Summer 2024 Plant.id Model Update

Plant.id went through an important machine learning model upgrade. The number of classes remains similar to the previous version with 35,000+ classes, of which there are almost 30,000 species. Significant improvements were reached in accuracy.

The accuracy on a weighted validation set (slightly changed compared to the last evaluation) has progressively increased through the versions of the Plant.id API model. The latest version 4.0.1 reached the highest accuracy for TOP 1, TOP 3, and TOP 10 categories, meaning that the correct answer was in the first answer, first three answers or first ten answers.

Next, we have analysed a GBIF sample consisting of 50,000 observations. GBIF, known for its high reliability and open-access plant data, served as the source for this analysis. When the data was categorized by region, the most significant improvements were made in the European dataset (a 6.0 % increase in TOP 1) and the North American dataset (a 3.6 % increase in TOP 1) compared to the model from October 2023.  Detailed results, including individual observations, can be found in this spreadsheet.

‍

Significant algorithmic improvements have been to the Similar images, which show the user the closest looking images from our database.

The number of classes also increased in the infraspecies (varieties) model, where mainly fruits and vegetables were added.