[dcherian]

Motivation

xpublish-tiles needs a vectorized "which cell contains each point" lookup. sel({"x": arr}, method="nearest") (0.2.4) wraps correctly on a periodic axis (global 1° grid, corners -180..180: 180.5, 539.5, -180.5 → 0, 359, 359), but:

• It has no bounds check. On a regional grid (corners -10..10, 20 cells), -100 → -90 and 180.5 → 190. A negative result is a valid numpy index, so a miss silently reads the wrong cell.
• NaN → 0 on all grids, and inf → 0 on a periodic axis (the NaN-to-int cast gives 0 on arm64).

API

RasterIndex.get_indexer(labels: Mapping[Hashable, ArrayLike]) -> dict[str, np.ndarray]

Follows pandas.Index.get_indexer. Keys of labels are coordinate names (as for sel). Returns int64 positions keyed by dimension name, with the input shape; -1 for a point outside the extent, NaN or inf.

• Rectilinear: each axis is looked up independently; X, Y or both may be given. On a periodic axis, the integer cell is wrapped modulo the period, then bounds-checked (so a subset of one period still gives -1 for a wrapped point outside it).
• Rotated/skewed: both coordinates are required, with the same shape; the full inverse affine is used, and a point is a miss on both dims if either position is out of range. Periods cannot be set on these transforms (from_transform already raises NotImplementedError), so no wrap is needed.

Edge convention

With p the corner-based pixel position (~index.transform()), a point is in cell i if i <= p < i + 1, i.e. floor(p). A cell includes the edge at its lower position and excludes the one at its higher position; for a north-up raster that is the top edge included, the bottom edge excluded, and the raster's bottom edge is outside. Away from cell edges this equals sel(..., method="nearest") (which uses np.round, round-half-to-even, at ties).

🤖 Generated with Claude Code