Key finding
The team combined deep learning on Sentinel-1 and Sentinel-2 imagery (Attention U-Net, 85% overall accuracy) with aboveground biomass data from ESA-CCI and the Hansen forest loss dataset to produce the first spatially explicit, crop-specific carbon emission factors for Colombia's five major biomes. The results reveal a deforestation emissions profile more nuanced than the simple pasture-dominates narrative.
Pasture expansion for livestock is the largest driver nationally, accounting for 65.1% of total emissions (90.57 MtC from 1.22 million hectares). But its dominance varies sharply by biome: in the Amazonas and Orinoquia regions, pasture accounts for over 84% and 89% of emissions respectively, while in the Andes region cocoa is the largest driver at 26.9%. Cocoa and coffee, despite their relatively small cultivation areas, exhibit surprisingly high emission factors (81.2 and 64.3 MgC per hectare respectively) because they expand into high-biomass forests. Smallholder croplands contribute up to 16.7% of regional emissions. Oil palm, often singled out in policy debates, plays a modest role nationally at 2.2%.
The pre-conversion carbon stocks vary substantially by biome. The Pacifico region has the highest at 101 MgC per hectare, followed by the Andes at 94, the Amazonas at 92, the Orinoquia at 83, and the Caribe at 67. Residual carbon after deforestation is consistently low across all biomes at 12 to 22 MgC per hectare, confirming substantial carbon depletion regardless of post-deforestation land use. Some of the highest crop-specific emission factors occur when crops expand into the most carbon-dense forests: rubber in the Pacifico reaches 123 MgC per hectare, and coffee in the Amazonas reaches 92 MgC per hectare.
What the research asks
Cocoa, coffee, pasture, oil palm: each commodity driving deforestation in Colombia has a different carbon footprint per hectare, but most carbon accounting still relies on generic IPCC Tier 1 default emission factors that ignore biome-specific variation and commodity-specific differences. The result is that carbon emission estimates for commodity-driven deforestation have high uncertainty, making it hard for companies, governments, and carbon markets to know whether their emissions figures are accurate. This study asked whether satellite imagery and deep learning could produce spatially explicit, commodity-specific carbon emission factors for Colombia's five major biomes, moving from generic default values to data-driven, locally relevant estimates.
What it finds
The team combined deep learning on Sentinel-1 and Sentinel-2 imagery (Attention U-Net, 85% overall accuracy) with aboveground biomass data from ESA-CCI and the Hansen forest loss dataset to produce the first spatially explicit, crop-specific carbon emission factors for Colombia's five major biomes. The results reveal a deforestation emissions profile more nuanced than the simple pasture-dominates narrative.
Pasture expansion for livestock is the largest driver nationally, accounting for 65.1% of total emissions (90.57 MtC from 1.22 million hectares). But its dominance varies sharply by biome: in the Amazonas and Orinoquia regions, pasture accounts for over 84% and 89% of emissions respectively, while in the Andes region cocoa is the largest driver at 26.9%. Cocoa and coffee, despite their relatively small cultivation areas, exhibit surprisingly high emission factors (81.2 and 64.3 MgC per hectare respectively) because they expand into high-biomass forests. Smallholder croplands contribute up to 16.7% of regional emissions. Oil palm, often singled out in policy debates, plays a modest role nationally at 2.2%.
The pre-conversion carbon stocks vary substantially by biome. The Pacifico region has the highest at 101 MgC per hectare, followed by the Andes at 94, the Amazonas at 92, the Orinoquia at 83, and the Caribe at 67. Residual carbon after deforestation is consistently low across all biomes at 12 to 22 MgC per hectare, confirming substantial carbon depletion regardless of post-deforestation land use. Some of the highest crop-specific emission factors occur when crops expand into the most carbon-dense forests: rubber in the Pacifico reaches 123 MgC per hectare, and coffee in the Amazonas reaches 92 MgC per hectare.
Why it matters
This paper has direct operational value for anyone building deforestation monitoring or carbon accounting systems for tropical supply chains. The Colombian government, companies sourcing from Colombia, and carbon project developers all need better than IPCC Tier 1 default emission factors to report emissions accurately. The study provides a reproducible framework for moving to Tier 2, combining open satellite data (Sentinel-1, Sentinel-2) with open biomass datasets and deep learning classification.
For EUDR compliance, the paper is especially relevant. The regulation requires companies to demonstrate that commodities were not produced on deforested land after December 31, 2020. But knowing that deforestation occurred is only part of the story. Companies also need to estimate the carbon emissions associated with that deforestation, and this paper shows how to do that at commodity-specific and biome-specific resolution. The finding that cocoa and coffee have higher emission factors than their small areas would suggest is a practical insight for companies sourcing these commodities from Colombia.
The methodology is fully based on open-access satellite data and open-source machine learning tools, meaning it can be replicated in other tropical countries. The authors note that their framework addresses essential gaps in earth-observation-based carbon emission quantification and provides clear, actionable baselines for monitoring policies aimed at promoting deforestation-free supply chains.