How to Add AI Image Classification to a Citizen Science Project: A Volunteer Coordinator's Checklist

Start with the project's recording objectives

Before introducing image classification, establish what the project intends to record. A reserve monitoring mammals, a school biodiversity survey and a volunteer butterfly inventory have different identification requirements. One may need reliable species-level records, while another may only need to distinguish animals from empty frames or assign observations to broad taxonomic groups. The required level of identification determines which images can be classified automatically and which need specialist attention.

Write a short identification policy that specifies the accepted taxonomic level, geographic scope, target species and treatment of uncertain records. For example, a project monitoring mammals across a woodland might accept a genus-level identification when the photograph does not show diagnostic features. This is preferable to forcing every volunteer to choose a species. The policy should also state who can approve a change to an existing identification.

• Define the target taxa and geographic recording area.
• Specify the minimum acceptable identification level.
• Identify records requiring specialist review.
• Assign responsibility for approving taxonomic changes.

Prepare volunteers and image collections

A classification system depends on the material it receives. Volunteers should receive instructions on image resolution, file formats, duplicate photographs, camera orientation and basic image handling. Preserve original files in a controlled project archive. Resizing or converting images may be necessary for some tools, but the processing copy should remain distinguishable from the source photograph.

Explain that an image can contain an animal without providing enough evidence for identification. A blurred silhouette, partial tail or distant bird may support an observation at a broad level but not a species-level label. Include examples in volunteer training and ask participants to practice assigning an uncertain status. A short exercise using 20 varied photographs can expose inconsistent interpretations before the project begins accepting larger submissions.

• Keep an unchanged copy of each original image.
• Publish accepted file formats and image requirements.
• Include examples of partial, blurred and obscured animals.
• Teach volunteers when to leave an identification unresolved.

Choose a data sheet that separates observations from predictions

A field data sheet template should distinguish information about the observation from information produced during classification. Observation fields describe what was recorded, where it occurred and when the image was collected. Classification fields describe the proposed taxon, the method used to obtain that label and the person or process responsible for accepting it.

A practical spreadsheet might contain record_id, image_filename, observation_date, site_code, observer_id, taxon_proposed, identification_level, classifier_name, confidence_value, reviewer_id and validation_status. Keep the original image identifier stable even if the filename changes during processing. For example, record CT-00482 can retain its identity when a processing copy is created or a species label is revised.

Use controlled vocabularies for fields such as validation_status. Suitable values might include pending, accepted, revised and unresolved. A blank field should not carry several possible meanings. It could indicate missing information, an unfinished record or a field that does not apply. Those states should be represented explicitly when they matter to subsequent analysis.

• Use a stable identifier for every observation.
• Separate proposed taxon from accepted taxon.
• Record classifier name and identification method.
• Use controlled values for validation status.
• Preserve the original observation date and site reference.

Set rules for reviewing automated classifications

A coordinator should decide in advance which classifications require human review. A high confidence value is not proof of a correct identification. It indicates how strongly a particular system supports its prediction under its own scoring method. Scores from different tools are not necessarily comparable, and some systems provide no meaningful confidence estimate at all.

One workable arrangement is to route every proposed species-level identification through a reviewer during the initial project phase. Broader classifications may receive a lighter review, depending on the consequences of an error. For example, a photograph classified as mammal may be sufficient for an introductory biodiversity inventory, while a proposed record of a locally rare species should receive specialist attention. The project should define these rules before volunteers encounter exceptional records.

Clean occurrence records before sharing or analysis

Species occurrence data cleaning should take place before records are exported to a biodiversity database or used in project summaries. Common problems include duplicate observations, inconsistent taxonomic names, invalid dates, missing coordinates, spelling variations and records assigned to incompatible identification levels. A photograph uploaded twice should not automatically become two independent biological observations.

Use an accepted taxonomic reference appropriate to the project's geographic and scientific context. Record the reference or version used, because scientific names and accepted classifications can change. Check that the reported taxon is consistent with the project's recording area, but treat an unexpected geographic record as a reason for investigation rather than automatic deletion.

A useful validation rule is to flag records with a species-level label but no supporting identification evidence or reviewer status. Another is to flag dates that fall outside the project survey period. These checks identify records for attention; they do not determine whether an unusual observation is biologically impossible.

• Identify duplicate image and observation identifiers.
• Standardize scientific names and taxonomic ranks.
• Flag incomplete dates and geographic information.
• Check consistency between identification level and evidence.
• Retain rejected and revised records in the audit history.

Establish responsibilities and a sustainable review schedule

Volunteer projects need a clear division of responsibility. Contributors can submit images and provide initial observations. Trained coordinators can manage incomplete records, while designated taxonomic reviewers handle difficult identifications and exceptional species. The person responsible for maintaining the data sheet should also control changes to field definitions and accepted vocabulary.

Start with a manageable review cycle. For example, a coordinator might examine submissions weekly, while a specialist receives a monthly batch of uncertain records. Record the date of each review and any change to the accepted identification. If the project grows, review workload should be assessed using the number of unresolved records, the frequency of corrections and the time required for specialist decisions.

Publish a short volunteer guide explaining how to submit images, how to interpret classification suggestions and how to report mistakes. Provide a contact route for disputed records. This makes the workflow repeatable when volunteer membership changes and prevents important identification decisions from remaining in private messages or individual spreadsheets.

Document the workflow before opening submissions

The project documentation should describe the classification tool, image preparation requirements, taxonomic reference, review responsibilities and data retention arrangements. Product capabilities, export formats and geographic coverage can change, so coordinators should confirm current documentation before adopting a particular system. Projects working across institutions should also agree on compatible identifiers and data fields before exchanging records.

A useful final preparation exercise is to take 30 example images through the complete submission and review process. Include clear animals, ambiguous photographs, duplicates and images without identifiable wildlife. Check whether each record receives an appropriate status, whether the original image remains accessible and whether the resulting spreadsheet can distinguish a machine suggestion from an accepted observation. Document the corrections needed and update the volunteer instructions before the first public submission period.

Questions readers ask