(a) Time series of CO2 mole fractions, both simulated and observed, at this CarbonTracker observation site. In the top panel, measured mole fractions (open black circles) are plotted along with CarbonTracker simulated values (filled light blue circles). Any observations that CarbonTracker can not assimilate successfully are shown as filled red circles.
(b) The time series of model residuals--the difference between the simulated and measured mole fractions--shown with dark green filled circles. These residuals should be uncorrelated in time, unbiased (i.e., have a mean of zero), and distributed normally. Also shown in the lower panel is the imposed model-data mismatch ("MDM", orange lines), which in part defines the rejection criterion (see documentation). Any model first guess value which is more than three times the MDM away from zero, after accounting for potential adjustments to the simulated value due to optimizing fluxes, is rejected by the optimization system. Rejected values, if there are any, are shown with filled red circles.
(c) Summary statistics for model residuals from to . Residuals are defined as simluated values minus observed values. "Bias" is the mean of the residuals; "std. dev." is the standard deviation, and RMSE is root-mean-square error. All quantities are in ppm of CO2. The time period over which these quantities are computed is set by the graphs (a) and (b).
(d) Seasonal histograms of residuals
(d) Seasonal histograms of the residuals for this dataset. See caption for
top figure for the definition of residuals. The left panel collects
all residuals for each northern hemisphere summer (June through
September); the right panel is the northern hemisphere winter
(November through April). Residuals before 1 Jan 2001 are excluded
from this analysis to avoid an effect of CarbonTracker burn-in from a
poorly-known initial CO2 distribution.
The tan color shows
the histogram of the residuals themselves; the blue lines and
statistics shown in blue text are a summary of the residuals
interpreted as a normal distribution. For this assimilated dataset, the assumed model-data mismatch
averaged over the seasonal span is shown in green (lines and text). The vertical scales are relative,
determined by the number of observations and how tightly they are
grouped, with the area under the histogram forced to unity.