COS 2-2 - Iteratively forecasting near-term terrestrial fluxes across a network of eddy-covariance towers

Monday, August 15, 2022

1:45 PM – 2:00 PM EDT

Presenting Author(s)

AH

Alexis Helgeson

Doctoral Student
Boston University, United States

Coauthor(s)

KZ

Katherine Zarada

Boston University, United States
CR

Cameron Reimer

Undergraduate Research Assistant
Boston University
Brookline, Massachusetts, United States
DZ

Dongchen Zhang

Boston University, United States
AD

Ankur Desai, PhD

Professor
University of Wisconsin, United States
BM

Bill Munger

Harvard University, United States
CF

Christina Francis

Columbia Univeristy, United States
XW

Xaun Wilson

Howard University, United States
MD

Michael C. Dietze

Boston University
Boston, Massachusetts, United States

Background/Question/Methods

There is a growing desire for accurate near-term forecasts of the carbon and water cycles. We used NOAA GEFS weather forecasts to predict carbon and water cycle pools and fluxes 35 days into the future using the Simplified Photosynthesis and Evapotranspiration Model (SIPNET). The forecast system then statistically assimilates net ecosystem exchange (NEE), latent heat (LE), and MODIS LAI, adjusting the model state variables used as initial conditions for the subsequent forecasts. We tested this forecast system at ten eddy-covariance sites that encompass a range of ecosystem types (forest, grassland, agriculture, semi-arid shrubland). Finally, we assessed the predictability of NEE, LE, and LAI as a function of forecast lead time, time of day, and season.

Results/Conclusions

Predictability was assessed by calculating forecast RMSE, Bias, and CRPS scores for NEE, LE, and LAI compared to observed values. The predictability of NEE, LE, and LAI all decreased as forecast lead time increases, but improved during the summer and winter months compared to fall and spring. Tower fluxes had lower RMSE and CRPS scores during the day compared to the night by a factor of 4 and 3 respectively. While bias oscillated from slightly positive during the night to slightly negative during the day. Overall error declines, as lead time increases, with increased numbers of data constraints. Building upon this initial analysis we will extend our prototype system to assimilate more data streams across a larger number of forecast locations.