Blog·

How to propagate measurement uncertainty into a landscape-scale biomass estimate

A wide view across a forested mountain ridge under a moody, cloud-streaked sky.

A biomass number without an error bar is a number a verifier can reject. If your project reports 142 Mg/ha above-ground biomass for a stratum and stops there, the next question in the review is always the same: how was that uncertainty estimated, and does it roll up correctly across the whole landscape? Here's the working method, the way a registry reviewer wants it built.

Where the error comes from

A landscape-scale AGB estimate carries uncertainty from several independent sources stacked on top of each other, and each layer has its own origin.

Start with plot measurement error: DBH tape slip, height estimation from the ground, crew-to-crew variation in how a borderline stem gets called. This is usually the smallest term if the plot network is well run, but it's rarely zero, and a registry reviewer expects to see it quantified, not assumed away.

Next comes allometric equation error, the gap between a tree's measured dimensions and its true dry-weight biomass. Published allometric equations carry their own published error terms (often a coefficient of variation in the 15-30% range depending on species and diameter class), and that error doesn't disappear when you apply the equation to a new stand. It propagates forward into every plot-level AGB value.

Then there's the regression or model error: the gap between what your canopy structure and spectral predictors say a pixel should hold and what the plot data actually measured there. This is the term that gets the most scrutiny in remote-sensing-based estimates, because it's the one unique to the modeling approach rather than inherited from field protocol. A regression that fits well in closed-canopy stands and poorly in young regrowth needs stratified error terms, not one pooled RMSE applied everywhere.

Last is spatial extrapolation error: the uncertainty added when you take a model fit on several hundred plots and apply it across tens of thousands of hectares the model never saw directly. Spatial autocorrelation in the residuals matters here. Pixels next to each other aren't independent observations, and treating them as if they were will understate your final confidence interval.

Running the propagation

The IPCC good practice guidance gives you two broad routes. Tier 1 error propagation combines variance terms by simple addition in quadrature, which is defensible when the sources are independent and the model is linear. In practice, biomass regressions against spectral and structural covariates are rarely linear across the full value range, and the error terms are rarely independent of each other, so Tier 2 (Monte Carlo) is the route most reviewers want to see for anything beyond a small, simple project.

A Monte Carlo propagation draws repeatedly from the distribution of each uncertainty source, the plot measurement error, the allometric CV, the regression residual distribution, and reruns the full AGB calculation thousands of times. The spread of outputs across those runs is your confidence interval, built from the actual joint behavior of the error sources rather than an assumption that they stack neatly. This is more computation, but it's the method that holds up when a second reviewer tries to reproduce your number.

What to put in the submission

A registry submission needs the mean AGB (or carbon stock) estimate, the 95% confidence interval expressed in absolute units and as a percentage of the mean, and a short account of which error sources were included. Leaving out the regression/model term because it's inconvenient to quantify is the single most common gap a technical reviewer flags, because it's usually the largest term in a remote-sensing-based estimate, not the smallest.

None of this is a substitute for a clean, well-documented field campaign. But the biomass layer itself, the spatial model that turns canopy structure and spectral response into an AGB surface, is what generates the regression error term you're propagating in the first place. An annually refreshed biomass layer built from wide-swath multispectral imagery gives you a consistent basis for that regression error across reporting cycles, instead of a new, incompatible error term every time the field campaign changes hands.

If your current error budget is missing a term, or built on a one-off field campaign that won't scale to the next reporting cycle, it's worth seeing what a repeatable biomass layer looks like for your landscape.