• Input data

    The ideal image size is 800 pixels for the small side, or otherwise 1280 pixels for the large side.

    Compression: jpg with a quality=90 is recommended

    Sending several images of the same individual plant (up to 5) do improve the results, even if they are of the same type of view (flower, fruit, leaf, etc.). Some types of view/organs, however, are more informative than others in the following order (best first): flower, fruit, leaf, entire plant, bark. So that it is often more efficient to submit several images of the most informative parts (e.g. several flowers and/or fruits) than trying to be exhaustive in the coverage of the different types. The bark view in particular is challenging. It is useful when there is no other visible part of the plant (e.g. deciduous trees during winter) but it leads to significantly lower accuracy than other views. Leaves are also often much less discriminative than flowers and fruits. Nevertheless, Pl@ntNet's AI algorithm automatically assigns different weights to the different types of view in accordance with those properties. You can thus submit images of different types without worrying too much about it. You can also submit the images without specifying the type of view. In that case, it will be automatically inferred by the AI.

    Whenever possible, you should use POST requests to improve processing speed

  • Rejection

    Images identified as non-plants will be rejected, and lead to a "404 Species not found" error. Should an inappropriate image be submitted, it will most likely be filtered by this mechanism. If you wish to manage the rejection of non-plant images by yourself, set "no-reject" parameter to true.

  • Information about the identification model

    Current model: Inceptionv3

    Methodology: Pl@ntNet's model is updated on approximately a two-monthly basis with a principle of non regression. Non regression is measured in terms of several indicators: accuracy, number of species recognized, respect of an acceptable response time, respect of an acceptable energy consumption, memory usage, etc. Accuracy is obtained by testing the model on a set of private datasets of different nature, from the easiest (e.g. common species from various families) to the most difficult (e.g. very similar species of complex and/or data deficient genera).

  • Additional data: GBIF species API

    To get more data about the species on the identification results, you can use the GBIF species API

    Some Pl@ntNet species have a gbif.id field.
    If they have, you can use it to load more data, for example: https://api.gbif.org/v1/species/5231190/vernacularNames