If you don't have primary-secondary attribution to help you, you can still implement some basic logic to decide who's on the losing end of ST_Difference. For example, at a T-shaped intersection, you might specify that the road that passes through the intersection will retain its buffer, while the road that ends at the intersection will have its buffer clipped.
To start, it would be helpful to have a function to express the "passes through tee" relationship. We can say that A passes through the A-B tee if:
A intersects B- The intersection of
A and B is an endpoint of B
- The intersection of
A and B is not an endpoint of A
Turning this logic into a function:
CREATE OR REPLACE FUNCTION passes_through_tee(geom_a geometry, geom_b geometry)
RETURNS boolean AS
$$
SELECT ST_Contains(endpoints(geom_b), ST_Intersection(geom_a, geom_b))
AND NOT ST_Contains(endpoints(geom_a), ST_Intersection(geom_a, geom_b))
$$ language 'sql';
However, endpoints isn't a PostGIS function, so we need to write it:
CREATE OR REPLACE FUNCTION endpoints(g geometry)
RETURNS geometry AS
$$
SELECT ST_Union(endpoints) FROM
(SELECT ST_Collect(ST_EndPoint(geom), ST_StartPoint(geom)) AS endpoints
FROM
(SELECT (ST_Dump(g)).geom) dump) subquery
$$ language 'sql';
Note that the ST_Dump subquery allows the function to work on MultiLineString geometries.
To apply these functions to our roads, we can first save the buffered geometries in a temporary table:
CREATE TABLE buffers AS
SELECT gid, geom, ST_Buffer(geom, 0.0001, 'endcap=flat join=round') as buffered_geom
FROM roads;
CREATE INDEX ON buffers USING GIST(geom);
Then, to pull out the clipped buffers:
SELECT gid, ST_Difference(buffered_geom,
COALESCE(
(SELECT ST_Union(buffered_geom) FROM buffers b
WHERE ST_Intersects(a.geom, b.geom)
AND a.gid != b.gid
AND CASE WHEN passes_through_tee(a.geom, b.geom) THEN FALSE
WHEN passes_through_tee(b.geom, a.geom) THEN TRUE
ELSE a.gid < b.gid END
), 'POLYGON EMPTY'::geometry))
FROM buffers a
Some example output from this query is below:
A couple of notes on the above:
- If we don't have a T-intersection (think end-to-end or four-way intersections) we fall back on
gid to pick a winner.
- The
COALESCE function substitutes POLYGON EMPTY for a null value, so that we get our original geometry back if there is nothing to clip against.
- Polygons will still overlap in cases where the buffers of two roads intersect but the roads themselves do not. This behaviour could be changed, but substituting
buffered_geom for geom in the ST_Intersects call.
Best Answer
I ended up using a workaround by forcing the labels into separate layers. First I duplicated the layer in question several times so that i had one copy for every label type required (ie. each unique value in an attribute). then I filtered each layer to just the necessary value and labelled it. In each layer the properties were set to display colliding labels, etc.
The benefit of all this was not just that the labels move around each other because they are in different layers, but i can also now label each value slightly differently instead of all the same text.
A bit klunky, and only really useable if there are not a ton of unique values, but it worked.