PlantCLEF 2025 Challenge

• PlantCLEF 2025 Challenge is a benchmark that tackles multi-species plant identification in high-resolution quadrat images using weak label supervision.
• It leverages a training set of 1.4 million individual plant images alongside 2,105 expert-annotated quadrat images to capture community-level biodiversity.
• Top solutions predominantly use tile-based inference and aggregation with vision transformers to bridge the gap between single-label training and multi-label test settings.

PlantCLEF 2025 was the LifeCLEF Lab at CLEF 2025 challenge on multi-species plant identification in vegetation quadrat images, formulated as a weakly labelled multi-label classification problem in which the goal was to predict all species present in a quadrat image using mainly single-label training data (Martellucci et al., 22 Sep 2025). The benchmark combined a new test set of 2,105 high-resolution multi-label images annotated by experts with a large training set of 1.4 million individual plant images, and it distributed organizer-provided vision transformer models pre-trained on this data (Martellucci et al., 22 Sep 2025). The challenge extended the multi-species vegetation-plot setting introduced in PlantCLEF 2024 and raised the best private score from 0.2873 in 2024 to 0.3648 in 2025, while still leaving the task far from solved (Martellucci et al., 22 Sep 2025, Goeau et al., 19 Sep 2025).

1. Ecological setting and challenge definition

PlantCLEF 2025 was motivated by a practical ecological bottleneck. Quadrat images are essential for ecological studies because they enable standardized sampling, the assessment of plant biodiversity, long-term monitoring, and large-scale field campaigns (Martellucci et al., 22 Sep 2025). In routine vegetation monitoring, botanists examine standardized plots—typically around 50 × 50 cm, and in the overview also described as plots of fifty centimetres or one square meter in size—and identify all species present (Martellucci et al., 22 Sep 2025). The challenge asked whether AI could identify all plant species visible in a single high-resolution quadrat image rather than requiring specimens to be photographed one by one (Martellucci et al., 22 Sep 2025).

This setting was technically distinctive within PlantCLEF. Earlier global-scale editions such as PlantCLEF 2022 and 2023 were observation-level species retrieval problems over 80,000 plant species, whereas PlantCLEF 2024 and 2025 focused on community-level species presence in vegetation plots (Goeau et al., 22 Sep 2025, Martellucci et al., 22 Sep 2025). The 2025 task therefore combined a label structure mismatch—training data are largely single-label, test data require multi-label output—with a domain gap between close-up individual-plant images and vertical or near-vertical quadrat scenes (Martellucci et al., 22 Sep 2025).

The ecological scenes themselves were difficult. Quadrat images were described as top-view, high-resolution, multi-species scenes with overlapping plants, variable density, clutter, partial visibility, shadows, blur, and strong differences in phenological stage (Martellucci et al., 22 Sep 2025). The overview reports about 8 species per quadrat on average, a density high enough that a score around 0.36 corresponded roughly to 4–5 correct predictions per image with relatively few false positives (Martellucci et al., 22 Sep 2025).

2. Data resources and supervision mismatch

The main training resource was a large subset of Pl@ntNet data focused on South-Western Europe (Martellucci et al., 22 Sep 2025). Its reported statistics were:

• 1,408,033 images
• 1,151,904 observations
• 7,806 species
• 1,446 genera
• 181 families (Martellucci et al., 22 Sep 2025)

The split provided for single-plant model development contained 1,308,899 training images, 51,194 validation images, and 47,940 test images, with splitting performed at the observation level so that all images from a single observation belonged to the same split (Martellucci et al., 22 Sep 2025). These splits were explicitly distinct from the hidden challenge test set of quadrat images (Martellucci et al., 22 Sep 2025).

The challenge test data consisted of 2,105 high-resolution quadrat images compiled from several expert-produced datasets in Pyrenean, Mediterranean, and South-Western European temperate floras (Martellucci et al., 22 Sep 2025). The official overview describes the test set as covering around 400 species (Martellucci et al., 22 Sep 2025). Participant papers, however, characterize the effective test flora as over 800 species or roughly 800 species (Gustineli et al., 8 Jul 2025, Herasimchyk et al., 14 Aug 2025). This suggests differing counting conventions across official and participant descriptions.

