DRB Geolocation Addon

The DRB geolocation add-on extracts the geolocation (named geometries + CRS) of a DRB node. Geolocation comes from two composed mechanisms:

• declarative descriptors: geoloc: RDF blocks attached to topic class URIs, loaded from installed TTL packages and Fuseki (the same double source as the core topic DAO);
• intrinsic providers: georeferencing carried by the data itself (raster affine transform or ground control points), read at access time, never declared.

The legacy cortex.yml YAML model is gone. There is no drb-metadata dependency: this add-on is the only geolocation source.

Model overview

The geoloc: descriptor vocabulary

Namespace http://knowledge-base.gael.fr/drb/addons/geolocation/, prefix geoloc:. This is deliberately not GeoSPARQL: GeoSPARQL predicates carry geometry data literals, while geoloc: carries extraction rules (XQuery programs run against a node). GeoSPARQL datatype URIs are reused as format terms, since they already type the extracted string exactly: geo:gmlLiteral, geo:wktLiteral, geo:geoJSONLiteral. CRS formats have no GeoSPARQL equivalent and live in the geoloc: namespace: geoloc:epsg, geoloc:wktCrs.

One building block: a geoloc:geometry block attached to a topic class URI. One block = one named geometry, exactly one source, and an optional CRS.

Block rules

• Exactly one source per block: geoloc:xquery (extraction), geoloc:value (constant), or geoloc:children (aggregate). A block declaring zero or several of these is skipped with a warning at load time — a malformed block never aborts loading globally.
• geoloc:format is mandatory for xquery/value sources and must be a registered geometry format term. An unknown or missing format also skips the block with a warning (the geoloc-validator catches this statically).
• CRS precedence: an explicit geoloc:crs sub-block wins over a CRS embedded in the format (GML srsName, GeoJSON implicit EPSG:4326), which wins over the default EPSG:4326.
• geoloc:children aggregate semantics: the pattern is a /-separated glob matched with fnmatch against each descendant's structural child names (the pattern's segment count bounds the walk depth); every matching child is geolocated recursively (its own intrinsic provider or its own descriptors); geoloc:select picks one named geometry from each child's set (default: every geometry of each child); every collected geometry is reprojected to EPSG:4326 before aggregation, and the aggregated result is always EPSG:4326. geoloc:aggregate selects the mode: geoloc:union (shapely unary union), geoloc:envelope (bounding box of the union, default), geoloc:collection (GeometryCollection).
• subClassOf inheritance: declarations are collected by walking the topic's subClassOf chain; on equal names, the most specific topic wins. Intrinsic providers enter the set with the lowest precedence, so a declarative block can override a provider-published name.
• Reserved name: the shipped raster provider publishes under bounds; family descriptors reference it with geoloc:select "bounds" in a geoloc:children block.
• Skip-and-warn: every rule above that a block violates causes that one block to be skipped (logged at WARNING), never a global load failure.

Example: XQuery source and explicit CRS

@prefix geoloc: <http://knowledge-base.gael.fr/drb/addons/geolocation/> .
@prefix geo:    <http://www.opengis.net/ont/geosparql#> .
@prefix sentinel-1: <http://knowledge-base.gael.fr/drb/sentinel-1/> .

# XQuery source: product footprint from the manifest
sentinel-1:product geoloc:geometry [
    geoloc:name    "footprint" ;
    geoloc:format  geo:gmlLiteral ;
    geoloc:xquery  """<XQuery returning the frameSet GML polygon>""" ;
] .

# Explicit CRS when the geometry value does not carry one
sentinel-1:product geoloc:geometry [
    geoloc:name    "scene-center" ;
    geoloc:format  geo:wktLiteral ;
    geoloc:xquery  """<XQuery returning a WKT POINT>""" ;
    geoloc:crs     [ geoloc:format geoloc:epsg ;
                     geoloc:xquery """<XQuery returning the code>""" ] ;
] .

