Compute nonoverlapping core regions of a set of polygons and use for intelligent suppression of boundary detections

[russelldj]

The goal of this is to make a set of potentially-overlapping polygons nonoverlapping. The definition for how to split is as follows: for any given point covered my multiple polygons, assign it to the polygon for which it is furthest interior from the nearest edge. The code so far only handles this for the pairwise case using the following steps:

• Identify the overlapping region
• Subtract this region from both input polygons to obtain the unambiguous regions
• Add additional points to the boundary using segmentize
• Create the veronoi tessellation to determine the region closest to each point
• Dissolve these regions based on which polygon boundary the point was from

I haven't rerun this recently, but I think it was having numerical issues.

[russelldj]

The goal of this work is to have a smart way to define the nonoverlapping "core" regions of a set of polygons. This can be used to filter detections such that each region is only predicted by the core of the corresponding tile. In practice, this approach is helpful because it allows the non-overlapping edges of tile to be retained, and closes #176. An example of the updated predictions with this new approach is shown below.

The core of the algorithm is a pairwise approach to splitting overlapping regions between two polygons repeated from the comment above

• Identify the overlapping region
• Subtract this region from both input polygons to obtain the unambiguous regions
• Add additional points to the boundary using segmentize
• Create the veronoi tessellation to determine the region closest to each point
• Dissolve these regions based on which polygon boundary the point was from

Then, for a set of polygons, this approach is run on each pair. For each input polygon, the intersection of all pairwise non-overlapping versions is taken, which defines the core area. An example of the output can be seen in this notebook and below.

This approach is then used in the geometric tree detector to remove tree tops detected only in the overlapping areas, not in all areas at the tile edges.