Acquisition conditions varied substantially. The quadrats could include wooden frames or measuring tape, could be more or less perpendicular to the ground depending on slope, and could contain shadows, blur, and weather-related quality changes (Martellucci et al., 22 Sep 2025). These properties mattered because the dominant training imagery emphasized individual plants or plant organs, whereas the test imagery emphasized mixed-species vegetation patches (Martellucci et al., 22 Sep 2025).

To support domain adaptation, organizers also provided an unlabeled pseudo-quadrat resource: 212,782 images from the LUCAS Cover Photos 2006–2018 archive, curated as high-resolution vegetation photos framed similarly to quadrats but not strictly limited to a 50 × 50 cm area (Martellucci et al., 22 Sep 2025). The overview notes that this complementary pseudo-quadrat dataset was not used by any of the teams that submitted working notes (Martellucci et al., 22 Sep 2025).

3. Evaluation methodology and official challenge resources

The official ranking criterion was an F1 score with transect-aware averaging (Martellucci et al., 22 Sep 2025). Per-image F1 was defined as

$$F1_i = \frac{2 \cdot \text{Precision}_i \cdot \text{Recall}_i}{\text{Precision}_i + \text{Recall}_i},$$

with

$$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$

The final challenge score averaged first within transects and then across transects:

$$F1_{\text{avg-transect}} = \frac{1}{T} \sum_{k=1}^{T} \left( \frac{1}{Q_k} \sum_{j=1}^{Q_k} F1^{(k)}_j \right)$$

where $T$ is the number of transects and $Q_k$ the number of quadrat images in transect $k$ (Martellucci et al., 22 Sep 2025). This choice was ecologically motivated because it prevented heavily sampled transects from dominating the final score (Martellucci et al., 22 Sep 2025).

The challenge was hosted on Kaggle (Martellucci et al., 22 Sep 2025). Metadata use and external data were allowed, but each such run had to be paired with an equivalent run using only the provided visual data, so that the contribution of metadata or external resources could be measured fairly (Martellucci et al., 22 Sep 2025).

A central official resource was the release of two organizer-provided Vision Transformer models, both based on ViT base patch-14 with registers, initialized from DINOv2 self-supervised pretraining on LVD-142M and then fine-tuned on the PlantCLEF individual-plant training data using timm (Martellucci et al., 22 Sep 2025). These were:

• ViTD2PC24OC (vit_base_patch14_reg4_dinov2_lvd142m_onlyclassifier), in which only the classification head was trained.
• ViTD2PC24All (vit_base_patch14_reg4_dinov2_lvd142m_onlyclassifier_then_all), initialized from the previous model and then fine-tuned on all layers (Martellucci et al., 22 Sep 2025).

On the held-out single-plant test split, ViTD2PC24OC reached Top-1 = 63.69 and Top-5 = 83.88, while ViTD2PC24All reached Top-1 = 75.91 and Top-5 = 92.82 (Martellucci et al., 22 Sep 2025). The official tiling-based baseline achieved a macro-averaged F1 of 0.21708 on the challenge benchmark (Martellucci et al., 22 Sep 2025).

4. Dominant methodological pattern: tiling, aggregation, and inference-time adaptation

The clearest 2025 pattern was the dominance of tile-based inference. The official overview states that all but two teams cut each high-resolution quadrat image into smaller tiles, ran a classifier on each tile, and then aggregated predictions across tiles (Martellucci et al., 22 Sep 2025). This was a direct response to the supervision mismatch: a model trained on images of individual plants was more effective when parts of a quadrat were made to resemble the training domain (Martellucci et al., 22 Sep 2025).

The second-place DS@GT system made this logic explicit. It combined ViTD2PC24All, $4 \times 4$ tiling, tile-level classification, visual-cluster priors from PaCMAP + K-Means, and geolocation filtering, all without additional training (Gustineli et al., 8 Jul 2025). A typical quadrat was about 2000 px wide, so a $4 \times 4$ partition produced tiles of roughly 500 px per side, closely matching the model’s $518 \times 518$ expected input size (Gustineli et al., 8 Jul 2025). In the DS@GT ablation, full-image inference with a plain ViT gave Private 0.00633 and Public 0.01157, whereas ViT, top-9, 4x4 gave Private 0.34420 and Public 0.30810 (Gustineli et al., 8 Jul 2025). This suggests that scale matching between tile size and receptive field was one of the central 2025 design principles.

