name: mesopotamia-marshes
version: "0.2"
category: change-detection
themes: [hydrology]
family: surface-water-extent-change
description: >
  Two-epoch surface-water extent of the Mesopotamian Marshes of southern Iraq
  from Landsat NDWI. Drained by canal and embankment after 1991 to roughly a
  tenth of their former area, then partially re-flooded from 2003 onward — a
  loss and a deliberate restoration in the same bounding box.

# WHAT THIS MEASURES, AND WHAT IT DOES NOT. READ BEFORE CITING A NUMBER.
#   NDWI > threshold is an OPEN-WATER PROXY, not a bathymetric or volumetric
#   measurement. Wet playa, salt crust, shallow turbid water and wet vegetation
#   sit near the threshold and cross it between epochs. The answer below is a
#   count of proxy-water pixels at 30 m (0.09 ha each), not a gauged surface
#   area, and it carries no uncertainty estimate.
#
#   No per-year axis. Each epoch is ONE median composite over its window, so the
#   answer is a two-epoch comparison, not an annual series. A genuine annual
#   series needs a per-year grouping reduction, and no shipped op provides one
#   (recipe-library tracker, defect 63).
#
#   UPSTREAM BUG (recipe-library tracker, defects 61/62). Some Landsat 7 (LE07)
#   assets on the Planetary Computer answer an HTML error page where a COG
#   should be, and the read dies as "Unknown byte order marker: [3C, 21]" (`<!`).
#   Confirmed 2026-08-17 on LE07_L2SP_036035_20221027_02_T1. There is no
#   platform filter to exclude L7 with: the source params carry none, and the
#   `query.platform` block the previous version of this recipe declared was
#   silently ignored. The only lever an author has is the calendar. Windows that
#   overlap 1999-2023 can pull L7 and can therefore hit this.
#
#   No month filter. The engine's source params carry no month selector, so the
#   seasonal windows below are calendar spans and off-season scenes do enter the
#   median. A median over many scenes is robust to that; a mean would not be.

settings:
  ndwi_threshold: 0.0        # NDWI > this → open water / flooded marsh
  cloud_cover_max_pct: 30    # eo:cloud_cover upper bound on the STAC search
  pixel_area_m2: 900         # 30 m Landsat pixel — multiply `count` by this for area


inputs:
  aoi:
    bbox: [46.4, 30.5, 47.9, 31.6]
    crs_target: EPSG:32638
    notes: >
      Al-Hammar and Central Marshes between the Tigris and Euphrates. Emergent
      reed beds read as vegetation, not water, so the proxy tracks OPEN water
      and misses vegetated marsh — a real limit on the restoration story.


  baseline_epoch:
    start: "1986-03-01"
    end: "1990-11-01"
    notes: Landsat 5 TM, before the post-1991 drainage works.

  recent_epoch:
    start: "2021-03-01"
    end: "2025-11-01"
    notes: Landsat 8/9 OLI, two decades into re-flooding.

  collection:
    uri: "@usgs/landsat/c2-l2"

layers:

  # ── SOURCE ──────────────────────────────────────────────────────────────────
  #
  # Bands are CANONICAL concept codes from registry/datasets/usgs/landsat/c2-l2/
  # _dataset.yaml (G = green, N = nir, QA = qa-pixel). Each distribution maps them onto its own
  # producer asset keys, so ONE band list is correct across TM, ETM+ and OLI —
  # which the previous SR_Bx spelling was not (SR_B5 is SWIR-1 on TM/ETM+ and
  # NIR on OLI, so the old files asked for a different physical band per sensor).
  #
  # The FUSED per-scene form (`scenes: true` + `expression:` + `mask_band:`) is
  # what makes the index correct: the distribution's declared scale/offset is
  # applied per band before the expression, fill DNs become NaN, and QA_PIXEL is
  # the platform mask band, excluded from the b1..bN binding. `temporal_reduce`
  # then takes the median across the surviving scenes.
  #
  # Do NOT collapse this to a bare `bands: [A, B, QA]` source and difference the
  # result. A source read emits the DECLARED BANDS (measured 2026-08-17: three
  # bands out for a three-band request), so `b2 - b1` there is raw band math,
  # not a normalized-difference change. It plans green and answers plausibly.

