Examples

Labeled dataset (KML annotations)

The most common case is to download satellite tiles and rasterize KML polygon annotations onto them.

Replace buildings.kml below with the path to your own KML or GeoJSON annotation file.

region:
  west: 4.883
  south: 52.371
  east: 4.896
  north: 52.378
  zoom: 17

tiles:
  source: google_satellite
  strip_rows: 4
  max_connections: 16
  policy: lenient

labels:
  path: buildings.kml
  label_field: null      # all polygons get class_id = 1
  all_touched: false

sampler:
  patch_size: 256
  stride: 0
  mode: grid
  edge_strategy: drop
  max_empty_ratio: 0.95

writer:
  staging_dir: ./output
  image_format: png

split:
  test_ratio: 0.20
  val_ratio: 0.10
  seed: 42
  strategy: stratified

---

Multiclass dataset (GeoJSON with class field)

When your GeoJSON has a property that identifies the class, pass it as label_field. Class IDs are assigned by encounter order starting at 1.

labels:
  path: landuse.geojson
  label_field: class      # e.g. "residential", "commercial", "park"
  all_touched: false

The returned class_map (visible in the manifest) tells you the exact name -> id mapping.

---

Unlabeled dataset (image-only)

Omit the labels section entirely. No Masks/ directory is written.

region:
  west: 74.274
  south: 31.568
  east: 74.327
  north: 31.613
  zoom: 19

tiles:
  source: esri_satellite

sampler:
  patch_size: 512
  stride: 0
  mode: grid
  edge_strategy: shift

writer:
  staging_dir: ./output
  image_format: jpg
  jpg_quality: 90

---

Semi-supervised dataset

Use labeled_ratios to generate labeled/unlabeled split files for semi-supervised training.

split:
  test_ratio: 0.20
  val_ratio: 0.10
  labeled_ratios: [0.05, 0.10, 0.20]
  seed: 42
  strategy: stratified

This writes 5/labeled.txt, 5/unlabeled.txt, 10/labeled.txt, 10/unlabeled.txt, and so on for each ratio, inside staging_dir/splits/. The directory name is the ratio as an integer percentage.

---

Custom tile source

Pass any XYZ URL with {z}, {x}, {y} placeholders via url_template instead of source.

tiles:
  url_template: "https://tile.example.com/{z}/{x}/{y}.png"
  strip_rows: 4
  max_connections: 8
  policy: lenient

---

Random patch sampling

Use mode: random to draw patches at random positions instead of a regular grid. Useful for large regions where you only need a sample.

sampler:
  patch_size: 256
  mode: random
  random_count: 500
  random_seed: 42
  edge_strategy: pad