Validating Satellite Biomass Estimates with Field Plots

Satellites do not measure carbon. They measure reflectance, backscatter and canopy height, and those quantities are converted to biomass by a statistical relationship built from ground measurements. That relationship is the weakest link in most remote-sensing carbon estimates, and it is the first thing a verifier examines.

Optical imagery records reflected light across spectral bands. It is excellent for detecting land-cover change and disturbance, and poor at estimating biomass directly, because reflectance saturates once canopy closes, so a moderately stocked forest and a heavily stocked one can look identical.

Synthetic aperture radar records backscatter, which relates to structure and moisture and penetrates cloud. That last property matters enormously in the tropics, where optical coverage is frequently unusable for months. SAR also saturates, at higher biomass levels than optical but still within the range of mature tropical forest.

LiDAR records height and vertical structure directly, and correlates far better with biomass than either of the above. Airborne acquisition is expensive; spaceborne LiDAR samples rather than maps continuously.

None of them observes a tonne of carbon. All of them require calibration.

The standard approach fits a model between sensor values and field-measured biomass at co-located plots, then applies that model across the wider area.

Three failure modes recur.

Extrapolation beyond the calibration range. A model fitted on plots spanning 50–200 t/ha will produce numbers for a 400 t/ha stand, and those numbers are extrapolation, not measurement. Verifiers ask for the calibration range; estimates outside it attract conservative treatment.

Transfer to a different forest type. A relationship built in lowland dipterocarp forest does not hold in mangrove or peat swamp. Structure, wood density and moisture all differ, and the model has no way to know.

Plot-pixel mismatch. A field plot may be a fraction of a pixel or span several. Geolocation error, edge effects and differing footprints all inject noise that is easy to overlook and hard to remove afterwards.

This distinction is where many projects lose credibility.

Calibration uses plots to build the model. Validation uses plots the model has never seen to test it. Reporting a fit statistic from the calibration data alone says almost nothing about predictive performance, as it merely describes how well the model memorised its training set.

Independent validation plots, held out from fitting, are what let you state an honest error. Where sample size is tight, cross-validation is an acceptable substitute, but it must be described as such.

Because plots serve two purposes, which are producing the estimate and calibrating the extrapolation, the sampling design must satisfy both.

That means covering the full range of conditions present, not just the convenient ones. A calibration set drawn only from accessible, uniform stands will not constrain the model where the area is most variable, which is exactly where the extrapolated estimate matters most.

It also means recording plot geometry and geolocation precisely enough to match sensor footprints, and recording measurement dates close enough to image acquisition that growth and disturbance do not confound the relationship.

Persistent cloud cover across much of the archipelago makes optical-only monitoring unreliable for large parts of the year, which is why SAR features heavily in tropical monitoring designs.

Peat and mangrove systems compound the problem in a different way: a large share of the carbon is below ground, entirely invisible to every sensor listed above. In those ecosystems, remote sensing can constrain above-ground change while saying nothing about the dominant pool, which is a limitation that must be stated rather than glossed.

Hero image: Birdseye view flight from Kuching to KLIA by Marufish, CC BY-SA 2.0, via Wikimedia Commons.

1. IPCC 2006 Guidelines Vol. 4 (AFOLU) — https://www.ipcc-nggip.iges.or.jp/public/2006gl/vol4.html (accessed 16 Sep 2026)

2. IPCC 2013 Wetlands Supplement — https://www.ipcc-nggip.iges.or.jp/public/wetlands/ (accessed 16 Sep 2026)

3. Kauffman & Donato (2012), CIFOR Working Paper 86 — https://www.cifor-icraf.org/knowledge/publication/3749/ (accessed 16 Sep 2026)

4. Verra methodologies and monitoring requirements — https://verra.org/methodologies/ (accessed 16 Sep 2026)