Data pipeline for machine learning with OpenStreetMap
JavaScript HTML Makefile
Latest commit 9d7745a Jan 6, 2017 @anandthakker anandthakker committed on GitHub Merge pull request #10 from developmentseed/image-removal
Add remove images as make command, separate out txt file creation


A pipeline to simplify building a set of training data for aerial-imagery- and OpenStreetMap- based machine learning. The idea is to use OSM QA Tiles to generate "ground truth" images where each color represents some category derived from OSM features. Being map tiles, it's then pretty easy to match these up with the desired input imagery.

Quick Start

The easiest way to use this is via the developmentseed/skynet-data docker image:

docker run -d -v /path/to/output/dir:/workdir/data developmentseed/skynet-data download-osm-tiles

docker run -d -v /path/to/output/dir:/workdir/data -e MapboxAccessToken=YOUR_TOKEN developmentseed/skynet-data

The first line downloads the OSM QA tiles to /path/to/output/dir/osm/planet.mbtiles. If you've already got that file on your machine, you can skip this.

The second builds a training set using the default options (Roads features from OSM QA tiles, images from Mapbox Satellite). To change the data sources, training set size and other options, send the relevant environment variables (see below) into docker run with -e VAR=value.


The make commands below work off the following variables (with defaults as listed):

# location of image files
IMAGE_TILES ?= "tilejson+$(MapboxAccessToken)"
# which osm-qa tiles extract to download; e.g. united_states_of_america
# location of data tiles to use for rendering labels; defaults to osm-qa tiles extract specified by QA_TILES
DATA_TILES ?= mbtiles://./data/osm/$(QA_TILES).mbtiles
# filter to this bbox
BBOX ?= '-180,-85,180,85'
# number of images (tiles) to sample
# define label classes output
# Filter out tiles whose ratio of labeled to unlabeled pixels is less than or
# equal to the given ratio.  Useful for excluding images that are all background, for example.
# set this to a zoom higher than the data tiles' max zoom to get overzoomed label images

Make a full training set according to these params with make all.



  • Install tippecanoe
  • Clone this repo and run npm install. (Note that this includes a node-mapnik install, which sometimes has trouble building in bleeding-edge versions of node.)

Sample available tiles

make data/sample.txt

This just does a simple random sample of the available tiles in the given mbtiles set, using tippecanoe-enumerate. For more intelligent filtering, consider using tippecanoe-decode to examine (geojson) contents of each tile.


Build label images: make data/labels/color or make data/labels/grayscale. Uses the CLASSES json file to set up the rendering of OSM data to images that represent per-pixel category labels. See classes/water-roads-buildings.json for an example. Rendering is with mapnik; see the docs for more on filter syntax.


Download aerial images from a tiled source: make data/images

Heads up: the default, Mapbox Satellite, will need you to set the MapboxAccessToken variable, and will cost you map views!


Preview the generated data by opening up preview.html?accessToken=<mapbox access token>&prefix=/path/to/data in a local web server.