Example: children aggregate composing the raster provider

# Children aggregate: composes the intrinsic provider, zero duplication
sentinel-1:product geoloc:geometry [
    geoloc:name      "measurements" ;
    geoloc:children  "measurement/s1*-grd-*.tiff" ;  # relative glob pattern
    geoloc:select    "bounds" ;      # named geometry taken from each child
    geoloc:aggregate geoloc:union ;  # union | envelope | collection
] .

Family descriptor packages (drbx-kb-geolocation-sentinel-1, -sentinel-2) are out of scope of this repository; naming follows the KB convention of hyphenated family identifiers (sentinel-1, sentinel-2, never sentinel1).

Python usage

from drb.topics import resolver
from drb.addons.geolocation import GeolocationAddon, GeoLocationSet

topic, node = resolver.resolve("<data_url>")

# the canonical addon interface
geoset = GeolocationAddon().apply(node, topic=topic)

geoset.names()               # declared/published names, no evaluation
footprint = geoset["footprint"]     # evaluated on first access, then memoized
footprint.transform("EPSG:2154")    # a new, reprojected GeoLocation
footprint.to_wkt()
footprint.to_geojson()               # a GeoJSON Feature (RFC 7946, WGS84)

geoset.transform("EPSG:2154").to_geojson()  # a GeoJSON FeatureCollection

GeoLocationSet is a lazy Mapping[str, GeoLocation]: a geometry is only extracted the first time its name is accessed, then cached. This matters because a geoloc:children aggregate may open many rasters through zip/S3 — apply() itself never triggers extraction, and one name failing to evaluate (GeolocationExtractionError, raised at access) does not prevent the other names in the same set from being read.

If no declaration and no provider apply to the node's topic chain, apply() raises GeolocationNotAvailable.

Provider extension point

Intrinsic providers publish geolocation without any RDF declaration, keyed by the format-topic class URIs they serve. Register a provider class through the drb.addon.geolocation.provider entry point group; the built-in RasterProvider (declared for drb:image and drb:jp2) is loaded the same way.

from drb.addons.geolocation import GeolocationProvider, GeoLocation

class MyProvider(GeolocationProvider):
    def topics(self) -> list[str]:
        return ["http://www.gael.fr/drb#my-format"]

    def names(self) -> set[str]:
        return {"bounds"}          # static, no node access

    def extract(self, node) -> dict[str, GeoLocation]:
        ...                        # called lazily, on first access only
        return {"bounds": GeoLocation("bounds", crs, geometry)}

[project.entry-points."drb.addon.geolocation.provider"]
my-provider = "my_package.geolocation:MyProvider"

The shipped RasterProvider reads node.get_impl(rasterio.DatasetReader) only — it is transport-agnostic and never opens files itself. It derives bounds from the affine transform and CRS when present (e.g. Sentinel-2 JP2 bands), falling back to the convex hull of the ground control points when only GCPs are available (e.g. Sentinel-1 GeoTIFF measurements), and raises GeolocationExtractionError when the raster carries neither. Intrinsic raster geolocation requires drb-driver-image at runtime (not a hard dependency of this package) to provide the rasterio node implementation.

Format extension point

Geometry and CRS value formats are registered functions, keyed by URI (GeoSPARQL datatypes for geometry, geoloc: terms for CRS). Register additional formats through the drb.addon.geolocation.format entry point group: each entry point resolves to a zero-argument callable that performs the registration.

from drb.addons.geolocation.formats import register_geometry_format

def register():
    register_geometry_format(
        "http://example.org/my-format", parse_my_format)
    # parse_my_format(value: str) -> (shapely.Geometry, pyproj.CRS | None)

[project.entry-points."drb.addon.geolocation.format"]
my-format = "my_package.geolocation:register"

Static validation: geoloc-validator

The geoloc-validator console script statically checks geoloc:geometry blocks in one or more Turtle files: unparsable Turtle, more than one or zero sources per block, and unknown/missing formats or aggregate modes are all reported without evaluating any XQuery.

geoloc-validator path/to/cortex.ttl [more.ttl ...]

Exit code 0 when every file is clean, 1 as soon as one problem is reported (one line per problem, on stderr).