Above-Ground Biomass Estimation: Methods Developed by the SMFM Project
Why Tropical Dry Forests Need Dedicated AGB Methods
Tropical dry forests cover vast areas of Latin America, Africa and Asia, yet they remain underrepresented in global carbon accounting. Their seasonal leaf loss, open canopies and heterogeneous structure make biomass retrieval harder than in humid rainforests. The SMFM Project therefore developed a dedicated workflow for above-ground biomass (AGB) estimation that combines multiple satellite data sources.

Unlike wet tropical forests, dry forests experience strong phenological cycles. A single satellite snapshot can therefore miss or overestimate woody material. The SMFM approach treats time as a key variable rather than a nuisance.
Optical Satellite Data for Canopy Cover and Phenology
Sentinel-2 and Landsat imagery form the optical backbone of the SMFM pipeline. Multispectral bands are used to derive vegetation indices such as NDVI and EVI, which track green-up and senescence across the year.
These indices are not converted directly into biomass. Instead, they feed a phenological model that separates green herbaceous material from persistent woody components. The SMFM project calibrates this separation with field plots in dry forest sites.
- Sentinel-2 at 10 m resolution for canopy dynamics.
- Landsat archive for long-term consistency.
- Vegetation indices corrected for soil background.
Radar and LiDAR for Structural Information
Optical data describe greenness but not volume. To capture woody structure, SMFM integrates Sentinel-1 C-band radar and, where available, GEDI LiDAR footprints. Radar backscatter responds to branches and trunks, while LiDAR provides canopy height samples.
The project uses a hierarchical fusion strategy. LiDAR footprints calibrate radar backscatter, and radar extrapolates structure across the landscape. This reduces dependence on sparse field inventories.
"Radar sees the wood, optics see the leaves, and together they see the forest," notes the SMFM technical brief.
Machine Learning Fusion and Validation
The final AGB map is produced by a random forest model trained on fused predictors. The SMFM workflow follows a fixed sequence of steps to ensure reproducibility.
- Harmonize all satellite bands to a common grid.
- Extract phenological metrics for each pixel.
- Fuse radar, LiDAR and optical features.
- Train and validate against field plots.
- Publish uncertainty layers alongside AGB.
Validation uses leave-one-site-out cross-validation, which is stricter than random splitting and better reflects real transferability.
Performance, Limits and Outlook
Across SMFM test sites, the fused approach achieved lower error than optical-only models. The table below summarizes typical results.
| Method | R² | RMSE (t/ha) |
|---|---|---|
| Optical only | 0.52 | 28.4 |
| Radar only | 0.58 | 25.1 |
| SMFM fusion | 0.71 | 18.7 |
Remaining challenges include cloud cover during the wet season and the scarcity of field plots in remote dry forests. Future SMFM work will test NISAR and BIOMASS mission data to further reduce uncertainty.