  source/ndwi-baseline-scenes:
    type: raster
    description: Per-scene NDWI, QA-masked, baseline epoch.
    uri: "@usgs/landsat/c2-l2"
    params:
      scenes: true
      bands: [G, N, QA]
      mask_band: QA
      expression: "(b1 - b2) / (b1 + b2)"
      bbox: [46.4, 30.5, 47.9, 31.6]
      temporal: ["1986-03-01", "1990-11-01"]
      cloud_lt: 30
      resolution: 30
      limit: 60

  source/ndwi-recent-scenes:
    type: raster
    description: Per-scene NDWI, QA-masked, recent epoch.
    uri: "@usgs/landsat/c2-l2"
    params:
      scenes: true
      bands: [G, N, QA]
      mask_band: QA
      expression: "(b1 - b2) / (b1 + b2)"
      bbox: [46.4, 30.5, 47.9, 31.6]
      temporal: ["2021-03-01", "2025-11-01"]
      cloud_lt: 30
      resolution: 30
      limit: 60

  # ── COMPUTE ─────────────────────────────────────────────────────────────────

  compute/ndwi-baseline:
    type: raster
    description: Median NDWI over the baseline epoch's scenes.
    compute:
      op: temporal_reduce
      inputs:
        time_series: { layer: source/ndwi-baseline-scenes }
      params:
        method: median

  compute/ndwi-recent:
    type: raster
    description: Median NDWI over the recent epoch's scenes.
    compute:
      op: temporal_reduce
      inputs:
        time_series: { layer: source/ndwi-recent-scenes }
      params:
        method: median

  compute/ndwi-pair:
    type: raster
    description: The two NDWI composites stacked into one 2-band raster (b1 = baseline, b2 = recent).
    compute:
      op: stack_concat
      inputs:
        layers:
          - { layer: compute/ndwi-baseline }
          - { layer: compute/ndwi-recent }

  compute/ndwi-change:
    type: raster
    description: >
      dNDWI = recent - baseline. Negative where marsh became dry ground;
      positive where re-flooding restored open water.
    compute:
      op: raster_calc
      inputs:
        raster: { layer: compute/ndwi-pair }
      params:
        expression: "b2 - b1"
        output_band: dndwi

  compute/water-baseline:
    type: raster
    description: >
      Proxy-water footprint. Water keeps the value 1; everything else becomes
      NaN (`b1 / 0.0`), which the stats kernel skips — so the answer layer's
      `count` is the water-pixel count.
    compute:
      op: raster_calc
      inputs:
        raster: { layer: compute/ndwi-baseline }
      params:
        expression: "where(b1 > ${settings.ndwi_threshold}, 1.0, b1 / 0.0)"
        output_band: water

  compute/water-recent:
    type: raster
    description: >
      Proxy-water footprint. Water keeps the value 1; everything else becomes
      NaN (`b1 / 0.0`), which the stats kernel skips — so the answer layer's
      `count` is the water-pixel count.
    compute:
      op: raster_calc
      inputs:
        raster: { layer: compute/ndwi-recent }
      params:
        expression: "where(b1 > ${settings.ndwi_threshold}, 1.0, b1 / 0.0)"
        output_band: water

  # ── ANSWERS ─────────────────────────────────────────────────────────────────

  answer/water-baseline:
    type: answer
    description: >
      Proxy open-water pixel count, baseline epoch. Area = count x 900 m².
    compute:
      op: summary
      inputs:
        raster: { layer: compute/water-baseline }
      params:
        stats: [count]

  answer/water-recent:
    type: answer
    description: >
      Proxy open-water pixel count, recent epoch. Area = count x 900 m².
    compute:
      op: summary
      inputs:
        raster: { layer: compute/water-recent }
      params:
        stats: [count]

outputs:
  water_baseline:
    layer: answer/water-baseline
    value_kind: answer
    description: Pre-drainage proxy open-water pixel count.
  water_recent:
    layer: answer/water-recent
    value_kind: answer
    description: Post-restoration proxy open-water pixel count. THE answer, read against the baseline.
  ndwi_change:
    layer: compute/ndwi-change
    value_kind: raster
    description: dNDWI raster, recent minus baseline (show-your-work).


license:
  recipe: Apache-2.0
  output_dataset: CC-BY-4.0
  upstream_data:
    - { source: "@usgs/landsat/c2-l2", license: usgs-public-domain }