TL;DR
The Pl@ntNet machine-to-machine (M2M) service provides programmatic access to AI models covering 85,000 plant species via a robust API. By enabling automated workflows that are up to 666 times faster than human-in-the-loop systems, it allows researchers to integrate high-accuracy plant identification directly into data pipelines for drones, agricultural robots, and large-scale social media analysis.
What Problem Does This Paper Address?
Monitoring global biodiversity is a massive undertaking that currently outpaces the availability of expert taxonomists. While AI-powered tools like the Pl@ntNet mobile app have revolutionized citizen science, they traditionally rely on "human-in-the-loop" workflows where a user manually uploads a photo.
This paradigm fails when faced with the "firehose" of data generated by modern ecological sensors—such as camera traps, drones, and autonomous robots—or when researchers want to retrospectively analyze millions of images from social media. There is a critical need for an interoperable, scalable service that can identify plants without human intervention, transforming raw imagery into structured ecological data at scale.
Methods
The authors present the Pl@ntNet API, a machine-to-machine service that exposes their identification engine via HTTPS requests.
- Core Architecture: The underlying models are currently based on vision transformers, a state-of-the-art architecture in computer vision.
- The
/identifyEndpoint: This is the central feature. It accepts up to five images of a single plant and returns a JSON response containing scientific names, common names, and confidence scores. - Taxonomic Intelligence: Unlike simple image classifiers, the service includes taxonomic post-processing. It aligns results with the POWO (Plants of the World Online) reference and provides identifiers for other major biodiversity databases like GBIF and IUCN.
- Automation Features: The API supports simultaneous requests (up to 20 per account), allowing for massive parallel processing. It also allows users to specify the "flora" (geographic context) or the "organ" (leaf, flower, fruit) to optimize accuracy.
Datasets
The service is backed by one of the largest curated botanical datasets in the world:
- Training Data: 18 million plant images annotated with curated labels, derived from a larger collaborative database of 30 million observations.
- Internal Benchmark: A test set randomly sampled from the 18 million curated images.
- External Benchmarks: To ensure generalizability, the authors evaluated the model on the Swiss Plant dataset (5,619 records) and the iNaturalist 2021 validation dataset (37,160 observations).
Key Insights and Ideas
- Speed as a Paradigm Shift: The transition from manual to M2M processing isn't just a minor improvement; it's a 666x speedup. This enables real-time applications, such as agricultural robots performing selective weed control, which would be impossible with manual workflows.
- Secondary Data Extraction: One of the most compelling use cases is identifying plants in photos where the plant isn't the primary subject. By analyzing images of insects (like the wasp in Figure 2), researchers can automatically reconstruct plant-pollinator interaction networks.
- The Resolution Bottleneck: A key finding is the sensitivity of vision transformers to image resolution. Performance dropped significantly on the iNaturalist dataset because average image sizes (500x350) were below the recommended 800x800 pixels.
- Collaborative Curation: The "secret sauce" of Pl@ntNet’s accuracy is its iterative trust rating mechanism, which weighs user expertise to validate observations before they enter the training set.
Key Results
- Scale: The service has processed over 84 million requests since 2018, with 40 million in 2025 alone. As shown in Figure 1, both the number of accounts and the query volume are experiencing exponential growth.
- Accuracy: The global flora model achieved a Top-1 accuracy of 0.886 on the internal test set. On the external Swiss Plant dataset, accuracy reached 0.864 (Top-1) and 0.963 (Top-5).
- Efficiency: In a reproducible example, the authors processed 2,424 images from iNaturalist on a standard laptop in just 30 minutes, successfully generating a dataset of 708 plant-pollinator interactions with a genus-level accuracy of 0.88.
Limitations and Future Work
- Image Quality: The authors emphasize that "garbage in, garbage out" applies here; images below 800x800 pixels significantly degrade taxonomic accuracy.
- Taxonomic Alignment: Aligning different datasets (like iNaturalist vs. Pl@ntNet) remains a challenge due to differing taxonomic references, though the use of POWO helps mitigate this.
- Future Directions: The authors envision a "new paradigm" where automated sensors (drones, camera traps) are directly connected to AI identification services, enabling truly real-time, global biodiversity monitoring.
Summary updated