DAI Labs publishes paper on an AI biomass estimation model that transfers to new forests with minimal field data
DAI Labs has published a paper presenting a new framework for estimating forest above-ground biomass (AGB) from satellite imagery and AI, achieving high accuracy with only a small amount of local field data.
This proposed ground-calibration approach is already at work in the field: DAI Labs is currently providing biomass mapping and data analytics for a large-scale forest restoration project in Africa, supporting the restoration of more than 200,000 hectares of degraded forest.
Read more about this project → Blog:DAI Labs Selected by Carbon Ventures Africa to Pioneer High-Precision Biomass
Accurate, spatially continuous measurement of forest biomass underpins the monitoring, reporting, and verification (MRV) systems that national GHG inventories, REDD+, and carbon credit markets rely on. Yet this has long presented a fundamental trade-off. Field inventories are highly accurate but spatially sparse and costly to scale. Satellite data can cover broad areas but satellite biomass estimation is primarily limited by signal saturation in dense canopies. And existing models struggle to transfer from one region to another without extensive local retraining.
DAI Labs’ framework addresses this trade-off with a two-stage approach:
- A single global model: A convolutional neural network (CNN) is trained once on a fused stack of four satellite data sources (Sentinel-2 optical imagery, Sentinel-1 C-band SAR, ALOS-2 PALSAR-2 L-band SAR, and a digital elevation model for terrain), using biomass reference data from NASA’s GEDI (Global Ecosystem Dynamics Investigation) mission, spanning multiple regions worldwide, as ground truth. Rather than retraining for each new region from scratch, the model is designed to learn transferable structure-biomass relationships that are not tied to any single region.
- A lightweight field-calibration workflow: Instead of retraining the model for every new landscape, a small number of local field plots (a minimum of 50 per site) is used to fit a Random Forest correction followed by a polynomial bias correction. This adapts the global model’s predictions to the specific species composition and terrain of a given region using only a handful of ground measurements.
To test the model in the field, we benchmarked it against the ESA CCI Biomass product, a widely used global reference dataset, across three regions in Malawi (Perekezi, Ntchisi, and Dzalanyama).
- ESA CCI Biomass (existing global product): weak agreement with ground measurements across all three regions, with R² ranging from -1.53 to -0.31 and RMSE from 32.5 to 48.9 Mg/ha.
- DAI Labs’ uncalibrated global model: performed similarly poorly on its own, with R² ranging from -2.1 to 0.16 and RMSE from 27.3 to 50.6 Mg/ha, in some cases underperforming ESA CCI Biomass itself.
- DAI Labs, after field calibration: performance improved consistently across all three regions to R² = 0.80 to 0.83, with RMSE between 12.5 and 12.8 Mg/ha, a clear improvement over both the uncalibrated model and ESA CCI Biomass.
At Perekezi, for example, calibration reduced RMSE from 48.9 Mg/ha (ESA CCI Biomass) and 41.9 Mg/ha (our uncalibrated model) down to 12.5 Mg/ha, roughly a 74% reduction relative to ESA CCI Biomass.
These results demonstrate that even in regions with very limited ground data, combining a globally trained model with a small local field-calibration effort can produce biomass maps accurate enough to support policy decisions and carbon credit issuance. This makes the approach particularly valuable in regions where field infrastructure and resources are limited, with applications extending beyond carbon accounting and REDD+ to forest management, biodiversity assessment, disaster risk (fuel load) evaluation, and agriculture and infrastructure monitoring.
DAI Labs is applying the AI capabilities developed through its custom AI development work to its own satellite-and-AI-based geospatial products, and the methodology presented in this paper forms a core part of that technical foundation.
Read the full paper here: Transferable Above-Ground Biomass (AGB) Estimation Model from Multi-Sensor Data with Sparse Field Calibration
Authors: Pann Thinzar Seint, Bryan Atwood, Subas Chhatkuli (DAI Labs, K.K.)
