Major model upgrade

Ondřej Veselý
November 24, 2023
6 min read
Share this post
We are excited to announce the latest upgrade to the model coverage and performance.

There is a more than 2.5x increase in the number of classes (this includes genus and species levels). Specifically:

Old model: 12,535 classes
New model: 33,325 classes

Below are the results of the test comparing the two models on the validation set weighted by taxa popularity. The three graphs show improvements in the Top 1, Top 3 and Top 10 suggestions.

To gain further insight into the model performance, we evaluated the performance of both models on a GBIF sample of 50,000 observations (dated to the first half of 2023), broken down by world regions. We chose GBIF because it is a highly trusted and reliable international platform that integrates diverse open-access data, including plants.

Of the observed plants, the new model was able to assign a class to 92.4% of them, contrary to the old model with 80.1%. The Top 1 accuracy has increased from 69.8% to 76.7%.

The graphs below show improvements across observations from different regions, with major improvements seen in South America, Oceania, Asia and Africa.

Detailed results with individual observations are available in this spreadsheet.

The following graphs shows the importance of metadata, specifically time and geographic coordinates.

There are other, more technical changes in the new model due to the changed infrastructure. For example, the new model tends to produce higher probability estimates than the old model.

We remain dedicated to ensuring that continues to be a cutting-edge plant identification service for users worldwide. The next major upgrade planned to September 2023 will focus at plant cultivars.

Share this post
By clicking “Accept All Cookies”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. View our Privacy Policy for more information.