From Snapshot Maps to Continuous Monitoring of Global Forest Carbon at 100 m Resolution (2000–2025)

Yan Yang, Ricardo Dalagnol, Wentao Lin, Nicholas Kwon, Zhihua Liu, Sassan Saatchi, Le Bienfaiteur Sagang

Accurate, high-resolution estimation of forest aboveground biomass (AGB) is critical for quantifying terrestrial carbon stocks and informing climate mitigation policy, yet existing global products remain constrained by coarse spatial resolution and limited temporal coverage that obscure sub-national-scale disturbances (Avitabile et al., 2016). This study presents a wall-to-wall global AGB product at 100 m native spatial resolution spanning 2000–2025, trained with a DenseNet architecture using independent AGB reference labels from ALS-derived AGB, field-plot and forest-inventory samples from the research community and selected regional/national inventories, and mangrove AGB training samples from global coastal regions. The predictor stack integrates Landsat surface reflectance time series (Chander et al., 2009; Roy et al., 2014; Zhu & Woodcock, 2012), ALOS PALSAR-1/2 L-band backscatter, GEDI and ICESat-2 canopy-height metrics used only as ancillary structural layers where valid 100 m aggregates are available, and topographic ancillary data (European Space Agency, 2021), employing spatial context windows (Kattenborn et al., 2021) to capture neighborhood-scale biomass gradients. Temporal harmonization is achieved through a buffered mask-aware correction strategy that excludes externally identified disturbance and activity areas from stable-pixel calibration, thereby preserving disturbance and regrowth signals while correcting broad cross-year biases. On a sample-level held-out validation/test set of approximately 1.01 million pixels compiled across eco-regions, the product achieves R² = 0.741, RMSE = 59.5 Mg ha⁻¹, and bias = −4.82 Mg ha⁻¹. Multi-decadal RGB change composites reveal spatially coherent carbon-dynamic patterns, including disturbance and recovery trajectories in southeastern Amazonia and regional biomass contrasts across major forest domains. These findings demonstrate that deep learning fusion of optical and SAR predictors with inventory-calibrated AGB reference labels and ancillary canopy-structure layers can support sub-kilometer global forest carbon monitoring, while spatial-block validation and public release of the processing provenance remain necessary before the product is used for formal reporting. A more detailed research paper will be published soon.