{"success":true,"data":{"author":"TGE Labs (Taylor Geospatial)","bucketAccessType":"public","bucketListable":true,"bucketName":"us-west-2.opendata.source.coop","bucketPrefix":"tge-labs/mind","category":"Earth Observation & Geospatial","cloudProvider":"aws","createdAt":"2026-11-16T15:39:54.060Z","description":"Precomputed global land embeddings from MIND (Matryoshka Implicit Neural Distillation), a location encoder that maps latitude and longitude to a 3,072-dimensional representation distilled from Earth observation teacher embeddings. The Matryoshka training means leading dimensions carry broad geographic patterns and later chunks add finer spatial variation, so users can truncate to 64, 256 or more dimensions for a downstream task without retraining. Files: mind.zarr (Zarr v3, all 3,072 dimensions on a 0.01 degree, about 1 km, EPSG:4326 grid over global land, chunked across space and embedding dimension, group 0 finest and groups 1-6 coarser overviews), mind_rgb.tif (PCA-to-RGB Cloud Optimized GeoTIFF for viewing), and mind_hex_r4.parquet / mind_hex_r5.parquet (GeoParquet H3 resolution 4 and 5 summaries of the first 64 dimensions). Read an area and a subset of dimensions with xarray.open_zarr and rioxarray (decode_coords='all'); values are stored encoded and the README shows decoding and ocean masking. Model weights are on Torch Hub for generating embeddings at arbitrary coordinates. Useful as covariates for species distribution, crop, socioeconomic and environmental regression. Licensed CC BY 4.0; cite the MIND paper.","displayTitle":"MIND Global Land Embeddings (Matryoshka Implicit Neural Distillation, 0.01 degree)","fuseMountBrowsable":true,"id":"9b7a779f-0d00-50aa-a92a-725b934f60cd","license":"CC-BY-4.0","licenseEvidenceUrl":"https://source.coop/tge-labs/mind","objectCountEstimate":1,"probeStatus":"ok","region":"us-west-2","schemaCard":{"accessHints":[{"language":"python-anonymous","snippet":"import boto3\nfrom botocore import UNSIGNED\nfrom botocore.client import Config\ns3 = boto3.client('s3', region_name='us-west-2', config=Config(signature_version=UNSIGNED))\ns3.list_objects_v2(Bucket='us-west-2.opendata.source.coop', Prefix='tge-labs/mind/', Delimiter='/')"},{"language":"python-local","snippet":"open('/mnt/archive/mind.zarr/0/embedding/c/0/1/14', 'rb')"}],"caveats":["Files are binary — use a domain-specific parser (e.g. metpy / pynexrad / pydicom) rather than text I/O."],"format":"binary","generatedAt":"2026-11-16T15:39:56.239Z","layout":{"example":"tge-labs/mind/mind.zarr/0/embedding/c/0/1/14","pattern":"{prefix}/{seg1}/{seg2}/{seg3}/{seg4}/{seg5}/{seg6}/{seg7}/{object}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"tge-labs/mind/mind.zarr/0/embedding/c/0/1/14","truncatedAt":512},{"bytes":"28b52ffd8058000040008c140026512b18a025cd18921a915655d38becdd9af5d2923d000000e0710a1f001e002a00b9ca355f79916f9eab50f97abd5eafae575f7df5d5575f7df5d5575f7df5034bbd06048ce338be7881e38a2f1cc7711c5fc4714d447c29bcc0f13ccff300859860669c42cc860504530e311ba61033e1a480704c21638c318861440c23440c881856abd5ea5507678938436797ceced2193a4be34c3897cc9c2574850467966e9ca51318124c2a1620960c1d33828197a881f18631f10bbb0112408070041482300803510884200c82300cc320088320088220088218630c555cb71d8d3ad5edaa70f6dc88799dc8b19a2e8a1e230c2b304685b5246d1e499171b612a8f998aa7f0a8dab2c0703d2c4c7ef171a46cc79c4c38be42bc116305916af0e2b98b2845a2b564b2eb0cb885582bbe13612d325440700c031d94032aa0de88ac13e550766e0bef4da3b73c200daf1ac18cd8944bbc752c721f3be149cf1482551572d364320ba966f8d0d06de09383e570625eb41bb88042fe5113e91433a15b1e8476c55e448453f724c6b3117592fe759cbb8c85d8e6d995991bb960935f7e28bcf29c76d8b6aa564877b15d5aacd79722f7ed86e92396191315d95635e453878fad9da2c7ac260ff3921ca85d285caab64d571b9b2f629f98dc974108c83601010460131954c3c4cf4be1813ee1b7b1493d88b","encoding":"hex","key":"tge-labs/mind/mind.zarr/0/embedding/c/0/1/15","truncatedAt":512}],"sampleObjects":["tge-labs/mind/mind.zarr/0/embedding/c/0/1/14","tge-labs/mind/mind.zarr/0/embedding/c/0/1/15","tge-labs/mind/mind.zarr/0/embedding/c/0/1/16","tge-labs/mind/mind.zarr/0/embedding/c/0/1/17","tge-labs/mind/mind.zarr/0/embedding/c/0/1/18"],"sidecars":[{"bytes":"# MIND global embeddings\n\n[Paper](https://arxiv.org/abs/2609.25454) · [Project page](https://research.taylorgeospatial.org/mind/) · [Code](https://github.com/taylor-geospatial/mind) · [Model weights](https://huggingface.co/taylor-geospatial/MIND)\n\nGlobal land embeddings from **MIND (Matryoshka Implicit Neural Distillation)**, a model that maps latitude and longitude to a 3,072-dimensional representation.