A dataset represents a group of one or more related STAC Collections. They group together any source imagery collections with the associated label collections to provide a convenient mechanism for accessing all of these data together. For instance, the bigearthnet_v1_source collection contains the source imagery for the BigEarthNet training dataset and, likewise, the bigearthnet_v1_labels collection contains the annotations for that same dataset. These two collections are grouped together into the bigearthnet_v1 dataset.

Radiant MLHub provides an overview of the datasets available through the Radiant MLHub API along with dataset metadata and a listing of the associated collections.

To list and fetch datasets, the Dataset class is the recommended approach, but there are also low-level client methods from radiant_mlhub.client. Both methods are described below.


The Radiant MLHub web application provides an overview of all the datasets and collections available through the Radiant MLHub API.


The objects returned by the Radiant MLHub API Dataset endpoints are not STAC-compliant objects and therefore the Dataset class described below is not a PySTAC object.

Discovering Datasets

You can discover datasets using the Dataset.list method. This method returns a list of Dataset instances.

>>> from radiant_mlhub import Dataset
>>> datasets = Dataset.list()
>>> for dataset in datasets[0:5]:  # print first 5 datasets, for example
>>>     print(dataset)
umd_mali_crop_type: 2019 Mali CropType Training Data
idiv_asia_crop_type: A crop type dataset for consistent land cover classification in Central Asia
dlr_fusion_competition_germany: A Fusion Dataset for Crop Type Classification in Germany
ref_fusion_competition_south_africa: A Fusion Dataset for Crop Type Classification in Western Cape, South Africa
bigearthnet_v1: BigEarthNet

The list() method also accepts tags and text arguments that can be used to filter datasets by their tags or a free text search, respectively. The tags argument may be either a single string or a list of strings. Only datasets that contain all of provided tags will be returned and these tags must be an exact match. The text argument may, similarly, be either a string or a list of strings. These will be used to search all of the text-based metadata fields for a dataset (e.g. description, title, citation, etc.). Each argument is treated as a phrase by the text search engine and only datasets with matches for all of the provided phrases will be returned. So, for instance, text=["maize", "rice"] will return all datasets with either "maize" or "rice" somewhere in their text metadata, while text=["maize rice"] will not match any datasets. The search text="land cover" will return all datasets with the phrase "land cover" in their text metadata.

Low-level client

The Radiant MLHub /datasets endpoint returns a list of objects describing the available datasets and their associated collections. You can use the low-level list_datasets() function to work with these responses as native Python data types (list and dict).

>>> from radiant_mlhub.client import list_datasets
>>> from pprint import pprint
>>> datasets = list_datasets()
>>> first_dataset = datasets[0]
>>> pprint(first_dataset)
{'id': 'umd_mali_crop_type',
'title': '2019 Mali CropType Training Data',

Fetching Dataset Metadata

You can fetch a dataset from the Radiant MLHub API based on the dataset ID using the Dataset.fetch method. This method returns a Dataset instance. Fetching returns the metadata but does not download assets.

>>> dataset = Dataset.fetch_by_id('bigearthnet_v1')
>>> print(
bigearthnet_v1: BigEarthNet

If you would rather fetch the dataset using its DOI you can do so as well:

dataset = Dataset.fetch_by_doi("10.6084/m9.figshare.12047478.v2")

You can also use the more general Dataset.fetch method to get a dataset using either ID or DOI.

from radiant_mlhub.client import get_dataset
# These will all return the same dataset
dataset = Dataset.fetch("ref_african_crops_kenya_02")
dataset = Dataset.fetch("10.6084/m9.figshare.12047478.v2")

Low-level client

The Radiant MLHub /datasets/{dataset_id} endpoint returns an object representing a single dataset. You can use the low-level get_dataset() function to work with this response as a dict.

>>> from radiant_mlhub.client import get_dataset_by_id
>>> dataset = get_dataset_by_id('bigearthnet_v1')
>>> pprint(dataset)
{'collections': [{'id': 'bigearthnet_v1_source', 'types': ['source_imagery']},
             {'id': 'bigearthnet_v1_labels', 'types': ['labels']}],
 'id': 'bigearthnet_v1',
 'title': 'BigEarthNet V1'}

Dataset Collections

If you are using the Dataset class, you can list the Collections associated with the dataset using the Dataset.collections property. This method returns a modified list that has 2 additional attributes: source_imagery and labels. You can use these attributes to list only the collections of a the associated type. All elements of these lists are instances of Collection. See the Collections documentation for details on how to work with these instances.

