine learning models can be applied to upscale all three GHG concentrations and fluxes from diverse catchments, revealing important spatio-temporal trends associated with catchment land use.