WorldView-3 satellite image tiles of wandering albatross breeding sites on South Georgia with citizen science annotations of individual birds, 2015-2022
GB/NERC/BAS/PDC/02157
Page Links
Jump To:
- Citation
- Access Data
- Constraints
- Basic Information
- Additional Information
- Locality
- Instrumentation
- Storage
Related Links
Summary
Abstract:
This dataset comprises of two components: i) Image data: 10,833 .png image tiles derived from 35 unique 31-cm resolution WorldView-3 satellite images. These images are of wandering albatross breeding sites on South Georgia during the late December - end of March breeding season in 2015-2022, and ii) Annotation data: .csv and .json files containing point annotations (in x, y pixel coordinates) identifying individual wandering albatrosses within the image tiles. These annotations were generated by citizen science volunteers as part of the Maxar GeoHive "Albatrosses from Space" campaign. Each image tile was reviewed independently by seven observers, and all unique observer annotations are included in the dataset. A separate set of annotations generated by expert observers are also included on a subset of image tiles for validation.
Very high-resolution satellite imagery can potentially be used to monitor large seabirds such as great albatrosses (Diomedea spp.) directly from space. These birds nest on remote islands, and while detectable in 31-cm satellite imagery, each individual appears as only a few white pixels, making manual analysis slow, costly, and subjective. This open-source dataset is intended to support the development of novel automated detection and counting methods for wandering albatrosses as well as other species visible in very high-resolution satellite imagery.
This dataset was collected and developed with support from the Darwin Plus scheme through the following grants: 1) DPLUS132: "Monitoring albatrosses using very high resolution satellites and citizen science", 2) DPLUS187: "Using satellite technology to monitor seabird populations at South Georgia". These grants supported both the implementation of the citizen science campaign and the and release of the annotated dataset.
Keywords:
South Georgia, citizen science, inter-observer variation, machine learning, satellite imagery, wandering albatross
Citation
Bowler, E., Attard, M., Phillips, R., & Fretwell, P. (2026). WorldView-3 satellite image tiles of wandering albatross breeding sites on South Georgia with citizen science annotations of individual birds, 2015-2022 (Version 1.0) [Data set]. NERC EDS UK Polar Data Centre. https://doi.org/10.5285/fd82803b-6764-4b50-a8ef-0e8729c07870
Access Data
GET DATA
PROJECT HOME PAGE
- https://www.bas.ac.uk/project/south-georgia-seabirds-from-space/
- https://www.bas.ac.uk/project/wildlife-from-space/albatrosses-from-space/
PROJECT HOME PAGE
REFERENCE MATERIALS
SOFTWARE PACKAGES
SOFTWARE PACKAGES
Constraints
| Access Constraints: | This dataset is under embargo until the publication of an associated paper. |
|---|---|
| Use Constraints: | The image tiles and citizen science annotations in this dataset is under the end user licence terms by Vantor: Internal Use License | Maxar. All derivatives (including data derivatives) must include the following copyright notice on or adjacent to the derivative: Vantor Products. Satellite image annotations 2026 Vantor. |
Basic Information
| Creation Date: | 2026-02-13 |
|---|---|
| Dataset Progress: | Planned |
| Dataset Language: | English |
| ISO Topic Categories: |
|
| Parameters: |
|
| Personnel: | |
| Name | UK Polar Data Centre |
| Role(s) | Metadata Author |
| Organisation | British Antarctic Survey |
| Name | Dr Ellen Bowler |
| Role(s) | Investigator, Technical Contact |
| Organisation | British Antarctic Survey |
| Name | Dr Marie R G Attard |
| Role(s) | Investigator |
| Organisation | British Antarctic Survey |
| Name | Dr Richard A Phillips |
| Role(s) | Investigator |
| Organisation | British Antarctic Survey |
| Name | Dr Peter T Fretwell |
| Role(s) | Investigator |
| Organisation | British Antarctic Survey |
| Parent Dataset: | N/A |
Additional Information
| Reference: | Attard, M. R. G., Phillips, R. A., Bowler, E., Davis, D. & Fretwell, P. T. (2026) Crowdsourcing for conservation: A citizen science approach to monitoring wandering albatrosses using satellite imagery. [in prep]. | |
|---|---|---|
| Quality: | Each satellite image tile was annotated by seven unique citizen scientists, who each received the same training and information prior to completing the task. Since there is a large amount of uncertainty in detection of wandering albatrosses, which only appear as approximately 4-5 pixel white dots, annotations from all observers are given in the dataset. This will allow exploration into machine learning approaches that can learn from and incorporate label uncertainty. A summary of the variance and agreement in these counts compared to ground-truthed data is provided by Attard et al. (2026). As an additional quality control measure, a subset of the imagery was annotated by seven remote sensing experts with specific expertise in identifying wildlife in very high-resolution satellite imagery. For this, a private campaign was conducted on the GeoHive platform, where the seven experts classified and labelled albatrosses in 543 image tiles from Bird Island (catalogue ID: 10400100066C1E00), Albatross Island (catalogue ID: 1040010029A1D400), and Prion Island (catalogue ID: 1040010029A1D400 and 10400100655C5200). Each tile was reviewed by all experts, producing seven classifications per tile. The same tutorial, help, and interface used in the crowd campaign were applied here. These expert annotations are provided in a separate .csv and .json file, allowing for comparison with citizen science data and potential use as a benchmark or validation set. |
