How image-based applications can measure plant biodiversity: A multi-specimen plot analysis based on Pl@ntNet citizen science platform.

Citizen science is playing an increasingly important role in biodiversity monitoring, providing unprecedented volumes of observations while engaging society in conservation efforts.

Yet many plant monitoring surveys still rely on expert-based field studies, which are time-consuming, costly, and difficult to scale up spatially and temporally. These constraints limit our capacity to keep pace with current rapid environmental changes. In this context, vision-based identification tools built on citizen science data such as Pl@ntNet, offer a promising opportunity to enable automated extraction of ecological information from images.

For this study, we developed a new workflow that integrates high-resolution plot imagery of annual vegetation communities, analysed with the Pl@ntNet model and filtered on the basis of phytosociological data to monitor plant communities at local scale.  While Pl@ntNet has traditionally been used for single-species observations, we adapt its capacities to multi-specimen plot images, extending the scope of citizen science tools to support ecological studies at larger scale.

We evaluate our approach on a collection of standardised plot images collected in the Mediterranean ecosystem and precisely annotated by botanical experts. Using the model outputs, we derived multiple biodiversity indicators and compared them to expert-based estimates, focusing on commonly used metrics such as species richness, Shannon and Simpson diversity indices, Raunkiaer life-form types and community-weighted means of functional traits.

To further understand model performance and  assess its reliability, we investigated how the accuracy of predicted indicators varied with ecological and methodological factors, including habitat type, seasonality, and image resolution. Discrepancies not explained by these factors can be attributed either to the intrinsic limitations of the model for which we will discuss avenues for future improvement, or, importantly, to observation bias, as botanical relevés are inherently subject to inter-observer error.

Our results demonstrate the potential of leveraging citizen-science infrastructures such as Pl@ntNet to generate scalable, repeatable biodiversity indicators from plot imagery. This approach offers a promising pathway for integrating citizen-science tools into local and national monitoring schemes, ultimately contributing to more efficient and inclusive biodiversity assessment workflows.