Wildlife Insights
We’re excited to announce the release of our latest AI model! Learn more about this latest model on our About our AI page and on Google Earth’s Medium blog.
A larger training dataset
We more than doubled the number of images in our training data. This model was trained on 35 million images - up from 15 million images in the last round of training.
The number of species in our training data also grew from 993 to 1,295. When including non-animal classes and higher taxonomic classes, our model is now trained to recognize 1,532 classes.
Check out the species table to see which species were part of the training dataset.
Improved accuracy across the board
Overall, the model is more confident when it makes a prediction and is able to correctly identify more images compared to the last model.
More blank images are caught
Our last model update showed big improvements in blank ID accuracy. This new model builds off those updates and identifies even more blank images. It’s also less likely to incorrectly identify an object as a blank, minimizing the possibility of removing valuable images of animals.
When testing on projects that weren’t part of the training process, the model caught 88.2% of blanks with an error rate of less than 2%, compared to 75.8% and 4% error rate in the previous model.
More human images are correctly classified
In part thanks to the introduction of a data cleanup process that filters mislabeled images in our dataset, the new model identifies many more human images.
When testing on projects that weren’t part of the training process, the model caught 96.7% of humans with an error rate of 2.2%, compared to 94.3% with a 11% error rate in the previous model.
More animals are identified
The model is more likely to identify an animal at the species level and is less likely to return an unknown, or “No CV Result”. This means there are less images that need close review.
When testing on projects that weren’t part of the training process, the number of images the model wasn’t confident about (“No CV Result”) was reduced to 3.6% from 7.6% in the old model.
Country-level Geofilters
We’ve added in a geofilter that blocks the AI model from predicting a species or higher taxonomic level if it doesn’t occur in the same country as the camera location. This means that if your camera location is in Brazil, the AI model won’t predict that an image shows a Malay tapir. If you see that the AI model is predicting species that shouldn’t occur in your country, let us know in a bug report.
In addition to these country level filters, we’ve also built in state filters for the United States. This will help avoid common issues like misidentifying a white-tailed deer as a mule deer in certain U.S. states.
Using the tools in Wildlife Insights to speed up data processing:
As a first step, we recommend reviewing images identified as Blank to remove the bulk of images in your dataset. Here’s how:
Go to the Identify tab in your project.
Apply the “Blank” status filter.
Scroll to the bottom of the page and select the drop down option to view 1000 images per page. (coming soon, this option will be placed at the top of the page)
Increase the thumbnail size to Medium or Large.
Review each thumbnail to look for objects or animals.
If all images are blank, you can use shortcut keys to select all images on the page (Ctrl+A or Cmd+A) and mark the images as blank.
If any image appears to have an object of interest, you can select that image, edit the ID, remove it from the Identify page, and return to reviewing for blanks.
From there, you can continue to filter for images by Deployment, Species or by “No CV Result” and review in a similar manner.