Overture — attribute pushdown with where= (live)¶

The bbox limits where you fetch; where= limits which rows, pushed down to the GeoParquet on S3 via DuckDB so only matching features leave the bucket. Here we keep only high-confidence places and a narrow set of columns (id/sources are retained automatically so the per-row license_id still works).

Setup¶

Imports and the area of interest. earthlens provides the unified EarthLens entry point; geopandas reads the GeoParquet that Overture writes.

In [ ]:

import geopandas as gpd

from earthlens.earthlens import EarthLens

import geopandas as gpd from earthlens.earthlens import EarthLens

Area of interest¶

A narrow bounding box over a single Times Square block, plus the output directory download() writes into.

In [ ]:

LAT_LIM = [40.757, 40.759]  # [south, north]
LON_LIM = [-73.987, -73.984]  # [west, east] — a Times Square block
OUT = "_overture_out"

LAT_LIM = [40.757, 40.759] # [south, north] LON_LIM = [-73.987, -73.984] # [west, east] — a Times Square block OUT = "_overture_out"

Fetch with attribute pushdown¶

Build the Overture request first. The bbox limits where you fetch; where= limits which rows, pushed down to the GeoParquet on S3 via DuckDB so only matching features leave the bucket. We keep only high-confidence places (confidence > 0.95) and a narrow set of columns (id/sources are retained automatically so the per-row license_id still works).

In [ ]:

overture = EarthLens(
    data_source="overture",
    dataset="places",
    variables=[],
    aoi=[LON_LIM[0], LAT_LIM[0], LON_LIM[1], LAT_LIM[1]],
    path=OUT,
    where="confidence > 0.95",  # pushed down to S3
    columns=["names", "confidence"],  # narrow projection
)

overture = EarthLens( data_source="overture", dataset="places", variables=[], aoi=[LON_LIM[0], LAT_LIM[0], LON_LIM[1], LAT_LIM[1]], path=OUT, where="confidence > 0.95", # pushed down to S3 columns=["names", "confidence"], # narrow projection )

Run the download. download() returns the list of written GeoParquet paths.

In [ ]:

paths = overture.download()
paths

paths = overture.download() paths

Verify the pushdown¶

Read the written parquet back and confirm every row satisfies the where= predicate, with the per-row license_id populated.

In [ ]:

gdf = gpd.read_parquet(paths[0])
print(len(gdf), 'high-confidence places')
print('all match the predicate:', bool((gdf['confidence'] > 0.95).all()))
gdf[['confidence', 'license_id']].head()

gdf = gpd.read_parquet(paths[0]) print(len(gdf), 'high-confidence places') print('all match the predicate:', bool((gdf['confidence'] > 0.95).all())) gdf[['confidence', 'license_id']].head()

where= is raw SQL evaluated by DuckDB against the parquet schema, so you can filter nested fields too — e.g. categories.primary = 'restaurant' for places, or height > 10 for buildings.