30 Day Map Challenge 2026 - Day 14 (Icon Map Slicer) - Borgesian map

Can a map ever contain everything?

For Day 14 of the 30 Day Map Challenge, the theme is Borgesian, after Jorge Luis Borges's one-paragraph story On Exactitude in Science (1946). In it, an empire's cartographers perfect their craft until they produce a map the size of the empire itself, matching it point for point. The next generation finds it useless, and leaves it to rot in the desert.

After Day 13's oceans, I wanted somewhere small enough to try mapping exhaustively. Skukuza Rest Camp, the Kruger National Park's mini-town in the bush, is about the right size: chalets, restaurants, a museum, a petrol station, a campsite and hundreds of trees, all inside one fence. I set out to put every feature OpenStreetMap knows about on one map, and to see what was still missing.

Explore the interactive report

  • Pan and zoom the map. The Features in view card and the bar chart count everything currently on screen. The count updates when you drag the map.
  • The bar chart is also the colour key. Each category's bar is filled with the colour its features have on the map. Click a bar to show only that category; click again to bring everything back.
  • Hover over anything for its name, category and type. Chalets show their booking code, such as BD3 148.
  • Lasso an area to count the features inside it.
  • What the open data misses lists the gaps I found, and Tree canopy cover gives one number for something the catalogue can't hold.

Interactive report: Day 14, Borgesian map

What the map doesn't show: anything OpenStreetMap doesn't have. That means no safari tents, no camping pitches, most of the communal kitchens and none of the trees. Hidden by design, the car parks and one paved square are still counted but not drawn, because the aerial imagery already shows them.

Starting with OpenStreetMap

All the features come from OpenStreetMap, downloaded with an Overpass query on 12 November 2026: every tagged node and way in a box around the camp, about 2.6 MB of JSON and 804 objects. A Python and DuckDB pipeline cleans, classifies and clips them, and writes one CSV for Power BI.

Choosing the boundary

The brief started as a one-kilometre square, but a square cuts through the camp and takes in staff housing, a solar plant and bush. The camp itself is in OpenStreetMap as a single closed way: Skukuza Rest Camp, tagged as both a fence and residential land. It encloses about 0.28 km², and everything is clipped to it.

It doesn't match the official SANParks camp map, and that turned out to be deliberate. The medical facility sits just outside the line. OpenStreetMap's edit history explains why: a mapper checked with reception, found the hospital can't be reached from the public part of the camp, and moved the boundary. Kruger Shalati, the train hotel on the old Selati railway bridge, is outside too: it's a separate hotel with its own entrance.

From tags to a catalogue

Turning 804 tagged objects into 385 features took four kinds of clean-up.

Clipping and dropping. Ward boundaries, the village name label and anything outside the fence go. Roads and paths that cross the fence are cut at it.

Chalets mapped twice. Each chalet is in OpenStreetMap as a point carrying its booking code, such as BE3 61, and again as an unnamed building outline. A point-in-polygon join merges them: 196 points fall inside their outline, and one sits 4 m away and takes the nearest. Two more chalets, 73 and 153, exist only as buildings named with their number, so a building named with nothing but a number becomes a chalet too.

Dots inside buildings. A café or a kitchen is often mapped as a point inside its own building. Where an unlabelled building holds exactly one point, the building takes that point's name and category and the point goes. Common kitchen, Kruger Station 3638 and the Museum Hut became buildings this way. Buildings holding several facilities keep their points: reception still shows its information desk, ATM, post office and toilets.

Names that say what a building is. Generic buildings named … Guest House, … Spa, … Conference … or … Kitchen are classified from the name. Two tagging quirks needed rules of their own. Skukuza Main Gate is tagged as a place name as well as a gate, so it was being dropped with the labels. And the only trace of the safari tents is an unnamed hotel point with the note Tented huts, which now gets that as its type.

The result is 385 features in 13 categories, led by Accommodation (207), Roads and transport (59), Paths, Buildings and Visitor services (22 each).

Gotcha: a missing tag drops the row

DuckDB returns NULL when you look up a tag a feature doesn't have, and NULL in a WHERE clause drops the row. My first filter, meant to remove chalet points only, removed every point without a tourism tag. Wrapping each lookup in coalesce(..., '') fixed it. A few checks after every run (unique ids, everything inside the fence, every row drawable) caught this and the gate.

One table, two kinds of row

Icon Map Slicer can filter the rest of the page by the map view: bind latitude and longitude to its slice settings, and every visual on the same table counts only what's on screen. That's the feature this map is built around.

The catch is how the visual decides what to draw. A row with a WKT shape and coordinates is treated as a point and skipped by the shapes layer. So the Features table carries two sets of map columns:

  • Map Geometry holds WKT for buildings, areas and lines, and an SVG icon for point features.
  • Point Latitude / Longitude are filled only for point features, which tells the map to draw the SVG as an image.
  • Latitude / Longitude on every row hold a representative point inside the feature, and that's what the map-view filter uses, so a building counts as in view even though it's drawn as a shape.

Icons drawn in SQL

The 51 point features are 26-pixel badges: a circle in the category colour with a white symbol for the type. There's a bench, a bin, a gate, an obelisk for memorials, WC for toilets, an envelope for the post office, and a tent for Tented huts. Each is a few lines of SVG generated in the pipeline, so changing a category colour recolours its icons and bars together.

