Fixing non-noded intersection problem using PostGIS

In my experience this problem is nearly always caused by:

1. High precision in your coordinates (43.231499999999996), combined with
2. Lines that are almost coincident but not identical

The "nudge" approach of the ST_Buffer solutions lets you get away with #2, but anything you can do to resolve these underlying causes, like snapping your geometry to a 1e-6 grid, will make your life easier. The buffered geometries are usually fine for intermediate calculations like overlap area, but you'll want to be careful about retaining them because they can make your close-but-not-quite problems worse in the long haul.

PostgreSQL's exception-handling capability lets you write wrapper functions to handle these special cases, buffering only when needed. Here's an example for ST_Intersection; I use a similar function for ST_Difference. You'll need to decide if the buffering and potential return of an empty polygon are acceptable in your situation.

CREATE OR REPLACE FUNCTION safe_isect(geom_a geometry, geom_b geometry)
RETURNS geometry AS
$$
BEGIN
    RETURN ST_Intersection(geom_a, geom_b);
    EXCEPTION
        WHEN OTHERS THEN
            BEGIN
                RETURN ST_Intersection(ST_Buffer(geom_a, 0.0000001), ST_Buffer(geom_b, 0.0000001));
                EXCEPTION
                    WHEN OTHERS THEN
                        RETURN ST_GeomFromText('POLYGON EMPTY');
    END;
END
$$
LANGUAGE 'plpgsql' STABLE STRICT;

Another benefit with this approach is you can pinpoint the geometries that are actually causing your problems; just add some RAISE NOTICE statements in the EXCEPTION block to output WKT or something else that will help you track down the problem.

Through a lot of trial and error I eventually realized that the non-noded intersection resulted from a self-intersection problem. I found a thread that suggested using ST_buffer(geom, 0) can be used to fix the problem (though it does make it a lot slower overall). I then tried to use ST_MakeValid() and when applied directly to the geometry before any other function. This seems to fix the problem robustly.

ipoint <- pg.spi.exec(
        sprintf(
            "SELECT 
                    %3$s AS id, 
                    st_intersection(ST_MakeValid(''SRID=''||st_srid(%2$s)||'';%4$s''::text), ST_MakeValid(''%5$s'', 0)) AS polygon 
            FROM %1$s 
            WHERE 
                ST_Intersects(ST_MakeValid(%2$s::text),ST_MakeValid(''SRID=''||st_srid(%2$s)||'';%4$s''));",
            arg1,
            arg2,
            arg3,
            curpoly,
            buffer_set$ewkb[1]
        )
    )

---

I've marked this as the answer as it seems to be the only approach that fixes my problem.

I ran into this same problem (Postgres 9.1.4, PostGIS 2.1.1), and the only thing that worked for me was to wrap the geometry with a very small buffer.

SELECT ST_Intersection(
    (SELECT geom FROM table1), ST_Union(ST_Buffer(geom, 0.0000001))
) FROM table2

ST_MakeValid didn't work for me, nor did the combination of ST_Node and ST_Dump. The buffer didn't seem to result in any degradation in performance, but if I made it any smaller I still received a non-noded intersection error.

Ugly, but it works.

Update:

The ST_Buffer strategy seems to work well, but I ran into an issue where it produced errors when casting the geometry to geography. For example, if a point is originally at -90.0, and it is buffered by 0.0000001, it is now at -90.0000001, which is an invalid geography.

This meant that even though ST_IsValid(geom) was t, ST_Area(geom::geography) returned NaN for many features.

To avoid the non-noded intersection problem, while maintaining valid geography, I ended up using ST_SnapToGrid like so

SELECT ST_Union(ST_MakeValid(ST_SnapToGrid(geom, 0.0001))) AS geom, common_id
    FROM table
    GROUP BY common_id;