Blog·

Model-based vs design-based inference for carbon stock estimation

Sunlight filtering down through a dense green hardwood forest canopy.

If you've sat through a methodology review and watched a project proponent get asked to justify their variance estimator, you already know this isn't an academic distinction. It decides whether an uncertainty figure in a monitoring report holds up or gets bounced back for revision.

Two different starting assumptions

Design-based inference treats the biomass on the ground as a fixed, unknown quantity. The randomness lives in how you selected your plots. If your sample is a probability sample, with known inclusion probabilities, you can use classical estimators (Horvitz-Thompson and its relatives) to get an unbiased estimate of the population total and a defensible variance, all without assuming anything about the relationship between biomass and any auxiliary variable. This is the approach foresters trained on sampling theory tend to reach for first, because it asks so little of you. No model to validate, no residual structure to argue about. Just a sound sampling frame and honest field measurement.

Model-based inference flips the randomness. Here you assume the biomass values themselves are a realization of some stochastic process, typically a regression relating canopy structure or spectral response to above-ground biomass at your field plots, then applied across the landscape. The uncertainty you report comes from the model's prediction error, not from how the plots were drawn. This is what lets a project generate a wall-to-wall biomass layer from imagery instead of a design-weighted total confined to the sampled locations. It's also what a reviewer should be checking hardest, because the variance estimate is now contingent on the model being correctly specified. A biased regression gives you a confident-looking number that's wrong in a way design-based estimators are structurally protected against.

Where this actually bites in a submission

The practical friction shows up in three places reviewers flag repeatedly.

First, pseudo-replication. A model fit on 40 plots and projected over 200,000 hectares has an effective sample size nowhere near what the landscape area suggests. A model-based variance that doesn't account for spatial autocorrelation in the residuals will understate uncertainty, sometimes badly.

Second, extrapolation beyond the range of the calibration data. If your field plots sit in the 50 to 150 Mg/ha range and the model is asked to predict stands at 300 Mg/ha, that's extrapolation, and the model-based variance formula generally doesn't warn you about it on its own. A reviewer has to check the covariate range by hand.

Third, model transportability across cycles. A regression calibrated against one year's plot campaign and reused on the next year's imagery without recalibration is a different statistical object than one refit each cycle, even if the methodology writeup doesn't flag the distinction.

The hybrid middle ground

In practice, most credible landscape-scale carbon accounting doesn't pick one camp outright. Model-assisted estimators, the generalized regression estimator is the classic example, use a model as an auxiliary tool to reduce variance while keeping the formal guarantees of the design-based framework. You still need a probability sample as your backbone, but the model (often built from remote sensing covariates) does the work of extending inference between plots. This is the version most registries are comfortable with, because it inherits design-based rigor for the variance claim while getting the coverage benefit of a model-based surface.

If you're the one evaluating a submission, the question to ask isn't "model-based or design-based," it's whether the proponent's variance estimator actually matches the inferential framework they claim to be using. A lot of reporting error comes not from bad models but from a design-based plot network bolted onto a model-based variance formula with nothing linking the two.

Project teams building out a biomass layer each reporting cycle often land here because refitting the regression and regenerating the surface every cycle, rather than reusing last year's model on new imagery, is one of the more straightforward ways to keep the model-based component honest. Biomass Mapping regresses canopy structure and spectral response to above-ground biomass from wide-swath multispectral imagery and refreshes the layer annually, which at least removes "stale model, new year" as a source of the drift reviewers end up chasing.

If your reporting cycle is due for a biomass layer that gets rebuilt rather than recycled, it's worth a look.