Data Requirements

Required data types

  • High-resolution orthophotos or satellite tiles (RGB minimum).
  • Optional but recommended: aligned DSM/CHM height data.
  • Pixel masks:
    • Instance masks for tree crowns.
    • Optional semantic masks for canopy/background.

Label format

Use one canonical format:

  • COCO instance segmentation (images, annotations, categories), or
  • Internal format converted to COCO during preprocessing.

Include:

  • image_id, category_id, segmentation polygons or RLE, area, bbox.
  • Annotation quality flag (high, medium, low).

Dataset size guidance

  • Pilot: 500-2,000 labeled tiles (512-1024 px).
  • Strong model: 5,000+ diverse tiles.
  • Ideal: multiple geographies, seasons, and sensor conditions.

Split strategy

  • Split by region/time, not random tile-only splitting.
  • Suggested split: 70% train, 15% validation, 15% test by geographic blocks.

Balancing and augmentation

  • Balance across dense forest, sparse cover, urban trees, shadows, mixed terrain, and season variation.
  • Use geospatial-safe augmentations:
    • horizontal and vertical flips,
    • 90-degree rotations,
    • mild brightness/contrast/haze,
    • optional blur and noise.
  • Avoid heavy geometric warps that break geospatial realism.

See also:

PreviousSAM Architecture Variants NextImplementation Plan