[Paper Review] Overview of LifeCLEF Plant Identification task 2019: diving into data deficient tropical countries

Hervé Goëau, Pierre Bonnet+1 more|Agritrop (Cirad)|Sep 23, 2025

Species Distribution and Climate Change1 references23 citations

TL;DR

The LifeCLEF 2019 Plant Identification challenge evaluated automated plant identification on data-deficient tropical flora using a 10k-species training set and a 742-item field-tested test set, comparing 26 DL systems across 6 teams to human experts; results show deep learning lags behind experts, with tropical flora being notably harder.

ABSTRACT

Automated identification of plants has improved considerably thanks to the recent progress in deep learning and the availability of training data. However, this profusion of data only concerns a few tens of thousands of species, while the planet has nearly 369K. The LifeCLEF 2019 Plant Identification challenge (or "PlantCLEF 2019") was designed to evaluate automated identification on the flora of data deficient regions. It is based on a dataset of 10K species mainly focused on the Guiana shield and the Northern Amazon rainforest, an area known to have one of the greatest diversity of plants and animals in the world. As in the previous edition, a comparison of the performance of the systems evaluated with the best tropical flora experts was carried out. This paper presents the resources and assessments of the challenge, summarizes the approaches and systems employed by the participating research groups, and provides an analysis of the main outcomes.

Motivation & Objective

• Assess automated plant identification performance on tropical, data-deficient flora regions (Guiana shield and Northern Amazon) using a new 10k-species training dataset.
• Compare automated system performance to tropical flora experts on a field-identified test set.
• Analyze data quality, noise effects, and the potential gains from data sources like herbarium images.

Proposed method

• Assemble a training set of 10k species from EoL and web sources, noting noise and duplicates.
• Provide a high-quality test set of 742 field-identified observations for evaluation.
• Evaluate up to 10 runs per team using Top1, Top3, Top5, and MRR metrics, against 5 expert annotations.
• Summarize participating methods, emphasizing CNN architectures (e.g., Inception-ResNet-v2/v4, DenseNet) and data augmentation.
• Compare system performance to human experts and analyze the impact of training data quantity and noise.

Figure 1: Regions of origin of the 10k species selected for PlantCLEF 2019: French Guiana, Suriname, Guyana, Brazil (states of Amapa, Para, Amazonas)

Experimental results

Research questions

• RQ1 How does automated plant identification perform on tropical data-deficient flora compared to tropical flora experts?
• RQ2 What is the impact of training data quality and noise (duplicates, non-plant images, herbarium drawings) on DL performance?
• RQ3 Can deep learning systems close the gap with experts when trained on large, noisy, but diverse datasets?
• RQ4 What roles do data augmentation, class priors, and additional training data (e.g., GBIF) play in system performance?
• RQ5 Is herbarium data a viable path to improve identification in data-poor tropical regions?

Key findings

• DL systems perform substantially worse than the best tropical flora experts on the test set (Top1 expert up to 0.675; median 0.376).
• The tropical flora task is notably harder than temperate flora, with expert top performance lower and larger gaps to machine predictions.
• Best automated system remains about half as accurate as the top expert, with a gap of around 0.365 in Top1 when comparing to experts.
• Noise and data quality (duplicates, non-plant images) significantly influence performance, especially for species with fewer training images; herbarium/drawings impact is inconclusive.
• Extending training data with herbarium records (GBIF/Herbarium digitization) shows potential gains, evidenced by corrected submissions achieving a top-1 accuracy of 41% in post-hoc evaluation.
• Complementary analyses show that a higher number of training images generally improves mean rank, while a high proportion of duplicates degrades results.

Figure 2: Scores between Experts and Machine