This paper discusses our submission to the PlantCLEF 2025 challenge, identical to PlantCLEF 2024, where the objective is to predict multiple plant species within vegetation plot images. The dificulty of this challenge stems from multi-class classification, domain shifts, and predictions on high resolution, evidenced by the macro-averaged F1 score not exceeding 0.37. The training data made available consist of more than 1.4 million single-class images, while the test plot set contains 2,105 high-resolution vegetation plot images taken from above. The plots exhibit multiple domain shifts including blurs, plant life cycle, occlusions (via organic or inorganic matter), and seasonal changes. Given the success of using patch-wise inference from last year's challenge, we opted to continue with this method while also exploring domain-aware pretext tasks to finetune the provided DinoV2 vision transformer to address the domain shifts but yielded limited performance, possibly due to using a limited subset of the training data. In contrast, we performed ablation studies using only the base model, focusing on post-hoc techniques including aggregation and filtering. Surprisingly, we found that we were able to post challenge score of 0.35 macro-average F1; surpassing all our model-centric attempts including all but first place. Relevant code and runs will be made available on GitHub
The PlantCLEF 2025 [
The primary diference between the 2025 and 2024 challenge lies in the expanded evaluation set,
which now includes 2,105 vegetation plots (up from 1,695). The dificulty, however, remains the same.
The most widely adopted and efective strategy has been to divide each high-resolution image into
non-overlapping patches, typically 64 or 16 per image although some teams opted for 4,9 and 25 [
Our own approach [
For 2025, we extended our investigation by exploring Self-Supervised Learning (SSL) with pretext training tasks aimed at improving the species-level representation learned by DINOv2. However, none of these modifications surpassed the base model’s performance due to a limited subset of data used for ifne-tuning the model. This result shifted our focus: rather than changing the model itself, we studied the post-hoc aggregation pipeline in depth. Our ablation experiments demonstrated that, with no model ifne-tuning, we could surpass our oficial submission and achieve a post challenge macro F1 score of 0.3518; a score that would have placed second overall.
The dataset is reused from the PlantCLEF 2024 [
The training set is a subset of the Pl@ntNet training data [
The test set consists of 2,105 high-resolution vegetation plot images (from 2000 to 4000 pixels per
side), significantly more as compared to PlantCLEF 2024 [
Following our work on PlantCLEF 2024, we chose to continue our work using the vision transformers
provided by the organizers [
A tiling approach, similar to multiple teams last year, was also implemented to infer species on the entire vegetation plot by splitting the image into 64 or 16 patches and inferring independently, and then aggregating them in post to get the overall results for the entire plot, as will be discussed in the Inference method section.
The dataset we used was the training set provided, however we chose to limit each class to a maximum of 50 images per class as to combat the class imbalance, as well as computational constraints. Although 50 is low for amount of samples, we employed a higher than normal learning rate to accelerate training, as will be discussed in our Architecture section.
Our base architecture consists of two parts; the backbone, which is the aforementioned pretrained DinoV2 model, as well as classifier heads using the hyperparameters described in Table 2. We opted to use pretext tasks as a method to finetune our models with two heads, one for species classification and the other for pretext labels, as shown in Figure 3. Both heads were just a simple linear layer. The classifier heads were trained first for 20 epochs while the backbone was frozen, then for the remaining 30 the full model was trained with the last two layers of the backbone unfrozen at a lower learning rate.
Pretext tasks, also known as auxiliary tasks, are intermediate tasks a model performs in tandem with the
main objective task . Typically employed in Self-Supervised Learning (SSL), it serves as to reinforce the
model’s internal representation of classes by evaluating diferent versions of an image which undergo
transformations, such as rotation [
SSL tasks have been used in the vegetation domain [
In this pretext task, we exploit the three primary colour channels: red, green, and blue, by selectively removing one of them from an image and training the model to predict which channel was eliminated. This yields four possible classes for the task: red removed, green removed, blue removed, and no modification. Only one channel is removed at a time per image.
