{
  "dataset": "CASID",
  "description": "Land-cover semantic segmentation labels: pixel value = class id in label PNGs (16-color palette PNG, unused entries rendered as gray i,i,i).",
  "classes": [
    {"id": 0, "name": "background", "rgb": [0, 0, 0]},
    {"id": 1, "name": "building", "rgb": [255, 0, 0]},
    {"id": 2, "name": "forest", "rgb": [85, 255, 0]},
    {"id": 3, "name": "road", "rgb": [255, 255, 0]},
    {"id": 4, "name": "water", "rgb": [0, 112, 255]}
  ],
  "class_order_source": "AerialSegTGRS paper: `background`, `building`, `forest`, `road`, `water`",
  "palette_note": "Label PNGs are palette-mode (P) 8-bit PNGs; reading np.array(img) yields class ids directly. Palette entries 5-15 are unused gray placeholders (i,i,i).",
  "provenance": "Unified 2026-10-15 from three legacy encodings (RGB-coded TIF labels, palette PNG labels with GID-style residue) using pixel-wise cross-checks: TIF/PNG double-version comparison on region 03 and comparison with the legacy cropped_label tiles used in the paper experiments."
}