23 Million Photos: “How Do You Keep Up?”
Stop for a second and try to visualize 23 million of something. The number of species on the planet? That’s roughly 8.7 million. The number of residents in Wisconsin? Nah, that’s not even 8 million! How about photos collected by the Snapshot Wisconsin project since 2016? Ding ding ding! (well – 23,706,425 photos to be exact, not like we are counting or anything…)
Up until recently, Snapshot Wisconsin volunteers were uploading roughly one million photos total per month – but this number is bound to increase after nearly doubling our volunteer base! When we share this statistic during trainings and presentations, we always know to expect the question, “How do you keep up with all of those photos!?”
“Well, it’s a little complicated”, is how we generally start the answer. In this blog post we will dive into how the project has been maintaining this vast amount of data so far, and exciting prospects for the future of Snapshot Wisconsin.
Filtering Photos for Zooniverse:
Zooniverse is a crowdsourcing platform hosting sites for a large variety of projects, including Snapshot Wisconsin. Here anyone with internet access can go online and classify images collected by trail cameras in the project. Before photos are sent to Zooniverse, Snapshot Wisconsin trail camera hosts have the opportunity to view and classify their own photos.
While the volunteers are required to identify and remove human photos, classifying blanks and animals is extremely helpful! Why is this? We conducted an analysis to determine the accuracy of single classifications made by volunteers in their MySnapshot accounts, versus consensus classifications made by volunteers on Zooniverse. We were able to identify species that volunteers are really great at classifying (e.g. deer, squirrels, raccoons, turkeys, etc.)! When these photos are classified by volunteers in their MySnapshot accounts, we do not send them to Zooniverse. Instead we take the volunteer’s classification as the “final classification” for the image, which helps cut back on the number of photos for which we rely on Zooniverse classifications.
Crowd Sourced Classifications on Zooniverse:
The remainder of photos are uploaded to Zooniverse in “seasons”, with each season containing anywhere from 30,000 to 50,000 images. Once a season wraps up, staff members can be caught scrambling around the office getting a new season ready to go. To prepare for each season we need to upload photos and their information, as well as manually review photos to ensure no humans or excessive number of blanks get uploaded. On average, it has taken roughly 3 months for a season to be fully classified and a new season to be uploaded, which generally includes a break that lasts a few weeks.
On Zooniverse, multiple volunteers will view and classify each photo to produce a consensus classification. Photos are viewed by upwards of 11 volunteers if a consensus isn’t reached early on; if a consensus is never reached the photos will go on to expert review by staff members. Since launching in 2016, Zooniverse volunteers have helped the project move through 9 full seasons of photos. Season 10 just launched last week! Across these seasons, the project has had over 6,000 registered volunteers participate and classify 2,253,244 images, or an average of 2,596 photos per day.
Moving Forward – Machine Learning to Classify Animals:
Recently a team of researchers created a computer model using machine learning that classifies images captured by trail cameras. Machine learning is the science of getting computers to learn and act like humans do, and improve their learning over time in autonomous fashion buy feeding them data and information in the form of observations and real-world interactions¹. In this case, the computer model was provided over 3 million images of animals, each that had already been classified by a human, to aid the computer in “learning” to determine which species is which in trail camera images.
The trained model was able to classify approximately 2,000 images per minute at 98% accuracy on images of species collected in the United States. The Snapshot Wisconsin team recently prepared a large set of classified images to further improve the model, and to potentially be incorporated into the project to help keep up with the booming number of images collected. This isn’t to say that volunteer classifications (MySnapshot or Zooniverse) would be replaced, but using automated classifications could help ramp up the speed at which the project produces viable data.
To view the scientific paper, Machine learning to classify animal species in camera trap images: application in ecology, visit this link: https://www.biorxiv.org/content/biorxiv/early/2018/06/13/346809.full.pdf
- Faggella, Daniel (2018, October 29). What is Machine Learning? Retried from https://www.techemergence.com
3 responses to “23 Million Photos: “How Do You Keep Up?””
Trackbacks / Pingbacks
- December 18, 2018 -
I noticed that National Geographic’s E-Naturalist app also uses very sophisticated machine learning. It is different from your project, in that people upload photos from their phone whenever they want to, instead of regular sampling at fixed sites, so it is wider but less systematic in its coverage.
Thanks for sharing, John! There are many ongoing efforts to train machine learning algorithms for trail camera pictures. Projects like ours will certainly benefit as the technology continues to improve.