Abstract
This article presents a formal correction to the data descriptor for the Penguin Watch project. Following a post-publication audit, a duplication error was identified in the original Dryad repository, leading to inflated image and camera counts. This correction provides updated metrics and confirms the integrity of the original technical validation.
Introduction
In large-scale citizen science projects like Zooniverse’s Penguin Watch, managing tens of thousands of images and volunteer classifications presents a significant data-wrangling challenge. Following the 2018 publication of the project's data descriptor, the authors discovered that two datasets (PETEc2014 and PETEd2013) deposited in the Dryad repository were duplicates.
This duplication led to inaccuracies regarding the total volume of data available, the number of cameras deployed, and the structure of the metadata. This correction ensures the scientific record reflects the actual scope of the project and provides a cleaned version of the dataset for the research community.
Methodology: Data Audit and Repository Update
The correction process involved a rigorous data audit and repository update. The authors identified specific duplicate folders within the Dryad repository and recalculated the total number of unique images and cameras.
Crucially, the authors verified that the technical validation performed in the original paper remained sound. Because the validation randomly selected images from the correct source folders rather than the duplicates, the reported accuracy of the volunteer classifications was not compromised. Finally, a corrected version of the dataset (v2) was deposited to Dryad, removing redundant folders and associated citizen science classification files.
Revised Dataset Metrics
The core of the Penguin Watch dataset consists of time-lapse imagery from Antarctica and the Southern Ocean, paired with "consensus click" data—volunteer-provided coordinates identifying penguins. The audit resulted in the following quantitative adjustments to the project's metrics:
- Total Image Count: Reduced by approximately 14.5%, from 73,802 to 63,070.
- Camera Count: Adjusted from 15 to 14.
- Folder Structure: The data is now organized into 22 image folders (down from 23) with 32 corresponding metadata and consensus folders (down from 34).
- Removed Files: Specific citizen science data files, including PETEd2013a_concl and PETEd2013b_concl, have been purged to eliminate redundancy.
Discussion: Data Hygiene and Integrity
This correction highlights the necessity of rigorous data hygiene, version control, and post-publication auditing for large open-access datasets. Even in high-impact journals, errors in data deposition can occur. The integrity of the original validation—confirmed because it utilized the correct source files—is a vital finding that preserves the utility of the existing research.
By providing a clear mapping of the errors and a direct link to the updated version (v36g.2), the authors maintain transparent provenance for the data, allowing the community to build upon a verified foundation.
Conclusion
Researchers who downloaded the Penguin Watch dataset prior to September 2019 should update their local copies to the new version to avoid training models on duplicate data or reporting inflated sample sizes. The corrected dataset is now available via Dryad at https://doi.org/10.5061/dryad.vv36g.2.
Summary updated