Supervised RF Classification (Sacramento Valley, CA)

# Platform Comparison: Supervised Classification -- Gold Spec
#
# Train a Random Forest classifier on Sentinel-2 imagery using
# ESA WorldCover as pseudo-labels. Sacramento Valley, CA.
# Ground truth: ~80% overall accuracy.
#
# Note: Training is a Python-only step (scikit-learn or GEE's RF).
# The Rust/WASM compute layer handles inference only.

name: supervised-classification
version: "1.0"
description: >
  Supervised land cover classification using Random Forest trained on
  Sentinel-2 bands with ESA WorldCover pseudo-labels. 4 classes:
  forest, crop, urban, water. Sacramento Valley, CA.
  Ground truth: ~87% overall accuracy (calibrated from GEE 2026-03-29).

settings:
  default_bbox: [-121.8, 38.4, -121.3, 38.8]
  default_crs: EPSG:4326

layers:

  # ============================================================
  # SOURCE LAYERS
  # ============================================================

  source/sentinel2:
    uri: stac://earth-search/sentinel-2-l2a
    type: raster
    description: >
      Sentinel-2 L2A summer 2023 composite, bands B2-B12.
      Sacramento Valley, CA.
    params:
      bbox: [-121.8, 38.4, -121.3, 38.8]
      datetime: "2023-06-01/2023-09-01"
      bands: [B2, B3, B4, B8, B11, B12]

  source/worldcover:
    uri: stac://earth-search/esa-worldcover
    type: raster
    description: >
      ESA WorldCover v200 (10m) as training labels.
    params:
      bbox: [-121.8, 38.4, -121.3, 38.8]

  # ============================================================
  # COMPUTE: PREPROCESSING
  # ============================================================

  compute/s2-masked:
    type: raster
    description: >
      Cloud-masked Sentinel-2 imagery.
    compute:
      op: cloud_mask_sentinel2
      inputs:
        data: { layer: source/sentinel2 }

  compute/s2-composite:
    type: raster
    description: >
      Median composite of cloud-masked S2 imagery.
    compute:
      op: temporal_reduce
      inputs:
        data: { layer: compute/s2-masked }
      params:
        reducer: median

  compute/labels:
    type: raster
    description: >
      WorldCover remapped to 4 classes: 1=forest, 2=crop, 3=urban, 4=water.
    compute:
      op: raster_calc
      inputs:
        lc: { layer: source/worldcover }
      params:
        expression: >
          where(lc == 10, 1,
            where(lc == 40, 2,
              where(lc == 50, 3,
                where(lc == 80, 4, 0))))

  # ============================================================
  # COMPUTE: CLASSIFICATION
  # ============================================================

  compute/classified:
    type: raster
    description: >
      Random Forest classification result (4 classes).
      Training uses stratified sample from WorldCover labels.
    compute:
      op: classify_rf
      inputs:
        features: { layer: compute/s2-composite }
        labels: { layer: compute/labels }
      params:
        n_trees: 50
        training_points: 500
        bands: [B2, B3, B4, B8, B11, B12]
        class_band: class
        stratified: true

  # ============================================================
  # RESULT: CLASSIFICATION ACCURACY
  # ============================================================

  result/accuracy:
    type: table
    description: >
      Overall classification accuracy vs WorldCover labels.
      Ground truth: ~87%.
    compute:
      op: raster_calc
      inputs:
        classified: { layer: compute/classified }
        labels: { layer: compute/labels }
      params:
        expression: "mean(where(labels > 0, classified == labels, nan))"