\nEarly embedding chunks capture broad geographic patterns, while later chunks add finer spatial variation.\nRetaining only the leading dimensions lets you choose the embedding width for your task without retraining the model.\n\n<p align=\"center\">\n  <img src=\"https://research.taylorgeospatial.org/mind/assets/fig-method.svg?v=377c226\" width=\"900\" alt=\"MIND architecture: nested teacher reconstruction during training and frozen embeddings for downstream prediction.\">\n</p>\n\nLinear heads reconstruct teacher embeddings at each supervised width during training.\nDownstream prediction uses the frozen encoder with truncation or the Chunked Penalty.\n\n<p align=\"center\">\n  <img src=\"https://research.taylorgeospatial.org/mind/assets/band-chunk-first.png\" width=\"49%\" alt=\"PCA map of chunk 0, dimensions 0–63.\">\n  <img src=\"https://research.taylorgeospatial.org/mind/assets/band-chunk-mid.png\" width=\"49%\" alt=\"PCA map of chunk 3, dimensions 256–511.\">\n  <br>\n  <img src=\"https://research.taylorgeospatial.org/mind/assets/band-chunk-last.png\" width=\"49%\" alt=\"PCA map of chunk 6, dimensions 2048–3071.\">\n  <img src=\"https://research.taylorgeospatial.org/mind/assets/band-full.png\" width=\"49%\" alt=\"PCA map of the full 3,072-dimensional embedding.\">\n</p>\n\n*The first three principal components map to RGB, with PCA fitted separately for each panel.\nCompare spatial patterns across panels rather than colors.\nExplore the maps on the [project page](https://research.taylorgeospatial.org/mind/#map-section).*\n\n## Available data\n\nThe embeddings cover global land on a 0.01° latitude–longitude grid in WGS 84 (EPSG:4326), approximately 1 km at the equator.\nThe 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.\n\n| File | Format | Contents |\n| --- | --- | --- |\n| [mind.zarr](https://source.coop/tge-labs/mind/mind.zarr) | Zarr v3 | All 3,072 embedding dimensions, with coarser spatial overviews. |\n| [mind_rgb.tif](https://data.source.coop/tge-labs/mind/mind_rgb.tif) | Cloud-optimized GeoTIFF | PCA-to-RGB visualization of the embeddings. |\n| [mind_hex_r4.parquet](https://data.source.coop/tge-labs/mind/mind_hex_r4.parquet) | GeoParquet | H3 resolution 4 summaries of the first 64 dimensions. |\n| [mind_hex_r5.parquet](https://data.source.coop/tge-labs/mind/mind_hex_r5.parquet) | GeoParquet | H3 resolution 5 summaries of the first 64 dimensions. |\n\n## Read embeddings for an area\n\nInstall the required packages:\n\n```bash\npip install rioxarray xarray \"dask[array]\" \"zarr>=3\" fsspec aiohttp\n```\n\nThe example below reads the first 64 dimensions for a small area in San Francisco, decodes the stored values, and sets ocean pixels to `NaN`.\nThe result is a NumPy array with shape `[64, height, width]`.\n\n```python\nimport rioxarray\nimport xarray as xr\n\nurl = \"https://data.source.coop/tge-labs/mind/mind.zarr\"\nn_dims = 64\n\nparams = xr.open_zarr(url, consolidated=True).isel(emb=slice(0, n_dims))\nds = xr.open_zarr(\n    url, group=\"0\", consolidated=True, decode_coords=\"all\"\n).isel(emb=slice(0, n_dims))\n\n# Select an area using longitude and latitude bounds.\naoi = ds.rio.clip_box(\n    minx=-122.50, miny=37.70, maxx=-122.30, maxy=37.90\n)\n\n# Decode stored values and mask ocean pixels.\nfeatures = aoi.embedding * params.emb_scale + params.emb_offset\nfeatures = features.where(aoi[\"mask\"] == 1).to_numpy()\n```\n\nChange the bounding box for your area and `n_dims` for your embedding width.\nGroup `\"0\"` contains the finest grid; groups `\"1\"` through `\"6\"` provide progressively coarser spatial overviews.\nEach group includes its own CRS and spatial transform; `decode_coords=\"all\"` exposes them to ","encoding":"utf8","key":"tge-labs/mind/README.md","name":"README.md","size":4991,"truncated":true}],"sourceObjectKey":"tge-labs/mind/mind.zarr/0/embedding/c/0/1/14","starterCell":"with open('/mnt/archive/mind.zarr/0/embedding/c/0/1/14', 'rb') as fh:\n    head = fh.read(1024)\nprint(len(head), head[:64])","starterSamplePath":"mind.zarr","totalObjects":"1 top-level prefixes"},"summary":"Precomputed global land embeddings from MIND (Matryoshka Implicit Neural Distillation), a location encoder that maps latitude and longitude to a 3,072-dimensional representation…","title":"MIND Global Land Embeddings (Matryoshka Implicit Neural Distillation, 0.01 degree)","updatedAt":"2026-11-17T17:10:30.713Z","visibility":"public","websiteUrl":"https://source.coop/tge-labs/mind"}}