How to Write Metadata for AI-Labeled Wildlife Images

Metadata must describe both the observation and its interpretation

An AI-labeled wildlife photograph contains at least two kinds of information. The first concerns the observation: what image was captured, where it was recorded, when it was taken and what biological evidence it contains. The second concerns interpretation: what a classifier suggested, how confident it was and whether a person subsequently accepted or changed the identification. Combining these into one species-name field removes an important part of the record's history.

This distinction matters in citizen science data quality. A volunteer may submit a photograph with a species identification, while an image classifier supplies another suggestion and a coordinator later accepts a third identification. These are separate events. A useful metadata structure preserves each contribution and its status. Future users can then determine whether a species record was machine-generated, volunteer-reported, independently reviewed or still unresolved.

Build a stable identification and observation record

Every image should have a persistent identifier that remains unchanged when the file is renamed, moved or processed. A project-specific identifier can be used when the original filename is not sufficiently unique. Store the original filename and file location as additional fields. Where appropriate, a checksum can help detect accidental changes to the image file. The identifier is the link between the photograph, its metadata and any later classification or validation events.

Observation fields should include the recorded date and time, location or approved spatial reference, observer identifier, project identifier and source of the observation. Add camera station information when the image comes from a camera-trap deployment. For volunteer photographs, distinguish the place where the photograph was taken from the place where the record was submitted or processed. These distinctions help prevent administrative information from being mistaken for ecological context.

• image_id: a stable identifier for the original image.
• original_filename: the source filename.
• observation_datetime: the recorded capture date and time.
• location: the approved geographic reference.
• observer_id: a project identifier for the contributor.
• project_id: the associated survey or recording project.

Keep machine predictions in separate fields

A classifier's output should be stored as a prediction record, not written over the original species identification. Useful fields include predicted scientific name, predicted common name, taxonomic rank, model name, model version, processing date and confidence score. Include the classifier's class identifier when one is available. If a system returns several candidate species, preserve their order and scores rather than retaining only the first result.

For example, an image might have a model prediction of red fox, Vulpes vulpes, with a confidence score of 0.87. A reviewer could subsequently identify the animal as a different canid species or determine that the photograph does not support species-level identification. The record should retain the original prediction and the corrected interpretation as separate entries. The numerical score should also retain its meaning and scale, since scores from different models are not necessarily comparable.

Use a clear validation status and record the evidence

Species record validation requires an explicit status that tells later users what has happened to an identification. A practical vocabulary might include automated suggestion, volunteer reported, awaiting review, human confirmed, revised and unresolved. Projects should define these terms before data collection begins. A label such as verified is too vague unless the project explains who performed the verification, what evidence was considered and what level of identification was accepted.

The validation record should include the reviewer identifier, review date, accepted taxon, taxonomic rank and decision reason. Where possible, it should also identify the evidence used, such as the original photograph, a comparison with diagnostic features or a supporting observation. A reviewer should be able to accept a genus-level identification without being forced to select a species. The accepted taxonomic resolution should reflect the available evidence.

• Awaiting review: the identification has not been assessed.
• Human confirmed: the reviewer accepts the stated taxon.
• Revised: the original identification has been changed.
• Unresolved: available evidence does not support a reliable identification.
• Rejected: the proposed observation or identification is not accepted.

Maintain taxonomic consistency across volunteer records

Common names can vary by language and region, and the same name may refer to different taxa. Scientific names provide a more consistent basis for exchange, but they also change as taxonomic knowledge develops. Store the scientific name, taxonomic rank and, where practical, the taxonomic authority or reference used. Keep the submitted common name as a separate field instead of replacing it with a standardized name.

A data-cleaning process should identify misspellings, obsolete names, duplicate records and incompatible taxonomic ranks. For example, one volunteer may submit a record as a species while another records only the genus. These entries should not automatically be treated as equivalent. A project can map both to a common taxonomic reference while preserving the original identification and its resolution. Record the date and reference used for any taxonomic reconciliation.

Protect location data and preserve the record history

Wildlife photographs can contain sensitive location information in EXIF metadata, embedded coordinates or filenames. Some projects may need to restrict precise locations for vulnerable species or sensitive sites. Define access levels for raw coordinates, public observation records and image files. If coordinates are generalized for publication, preserve the original value in an appropriately restricted system when project permissions allow. Record the generalization method so that later users understand the spatial precision of the published record.

Metadata should also document changes over time. A revised species name, corrected timestamp or adjusted location should generate a traceable update rather than silently overwriting the earlier value. This is especially important when records are exported between spreadsheets, biodiversity databases and volunteer platforms. Before sharing a dataset, check that identifiers remain unique, dates use a consistent format, coordinates have a stated reference system and every accepted identification has a defined validation status.

A practical metadata example for a volunteer project

Consider a volunteer project receiving a photograph of a small mammal. The contributor records the image identifier, capture date, approximate location and submitted identification. An image classifier suggests a species with a confidence score of 0.91. A coordinator reviews the original photograph and accepts the genus but leaves the species unresolved. The final record should retain the submitted name, automated prediction, confidence, accepted genus-level identification and reviewer decision.

A spreadsheet can support this workflow if each column has a documented definition and each row has a stable identifier. Larger projects may benefit from separate observation, prediction and validation tables connected through that identifier. Before the next volunteer submission period, prepare a data dictionary, agree on the allowed validation statuses and add a field for the source of every identification. These three steps provide a workable foundation for consistent biodiversity records.

Questions readers ask