The third-place Chlorophyll Crew system also treated multi-label prediction as an inference-time construction over local evidence rather than as end-to-end scene parsing. It used a frozen DINOv2 ViT-B/14 backbone with multiple heads for species, genus, and family, combined with multi-scale non-overlapping tiling, border cropping to remove non-plant artifacts, one-species-per-tile constraints, and dynamic threshold optimization based on mean prediction length (Herasimchyk et al., 14 Aug 2025). Their best-performing submissions used scales 4 and 5, meaning $4 \times 4$ and $$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$0 tilings, and dynamically targeted a mean prediction length around 4 species per image (Herasimchyk et al., 14 Aug 2025).

A more divergent fifth-place strategy came from ADAM, which replaced standard tiling by prototype-guided zero-shot segmentation (Filho et al., 23 Dec 2025). The method extracted DINOv2 embeddings from the training set, formed $$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$1 K-Means centroids as prototype targets, trained a narrow ViT on test-image patches to reconstruct a $$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$2 prototype matrix, and then used attention maps as a proxy localization mechanism before classification (Filho et al., 23 Dec 2025). Its central empirical lesson was that context matters much more than isolated patch classification: direct patch-wise classification was weak, while $$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$3 contextual grid assembly around high-attention patches was much stronger (Filho et al., 23 Dec 2025).

5. Leaderboard, participation, and representative systems

PlantCLEF 2025 was organized on Kaggle and drew 540 initial entrants, 55 participants grouped into 38 teams, 659 runs, and 10 teams that submitted working notes (Martellucci et al., 22 Sep 2025). The top of the challenge was methodologically concentrated but not uniform.

Rank	Team	Characteristic strategy
1	TheHeartOfNoise	Image preprocessing, JPEG recompression, quality and YCbCr subsampling, tiling
2	DS@GT-LifeCLEF	$$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$4 tiling, species-frequency aggregation, geolocation filtering, region-specific Bayesian reweighting
3	Chlorophyll Crew	Multi-head species/genus/family ViT, multi-scale tiling, dynamic thresholding
4	webmaking	Multi-scale tiling, GroundingDINO + SAM, seasonal filtering, ecological adjustment, cross-year aggregation
5	ADAM	Prototype-guided zero-shot segmentation and contextual patch-grid classification

The winning system, TheHeartOfNoise, was notable because it pushed low-level image preprocessing unusually far. According to the overview, the participant focused heavily on JPEG recompression and the tuning of quality and YCbCr subsampling, and these choices substantially improved F1 by reducing the train-test distribution gap and acting as a form of regularization (Martellucci et al., 22 Sep 2025). The official overview reports a best private leaderboard score of 0.3648, and also states that the highest purely visual performance, 0.35016, was obtained by TheHeartOfNoise using test-time image preprocessing and multi-scale tiling with nine different sizes (Martellucci et al., 22 Sep 2025).

The DS@GT system achieved second place with a best private macro-F1 of 0.34834 (Gustineli et al., 8 Jul 2025). Its best variant, ViT + PRIORS, re-weighted tile predictions using cluster-specific priors estimated by averaging model posterior vectors over 3 K-Means clusters formed from [CLS] embeddings of the test images (Gustineli et al., 8 Jul 2025). A related geolocation filter reduced the candidate label set from 7,806 to 4,981 species by retaining only species whose nearest geotagged observation was within France, Spain, Italy, or Switzerland (Gustineli et al., 8 Jul 2025).

Chlorophyll Crew, the third-place team, reported 2nd place on the public leaderboard and 3rd place on the private leaderboard (Herasimchyk et al., 14 Aug 2025). Their best private score among listed submissions was 0.34575 for a 5h1l model with target mean length 4.0, max length $$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$5, tiling scales 4, 5, and crop 10%, while the best private score among their five selected final submissions was 0.33655 (Herasimchyk et al., 14 Aug 2025). The system’s main contributions were taxonomy-aware multi-head prediction and explicit calibration of prediction set size through mean-length-controlled thresholding (Herasimchyk et al., 14 Aug 2025).

The webmaking system, ranked 4th, combined multi-scale tiling, test-time augmentation, GroundingDINO + SAM to down-weight non-vegetal regions, seasonal filtering using GBIF occurrence data, ecological adjustment using Ellenberg Indicator Values for Europe (EIVE), and cross-year aggregation (Martellucci et al., 22 Sep 2025). It was one of the strongest examples of ecological prior integration in the challenge (Martellucci et al., 22 Sep 2025).

The ADAM system, ranked 5th, achieved $$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$6 on the private leaderboard (Filho et al., 23 Dec 2025). Its prototype-guided attention model showed that attention-based localization could be competitive, but only when classification restored enough local context through $$\text{Precision}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FP}_i}, \qquad \text{Recall}_i = \frac{\text{TP}_i}{\text{TP}_i + \text{FN}_i}.$$7 contextual patch grids (Filho et al., 23 Dec 2025).

6. Recurring themes, misconceptions, and unresolved problems

Several methodological conclusions recur across the official overview and the working notes. The first is that PlantCLEF 2025 was not mainly a problem of inventing a new classifier, but of adapting an existing plant recognizer to a radically different test domain (Gustineli et al., 8 Jul 2025). Tiling, aggregation, cropping, priors, and low-level preprocessing repeatedly mattered more than architectural novelty alone (Martellucci et al., 22 Sep 2025, Herasimchyk et al., 14 Aug 2025).

A common misconception was that the problem could be addressed by ordinary whole-image plant classification. Participant ablations strongly contradicted that view. In DS@GT, full-image inference was almost useless, whereas quadrat tiling produced the decisive performance jump (Gustineli et al., 8 Jul 2025). The overview likewise reports that most successful teams converged on some form of tiled high-resolution inference (Martellucci et al., 22 Sep 2025).

A second misconception concerned “zero-shot” claims. Both DS@GT and ADAM described their methods as training-free or zero-shot relative to the 2025 task, but both relied on organizer-provided DINOv2-based ViT models already fine-tuned on PlantCLEF data from earlier training resources (Gustineli et al., 8 Jul 2025, Filho et al., 23 Dec 2025). In other words, the systems were zero-additional-training for the quadrat task, not zero-shot in the absolute sense.

The challenge also exposed limits of segmentation and detection. The overview concludes that vegetation-focused filtering did not consistently yield strong measurable gains, and that conventional object detection was poorly matched to dense vegetation because quadrat scenes have no clear object boundaries, strong overlap, many indistinct plant instances, and very fine-grained species differences (Martellucci et al., 22 Sep 2025). This was one reason why systems based on detection-style localization were not dominant, despite the apparent naturalness of localizing plants in cluttered scenes (Martellucci et al., 22 Sep 2025).

Public/private divergence was another central issue. The official overview reports large leaderboard shifts: TheHeartOfNoise moved from 3rd public to 1st private, webmaking fell from 1st public to 4th private, and DS@GT PlantCLEF rose from 9th public to 2nd private (Martellucci et al., 22 Sep 2025). This established leaderboard overfitting as a real concern and partly explains why many high-scoring systems emphasized robust priors and conservative aggregation rather than purely public-score-driven tuning.

Substantial headroom remained. The overview identifies open directions such as multi-label models that process full-resolution quadrat images end-to-end, robustness to low-level encoding and compression differences, stronger integration of ecological priors and structural priors, and better domain adaptation between individual-plant and quadrat-image domains (Martellucci et al., 22 Sep 2025). Participant papers added more specific proposals, including sliding-window inference, adaptive receptive fields, token merging, self-training on pseudo-labels, and CNN/ViT ensembling (Gustineli et al., 8 Jul 2025). PlantCLEF 2025 therefore marked both a major empirical advance and a clear statement that weakly supervised plant community recognition in quadrat imagery remained an open research problem.