Technology Architecture

A scientific product page detailing our remote sensing infrastructure, machine learning SOC estimation pipelines, and end-to-end MRV verification architecture.

[ SPECTRAL SENSING MATRIX ]

Remote Sensing Infrastructure

We ingest daily Sentinel-2 L2A atmospheric-corrected imagery across 13 spectral bands. Our pipeline extracts six critical channels for soil organic carbon estimation.

BANDWAVELENGTHNAMEAPPLICATION

B02490 nmBlueAtmospheric correction, water body detection

B03560 nmGreenVegetation vigor assessment, peak reflectance

B04665 nmRedChlorophyll absorption depth measurement

B08842 nmNIRBiomass density, NDVI calculation base

B111610 nmSWIR-1Soil moisture detection, mineral mapping

B122190 nmSWIR-2Clay content estimation, SOC proxy index

[ ML INFERENCE ENGINE ]

SOC Estimation Models

Our ensemble machine learning models are calibrated against 72 physically verified, deep-core baseline assays with spatial leave-one-block-out cross-validation to guarantee scientific precision.

R²0.847

Coefficient of determination on held-out spatial blocks

RMSE0.23%

Root mean square error of SOC concentration prediction

Bias-0.012%

Systematic prediction bias across all validation folds

[ MODEL ARCHITECTURE ]

MODEL: Gradient Boosted Machine Learning Ensemble

FEATURES: 47 spectral + 12 terrain + 8 climate covariates = 67 input features

TRAINING: 72 baseline soil cores × 24 temporal composites = 1,728 high-fidelity calibration points

VALIDATION: Spatial Leave-One-Block-Out (LOBO) with 5 km block size

EXPLAINABILITY: SHAP feature importance + partial dependence plots per region

[ PIPELINE ARCHITECTURE ]

End-to-End Data Pipeline

Data Ingestion

ACTIVE

Raw geospatial datasets are pulled from ESA Copernicus Open Access Hub, NASA EarthData, and ISRIC APIs into our cloud-native spatial data lake.

Sentinel-2 L2A tilesSRTM DEM elevationERA5 climate reanalysisISRIC SoilGrids baseline

Feature Engineering

ACTIVE

Multi-temporal spectral indices, terrain derivatives, and climate covariates are computed per 10m pixel grid cell across all pilot regions.

NDVI / NDWI indicesSWIR ratio transformsTopographic wetness indexTemporal composites

Model Inference

ACTIVE

Gradient boosting models calibrated against 72 high-fidelity, deep-core soil assays predict SOC concentration at 10m resolution with calibrated confidence intervals.

Machine learning ensembleSpatial cross-validationUncertainty quantificationSHAP explainability

Validation & MRV

ACTIVE

Model outputs are validated against held-out physical cores, audited by accredited verifiers, and submitted to carbon registries for credit issuance.

Physical soil assaysThird-party auditorsAnnual re-verificationRegistry submission

[ FUTURE INFRASTRUCTURE ]

APIs & Platform Services

VELSTROM is building toward an open agricultural intelligence API layer — enabling third-party researchers, governments, and enterprises to access calibrated SOC estimations, spectral analytics, and climate data pipelines.

[ PLANNED API ENDPOINTS ]

GET /api/v1/soc-estimate?lat=&lon=&depth=

GET /api/v1/spectral-index?region=&band=

POST /api/v1/mrv-submit

GET /api/v1/pilot-regions

STATUS: IN DEVELOPMENT // ETA: Q3 2026