Stock di Carbonio

Land Usev01.20 credits/per_km2

Responsible:Niccolò Calandri

Carbon Stock — Random Forest Downscaling

Overview

The Carbon Stock layer estimates terrestrial carbon density in tonnes of carbon per hectare (tC/ha) by combining a global baseline carbon map with high-resolution remote sensing variables through Random Forest machine learning downscaling.

The baseline dataset is derived from Walker et al. (2022), which provides a global map of carbon density for the year 2016 at ~300 m resolution. This baseline captures above-ground biomass, below-ground biomass, soil organic carbon, and dead organic matter. To achieve higher spatial resolution and temporal relevance, the baseline is downscaled using current satellite-derived features:

NDVI from Sentinel-2 (10 m) — proxy for vegetation density and health
Land Cover from Dynamic World (10 m) — categorical land use classification
Elevation from SRTM (30 m) — topographic influence on carbon storage

The resulting maps enable site-level carbon stock assessment for carbon credit verification, baseline carbon inventories, reforestation monitoring, and ESG reporting.

Calculation Methodology

Random Forest downscaling pipeline:

Baseline extraction — The Walker et al. 2022 carbon density raster is loaded from S3 and clipped to the area of interest
Feature engineering — For each pixel:
- NDVI value (median composite over the analysis period)
- Land cover class from Dynamic World
- Elevation from SRTM DEM
Training — A Random Forest regressor is trained on the relationship between baseline carbon values and the auxiliary features at the baseline resolution
Prediction — The trained model predicts carbon density at the higher resolution of the auxiliary features
Validation — Cross-validation metrics (R², RMSE) are computed to assess prediction quality

Carbon pools included:

Pool	Description
Above-ground biomass	Living vegetation (trees, shrubs, crops)
Below-ground biomass	Root systems
Soil organic carbon	Organic carbon in top 30 cm of soil
Dead organic matter	Litter, deadwood

Output unit: tC/ha (tonnes of carbon per hectare)

Data Sources

Code	Name	Provider	Type	Coverage	Resolution	Reference Year
Walker 2022	Global Carbon Density	Walker et al. / S3	Raster (COG)	Global	~300 m	2016
`COPERNICUS/S2_SR_HARMONIZED`	Sentinel-2 L2A	ESA/Copernicus	Multispectral	Global	10 m	2017 — present
`GOOGLE/DYNAMICWORLD/V1`	Dynamic World	Google/WRI	Land cover	Global	10 m	2015 — present
`USGS/SRTMGL1_003`	SRTM DEM	NASA/USGS	Elevation	60°N–56°S	30 m	2000

Quality Range and Levels

Indicator	Unit	Range	Inverted
`carbon_stock`	tC/ha	[0, 10, 30, 80, 150, 300]	No

Inverted = No: a higher value indicates greater carbon storage.

Level	Thresholds	Interpretation
E (Very Low)	< 10 tC/ha	Bare soil, urban, desert
D (Low)	10–30 tC/ha	Sparse vegetation, degraded land
C (Moderate)	30–80 tC/ha	Grasslands, croplands, young forests
B (High)	80–150 tC/ha	Mature temperate forests, dense shrublands
A (Very High)	> 150 tC/ha	Tropical forests, peatlands, old-growth forests

Technical Notes

The baseline (Walker 2022) represents conditions circa 2016; significant land use changes since then may not be fully captured
Random Forest downscaling assumes that the relationship between carbon and auxiliary variables is spatially consistent within the analysis area
NDVI saturation in very dense forests may limit discrimination of high carbon stocks
Soil organic carbon estimates carry higher uncertainty than above-ground biomass estimates

References

Walker, W.S., Gorelik, S.R., Cook-Patton, S.C. et al. (2022). "The global potential for increased storage of carbon on land". Proceedings of the National Academy of Sciences, 119(23), e2111312119.
Spawn, S.A., Sullivan, C.C., Lark, T.J., Gibbs, H.K. (2020). "Harmonized global maps of above and belowground biomass carbon density in the year 2010". Scientific Data, 7, Article 112.
Breiman, L. (2001). "Random Forests". Machine Learning, 45(1), pp. 5–32.

Data Sources & References

Sources

WRD_CARBN_22

Methodology

Lo stock di carbonio viene stimato attraverso una tecnica di downscaling spaziale che combina una mappa di riferimento con variabili ausiliarie ad alta risoluzione. La mappa di riferimento è il dataset Walker et al. 2022 (carbon density anno base 2016), disponibile su AWS S3 come Cloud Optimized GeoTIFF, con risoluzione di circa 100 metri. Il downscaling utilizza un modello Random Forest addestrato su: 1) NDVI da Sentinel-2 per il periodo richiesto (proxy della biomassa vegetale attuale), 2) Land Cover da Dynamic World (stratificazione per tipo di vegetazione), 3) Elevazione da SRTM DEM (influenza su temperatura e precipitazioni). I dati di training vengono scaricati alla risoluzione MODIS (~100m), mascherando valori invalidi e corpi idrici. Il modello RF viene configurato con 100 estimatori e profondità massima 15. La predizione viene poi applicata alle variabili ausiliarie scaricate ad alta risoluzione (25-50m). L'output è espresso in tonnellate di carbonio per ettaro (tC/ha), con range tipico 0-300 per la maggior parte delle aree, fino a 500 per foreste tropicali dense.