"Uncertainty" in a carbon report is usually one number, and it usually represents one of five distinct failure modes. They compound differently, they are detected differently, and only the first is reduced by collecting more data, which is where most of the improvement budget goes.
The Five
| Type | What it is | Reduced by |
|---|---|---|
| Data | Measurement and sampling error in the inputs | More or better measurements |
| Parameter | Uncertainty in fitted coefficients | Better calibration data |
| Structural | The model form is wrong | A different model |
| Transferability | Model applied outside its calibration domain | Local validation |
| Extrapolation | Applied beyond the observed range | Nothing — only bounded by honesty |
Why More Plots Only Fixes One
Adding plots reduces sampling error. It has no effect on the other four.
If the allometric equation is structurally unsuited to the growth form present, measuring a thousand trees with it produces a very precise wrong answer and precision without accuracy is worse than neither, because the narrow confidence interval implies a reliability that is not there.
This is the most common misallocation in carbon project budgets: money spent on the term that is easiest to quantify rather than the term that is largest.
Transferability Is the Under-Recognised One
A model calibrated on Amazonian forest applied to Kalimantan peat swamp is being transferred across a domain boundary. It may perform well. The point is that nobody has checked, and the uncertainty from that transfer appears in no standard calculation.
Detecting it requires local validation data — a modest set of local measurements compared against model predictions. Not enough to recalibrate, just enough to see whether the model is systematically high or low.
Projects that collect a small validation sample frequently discover a consistent bias they were otherwise propagating through every plot. It is among the highest-value small investments available in forest measurement.
Structural Error Hides From Statistics
A model can fit its calibration data well and have the wrong form. The fit statistics will look good, because they measure agreement with the data the model was fitted to.
The diagnostic is residual behaviour: if errors are systematically positive at one end of the range and negative at the other, the functional form is wrong regardless of how high the correlation is.
This matters in growth modelling especially. A curve fitted to juvenile growth data will fit that data beautifully and mis-specify the deceleration, because the deceleration was not in the data.
Extrapolation Cannot Be Fixed
Applying a model beyond its observed range produces uncertainty that cannot be quantified from within the model, because there is no data to quantify against.
The only honest responses are to avoid it, or to state clearly where it occurred and treat the resulting figures as indicative rather than accounted. Projecting a growth curve twenty years past the last measurement and reporting the result to three significant figures is a false-precision claim.
What to Do With This
For each model in the chain, ask which of the five dominates. In a typical ARR project the answer is structural or transferability error in the allometric model, not data error in the plots.
Then spend accordingly. A modest local validation exercise addressing transferability often buys more certainty than doubling the plot count and, unlike extra plots, it can reveal that the estimate was biased rather than merely imprecise.
Turn uncertainty into a plan
Sampling design and growth assumptions decide how much measured carbon survives conservative deduction. TREEO Carbon Simulator models sequestration scenarios and biomass growth before the field season.