>>> len(first_dataset.collections)
>>> len(first_dataset.collections.source_imagery)
>>> first_dataset.collections.source_imagery[0].id
>>> len(first_dataset.collections.labels)
>>> first_dataset.collections.labels[0].id


There are rare cases of collections that contain both source_imagery and labels items (e.g. the SpaceNet collections). In these cases, the collection will be listed in both the dataset.collections.labels and dataset.collections.source_imagery lists, but will only appear once in the main ``dataset.collections`` list. This may cause what appears to be a mismatch in list lengths:

>>> len(dataset.collections.source_imagery) + len(dataset.collections.labels) == len(dataset.collections)


Both the class methods and the low-level client functions accept keyword arguments that are passed directly to get_session() to create a session. See the Authentication documentation for details on how to use these arguments or configure the client to read your API key automatically.

Downloading Datasets

The dataset downloader offers download of STAC catalog archives, linked dataset assets, as well as partial downloads with filtering options.

  • Robustness
    • Asset download resuming.

    • Retry and backoff for http error conditions.

    • Error reporting for unrecoverable download errors.

  • Performance
    • Scales to millions of assets.

    • Multithreaded workers: parallel downloads.

  • Convenience
    • STAC collection_id and item asset key filter

    • Temporal filter

    • Bounding box filter

    • GeoJSON intersection filter

Download All Assets

The most basic usage is to fetch a dataset, and then call it’s download method. The output directory is the current working directory (by default).

>>> from radiant_mlhub import Dataset
>>> nasa_marine_debris = Dataset.fetch_by_id('nasa_marine_debris')
>>> print(nasa_marine_debris)
nasa_marine_debris: Marine Debris Dataset for Object Detection in Planetscope Imagery
nasa_marine_debris: fetch stac catalog: 258KB [00:00, 412.53KB/s]
INFO:radiant_mlhub.client.catalog_downloader:unarchive nasa_marine_debris.tar.gz ...
unarchive nasa_marine_debris.tar.gz: 100%|████████████████████| 2830/2830 [00:00<00:00, 5772.09it/s]
INFO:radiant_mlhub.client.catalog_downloader:create stac asset list (please wait) ...
INFO:radiant_mlhub.client.catalog_downloader:2825 unique assets in stac catalog.
download assets: 100%|██████████████████████| 2825/2825 [03:27<00:00, 13.62it/s]
INFO:radiant_mlhub.client.catalog_downloader:assets saved to nasa_marine_debris

Download STAC Catalog Archive Only

If you want to inspect the STAC catalog or write your own download client for the assets just pass the catalog_only option to the download method:

sen12floods: fetch stac catalog: 2060KB [00:00, 127903.52KB/s]
INFO:radiant_mlhub.client.catalog_downloader:unarchive sen12floods.tar.gz...
unarchive sen12floods.tar.gz: 100%|█████████████████████████████████████████| 22278/22278 [00:01<00:00, 14284.65it/s]
INFO:radiant_mlhub.client.catalog_downloader:catalog saved to /home/user/sen12floods


The Python logging module can be used to control the verbosity of the downloader. The default log level is INFO.

  • Turn on WARNING level to see fewer log messages.

  • Set DEBUG level to see more messages. This includes verbose HTTP-level log messages.

>>> import logging
>>> logging.basicConfig(level=logging.DEBUG)
DEBUG:radiant_mlhub.client.catalog_downloader:(thread id: 123145809592320) -> .../nasa_marine_debris/nasa_marine_debris_labels/nasa_marine_debris_labels_20170326_153234_0e26_17069-29758-16/pixel_bounds.npy
DEBUG:urllib3.connectionpool:Starting new HTTPS connection (1):
DEBUG:urllib3.connectionpool: "HEAD /mlhub/nasa-marine-debris/labels/20181031_095925_103b_32713-31765-16.npy HTTP/1.1" 200 0
(omitted many log messages here)

Output Directory

The output directory is by default, the current working directory. The output_dir parameter takes a str or pathlib.Path. It will be created if it does not exist.

