Optimizing aboveground biomass estimation in novel restoration systems through remote sensing and field data fusion

Accurate assessment of aboveground biomass (AGB) stocks is critical for monitoring nature-based projects, yet field measurements and remote sensing each have limitations when applied independently. Field measurements are accurate but costly and often lack spatial representativeness, while global deep learning models offer broad coverage but can perform poorly in heterogeneous and novel restoration systems.

We propose a fusion model method that merges AGB estimates from a global Convolutional Neural Network (CNN) model with a field-driven layer generated by Gaussian Process Regression (GPR). The global CNN, trained on airborne and spaceborne LiDAR, ingests optical (Sentinel-2 at 10 m; Landsat-8/9 at 30 m) and SAR (Sentinel-1 and PALSAR-2) data to directly predict AGB. Field plots play a dual role: beyond validating the CNN, 6-fold cross-validation splits them into a calibration subset used as labels to train the GPR (with coordinates and Sentinel-2 spectral data as inputs for spatial generalization), and a validation subset evaluating the fused GPR-CNN map. The CNN and GPR estimates are combined through inverse-variance weighting, producing a recalibrated AGB map with explicit pixel-level uncertainty. We evaluated this approach across 62 secondary forest and agroforestry sites of varying ages (5–48 years) covering 268.8 ha in Pará State, Brazil, using 152 field measurement plots.

The global CNN model performed poorly on these heterogeneous sites (R² = 0.24, MAE = 42.32 tDM/ha, RMSE = 53.16 tDM/ha), reflecting training data dominated by mature forests. The fusion model substantially improved accuracy (R² = 0.45, MAE = 23.38 tDM/ha, RMSE = 30.03 tDM/ha). Even with only 25% of available field measurements (approximately 2% of total area), MAE decreased by 17% relative to the CNN model alone.

This approach offers a scalable, cost-effective solution for carbon stock monitoring in restoration and agroforestry projects.