|
| Lineage/Methodology: | Thirty five WorldView-3 satellite images (31-cm resolution) were selected, captured in late December-March (incubation period) between 2015 and 2022. These images cover 24 wandering albatross breeding sites across South Georgia. Only breeding sites reported to have at least five breeding pairs, as recorded in the 2003/04 and 2014/15 ground censuses, were included. The selected satellite images were split into 11,839 unique tiles, each measuring 150 m x 150 m with a 5 m overlap to ensure any birds on the tile edges were included. These tiles were classified and annotated though a Vantor GeoHIVE citizen science campaign, involving 639 volunteers, with each tile independently reviewed by seven observers. Citizen scientists received the same training and instructions on the GeoHIVE platform. They were asked to visually inspect each image tile and classify it as either 1) having albatross present, 2) having no albatross present, or 3) being a poor-quality image (e.g. to cloudy or dark to determine albatross presence or not). If albatross presence was identified, annotators were asked to place a point marker on any white dot they believed to be a wandering albatross. A summary of the training is available in Attard et al. (2026) After the campaign annotations were post-processed, to clean and standardise the outputs. GeoHive partners filtered "bad workers" (e.g., annotations generated by bots which were evidently poor quality). Subsequent to this, tiles which had no longer been reviewed by exactly seven observers were removed to maintain consistency. In some cases this included images which had received more than seven views. After this post-processing a total of 10,833 image tiles with annotations remained, which are included in this final dataset. Here image tiles are provided in non-georeferenced RGB .png format, in accordance with Vantor's license policy (https://www.maxar.com/legal/internal-use-license). Image and annotation information are included as separate .csv files, which can be linked using the "image_id" identifier. Image files contain metadata for each satellite image tile, including: image id, file name, image acquisition date, off nadir angle, sensor name, target azimuth, cloud cover percentage, and the name of the breeding site captured in the image. Note that cloud cover percentage is calculated by Vantor's in house algorithm and is assigned to each entire satellite image scene, rather than each image tile. A catalogue ID is also included, so the original satellite image can be identified and purchased from Vantor if required. Finally, tile-level metadata from the citizen science campaign is included. This includes the total number of annotators ("num_reviewers", n=7), and the number of annotators who identified the tile as 1) having albatross present ("num_feature_present"), 2) having no albatross present ("num_no_feature_present"), and 3) being a poor-quality image ("num_poor_image"). The annotations .csv file includes: the image id the annotation was made on, the user id for the annotator, the point annotation as x,y pixel coordinates, and the feature class name. The image and annotation information is additionally formatted into nested .json format, ready for use in common machine learning algorithms. We hope this will support the development of automated detection systems including methods for integrating uncertainty from multi-annotator labels. |
|
Locality
| Temporal Coverage: | |
|---|---|
| Start Date | 2015-01-10 |
| End Date | 2022-01-17 |
| Location: | |
| Location | South Georgia Island |
| Detailed Location | N/A |
Instrumentation
| Data Collection: | Satellite sensor: Imagery data was collected using the Maxar WorldView-3 optical satellite sensor, which proves 31-cm resolution multispectral imagery. Annotation platform: Satellite images were processed by Maxar, including dividing images into image tiles. The image tiles were hosted and annotated on the Maxar GeoHive platform (https://geohive.maxar.com/). Citizen scientists accessed the website on their own device. They classified each image tile as 'Albatrosses Present', 'No Albatrosses Present' or 'Poor image'. For image tiles classified as 'Albatrosses Present', citizen scientists were asked to then annotate those images using point markers. Once completed, the citizen scientist was shown the next image tile to evaluate. The platform allowed users to zoom, sharpen, and brighten image tiles to aid visual inspection, though the dataset itself contains the original, unadjusted imagery. Post-processing software: The original Geotiff image tiles were converted to non-georeferenced RGB .png format in compliance with Maxar's licence policy. Annotation shapefiles were converted to x,y pixel coordinates per image tile. All post-processing was conducted using python. The scripts for image and annotation processing are available at: https://github.com/EllieBowler/wandering-albatross-worldview-data-paper. Detailed locations Bird Island, Demidov isthmus, Invisible Island, Mollyhawk Island, Numez Peninsula, Trollhul, Frida Hole, Outer Lee, Proud Island, Cape Rosa, Tidespring Island, Granat Point, Kupriyanov Island outer, Kade Point, Weddell Point, Crescent Island, Inner Lee, Coal Harbour, Saddle Island, Cape Alexandra, Prion Island, Albatross Island |
|---|
Storage
| Data Storage: | 10,833 satellite image tiles (.png); 4.16 GB Citizen science annotations (.csv); 3.66 MB Citizen science image list (.csv); 1.95 MB Expert annotations (.csv); 0.4 MB Expert annotations image list (.csv); 0.9 MB Citizen science annotations formatted for machine learning (.json); 15.7 MB Expert annotations formatted for machine learning (.json); 1.2 MB Satellite images are 31-cm resolution WorldView-3 data products. |
|---|