GeoZarr

GeoZarr is built from three Zarr conventions. topozarr writes all three as group attributes, and each one answers a single question about the store. Together, they give a web map like zarr-layer what it needs to place the data and pick a resolution without reading any coordinate arrays.

In the store, the root group describes the whole pyramid. Each level group repeats proj:code, proj:wkt2, and zarr_conventions. The figure shows only the key attributes. The root also carries proj:wkt2, spatial:transform, spatial:shape, and spatial:registration.

multiscales

Spec: zarr-conventions/multiscales v0.1

The multiscales convention lists the levels in layout. Each entry after "0" names the level it was derived_from (the previous one), the resampling_method, and a transform whose scale is the ratio between adjacent levels along each spatial dimension (2 by default, or 4 for factors=[1, 4, 16]). Each entry also carries its own spatial:transform and spatial:shape, which give the absolute pixel size.

scale and translation cover every dimension of the array, in the dataset's dimension order. For [time, lat, lon], they are [1, 2, 2] and [0, 0, 0]. translation is always zero because every level shares the same corner origin in spatial:transform, as the spatial convention recommends.

proj

Spec: zarr-conventions/proj v0.1

The proj convention sets the CRS with proj:code and proj:wkt2. It applies only to a group's direct child arrays. Level arrays are grandchildren of the root, so topozarr also tags each level group.

spatial

Spec: zarr-conventions/spatial v0.1

The spatial convention maps array indices to coordinates. spatial:transform is a six-number affine transform. Because each level gets only one transform, this works only for 1-D, uniformly spaced coordinates.

spatial:dimensions names the Y and X dimensions, in that order. Other dimensions, like time, are left alone.

topozarr always writes spatial:registration: "pixel", which means cell edges, not cell centers, line up with spatial:bbox.