Slow convex hull computation for very fragmented rasters

Rasters with lots of small holes are currently taking up to 20 minutes to ingest (10980 x 10980), with most of the time spent computing the convex hull of the footprint.

Possible remedies:

1. Pre-filter raster data with raster.features.sieve to remove small-scale variability. This works well (cuts computation time to 30s), but what should the sieve size be?
2. Replace shapely.ops.unary_union with shapely.geometry.MultiPolygon, since we don’t need to compute a real union to get the convex hull of the dataset (cuts computation time to 3min).
3. Use the chunked version of compute_metadata, which is considerably faster while computing the convex hull (cuts computation time to 1min).
4. Find a cheaper way to compute the convex hull of the raster, e.g. via scipy.spatial.ConvexHull. There should be no reason to polygonize the entire raster in order to get the convex hull. However, I would prefer to avoid introducing SciPy as a dependency if we can.

[dionhaefner]

By definition, only the first and last element of each row and column can contribute to the convex hull of an image. So by prefiltering the mask that is polygonized to these elements, the computational complexity is down from O(n^2) to O(2n). I got a working prototype, which cuts computation time to about 6s, at the expense of slightly higher memory consumption. That should be good enough for now.

[j08lue]

💯 💯