Results from the Reflex experiment

1
Results from the Reflex experiment

• Mathew Williams, Andrew Fox and the Reflex team

2
Reflex Objectives

• To compare the strengths and weaknesses of
  various model-data fusion techniques for
  estimating carbon model parameters and predicting
  carbon fluxes.
• To quantify errors and biases introduced when
  extrapolating fluxes (and related measurements)
  made at flux tower sites in both space and time,
  using earth observation data and models
  constrained by model-data fusion methods.

3
Participants
4
Protocol

• Inputs
• Daily meteorological drivers
• Initial C stocks
• Daily NEE (gaps) and LAI (sparse)
• Some synthetic, some observed
• A simple C model
• Outputs
• Full C flux and stock estimates with uncertainty
• Parameter estimates with uncertainty

5
Identifying sources of error

• Synthetic experiment - deals with observational
  and algorithmic error (and user error).
• Real experiment adds model error.

6
(No Transcript)
7
(No Transcript)
8
Algorithm mean cumulative NEE
9
Flux estimation - synthetic
Uncertainty on retrieval of cumulative/integrated
C dynamics. Time series of monthly means (shows
uncertainty between algorithms from range of
means) for deciduous (top) and evergreen (bottom)
synthetic experments.
10
Flux estimation - uncertainty
Range of confidence intervals on retrieval of
cumulative/integrated C dynamics. Time series for
deciduous (top) and evergreen (bottom) synthetic
experments.
11
Flux estimation - observed
12
Flux estimation
13
Other analyses

• Identification of parameter correlations from
  parameter error covariance matrices
• Eigenvector analysis
• Taylor diagrams (bias, phase, variability)
• Test C state dynamics with CLs
• Compare with gap-filled, use CLs

14
Questions

• Can parameters and their uncertainties be
  effectively determined?
• We show different levels of uncertainty (DC)
• Parameter figures, eigenvectors, CM (CT MvW)
• Synthetic v true comparison (ET)
• Can the full C cycle be described and forecast?
• CLs on predictions for all years, fluxes, stocks
  (ZL, AF)
• Taylor diagrams, Chi-squared test on years 1, 2,
  3 (TQ, DR)
• Gap filled estimates (AR)
• Cumulative uncertainty on NEE predictions (MW)
• GPP, Re, NEE predictions and uncertainty (AR)
• What is the relative contribution to errors from
  observations, algorithms and model structure?

15
Lessons for LSM calibration

• Synthetic studies can show how data density and
  error can contribute information
• A variety of DA methods show promise
• Best constrained parameters are not intuitive
• Difficult to identify model structural errors