Downscaling of Past and Future Climate Forcing
Authors
Utkarsh Mital*, Dipankar Dwivedi, Helen Weierbach, Carl Steefel (CISteefel@lbl.gov)
Institutions
Earth and Environmental Sciences Area, Lawrence Berkeley National Laboratory, Berkeley, CA
URLs
Abstract
An accurate characterization of snowpack water content is necessary to quantify water availability. Lidar technology can quantify snowpack at high resolution scales of ~1m and finer but the frequency of these observations is very low. In this study, scientists developed a machine learning framework based on random forests, which is capable of estimating snowpack at time points when lidar observations via Airborne Snow Observatories are not available. The framework approximates the physical processes governing accumulation and melt, as well as snowpack physical characteristics. Fifteen different variables were used, derived from precipitation, temperature, surface reflectance, elevation, and canopy. The framework was implemented in the Rocky Mountains of Colorado and the snowpack estimates were found to be more accurate than those of the Snow Data Assimilation System (SNODAS), a respected industry standard. Specifically, the coefficient of determination using the new framework was 0.57, which is a significant improvement from 0.13 obtained using SNODAS. Additionally, the effects of different types of variables for modeling snowpack were investigated by developing three different models. Precipitation and temperature were found to be more important than surface reflectance. This research provides a way to expand the applicability of costly lidar data, which helps to improve snowpack estimation that is critical for water resource management (Mital et al. 2022).
A second effort involves taking future climate data from the Coupled Model Intercomparison Project 6 (CMIP6) to model future global climate using shared socioeconomic pathways (SSPs). Five different SSP scenarios project global warming ranging from 1.4°C (best) to 2.4°C (worst) by the end of the century. Although the CMPI6 model resolution is >100 km with daily time data, higher resolution CMPI6 downscaled datasets do exist, such as the NASA Downscaled CMPI6 dataset with 25-km spatial resolution. However, the integrated surface-subsurface Advanced Terrestrial Simulator (ATS) model requires downscale model outputs <1 km. Using the NASA Downscaled CMPI6 dataset, a modified U-Net architecture is being developed to learn statistical relationships between coarse resolution (25 km) and fine resolution climate (800 m). Efforts are underway to incorporate these high-resolution climate data into Amanzi-ATS to understand future water availability in the Rocky Mountains of Colorado.
References
Mital, U., et al. 2022. “Modeling Spatial Distribution of Snow Water Equivalent by Combining Meteorological and Satellite Data with Lidar Maps,” Artificial Intelligence for the Earth Systems 1(4), e220010.