Datasetpublic

calebrob6/geobenchv1-webdataset

GeoBench V1 - pickle-free sharded edition This is a data-only metadata conversion of the six classification datasets in isaaccorley/geobenchv1-webdataset, pinned at revision cb847e5ff87a2c8f00064d631ed7b6a39a084c68. The original benchmark is GEO-Bench. The image arrays, sample IDs, labels, shard boundaries, and train/validation/test partitions are unchanged. Sample metadata is stored as JSON instead of pickle. No pickle deserialization is needed to read this edition.… See the full description on the dataset page: https://huggingface.co/datasets/calebrob6/geobenchv1-webdataset.

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes93downloads

Dataset Card

GeoBench V1 - pickle-free sharded edition

This is a data-only metadata conversion of the six classification datasets in isaaccorley/geobenchv1-webdataset, pinned at revision cb847e5ff87a2c8f00064d631ed7b6a39a084c68. The original benchmark is GEO-Bench.

The image arrays, sample IDs, labels, shard boundaries, and train/validation/test partitions are unchanged. Sample metadata is stored as JSON instead of pickle. No pickle deserialization is needed to read this edition.

Dataset	Samples	Shards	Classes	Task
m-bigearthnet	22,000	22	43	Multilabel classification
m-brick-kiln	17,061	18	2	Classification
m-eurosat	4,000	4	10	Classification
m-forestnet	8,446	9	12	Classification
m-pv4ger	13,812	14	2	Classification
m-so2sat	21,964	22	17	Classification
Total	87,283	89

Format

Each dataset directory contains its original partition/statistics JSON files, original README and LICENSE, and shard_*.tar files. Each sample has two archive members:

text

<sample_id>.bands.npz
<sample_id>.meta.json

The .bands.npz payloads are copied byte-for-byte from the source. Their embedded arrays are numeric and can be loaded with numpy.load(..., allow_pickle=False).

The JSON metadata contains label (a class ID or multilabel vector), bands_order (ordered source-band names), and per-band transform and crs values. Affine transforms are numeric lists, CRS values are strings, and missing georeferencing is null. Each JSON record includes an _conversion notice describing the format change.

The unused task_specs.pkl sidecars are omitted. Python band-description objects, date objects, and auxiliary per-band records are not included in the JSON schema; source-band names, including temporal suffixes, are preserved. Original label maps and label statistics are retained wherever the source supplies them.

Reading with torchgeo-bench

Use a torchgeo-bench version that supports V1 .meta.json shards. Legacy pickle-only readers are not compatible with this metadata format. Download explicitly with the Hugging Face CLI into a fresh directory:

bash

hf download calebrob6/geobenchv1-webdataset --repo-type dataset \
  --include "m-eurosat/*" --local-dir data/classification_v1.0_wds
torchgeo-bench run model=rcf "dataset.names=[m-eurosat]"

SHA256SUMS contains one checksum per tar archive. For a complete download, run sha256sum -c SHA256SUMS from the repository root.

Licenses and attribution

The original per-dataset LICENSE and README files are preserved verbatim, including their attribution and source links. The format conversions are distributed under the same applicable dataset licenses; there is no blanket license that replaces those terms.

Dataset	Upstream license
m-bigearthnet	Community Data License Agreement - Permissive 1.0
m-brick-kiln	CC BY-SA 4.0
m-eurosat	CC BY 4.0 for imagery; MIT for labels
m-forestnet	CC BY 4.0
m-pv4ger	MIT
m-so2sat	CC BY 4.0

All tar archives have been modified to replace pickle metadata with JSON. Image-array payloads and benchmark partitions have not been modified. See each dataset's original notices for its creators and applicable licensing details.