README
MIND global embeddings
Paper · Project page · Code · Model weights
Global land embeddings from MIND (Matryoshka Implicit Neural Distillation), a model that maps latitude and longitude to a 3,072-dimensional representation. Early embedding chunks capture broad geographic patterns, while later chunks add finer spatial variation. Retaining only the leading dimensions lets you choose the embedding width for your task without retraining the model.
Linear heads reconstruct teacher embeddings at each supervised width during training. Downstream prediction uses the frozen encoder with truncation or the Chunked Penalty.
The first three principal components map to RGB, with PCA fitted separately for each panel. Compare spatial patterns across panels rather than colors. Explore the maps on the project page.
Available data
The embeddings cover global land on a 0.01° latitude–longitude grid in WGS 84 (EPSG:4326), approximately 1 km at the equator. The Zarr store is chunked across space and embedding dimensions, so you can read an area and a subset of dimensions without downloading the full dataset.
Read embeddings for an area
Install the required packages:
The example below reads the first 64 dimensions for a small area in San Francisco, decodes the stored values, and sets ocean pixels to NaN.
The result is a NumPy array with shape [64, height, width].
Change the bounding box for your area and n_dims for your embedding width.
Group "0" contains the finest grid; groups "1" through "6" provide progressively coarser spatial overviews.
Each group includes its own CRS and spatial transform; decode_coords="all" exposes them to rioxarray.
To generate embeddings directly from coordinates, see the Torch Hub example on the project page.
License
The data in this collection are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).
If you use these embeddings in your work, please cite: