Study on identification accuracy: Flora Incognita reaches 98.8 %

/in /by

About the study

A scientific study was recently published which tested five plant identification apps in the UK with regard to their identification accuracy. Flora Incognita was not included, but the images taken and analysed for the study were part of the publication. This gave us the opportunity to test our app with this independent image dataset.

The (first) result

We identified the images with Flora Incognita and compared our results with the species names that the authors expected. In this first run, Flora Incognita was correct for more than 90% of the images, already outperforming the other apps. In the next step, however, we wanted to find out why Flora Incognita (allegedly) provided incorrect species names for the remaining 10% of the images. We therefore manually checked these questionable images again in collaboration with external botanists. It turned out that Flora Incognita was only wrong in a few cases.

Reasons for ‘ false’ identification

In most cases, Flora Incognita identified correctly, and there were other reasons why the proposed species name was not the one the authors expected:

1) Taxonomy. There are several taxonomies existing in parallel, sometimes with slightly different concepts, but all valid – leading in some cases to different names or different taxonomic ranks (species vs. subspecies) for the same plant species.

2) Identification errors. Sometimes the expected species was wrong, and the app result correct.

3) Biology. Sometimes it is not possible to categorise the result of an app identification as correct or incorrect because the species in the picture cannot be identified to species level with certainty even by experts.

Conclusion

We would like to encourage you not to immediately say: ‘The app is wrong’ if you get an unexpected identification result. Next time, ask yourself: Do I know this species under a different name? Was the picture really suitable for identifying the species? Couldn’t it be the species suggested by the app? Because in our new publication, we show that Flora Incognita correctly identified 98.8 % of the species in the entire data set after a thorough investigation into the causes of errors in the first run.

Read the paper here: More than rapid identification—Free plant identification apps can also be highly accurate

Rzanny, M., Bebber, A., Wittich, H. C., Fritz, A., Boho, D., Mäder, P., & Wäldchen, J. (2024). More than rapid identification—Free plant identification apps can also be highly accurate. People and Nature, 00, 1–4. https://doi.org/10.1002/pan3.10676

New publication: Does an artificial intelligence recognize the leaf shape?

/in /by

Leaf shape and AI
The automatic identification of plant species using artificial intelligence (AI) is largely based on the analysis of leaf shapes. However, how these complex structures are interpreted by the AI is often difficult to understand and represents a so-called ‘black box’. In particular, the high variability of leaf shapes within a species, as can be observed in Ranunculus auricomus buttercup, poses a challenge for AI.

Geometric morphometrics
In order to evaluate and better understand the performance of AI-based identification systems, we used geometric morphometrics as a tool of explainable AI (eXplainable AI or XAI). This method enables a detailed quantitative description of shape variations and thus allows a direct comparison between the results of the AI and an established morphometric method.

Experiment
Our focus was on recognizing regional population differences in the golden buttercup. We investigated whether the AI is able to recognize subtle morphological differences between several populations based on basic smartphone images. The results show that the AI can reliably identify characteristic features of the leaves even with complex image backgrounds. The agreement with the results of geometric morphometrics confirms the performance of the AI and underlines the potential of this technology for automated phenotyping in plant research.

Significance of this research
Our study provides important insights for the use of AI in plant identification and opens up new perspectives for biodiversity research. The combination of AI and modern morphometric methods enables a comprehensive analysis of plant traits and thus contributes to a better understanding of the evolution and adaptation of plants to different environmental conditions.

The publication is now freely available: Hodač, L., Karbstein, K., Kösters, L., Rzanny, M., Wittich, H.C., Boho, D., Šubrt, D., Mäder, P. and Wäldchen, J. (2024), Deep learning to capture leaf shape in plant images: Validation by geometric morphometrics. Plant J. https://doi.org/10.1111/tpj.17053

Title image: Ranunculus auricomus, Ladislav Hodač