Layers around the data

  • Aerial imagery. South Africa's national aerial photography from CD:NGI, 25–50 cm, served as tiles by OpenStreetMap South Africa, sits under the data at 15% transparency. At zoom 19 every thatched rondavel is visible, which makes the gaps visible too.
  • The fence is a fixed GeoJSON overlay embedded in the report, not a data row, so clicking a bar never hides it.
  • A quieter basemap. Icon Map's light style, muted, with point-of-interest and place labels switched off, so the basemap doesn't label things the data doesn't hold.

Measuring what isn't mapped: tree canopy

The imagery shows Skukuza is a camp under trees, and OpenStreetMap maps none of them inside the fence. Detecting individual trees was too uncertain: crowns merge, and the round thatched roofs look like trees from above. Canopy cover was a fairer measure.

A short script fetches the 154 zoom-19 tiles covering the fence (0.27 m pixels), masks out the building outlines, and scores each pixel on an excess-green index. Otsu's method then picks the threshold between canopy and dry ground. The fence measured on the imagery came to 279,974 m² against OpenStreetMap's 278,552 m², within 0.5%, so the two line up. About 40% of the camp (111,398 m²) is under tree canopy. Checked against the imagery, the mask catches the crowns and leaves out the roofs, car parks and bare ground. It misses some small, shaded crowns, so it's more likely an underestimate than an overestimate.

Design choices

  • The chart is the legend. One colour per category, and the bar chart's bars use the same colour measure as the map.
  • Draw what helps, count everything. Car parks and one paved square aren't drawn, because the imagery shows them and the grey fill only covered it, but they're still counted. The campsite is a solid outline with no fill, so the space where the tents should be stays visible.
  • Contrast checked. The basemap tool's contrast check flagged two colours as too faint, so Utilities and Buildings are a shade deeper.
  • Everything at once. I built zoom tiers first, with four GeoJSON layers revealing more detail as you zoom in. Showing everything at once makes the point better: the density is the argument.

What the map reveals

We can map a lot. OpenStreetMap holds 385 features inside a fence of 0.28 km²: 197 chalets, 12 benches, 3 gates and 2 bins, 10 heritage features from bronze sculptures to a clock tower, and the utilities behind them.

We still miss things. OpenStreetMap knows about 97% of the numbered chalets (197 of about 203), but none of the 21 safari tents and no camping pitches. That leaves 3 communal kitchens of about 15, 6 toilets against about 8 ablution blocks, and no car wash, key box or bird hide. Campers are who the open map misses.

Some of reality doesn't reduce to objects. About 40% of the camp is under tree canopy, and OpenStreetMap records none of it. Canopy isn't a feature someone forgot to add. It's a quality of the place that a catalogue of points and outlines can't hold, however complete it gets.

These are mapped records, not an inventory. A visual comparison was done with the SANParks camp map, which isn't traced or shown. A feature missing from OpenStreetMap isn't missing from Skukuza, and the canopy figure is an estimate from imagery of an unstated date.

What I learned from Day 14

1. Check what a boundary means. The fence I nearly "fixed" was right. It follows public access, and the edit history said so.

2. Count everything, draw what helps. Hiding a feature on the map doesn't have to remove it from the data. The car parks still count, the map is just clearer.

3. Make the chart the legend. When the bars are coloured like the map, the reader learns the colours by reading the numbers.

4. Split columns, not tables. A row can only play one role on the map, but the same row can carry a shape for one layer and a point for the filter.

5. Some things aren't objects. Canopy, shade and character don't fit a feature table. One honest number on a card says more than 2,000 guessed trees.

Final thoughts

Borges's empire failed because it confused completeness with usefulness. Day 14 makes three arguments at once: we can map a lot, we still miss things, and some of the territory never reduces to mapped objects. The same applies to any estate, campus or site you manage. Catalogue what you can, mark plainly what's missing, and measure the qualities that a list of assets can't hold. The useful map isn't the one with everything on it, but the one that shows where everything stops.

Sources and attribution

  • Features: © OpenStreetMap contributors, ODbL 1.0. Extracted with Overpass on 12 November 2026. Camp boundary: OSM way 520976645, Skukuza Rest Camp.
  • Aerial imagery: Chief Directorate: National Geo-spatial Information (CD:NGI) national aerial photography, 25–50 cm, tiles via OpenStreetMap South Africa. NGI places no stated restrictions on use; credited as "Aerial imagery © CD:NGI". Capture date not published. Accessed 13 November 2026.
  • Gap audit reference: SANParks Skukuza camp map, used only as a private checklist. It isn't reproduced, traced or used as a source for OpenStreetMap edits.
  • Story: Jorge Luis Borges, Del rigor en la ciencia (On Exactitude in Science), 1946.

Methodology note: Features are OpenStreetMap nodes and ways inside the Skukuza Rest Camp fence, clipped to it with DuckDB spatial; relations and place labels are dropped. Chalet points are joined to their building outlines by point-in-polygon (nearest outline within 10 m otherwise). A point inside a generic building with no other points is merged into the building; generic buildings are classified from their names where the name states a use. Categories are assigned from OpenStreetMap tags; each has one colour. Car parks and one paved footway area are counted but not drawn. Coverage figures compare OpenStreetMap with a visual review of the SANParks camp map; numbering gaps not on that map are assumed not to exist. Tree canopy is estimated from 154 zoom-19 CD:NGI tiles (0.27 m pixels) with an excess-green index (2g − r − b on chromatic coordinates), an Otsu threshold computed inside the fence and a 5 × 5 majority filter, with OpenStreetMap building and chalet outlines removed; canopy cover is canopy pixels divided by fence pixels (39.8%). Shaded crowns are missed and green lawns may be included.