A process-informed deep learning biosphere model for ecosystem carbon flux partitioning and beyond

Abstract. Ecosystems exchange carbon dioxide with the atmosphere through photosynthesis, which removes carbon dioxide, and respiration, which releases it. Eddy covariance flux towers measure only the net ecosystem exchange (NEE), which must be partitioned into gross primary production (GPP) and ecosystem respiration (R_eco) to understand ecosystem carbon cycling. This partitioning problem is structurally underdetermined and is traditionally resolved by fitting simplified, pre-defined functional forms separately to daytime and nighttime data, repeatedly re-optimized in short time windows and assuming spatially homogeneous land cover within the tower footprint.

We present pyVPRNN, a process-informed deep learning framework that unifies partitioning, mechanistic interpretation, and spatial upscaling within a single, jointly-trained model. High-resolution Sentinel-2 satellite observations of vegetation indices and land cover are combined with half-hourly flux footprints, to train a single model that resolves heterogeneous vegetation density and mixed land-cover classes within the footprint and learns the relationships governing photosynthesis and respiration directly from data, rather than prescribing them.

Applied to observations from 32 European eddy-covariance stations, pyVPRNN exceeds the performance of established daytime-based partitioning at most sites, and avoids uncertainties and biases that arise when nighttime-based partitioning is extrapolated to daytime conditions. Unlike the traditional partitionings, pyVPRNN provides a single, temporally consistent model rather than a sequence of independently re-fit and potentially inconsistent curves, at some cost to short-term adaptivity at frequently disturbed or managed sites. pyVPRNN performs particularly well at forest sites and undisturbed grasslands, and its spatiotemporal design further enables separation of the underlying ecosystem signal from footprint-driven measurement variability.

Using simulation experiments and observations from two eddy covariance sites, we further show that explainable AI techniques applied to pyVPRNN recover ecologically meaningful responses of photosynthesis and ecosystem respiration to environmental drivers such as temperature. Because its inputs consist solely of satellite observations, land cover, and meteorology, without any footprint- or tower-specific information at inference time, the model also provides a direct route to spatial upscaling of carbon fluxes beyond the tower footprint.

pyVPRNN is implemented within the open-source pyVPRM framework, enabling convenient application and extension across sites and biomes.