In the face of global climate change it is critical that accurate predictions can be made about future concentrations of CO2 in earth’s atmosphere. Heterotrophic soil respiration (HSR) constitutes 35% of global CO2 emissions. An improved capacity to model emissions specifically by ecosystem and temperature would help investigators to more accurately predict future atmospheric CO2 in order to craft public policy that curtails anthropogenic greenhouse gas emissions to a safe, scientifically supportable level. Current modeling assumes a flat value for HSR across ecosystems and time, despite the fact that HSR is known to increase with rising temperatures and vary with ecosystem. We address this shortcoming by predicting HSR with a consideration of temperature sensitivity (Ea) and the possibility of a positive feedback with global climate change using Machine Learning (ML) techniques. Additionally, we investigated the robustness of a novel factor- mycorrhizal fungal dominance- as a predictor of HSR and Ea. Predictions based on mycorrhizal fungal dominance may allow for large-scale models with improved ecosystem specificity. Project data was collected in the summer of 2017 at three locations in the eastern United States. It is both noisy and multi-dimensional, making it difficult to apply traditional statistical tools which assume that data conform to a specific distribution or have some amount of linearity. Thus, data-driven ML techniques interact with our multi-factor, highly correlated data to create more performant models than are currently available. Transformation of data by binning also yielded improved performance when a larger array of ML algorithms was applied. Ultimately, we were able to produce models that predicted intervals of Ea with accuracies, scored by percentage of correctly identified classes, ranging from 0.77 to 0.95 and which utilized HSR, mycorrhizal fungal dominance, Soil % Nitrogen, Carbon, and Water, and location information.