# output_dir as string'/tmp')

# output_dir as Path object
from pathlib import Path / 'my_projects' / 'ml_datasets')

Large Dataset Performance

Let’s try a bit larger dataset (tens of thousands of assets). After downloading the complete dataset, we’ll explore all of the options for filtering assets. Filtering lets you limit the items and assets to those you are interested in prior to downloading.

This download example was run on a compute-optimized 16-core virtual machine in the MS Azure West-Europe region. You would likely experience slower download performance on your machine, depending on number of cores and network bandwidth.

>>> sen12floods = Dataset.fetch_by_id('sen12floods')
>>> %%time
sen12floods: fetch stac catalog: 2060KB [00:00, 127699.36KB/s]
INFO:radiant_mlhub.client.catalog_downloader:unarchive sen12floods.tar.gz...
unarchive sen12floods.tar.gz: 100%|█████████████████████████████████████████| 22278/22278 [00:01<00:00, 14239.53it/s]
INFO:radiant_mlhub.client.catalog_downloader:create stac asset list...
INFO:radiant_mlhub.client.catalog_downloader:39063 unique assets in stac catalog.
download assets: 100%|███████████████████████████████████████████████████████████| 39063/39063 [06:26<00:00, 101.06it/s]
INFO:radiant_mlhub.client.catalog_downloader:assets saved to /home/user/sen12floods

CPU times: user 11min 44s, sys: 2min 15s, total: 14min
Wall time: 6min 40s

15GB of assets were downloaded into the sen12floods/ directory. You may not necessarily want to download all of the assets in a dataset. In the following sections, all the filtering options are explained.


Download filters may be freely combined, except bbox and intersects which are independent options.

Checking Dataset Size

Consider checking the dataset size before downloading.

>>> dataset = Dataset.fetch('nasa_marine_debris')
>>> print(dataset)
nasa_marine_debris: Marine Debris Dataset for Object Detection in Planetscope Imagery
>>> print(dataset.stac_catalog_size)  # OK the STAC catalog archive is only ~260KB
>>> print(dataset.estimated_dataset_size)  # OK the total dataset assets are ~77MB

Filter by Collection and Asset Keys

To download only the specified STAC collection ids and STAC item asset keys, create a dictionary in this format and pass it to the collection_filter parameter:

{ collection_id1: [ asset_key1, asset_key2, ...], collection_id2: [asset_key1, asset_key2, ...] , ... }

For example, using the sen12floods dataset, if we only wanted to download four bands of the source imagery:

my_filter = dict(
    sen12floods_s2_source=['B02', 'B03', 'B04', 'B08'],   # Red, Green, Blue, NIR
    sen12floods_s2_labels=['labels', 'documentation'],

Filter by Temporal Range

To download only STAC assets within a temporal range, use datetime parameter to specify a datetime range (tuple of datetime objects), or a single datetime


from dateutil.parser import parse
my_end_date=parse("2019-04-07T00:00:00+0000"), my_end_date))

Filter by Bounding Box

To download only STAC assets with a spatial bounding box, use the bbox parameter to specify a bounding box in lat/lng (CRS EPSG:4326). This performs a spatial intersection test with each STAC item’s bounding box.

my_bbox = [-13.278254, 8.447033,
           -13.231551, 8.493532]


The bbox filter may not be used with the intersects filter (use one or the other).

Filter by GeoJSON Area of Interest

To download only STAC assets within an area of interest, use the intersects parameter. This performs a spatial intersection test with each STAC item’s bounding box.

import json
my_geojson = json.loads(
        "type": "Feature",
        "geometry": {
            "type": "Polygon",
            "coordinates": [


The intersects filter may not be used with the bbox filter (use one or the other).

Error reporting

Any unrecoverable download errors will be logged to {output_dir}/{dataset_id}/err_report.csv and a Python exception will be raised.

Appendix: Default Filesystem Layout of Downloads

STAC archive file:


Unarchived STAC catalog:


Collection, Item and Asset layout:


Common Assets, ex: documentation.pdf are saved into a _common directory instead of duplicating them for many items:


Asset Database:


Error Report:



The mlhub_stac_assets.db file is an artifact which may be safely deleted to free up disk space.