全球变化知识资源中心

The terrestrial biosphere strongly modulates atmospheric CO2 mixing ratios, whose inexorable rise propels anthropogenic climate change. Modeling and mechanistically understanding C uptake by the terrestrial biosphere are thus of broad societal concerns. Yet despite considerable progress, scaling up point observations to landscape and larger scales continues to frustrate analyses of the anthropogenically perturbed global C cycle. While that up-scaling is our overarching motivation, here we focus on one of its elements, modeling C uptake at a given site. We devise a novel artificial neural network (ANN)-based model of C uptake at Harvard Forest that combines locally observed and remotely sensed variables. Most of our model predictors are those used by an established ecosystem C uptake model, the Vegetation Photosynthesis and Respiration Model (VPRM), easing comparisons. To those, we add observed cumulative antecedent precipitation and soil temperature. We find that model errors are much larger in winter, indicating that better understanding and modeling of respiration will likely discernibly improve model performance. Comparing the ANN and VPRM results reveals errors attributed to unrealistic treatment of temperature in the VPRM formulation, indicating that better representation of temperature dependencies is also likely to enhance model skill. By judiciously comparing VPRM and ANN errors we thus overcome ANNs' notoriety for concealing the mechanisms underlying their predictive skills. We demonstrate their ability to identify outstanding ecosystem science knowledge gaps and particularly fruitful corresponding model development directions, improving site specific and up-scaling flux modeling and understanding of the climate impacts of the northern forest.

Plain Language Summary Anthropogenic climate change due to atmospheric carbon dioxide (CO2) buildup reflecting imbalances between emissions and uptake is a key challenge. Unfortunately, understanding photosynthetic C uptake by land biomes, our focus, is incomplete. The central source of insights into this uptake is a network of exchange measuring towers. However, this sparse network undersamples the ecosystems they strive to represent. Better understanding land C uptake thus depends on vegetation models that can upscale local C uptake to regional and global scales. To this end, we develop an artificial neural network (ANN) model of C uptake in a northern mixed forest measurement site in Massachusetts which reduces errors by the equivalent of the emissions of 5.5 million Americans. This ANN can be readily applied to any observed terrestrial C uptake record at a particular biome, or to all simultaneously. Coupling it to a geographically explicit model for filling spacetime gaps thus achieves the quest for a spatiotemporally complete land C uptake modeling. Our results can also guide future model development by identifying error sources. For example, we identify the representation of respiration and temperature dependence as promising avenues for future research, demonstrating the unique role of ANNs in improving land C uptake modeling.