Reducing Uncertainties in Coupled Carbon-water Cycle Predictions in the Southeastern United States Using Perturbed Parameter Experiments and Machine Learning Approaches
Type of DegreeMaster's Thesis
Forestry and Wildlife Science
Restriction TypeAuburn University Users
MetadataShow full item record
Carbon-water cycle interactions are considerably uncertain in the climate models. Parameter sensitivity studies are extensively utilized to reduce uncertainty in hydrological predictions. However, because of its high computation requirement, parameter sensitivity study has limited in-roads in climate modeling applications. A new National Center for Atmospheric Research – National Ecological Observatory Network (NCAR-NEON) simulation system provides a renewed opportunity to investigate parameter sensitivity in climate modeling applications. Two main reasons are: (1) the NCAR-NEON tool is ultra-computationally efficient, and (2) the availability of quality controlled and high-frequency NEON observations. Using the NCAR-NEON system, we perform parameter perturbation experiments at three flux tower sites utilizing a subset of 30 parameters affecting carbon-water cycle processes. We use the result to develop a land surface model Gaussian process regression emulator. About 244,800 years of CLM5 simulations are realized, including 200 years of spin-up runs and four years of production runs for each of the 400 parameter sets and at three sites (3 sites x 400 parameter sets x 204 years). The best-performing parameter set is selected as the optimized one that improved CLM5 performance in simulating carbon-water cycle processes from 0.37 to 0.78 on a normalized score from 0 to 1. Sensitivity analysis shows that photosynthetic parameters significantly influence carbon-water cycle interactions. The optimized parameters were examined at a regional scale using a high-resolution CLM configuration and forcing data. Upscaling of point-scale calibration at the regional scale shows mixed results, suggesting the need for separate parameter calibration for each plant functional type distribution. In part 2, we developed a Gaussian process regression emulator to mimic the behavior of CLM5 and provide predictions based on given inputs. The NCAR-NEON system, although computationally efficient, faces limitations due to the initialization process and resource consumption, limiting thorough parameter exploration for better alignment with observations. To overcome this barrier, we train an emulator that requires no initialization and is fast and efficient in mimicking land surface process behavior. The emulator improved CLM5 performance from 0.78 to 0.84.