Large-scale regionalization of water table depth in peatlands optimized for greenhouse gas emission upscaling

Abstract. Fluxes of the three main greenhouse gases (GHG) CO₂, CH₄ and N₂O from peat and other soils with high organic carbon contents are strongly controlled by water table depth. Information about the spatial distribution of water level is thus a crucial input parameter when upscaling GHG emissions to large scales. Here, we investigate the potential of statistical modeling for the regionalization of water levels in organic soils when data covers only a small fraction of the peatlands of the final map. Our study area is Germany. Phreatic water level data from 53 peatlands in Germany were compiled in a new data set comprising 1094 dip wells and 7155 years of data. For each dip well, numerous possible predictor variables were determined using nationally available data sources, which included information about land cover, ditch network, protected areas, topography, peatland characteristics and climatic boundary conditions. We applied boosted regression trees to identify dependencies between predictor variables and dip-well-specific long-term annual mean water level (WL) as well as a transformed form (WL_t). The latter was obtained by assuming a hypothetical GHG transfer function and is linearly related to GHG emissions. Our results demonstrate that model calibration on WL_t is superior. It increases the explained variance of the water level in the sensitive range for GHG emissions and avoids model bias in subsequent GHG upscaling. The final model explained 45% of WL_t variance and was built on nine predictor variables that are based on information about land cover, peatland characteristics, drainage network, topography and climatic boundary conditions. Their individual effects on WL_t and the observed parameter interactions provide insight into natural and anthropogenic boundary conditions that control water levels in organic soils. Our study also demonstrates that a large fraction of the observed WL_t variance cannot be explained by nationally available predictor variables and that predictors with stronger WL_t indication, relying, for example, on detailed water management maps and remote sensing products, are needed to substantially improve model predictive performance.