The underlying motivation is to simulate chromatic degradation or variation that may occur in real-world vegetation imagery due to environmental conditions such as lighting, seasonal changes, or plant life cycle stages. By forcing the model to distinguish which colour component is missing, we hypothesize that it will learn more robust internal representations that are invariant to certain types of colour-based distortions.
Building on the RGB elimination task, we also explore a secondary chromatic representation: the HSV colour space: Hue, Saturation, and Value. This colour space is often more aligned with human perception of colour properties. Hue refers to the dominant wavelength of a colour (e.g., red, green, blue), essentially defining its "type." Saturation indicates the intensity or purity of the colour, where higher values correspond to more vivid colours. Value represents brightness, with lower values producing darker tones, and zero resulting in black, though for our runs we set a limit of 0.25 to avoid a completely black image.
Here we selectively eliminate one of the three HSV components at a time, setting the corresponding channel to zero and train the model to predict which component has been removed. The objective is not to alter colour directly (as with RGB), but to manipulate properties such as darkness and vibrancy. Our motivation is that by exposing the model to colour variants that simulate environmental efects like shadows (low value) or faded pigmentation (low saturation), it may develop more invariant and generalizable feature representations of plant species under diverse conditions.
The direct inverse of HSV elimination, where the value of the channels are multiplied by 1.5 to increase Hue, Saturation and Value.
For this pretext task, we apply varying levels of contrast enhancement to the input image and task the model with identifying the applied contrast level. The contrast adjustment is implemented via a simple scaling of pixel intensities, where the image is multiplied by a factor ranging from no change (1.0×) to a significant boost (up to 1.5 ×, i.e., +50% contrast). The task includes four possible classes: unchanged, +10%, +25%, and +50%.
This task aims to teach the model to become sensitive to intensity-based variations that may occur in real-world scenarios such as overexposure, harsh lighting, or high reflectance from leaves or soil. By learning to recognize plant structures under diferent contrast levels, the model can develop robustness to varying imaging conditions and enhance its feature extraction across heterogeneous lighting environments.
In line with our previous approach from PlantCLEF 2024, we continue to use patch-wise inference as
the primary method for prediction. Each vegetation plot image is first divided into multiple smaller
patches, both 64 and 16 per plot. These patches are then passed individually through the model to
generate class-wise predictions. The outputs are aggregated using a Bayesian Model Averaging [
(|) × () ∑︀ =1 (|) × ()
Here ( | ) is the likelihood measuring how well explains the data , () is the prior, representing the assumed probability of before seeing any data, and the denominator is the sum of all models product of their respective likelihoods and priors.
An important distinction to make is that though BMA is used for ensemble models, we applied its weighting calculations for our case, substituting model performance for prediction confidence.
This aggregation method is particularly useful in our case, as it inherently handles uncertainty across patches. Moreover, due to the multi-label nature of the task, we compute a posterior probability for each class independently across patches, assuming conditional independence between them. Figure 6 illustrates our overall inference for an image in detail.
This method remains unchanged from our 2024 submission [
Similarly we also incorporated using z-score as the threshold. Equation 2 describes the z-score, which is a measure of how many standard deviations a data point is from the mean of a set. = −
Where is the data point, is the mean and is the standard deviation of the set. Typically used to find outliers, we implemented this calculation to find any classes that are significant enough to be considered as present in a plot and as such we set our threshold to 2.
Bayesian Model Averaging (BMA) typically combines multiple models by weighting their outputs according to their posterior probabilities given the observed data. It consists of two key components: the likelihood of each model given the data , and the prior probability of the model before observing the data.
likelihood of a patch :
As stated in our previous working notes [
1 × ∑︀
1 =1 ×
(|) = ∑︀=1 (|) = ∑︀ =1
Leaving the final probability of a patch as a fraction of the sum of all likelihoods of all patches. We chose 3 metrics to calculate the likelihood of the patches, namely variance, entropy, and plant percentage. Each method was used one at a time, using each method to assign a weight to each patch in a plot.
