Global 1km embeddings from the Matryoshka Implicit Neural Distillation (MIND) model. MIND is a compact geographic implicit neural representation (INR) with adjustable spatial scale.
The encoder is a neural network which takes in a lat/lon coordinates and produces 3k dimensions partitioned into chunks where earlier dimensions contain coarse features and later dimensions contain local high-resolution detail.
For more information see https://research.taylorgeospatial.org/mind/
Global land embeddings from MIND (Matryoshka Implicit Neural Distillation), a model that maps latitude and longitude to a 3,072-dimensional embedding.
Early chunks capture broad geographic patterns, while later chunks add finer variation. You can choose the embedding width for your task without retraining the model.
The global grid has 0.01° spacing (approximately 1 km) in WGS 84 (EPSG:4326). 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 store.
This example reads the first 64 dimensions for a small San Francisco area, decodes the stored values, and masks ocean pixels.
Requires rioxarray, xarray, dask[array], zarr>=3, fsspec, and aiohttp.
Change the bounding box for your area and 64 in both selections for your embedding width.
Group "0" is the finest grid; groups "1" through "6" provide coarser overviews.
Paper · Project page · Code · Model weights