inconsistant spatial filtering for null or empty geometries

[Even Rouault]

Nyall Dawson pointed to me in 
https://github.com/qgis/QGIS/pull/5501#issuecomment-340934947 about a weird behaviour 
of OGR spatial filtering regarding null and empty geometries.

To make it concrete, consider the following CSV file

id,WKT
1,"POLYGON EMPTY"
2,
3,"POLYGON ((0 0,0 1,1 1,0 0))"


$ ogrinfo test.csv -al -spat 0 0 1 1 -q

Layer name: test
OGRFeature(test):1
  id (String) = 1
  WKT (String) = POLYGON EMPTY
  POLYGON EMPTY

OGRFeature(test):2
  id (String) = 2
  WKT (String) = 

OGRFeature(test):3
  id (String) = 3
  WKT (String) = POLYGON ((0 0,0 1,1 1,0 0))
  POLYGON ((0 0,0 1,1 1,0 0))

$ ogrinfo test.csv -al -spat 1 1 2 2 -q

Layer name: test
OGRFeature(test):2
  id (String) = 2
  WKT (String) = 

OGRFeature(test):3
  id (String) = 3
  WKT (String) = POLYGON ((0 0,0 1,1 1,0 0))
  POLYGON ((0 0,0 1,1 1,0 0))

So features with null geometries (FID 2) are always selected. And for a feature with an empty 
geometry (FID 1), it is selected by the -spat 0 0 1 1 filter since internally the envelope of 
POLYGON EMPTY is computed as xmin=xmax=ymin=ymax=0, but not by -spat 1 1 2 2

If ingesting this file into PostGIS and doing the same test, we get:

$ ogrinfo pg:dbname=autotest test -spat 0 0 1 1 -q

Layer name: test
OGRFeature(test):3
  id (String) = 3
  wkt (String) = POLYGON ((0 0,0 1,1 1,0 0))
  POLYGON ((0 0,0 1,1 1,0 0))

$ ogrinfo pg:dbname=autotest test -spat 1 1 2 2 -q

Layer name: test
OGRFeature(test):3
  id (String) = 3
  wkt (String) = POLYGON ((0 0,0 1,1 1,0 0))
  POLYGON ((0 0,0 1,1 1,0 0))

So null or empty geometries are never selected when a spatial filter is set. This last behaviour 
seems to be the most logical for me, and I'd propose we adopt it for non-SQL based drivers 
as well.
(note: the current behaviour dates back to the initial introduction of spatial filtering in OGR 
per https://trac.osgeo.org/gdal/changeset/7191 , 13 years ago )

[Jukka Rahkonen]

Not directly related but something to keep in mind - MapServer by default adds the &&[bounding box] filter which is OK for WMS but with WFS it means that features without geometries are never selected. I do not know how the QGIS WFS server behaves.

Depending on the use case features without geometry/with empty geometry can make sense and in that case the software should not automatically add spatial filter. This ticket is about a case where user explicitly gives a spatial filter and I agree that then empty geometries should not get selected.

[Even Rouault]

You can avoid MapServer WFS code to emit the spatial filter with "wfs_use_default_extent_for_getfeature" "no" in LAYER.METADATA. This is currently only honoured in practice by the MapServer OGR backend. But... both the OGR and PostGIS MapServer backends will anyway skip features with null geometries. So there would be work needed to support such cases, and this OGR change shouldn't have direct impact on this behaviour

[Even Rouault]

In 40613:

OGRLayer::FilterGeom(): make sure a feature with a null or empty geometry never matches a spatial filter (fixes #7123)

[Even Rouault]

In 40614:

Adjust test result due to refs #7123 changes

[Even Rouault]

In 40615:

Adjust test result due to refs #7123 changes