Variance is a statistical term that describes the spread of a given data based on the standard deviation and mean of the set. Typically, the higher the variance, the higher the spread and the more skewed the set is.
2 = ∑︀=1( − ¯)2
− 1
Equation 6 describes variance where 2 is the sample variance, is the number of samples and ¯ is the mean of all the samples. Hence variance was used as a proxy for how ’confident’ the model is; a higher variance suggests a few classes score disproportionately higher than the rest, therefore the variance of the set as a whole would increase, and we can assume the model is confident. Likewise, if the model is confused and produces a probability distribution that is less exaggerated or more spread out, we assume the model is not confident.
Typically variance values are in the range of 0 to 1, and in our case typically less than 0.01. To amplify the score we used the absolute common log (log10) of the variance to get a range that is easier to work with as in Equation 7.
= |10(2)|
Another challenge would be mapping the variance to a usable range. while normalizing all the values could be an option, we opted to use a custom curve that would penalize higher values less as shown in Equation 8
√︂ + 0.5 − = (8)
+ 0.5
Where is the maximum absolute log of variance across all patches and is the absolute log of variance of a given patch
This would provide a mapping function from the absolute log of variance to a confidence score, which
could then be used as a likelihood for the patch.
3.3.4. Entropy
Entropy, or Shannon Entropy [
=1
The entropy of the patch’s distribution is then used directly as the likelihood.
In an attempt to classify a patch as truly plant based, we developed a small regression model that would
give a score from 0 to 1. The goal was to have the model distinguish between organic (plants) and
inorganic (rulers, wooden planks, rocks, dirt) and produce a score for the patch as a whole. The training
set was 3000 images taken randomly from all the patches across all plots, specifically where each plot is
split into 64 patches. Using a custom labelling software, shown in Figure 7, a patch’s plant percentage
was estimated by overlaying an 8 by 8 grid onto the patch, and calculating the percentage of squares
contain plants.
(6)
(7)
(9)
3 models were tested, namely ResNet50 [
Due to overlapping commitments within the limited competition timeline only 5 runs were oficially submitted. Originally there were a planned 17 including the oficial runs, however the remaining 12 were submitted after the deadline and will be included but noted as unoficial scores.
The evaluation metric used was macro-averaged F1 score per sample, to balance both false positives (incorrectly predicting species that are not present) and false negatives (not predicting species that are present), using Precision and Recall. Equation 10 describes how macro-averaged 1 is calculated Where 1 is calculated by: 1 = 1 ∑︁( 1 ∑︁ 1 )
=1 =1 1 = 2 × ×
+
Precision (Equation 12) serves to calculate how many true positives were predicted from a full set of true positives and false positives , while Recall (Equation 13) relates to how many true positives were predicted from a full set of true positives and false negatives = =
+
+
Two scores were produced: a public score that was calculated on 11% of the test data during the competition, and the private score which was shown for all teams at the end of the competition using the remaining 89%.
Given the 4 pretext tasks discussed and an additional base configuration where the model was taken as-is, and the 3 methods of calculating likelihood, 15 runs were produced. We also submitted a baseline using the base configuration and only using 64 patches per plot, whereas all the runs used an both 64 and 16 patches per plot, as well as another baseline which utilised of simple aggregation, which added (10) (11) (12) (13) up all the class probabilities across all plots and applied the threshold. Table 5 shows the results of the 17 runs.
Among our oficial submissions, the highest-scoring model, ironically, was our baseline, placing 7th overall. Our other runs placed 13th through 15th, with the lowest being dinov2_contrast_variance at 18th. Due to limited time and resources, we were unable to submit our strongest runs during the competition window. Frustratingly, these also turned out to be simple base models. One such post-hoc run using straightforward aggregation with the provided ViT model achieved a macro-F1 score of 0.33856, which would have outperformed the oficial 3rd place. The strongest unoficial result came from dinov2_base_variance, scoring 0.34318, again placing at 3rd had it been submitted.
