The video recording is available here: Camtrap DP on Vimeo
Here is a transcript of the Q&A:
The coordinates shared are the ones of the camera, not the organism observed. How are the coordinates from Camtrap DP datasets interpreted by GBIF?
The location of the camera is indeed the location of an occurrence. Typically cameras won’t detect much further than 30 meters, maybe 100 meters under particularly good conditions (like an open field). So the occurrence is very close to the camera location. The detection distance can be indicated in detectionDistance.
A bigger contributor to uncertainty are generalized or rounded, for example, if you would not like to share the precise location of the camera. You can indicate how much you rounded in the metadata, in coordinatePrecision. The sum of both uncertainties (and errors on the GPS measurement in the field) can be expressed in coordinateUncertainty. That field is mapped to Darwin Core Quick Reference Guide - Darwin Core in Darwin Core.
The R package camtrapdp has a function round_coordinates() that will set both coordinatePrecision and coordinateUncertainty based on how much you round. The same R package is used by GBIF to convert published Camtrap DPs to Darwin Core.
How can I familiarise myself with the Camtrap DP format? Are there templates available?
Note that camera trap data can be shared in different formats, using Camtrap DP is not the only way.
With that in mind, strongly encourage you to check the camera trap publishing guide: Reyserhove L, Norton B & Desmet P (2023) Best Practices for Managing and Publishing Camera Trap Data. GBIF Secretariat: Copenhagen. Best Practices for Managing and Publishing Camera Trap Data
See Best Practices for Managing and Publishing Camera Trap Data, the best option is to use an existing system, many of which are free then upload the Camtrap DP to the IPT. You can also find an exemplar dataset here: Example dataset - Camtrap DP.
How are the scientific names generated?
Some camera trap data management systems have built-in automated species identifications (based on the images). Then the species can be manually verified by someone.
Would the recommendation be to preprocess existing datasets?
You might have identifications generated in various ways (AI, manual). You should generate a consensus identification. You have to preprocess your data otherwise you will have several occurrence records with conflicting information. You could use for example a filter to exclude records identified by an AI algorithm below an 80% certainty threshold.
There was a lot of mapping in the presentation, would you say that it was the most time-consuming aspect of publishing the dataset?
The time consuming part for Cecilie was fixing the unexpected issues after publishing. In this case, the system couldn’t ingest the data because of duplicate identifiers. Cecilie had to go back and figure out what was the issue and how to fix it.
Note that the Python library frictionless-py can be used to validate the archive prior to data publishing. See Camtrap DP. It will check that your archive is structurally sound, on your own computer. It won’t give you warning if an unexpected species is present but it will tell you if you have duplicate identifiers for example.
We hope that mapping all the data in the IPT wouldn’t be a typical workflow (although it might be needed for legacy data). Most publishers would download a Camtrap DP from a camera trap management system and upload it to the IPT. The IPT would make some checks before publishing it on GBIF. We are in the process of improving the documentation.
When trying to publish such a dataset in the past year, there were two challenges: the new Camtrap DP and the new IPT at the time (version 3.0). You could publish this dataset with the latest version of the IPT but this isn’t the IPT version I have and I am not able to update it yet. So maybe if you have issues publishing camtrap DP, it might be because your IPT version is too old.
Thank you for reporting so many issues on the IPT. These are the challenges of being an early adopter. We really appreciate your input and feedback.
Are the issues associated with Camtrap DP in previous IPT versions concerning the generation of the Camtrap DP archive? Would someone be able to upload a well-formatted archive to an older IPT and publish it?
No, it wouldn’t work. You would need to update the IPT to the latest version.
If we have camera trap data on GBIF that were published prior to the development of Camtrap DP, should we reformat the data to Camtrap DP? What would be the recommendation?
It depends on what you would like to do with the data. If you have a lot of data that was lost when mapping to Darwin Core, you could consider republishing the data as Camtrap DP. If not, it is fine to leave the data as they are.
Note that there is no way to currently change the format of the data directly in the IPT.
Eventually, we would like to make a view for users to discover and download Camtrap DP data. However, it would be difficult to highlight camera trap data that weren’t shared as Camtrap DP.
Related to image annotation: I have a data provider making annotations of photos taken from an unmanned underwater vehicle. Often, these photos often don’t allow for species-level identification but this person knows the region and what species are likely to be found (for example, there is only one species of pink starfish in Antarctica). Also divers work with them so the provider is pretty confident in their identification even though the photos alone might not be enough evidence. How can they express these levels of confidence when publishing this type of data?
In Camtrap DP, there is a field called classificationProbability which allows providers to convey how certain they are about the identification. There is no similar field in Darwin Core. There is something called identificationVerificationStatus but it isn’t equivalent. It recommends the use of a standard for herbaria.
It would be a good field to have in Darwin Core and GBIF.
