Estimating Above-Ground Biomass from Remote Sensing

Updated 2026-11-02 · 4 min read · By the akar.earth specialist team

Remote sensing biomass estimation explained: which sensors work, why field plots are needed, how saturation and error behave in Indonesian forests.

Remote sensing biomass estimation predicts above-ground biomass (AGB) by training a statistical or machine-learning model on field plots, then applying it to satellite or airborne data across the whole project area. The satellite never weighs a tree: the accuracy of the map depends on the field plots, the allometric equation behind them, and how well the sensor responds to biomass in your forest type.

How the estimate is built

Every credible remote sensing biomass workflow follows the same chain. Break any link and the uncertainty grows.

1. Measure trees in field plots. Diameter at breast height (DBH), height on a subsample, species for wood density. See plot sampling design.
2. Convert trees to biomass with an allometric equation, for example the Chave et al. (2014) pantropical model using diameter, height and wood density.
3. Sum to plot AGB in tonnes of dry matter per hectare, then apply a carbon fraction and the 44/12 ratio to get tCO2e.
4. Extract remote sensing predictors for each plot: spectral bands and indices, radar backscatter, canopy height, texture.
5. Fit and validate a model (regression, random forest, gradient boosting) with held-out plots or cross-validation.
6. Predict wall to wall, then report the total and its uncertainty, not just a map.

The plots must be accurately geolocated. A plot that sits 15 m off its true location can be matched to the wrong pixels and degrade the whole model. That is why we use RTK GNSS for plot centres.

Which sensors respond to biomass

Optical data from Sentinel-2 is the workhorse for land cover, but its relationship with biomass flattens as canopies close. In lowland dipterocarp forest in Kalimantan or Sumatra, two stands with very different biomass can look nearly identical in reflectance. Structure-sensitive data, such as GEDI lidar and canopy height models, is what breaks that saturation.

Choosing the approach for your project type

ARR (tree planting on non-forest land). Biomass starts near zero and grows slowly. Optical indices track early canopy closure well, which is why VM0047’s area-based approach uses a remotely sensed vegetative stocking index with matched control plots. But the credited tonnes still rest on field measurement. See the VM0047 area-based approach.

REDD+ and IFM (standing forest). Carbon stocks are high and the main question is loss or degradation. Here, remote sensing is better at mapping activity (clearing, logging roads, fire scars) than at mapping stock. Stocks per stratum usually come from plot inventories, and activity data from change detection.

Agroforestry and smallholder mosaics. Plots are small, often under 1 ha, and mixed with crops. 10 m pixels mix tree and non-tree cover. Drone imagery or tree-level field measurement is often the only way to get reliable per-farm estimates.

Mangroves. Above-ground biomass maps exist, but most mangrove carbon sits in the soil. Remote sensing does not see it. See mangrove soil carbon.

Common mistakes

• Using a global biomass map as the answer. Global products are useful priors but carry large local errors and are not calibrated to your forest type.
• Too few plots, or plots only near roads. The model learns road-edge forest and predicts the interior badly.
• Ignoring saturation. A model trained on degraded forest will underpredict intact forest, and vice versa.
• Mismatched dates. Plots measured in 2026 matched to 2022 imagery in a landscape with active clearing.
• Reporting pixel accuracy, not total uncertainty. Verifiers care about the confidence interval on project-level carbon, which needs proper propagation. See uncertainty deductions.
• No independent validation. Training and testing on the same plots inflates apparent accuracy.

What good evidence looks like for a verifier

A validation and verification body (VVB) will want to retrace the chain from tree to total. Prepare:

• A plot register with RTK coordinates, dates, field crew and raw tree data.
• The allometric equation, wood density source and carbon fraction, with justification.
• Imagery IDs and acquisition dates, pre-processing steps (cloud masking, terrain correction) and code versions.
• Model specification, training/validation split and accuracy metrics (RMSE, bias) in tonnes per hectare.
• An uncertainty calculation at the stratum and project level, matching the methodology’s required confidence level.
• Maps and tables that reconcile: the sum of the map equals the reported total.

Indonesia and Southeast Asia specifics

Persistent cloud over Sumatra, Kalimantan and Papua limits optical time series, so radar and multi-date compositing matter. Steep terrain in Sulawesi and Papua distorts radar and needs terrain correction. Peat swamp forest has distinctive structure and needs its own stratum. And smallholder landscapes in Java, Bali and Nusa Tenggara change parcel by parcel, which favours finer resolution.

If you need a biomass model that a VVB can audit, our carbon modelling service combines field campaigns with remote sensing. The wider workflow is covered on the dMRV hub.