Operationalv0.7.0 · SunsetSentinel-2 L2A · STAC · COG
TERRAEarth
Observation
  1. Home
  2. Products
  3. Land cover

LC-01Product sheet

Land cover

TERRA classifies land cover over an area of interest from a Sentinel-2 L2A time series, and gives every pixel a class and a confidence.

The studio board: an area in Tocantins classified on the globe, with land-cover and NDVI callouts, beside the run built as a graph of wired cards
FIG. 01An area in Tocantins classified by the Temporal Transformer over eleven scenes, read on the globe, with the run that produced it wired as cards.

Five classes, one protocol

Three model paths run over the same area protocol, and all of them emit the same five MapBiomas classes: forest formation, agriculture–pasture mosaic, non-vegetated area, soybean and other temporary crops.

Reading a season of scenes rather than one date is what lets the classifier tell apart crops that look alike on any single day. It is also how it gets past the cloud that covers part of most Sentinel-2 scenes.

What comes out

  • A class overlay and a confidence overlay, as PNG and GeoTIFF
  • Vegetation-index series (NDVI, EVI, SAVI) and a phenology summary
  • MapBiomas land cover and use over the same area, where it falls in Brazil
  • The share of each class, in hectares

Three model paths

ModelWhat it readsArtifact
Spectro-temporal Random Forest
Default		80 features per pixel: band statistics, NDVI/EVI/SAVI temporal descriptors, 22 raw NDVI dates300 trees at depth 20
Temporal Transformer
Optional		Six bands padded to 22 dates, mean-pooled over timePyTorch model
Prithvi-EO 2.0 300M
Optional		One acquisition: frozen foundation-model embeddings with a Random Forest headWeights fetched on first run
Two classified areas on one board, each with its own land-cover callout, on the globe and in the 3D viewport
FIG. 02Two areas on one board, each with its own class shares, side by side on the globe and in the 3D viewport.

!Current constraints

  • The legend is fixed. The current models cannot predict pasture, savanna or any other MapBiomas code.
  • The current models were fitted for western Paraná study areas; adapting them to other ground is in research.
  • Class 41 is a residual bucket: high overall accuracy does not imply fine crop identity.
  • Areas in hectares are pixel counts, uncorrected for classification error.
  • Prithvi takes the middle scene of the window and discards the rest, so widening the period changes which acquisition is read rather than how many.
All flight rules