PEER-REVIEWED EARTH OBSERVATION METHODOLOGY

Algorithmic pipeline, sensor physics, and machine learning specifications

GeoClass integrates multi-spectral optical surface reflectance from Copernicus Sentinel-2 MSI with dual-polarization C-band microwave backscatter from Sentinel-1 SAR. Ground truth labels derive from Dynamic World near-real-time global land cover predictions, classified via an on-the-fly Random Forest or Spatial Contextual architecture.

Spatial Ground Resolution
10 meters / pixel
Sentinel-2 & Sentinel-1
Optical Band Coverage
10 Diagnostic Bands
VNIR to Shortwave Infrared
SAR Radar Frequency
5.405 GHz C-Band
All-weather cloud penetrating
Classification Topology
9 LULC Classes
Dynamic World Harmonized
SYSTEM ARCHITECTURE

End-to-End Remote Sensing Architecture

GeoClass executes all heavy computation serverless-side on Google Earth Engine supercomputing clusters. Client requests specify an Area of Interest (AOI) polygon, a temporal observation window, cloud cover thresholds, and masking preferences.

1Atmospheric Correction and Quality Screening

Raw Sentinel-2 Level-2A surface reflectance granules undergo rigorous masking using both the Scene Classification Layer (SCL) and QA60 opaque cloud and cirrus bitmasks. Cloud shadows (SCL 2), defective pixels, and saturated artifacts are eliminated prior to temporal compositing.

2Temporal Median Reducer & Seasonal Compositing

Pixels are aggregated over the specified date window using a median reducer. This rejects intermittent cloud fringes, sensor anomalies, and transient haze while preserving persistent land cover reflectance signatures. Optional seasonal partitioning isolates dry or wet phenological cycles.

3Feature Extraction: Spectral Indices & SAR Fusion

On top of 10 optical surface reflectance bands (B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12), mathematical diagnostic indices (NDVI, MNDWI, NDBI, NBR) are computed and added. When enabled, dual-polarized Sentinel-1 C-band SAR backscatter features (VV, VH, and VV/VH cross-polarization ratio) are fused directly into the feature stack.

4Supervised Machine Learning Training & Inference

Ground truth pixels are sampled stratified across 9 Land Use / Land Cover classes using Dynamic World modal predictions. A Random Forest ensemble of decision trees is trained on the fly on Earth Engine, classifying every pixel within the AOI at 10-meter spatial resolution.

Harmonized 9-Class LULC Taxonomy

Class IDClass NameColor ValueDiagnostic Spectral & SAR Characteristics
0Water#419BDFHigh NIR absorption, high MNDWI (>0.1), specular microwave reflection (very low VV/VH backscatter).
1Trees / Forest#397D49High red edge and NIR reflectance (NDVI >0.6), high cross-polarized VH microwave volume scattering.
2Grass#88B053Moderate NDVI (0.3 to 0.6), low surface roughness, lower VH return than mature forest canopy.
3Flooded Vegetation#7A87C6Mixed water/canopy response, double-bounce microwave scattering between water surface and emergent stalks.
4Crops#E49635Periodic phenological variations in NDVI, geometric field parcel structures, moderate SAR backscatter.
5Shrub & Scrub#DFC35ALow-stature woody vegetation, intermediate NIR reflectance between grass and closed canopy.
6Built-Up / Urban#C4281BHigh NDBI (>0.05), high SWIR reflectance, corner-reflector dihedral microwave bounce (intense VV return).
7Bare Ground#A59B8FFlat spectral response across visible to NIR, low NDVI (<0.1), surface roughness driven SAR response.
8Snow & Ice#B39FE1Extremely high visible reflectance, sharp absorption in SWIR bands B11 and B12.

Multi-Mission Earth Observation Constellations

GeoClass bridges European Space Agency (Copernicus) and NASA / USGS constellations. Through our STAC catalog integration and Earth Engine bindings, users can dynamically switch optical baselines between Sentinel-2 and Landsat 8/9, or fuse Sentinel-1 synthetic aperture radar.

ConstellationOperatorSensor TypeSpatial ResolutionTemporal RevisitAtmospheric ResiliencePrimary Role in GeoClass
Copernicus Sentinel-2ESA / European UnionMSI (Multi-Spectral 13 Bands)10m / 20m5 days (Constellation 2A + 2B)SCL & QA60 Cloud ScreeningHigh-resolution optical classification, 10m spectral indices (NDVI, NDRE)
USGS / NASA Landsat 8 & 9USGS / NASAOLI / OLI-2 + TIRS (11 Bands)30m (Optical) / 100m (Thermal)8 days (Combined 8 + 9)QA_PIXEL Bitmask ScreeningDecadal historical continuity, 30m Level-2 Tier-1 surface reflectance cross-validation
Copernicus Sentinel-1ESA / European UnionC-SAR (Active Microwave 5.405 GHz)10m (IW Ground Range Detected)6-12 days100% Cloud-Penetrating (All-Weather)Cloud-penetrating radar backscatter fusion (VV, VH), canopy moisture and structural roughness