Transcript Towards a robust, generalizable non

Towards a robust,
generalizable non-linear
regression gap filling
algorithm (NLR_EM)
Ankur R Desai – National Center for Atmospheric Research (NCAR)
Boulder, Colorado, USA
University of Wisconsin, Atmospheric & Oceanic Sciences, Madison, Wisconsin, USA
Pennsylvania State University, Meteorology, University Park, Pennsylvania, USA
Bruce D Cook – University of Minnesota, Forest Resources
St. Paul, Minnesota, USA
Kenneth J Davis – Pennsylvania State University, Meteorology
University Park, Pennsylvania, USA
Gap Filling Workshop
18 Sept 2006
Max-Planck BGC, Jena, Germany
Goals







Simple, fast, general GPP/RE and gap-filling
estimation for eddy flux NEE
Theoretically meaningful parameters
Statistically valid regression
Flexible moving window regression
Temperature / PAR forcing only
Used at ChEAS Ameriflux sites
Code written in IDL, available to all
Primary references



Desai, A. R., P. Bolstad, B. D. Cook, K. J. Davis, and E. V.
Carey, 2005: Comparing net ecosystem exchange of carbon
dioxide between an old-growth and mature forest in the
upper Midwest, USA. Agric.For.Meteorol., 128, 33-55 (doi:
10.1016/j.agrformet.2004.09.005).
Cook, B. D., K. J. Davis, W. Wang, A. R. Desai, B. W.
Berger, R. M. Teclaw, J. M. Martin, P. Bolstad, P. Bakwin,
C. Yi, and W. Heilman, 2004: Carbon exchange and venting
anomalies in an upland deciduous forest in northern
Wisconsin, USA. Agric.For.Meteorol., 126, 271-295
(doi:10.1016/j.agrformet.2004.06.008).
Eyring, H., 1935: The activated complex in chemical
reactions. J.Chem.Phys., 3, 107-115.
Sites that use NLR_EM

http://cheas.psu.edu

Sites that use it: Sylvania Old-growth, Lost Creek wetland, WLEF 447-m
tall tower (3 levels), Willow Creek upland. Others only site-to-site comp.
Algorithm highlights



1 parameter set per day for ER and GPP
30-60 day moving window: size increases until
200 good half-hourly points – all user definable
One-tailed t-test for parameter fit
if confidence < 0.90, replace ER/GPP with daily
mean ER/GPP over window
 mostly occurs in winter


Monte-Carlo random gap generator to compute
sensitivity of filling to gaps - reported for sites
Respiration


Use nighttime u*-screened NEE and 5 cm soil
temperature (can use air temp instead)
Regress against Eyring equation:


similar to Arrhenius but more accurate description
of reaction activation energies by including entropy
For regression, total carbon content not needed
Respiration

Gibbs free energy:

Regress with linear form of equation:
Example of ΔG++

From Cook et al (2004)
GPP

Simple 2 or 3 parameter equation:
Can relate b1/b2 to Amax and quantum yield
 b3 can be included as an intercept



T-test failure replaces b1 with mean GPP
Levenberg-Marquardt non-linear regression
Example of b1/b2

From Cook et al (2004)
Error estimation



Simple Monte Carlo estimate of error induced
by gap-filling
100 sets of random 10-40 artificial gaps of
lengths 30 minutes – 5 days
Error reported as standard deviation and range
of NEE across 100 sets (e.g., Desai et al., 2005;
Desai et al., in press, Ag. For Met.)
Other notes

Can use filled or non-filled met data
ChEAS filled met relies on cluster of met sites
 Without filled met data, mean diurnal values and
interpolation used to fill either met or flux





Can use other ER/GPP equations
Day/night determined by sunrise/set at lat/lon
and by a low PAR criterion
Algorithm has been used across Ameriflux in a
Modis GPP – flux tower evaluation project
Filling NEE = ER - GPP
Next steps

Harder to use well in non-temperate sites
Exploration of phenologically controlled windows
 Not good for moisture-limited sites



Cross-site Gibbs free energy comparisons
require total carbon content, but has promising
use for examining ER parameter spatial var.
Interested in understanding model bias in gap
filling and exploring energy activation across
sites used in this study
Advertisement

GPP/ER intercomparison
Tuesday, 13:00
 Most gap-filling methods can produce GPP/ER


How variable are they?




Across methods / sites
Due to gaps in NEE
Can we infer ecosystem parameters?
I have most of these data, but not all
Send them in to me
 or else
