Machine learning and model training

Pl@ntNet is a collaborative application dedicated to plant identification, used by millions of people worldwide. Artificial intelligence (AI), and more specifically deep learning, is at the heart of Pl@ntNet’s operation.

The principle of deep learning and images

Deep learning is a branch of artificial intelligence where “neural networks” learn to recognize patterns in data, here images. By analyzing millions of photos of plants, a computer model can learn to differentiate species based on their visual characteristics: leaf shape, flower color, etc.

Pl@ntNet relies on millions of images collected by users and experts. These photos are sorted, validated, and organized to train deep learning models. The more images and diversity there are, the more accurate the model becomes!

Recent progress thanks to advanced models

Today, Pl@ntNet uses a cutting-edge technology called Vision Transformer (ViT). These models, such as the one from DINOv2, make it possible to take advantage of the large quantities of available images while improving the recognition of rare or poorly photographed plants.

In addition to being trained on photos provided by the community, the model is evaluated on very specific test images. These images, often from experts and not accessible online, serve to ensure that the system is reliable, even for plants that are difficult to identify.

The importance of challenges and tests

Pl@ntNet organizes the PlantCLEF challenge every year, an event where researchers and engineers test the latest advances in plant identification. These competitions allow different approaches to be compared and the performance of the models to be improved. For example, at PlantCLEF2022, a challenge based on 4 million images showed that new architectures, such as Vision Transformers, surpass older systems based on convolutional neural networks.

Internally, Pl@ntNet also carries out rigorous tests. “Sanctuary” sets of images, selected for their quality and rarity, make it possible to verify that the model correctly recognizes complex or under-represented species.

Why all this is important

Thanks to these advances, Pl@ntNet is not just a practical tool for botany enthusiasts. It contributes to the preservation of biodiversity by helping to document plants on a global scale. Each photo shared enriches the database and strengthens the model’s capabilities, thus creating a virtuous cycle for science and nature.

With this technology, everyone can participate in better understanding and protecting the plants around us!