How the identification model works
The Pl@ntNet identification model is designed to help users identify plants from images. It relies on a combination of artificial intelligence (AI) and human contributions to achieve the best possible results. Here’s how it works:
1. Taking photos
The user takes one or more images of the plant they wish to identify. These images ideally illustrate different parts of the plant such as leaves, flowers, fruits and bark. The clearer and more varied the images, the more likely the model will be able to make an accurate identification.
2. Initial analysis by AI
Pl@ntNet’s artificial intelligence is based on image recognition models capable of comparing photos sent by users to a vast database of plant species. Initially, Pl@ntNet used convolutional neural networks (CNNs), a classic deep learning method particularly effective for processing images. CNNs analyze images by extracting important visual features, such as leaf shape, flower color, or stem texture. These features are then used to suggest similar species.
However, CNNs have some limitations and Pl@ntNet has made a major technological leap by adopting Vision Transformers (ViT). Unlike CNNs, which process images by locating and analyzing parts of the image from small windows, Vision Transformers process the entire image using an attention-based approach, a method inspired by models used in natural language processing.
Vision Transformers are particularly well-suited to analyzing complex images and fine details on plants, such as the precise shape of leaves or the configuration of flowers. This approach allows the model to be more accurate and robust, especially when it comes to identifying plant species that share similar characteristics but differ in subtle details. Vision Transformers are therefore a major advance in the accuracy and reliability of identifications made by Pl@ntNet’s AI.
Thanks to this evolution, the AI model has become more powerful and capable of processing more complex images with better accuracy. This allows Pl@ntNet to offer even more reliable plant identifications, even for species that are difficult to distinguish visually.
3. Community validation
However, AI alone is not enough and the participation of the Pl@ntNet community in validating automatic identifications is essential. Users can review the suggestions made by the AI and validate or correct the identifications. If several users confirm an identification, it becomes more reliable.
4. Continuous model improvement
Each time a user validates an identification or corrects an error, this information is used to improve the AI model. In other words, the AI learns from its mistakes thanks to human contributions, which increases the accuracy of future identifications.
5. Use of taxonomic databases
The identification model uses taxonomic databases to verify that a plant name is valid. These databases contain reliable information about species and help avoid errors in identifications.
6. Identification reliability
The model also calculates a confidence score for each identification. This score reflects how confident the AI is in its identification. The higher the score, the more reliable the identification. If the identification does not reach a certain confidence threshold, it will probably be rejected or require further review.
7. Contribution to the database
When the identification is validated, it is added to the Pl@ntNet database. This data is then used to enrich the platform and contribute to scientific research on biodiversity.
8. Interaction with other users
In addition to automatic identification, Pl@ntNet allows users to vote to validate or reject identification suggestions made by others. This creates a collaborative process that constantly improves the quality of identifications.
In summary
The Pl@ntNet identification model works thanks to a combination of artificial intelligence, community reviews, and taxonomic databases. Each observation is first analyzed by the AI, then validated or corrected by the community. The AI learns from these interactions to become more accurate over time, thus creating a collaborative and evolving system for plant identification.