Agent guidance — Latvia Lauki Parcels

Latvia field boundaries in the fiboa schema, 1 edition (2026). Every claim below is quoted from the source, the converter, or measured from the published files; each query was run before it was written down, and its output follows it as comments.

Access

• Latest edition, stable path: https://data.source.coop/ftw/harmonized-field-data/lv/latest/lv.parquet
• One edition: https://data.source.coop/ftw/harmonized-field-data/lv/year=<year>/<file>.parquet, e.g. https://data.source.coop/ftw/harmonized-field-data/lv/year=2026/lv.parquet
• All editions (hive partitioned): s3://ftw/harmonized-field-data/lv/year=*/*.parquet — the S3 form of the same prefix through the Source Cooperative proxy, because * needs a listing that plain https does not provide. In DuckDB: CREATE SECRET sc (TYPE s3, PROVIDER config, ENDPOINT 'data.source.coop', URL_STYLE 'path', REGION 'us-west-2'); then read_parquet(glob, hive_partitioning = true) adds the year column. No credentials are needed.
• PMTiles for maps: https://data.source.coop/ftw/harmonized-field-data/lv/year=2026/lv.pmtiles, layer lv; MapLibre styles in styles/.

Quirks that produce silently wrong answers

• CRS is EPSG:3059, not WGS84. ST_Area/ST_Distance return units of that CRS; transform with ST_Transform if you need lon/lat, or use metrics:area.
• metrics:area is in square metres (source column area, hectares × 10 000; where the source value is missing or 0 the converter computed it from the geometry, in EPSG:6933 when the CRS is not metric). Divide by 10 000 for hectares.
• year is the edition, not the observation date. It is the year of the source publication (the converter variant). determination:datetime, where present, is the source's own date for a field.
• id is only guaranteed unique within one edition (fiboa requires uniqueness per file; it is the source column OBJECTID). Whether an id persists across editions is not verified here; do not join editions on it without checking.
• hcat:code is hierarchical. The first 4/6/8 digits are increasingly specific crop groups; compare prefixes, not equality, to aggregate (see the crop query below). Source crops without a mapping in the converter's HCAT table (lv_2021.csv) have NULL.
• Some fiboa properties are not columns. Values constant for the whole file are stored once in the GeoParquet collection key-value metadata: admin:country_code = LV, crop:code_list = https://raw.githubusercontent.com/maja601/EuroCrops/refs/heads/main/csvs/country_mappings/lv_2021.csv (2026 edition). Read them with parquet_kv_metadata() in DuckDB or pyarrow.parquet.ParquetFile(f).schema_arrow.metadata[b'collection']; they differ per edition where the source does.

Tested queries

Fields and hectares per edition, through the partition glob:

Largest crop groups in the latest edition (HCAT level 3 = first 6 digits):

Fields around a point, transforming the point into the data's CRS instead of the data into WGS84:

Related collections

Every collection in this catalog shares the fiboa core columns, so the same queries work across countries; s3://ftw/harmonized-field-data/*/latest/*.parquet with union_by_name = true reads the newest edition of all of them (see the catalog AGENTS.md).

Structure

Assets and structural links resolve relative to the object that carries them; there is no self link. Source: this collection is generated by tools/catalogize.py in the catalog repository — fix documentation there.