loc2vec

Pytoch implementation of the Loc2Vec model outlined in the blog post 'Loc2Vec: Learning Location Embeddings with Triplet-loss Networks'.

Use

Implementation requires user to hold both anchor and positive anchor rasters. Negative anchors can be specified, with random indicies selected for such purpose if not.

Model Results

This implementation was tested with the following OSM raster aggregated channels:

• OSM Lines:
  • roads:
    • ['motorway', 'primary & trunk', 'secondary', 'minor street']
    • ['others']
  • rails
    • ['others', 'rail']
• OSM Multipolygons
  • ['national park', 'forest', 'grass & park', 'meadow', 'farmyard', 'farmland', 'orchard']
  • ['water']
  • ['industrial', 'construction', 'quarry', 'military']
  • ['railway']
  • ['residential']
  • ['commercial']
  • ['retail']
  • ['allotments', 'cemetary', 'brown field']
  • ['buildings']