HackPrinceton models_index.csv: 1,374 rows without lat/lon and 14 outliers (7 'Los Angeles, CA' models in India, 6 'Houston' in Austin) that break map extents
Symptom: after loading the HackPrinceton 2026 Matterport models_index.csv (14,944 rows) into GeoPandas/GeoParquet and querying the extent with DuckDB, you get lat 11.25..40.68, lon -118.70..77.62 instead of a box around Houston and Los Angeles. Auto-fit maps zoom out to the whole planet, and H3/hexbin or nearest-neighbour code gets a few nonsense cells.
What's in the file:
- 1,374 rows have empty lat/lon and an empty latlon_source (the other values are model 11,708 and costar 1,862). dropna(subset=['lat','lon']) leaves 13,570 rows.
- 14 rows with coordinates fall outside both metros, all with latlon_source = model: 7 with city = Los Angeles, state = CA at about (13.0623, 77.6218) and (11.2507, 76.0356), which are in India; 6 with city = Houston, state = TX at (30.2866, -97.7483) and (30.3058, -97.6622), which is Austin; and one with state NY at (40.6833, -73.9616), Brooklyn. Their stored location doesn't match the city.
- city isn't normalised: you'll see Los Angeles, Los Angeles (7 rows) next to Los Angeles, plus many suburbs (Culver City, Pasadena, Webster, Nassau Bay...). Group by state or by a spatial join if you want 'Houston vs LA'.
- In pandas, costar_property_id, homes_listing_id and drone_photos are float64 because of blanks, so an id prints as 12345.0. Read them with dtype={'costar_property_id': 'Int64', 'homes_listing_id': 'Int64'} before joining to CoStar data. DuckDB infers them as BIGINT.
Fix: filter to the two metro boxes, and use the address only as a fallback:
df = pd.read_csv('models_index.csv', dtype={'costar_property_id':'Int64','homes_listing_id':'Int64'})
houston = df.lat.between(29.0, 30.5) & df.lon.between(-96.2, -94.6)
la = df.lat.between(33.3, 34.9) & df.lon.between(-119.2, -117.4)
df = df[houston | la] # 13,556 of the 13,570 rows with coordinatesAll 232 splat models (splat == 'yes') have coordinates and fall inside the two metros, so splat-only work isn't affected.