Add image segmentation for object-based raster analysis

[brendancol]

Motivation

Object-based image analysis (OBIA) is a standard workflow in land cover mapping and remote sensing. The typical pipeline is: segment the image into regions, compute statistics per region, then classify regions. xarray-spatial has nothing for the segmentation step, so users can't do OBIA without leaving the library.

Scope

Segmentation algorithms:

• Superpixel segmentation (SLIC or similar) that works on multi-band rasters
• Watershed segmentation from gradient/edge images
• Region growing from user-supplied seed points

Segment-level statistics:

• Mean, std, min, max of spectral values per segment
• Shape metrics per segment: area, perimeter, compactness, elongation
• Output as a segment attribute table (pandas/xarray Dataset)

Implementation notes

• Output should be an integer-labeled raster, consistent with what regions() already produces. This lets existing zonal stats functions work on the segments directly.
• The hard part is dask compatibility. Segmentation is inherently a global operation since segments can span chunk boundaries. Realistic options:
  • Run segmentation on the full array (fine for moderate sizes, and what most users actually do)
  • Two-pass approach: segment each chunk independently, then merge segments that touch at boundaries
  • Keep segmentation as a non-dask operation, but make the downstream stat extraction dask-friendly
• For GPU support, SLIC and watershed both have known CUDA implementations. Region growing is trickier on GPU but possible with iterative kernel launches.
• Multi-band input means the segmentation functions need to accept a 3D (band, y, x) DataArray or a list of 2D DataArrays.