Secondary Importance Processing

Nominatim geocoder can make use of a so called 'secondary_importance' raster table for ranking results. See Importance in the Nominatim documentation.

This repository creates such a file.

Overview

We create a Postgres raster table covering the planet with about 250m pixels (equal to zoom level 14 of a typical OpenStreetMap). Each pixel stores an importance score between 0 and 65535 (pixeltype 16BUI). Output is a single *.sql.gz file. Optional a GeoTIFF file for visualization.

All processing happens inside Postgresql database.

Input is an existing Nominatim planet database install with at least admin import style and Wikipedia file import.

For each pixel in the raster we look at all places (which cover an area, usually relations) and their importance (either from the placex.importance column or calculated from the place type) and set the highest importance we find.

Background on ranking

Nominatim can import an optional 'secondary_importance' raster table. https://nominatim.org/release-docs/latest/customize/Importance/ The import features was added with PR 2779 and PR 2835.

A previous project in 2022 tried to use OpenStreetMap map tile views to create a raster. See https://www.openstreetmap.org/user/tareqpi/diary/399922 One challenge was that some small websites were loading a map, for example on their homepage, and made a region seem more popular than expected.

Secondary importance score is used when search results have otherwise same importance, for example villages, street, parks.

Data sizes

Input

About 6 million places with rank_search < 25. 99.75% have an importance set. There are 11385 unique importances.

SELECT count(*) FROM placex
WHERE rank_search < 25
  AND ST_GeometryType(placex.geometry) IN ('ST_Polygon', 'ST_MultiPolygon');

 rank_search |  count
-------------+---------
           2 |       63
           4 |      217
           6 |      147
           7 |        4
           8 |     3022
          10 |     5098
          11 |        4
          12 |    41279
          13 |        2
          14 |    45530
          15 |        5
          16 |   229924
          17 |    76887
          18 |   162716
          19 |   104623
          20 |   872076
          21 |    24947
          22 |  3097606
          24 |  1139306
          25 |   322725
          26 |   764490
          27 |    34801
          28 |      396
          30 | 99555017

Processing time

We split the world into tiles. For each tile we lookup the places in its extend, then group them by their importance. Each layer is then written to a temporary table. The temporary table is then merged into a single layer by looking at the highest pixel value.

tile size | time  | tiles  | with country | no places in tile
----------+-------+--------+--------------+-------------------
50x50     | 2:52h | 107584 |        33450 |             10051
100x100   | 1:57h |  26896 |         9191 |              2039
200x200   | 2:18h |   6724 |         2684 |               459
400x400   | 3:42h |   1681 |          812 |               140

Output

• 8.1M secondary_importance.sql.gz (832M uncompressed, 99% compression)

• 21M secondary_importance.tiff (LZW compression, likely 512M uncompressed)

Visualize

Development

You can check the size of a pixel with the https://tools.geofabrik.de/map/?type=Geofabrik_Standard&grid=1#14/69.6486/18.9507 tool.

Further optimization

• It should be possible to exclude more tiles from processing. We already exclude more than 50%. Still another 30% are later seen empty.

• A temporary table might not be needed.

• We use ST_MapAlgebra, the expression version, which takes two raster. There's also ST_MapAlgebra which takes a

• We could skip some process for a temporary table with 0 or 1 layers.

• Places might cover a full a tile. If so then any places with lower importance could be disregarded.