Interpretation protocol for the GRAC reference sample (blinded reference interpretation; Section 2.3)

Date of interpretation: 9 October 2026.

Unit of assessment
  The 20 m cell of the GRAC grid (EPSG:3035) outlined in red at the centre of each image.

Evidence shown for each location
  1. Sentinel-2 L2A true-colour composites (bands 4, 3, 2; per-pixel median of observations from 1 May to
     30 September, after removing pixels flagged by the scene classification as cloud, cloud shadow, cirrus,
     snow, saturated or no data) for 2018, 2021 and 2024, 400 m x 400 m, north up.
  2. The monthly median NDVI of the cell (mean of the 10 m pixels inside it) for January 2017 to December 2025,
     from the same masked Sentinel-2 observations; months without a valid observation are left blank.
  The interpreter is not shown the sampling stratum, the GRAC codes, any other land-cover product or the
  coordinates; locations are presented in random order with neutral identifiers (G001-G250).

What is recorded
  For each of 2018, 2021 and 2024: is the cell grassland? yes / no / uncertain, using the definition of the
  Copernicus HRL Grasslands GRA layer (Product User Manual v2.3, reference [21]): herbaceous vegetation with at
  least 30% cover dominated by graminoids, with scattered trees or shrubs covering at most 10%, and not tilled
  in the reference year or the preceding years. Arable land (including grass or fodder crops in rotation),
  vegetation dominated by shrubs, trees or tall forbs, sealed or built surfaces, solar arrays, bare ground and
  water are non-grassland.
  For each period (2018-2021 and 2021-2024): did the cover of the cell change? no / yes / uncertain. A cell
  whose cover is the same throughout a period is recorded as unchanged even when its class is uncertain.
  For each year also the cover type (grassland, arable, shrub/woody, forest, built/solar, bare, water, mixed)
  and a short note on the evidence.

Decision cues
  - Arable or tilled: a growing-season NDVI minimum of about 0.25 or lower outside December-February, a single
    sharp crop peak followed by harvest, bare or ploughed soil in the composite, regular field geometry.
  - Grassland: continuous herbaceous cover without bare-soil periods; NDVI typically 0.4-0.8 in spring and
    lower in summer drought; mowing produces short dips with recovery in the same season.
  - Woody: dark, textured canopy and NDVI that stays high (about 0.6 or more) through August-October; shrub
    encroachment is non-grassland when woody cover clearly exceeds about 10% of the cell.
  - Built, solar or bare: low NDVI throughout the year with grey, white or dark regular structures.
  - December-February values are not used as evidence of tillage (snow, frost, dormancy).
  - A year is labelled 'uncertain' when the image and the NDVI series are insufficient or contradictory, or
    when the cell straddles cover types so that no type clearly dominates.

Derived reference events (computed afterwards, not by the interpreter)
  loss 2018-2021 = change recorded and grassland in 2018 and not in 2021 (no loss if the cover was recorded as
  unchanged); loss 2021-2024 likewise; grassland loss 2018-2024 = either. Gain is defined in the same way in the opposite direction. A location is
  excluded from an estimate if a label needed for that estimate is 'uncertain'.

Notes on the evidence and the procedure (added after the interpretation, 9 October 2026; the labels were not changed)
  - Cloud masking: the composites and the NDVI series keep only pixels whose scene classification (SCL) is 4
    vegetation, 5 not vegetated, 6 water or 7 unclassified; dark-area pixels are therefore also masked, in addition
    to the classes listed above. R1_GRAC_evidence_GEE.js reproduces the panels exactly (checked on G001 2018 image
    pixels and on the 2018 NDVI values of G001, G002 and G150).
  - NDVI 2025: the interactive computation exceeded Earth Engine memory limits for 2025 except for the last 50
    locations (G201-G250); for G001-G200 the series used ends in December 2024. The input actually used is
    R1_GRAC_NDVI_monthly.csv (rounded to three decimals).
  - Presentation: locations were shown in batches, each row with the identifier, the three image chips and a chart of
    the monthly series; the March-November values of each year (x 100) were also listed as numbers. Nothing else was
    displayed. One image (G151, 2024) could not be generated; that year was judged from the NDVI series.
  - Tillage rule: the GRA definition excludes land tilled in the reference year or the preceding years (six years
    in the definition table of the Product User Manual v2.3, five in its description of the tillage mask). Because
    the series starts in 2017, earlier tillage could not be detected. One location (G244) was labelled
    as loss only because of this rule (ploughed in 2020, grass-like cover from 2021); R1_GRAC_accuracy.py reports a
    variant in which it is treated as unchanged grassland.
  - Blinded procedure: the evidence panels and listings used for the interpretation showed only the neutral
    identifiers (G001-G250), without access to the sampling strata, GRAC classes, other land-cover products
    or coordinates. Each location was read once.
  - Interpreter and recording: one interpreter (M.T.) inspected the evidence of every location and decided each
    label. The panels were presented in batches, and the labels and notes recorded, in one browser session with
    an AI assistant (Claude, Anthropic) instead of separate Code Editor tabs; every recorded label was checked
    by the interpreter.
