New paper: Sequential deep learning models offer only modest gains for upscaling land–atmosphere fluxes. The main bottlenecks are spatial sampling bias, which amplifies regional uncertainties and affects model sensitivity, and feature selection, which has large leverage on results.