Given the success of our best, albeit unoficial, run of dinov2_base_variance, we opted to explore the parameters that may increase its performance. For this run we would like to emphasise that no model tuning was done, and the only parameters were related to post processing the model’s prediction. All results are derived using the pre-trained model provided by the organisers and the variation lies solely in the aggregation of patch-level predictions.
For every plot, both 64 and 16 patch per plot predictions were obtained and aggregated using our Bayesian Model Averaging (BMA) framework. In this setup, we employed variance as the model likelihood, motivated by its consistent performance; it outperformed both entropy and plant-percentage based measures in 4 out of 5 models, with only a marginal exception (dinov2_rgb_elim_entropy, outperforming by 0.00021).
Our motivation was straightforward: since all finetuned models underperformed the baseline, we hypothesized that substantial gains could be made by tuning post-processing alone, treating the model as fixed and optimizing everything around it.
Accordingly, we explored a range of hyperparameters within the aggregation pipeline to better understand their influence on performance.
We first identified the inputs to the pipeline. In prior runs we relied on prediction probabilities after softmax has been applied, which normalises all logits, across 7,806 classes, from 0 to 1. However we argued that if the disparity between the probabilities of the highest predicted class and the lowest predicted class is too large, it may exaggerate some predictions, particularly in the lower end of top-, to be included into the final scores which may be false negatives.
To address this, we experimented with using raw logits directly. Instead of applying softmax across all classes, we first sorted the raw logits per class in descending order and then applied softmax to a truncated subset. This allowed for finer control over the distribution of logits within the top- classes without distorting the relative magnitude of the raw outputs. 6.1.2. K values Following the adjustment to logits, we examined the efect of varying the number of top predictions () considered per plot. Under the assumption that typical patches may contain fewer than 10 distinct species, we hypothesized that limiting could reduce prediction noise. All our previous runs used a k-value of 100. We evaluated values of 100, 500, 1,000, and 5,000 6.1.3. Z-score Finally, we investigated the z-score threshold used in the final selection step. While a threshold of 2.0 was initially chosen as a conventional outlier cutof, this value was largely arbitrary. We suspected that adjusting this threshold might allow borderline-relevant classes to be included in the final predictions.
To evaluate this, we varied the threshold from 1.0 to 2.0 in increments of 0.1. This allowed us to assess whether loosening the strictness of the filter could yield improvements without introducing excessive noise.
Another hyperparameter does exist regarding z-score, pertaining to how many values out of the final set are considered. Our previous runs use 100 (similar to -value) to calculate the z-scores, however for these experiments we opted to keep it at 100 regardless if is higher than 100.
To assess the impact of the identified hyperparameters, we conducted a full permutational sweep across all configurations, resulting to 88 distinct runs, with top 10 shown in Table 6.. The following section outlines key performance trends and notable takeaways.
The ablation results were surprising; not only did the best-performing configuration exceed expectations, but it also significantly outperformed the second-ranked configuration by a margin of 0.00639 in macro-F1 score, and was only 0.01351 behind the top-performing entry overall. Notably, these results were achieved without any model fine-tuning or architectural changes; purely through adjustments to post-processing parameters. This highlights the potential of post-hoc methods, particularly when starting from a strong pretrained model.
Among the hyperparameters, the z-score threshold had the most pronounced efect. The performance curve was smooth and showed a clear optimum around z = 1.6 as shown in Figure 9, validating the idea that threshold tuning alone can yield substantial gains.
In contrast, varying the -value (i.e., the number of top predictions considered) showed less consistent trends as seen in Figure 10. While higher -values (e.g., = 5000) produced some of the top-performing runs, suggesting increased stability, other top-10 runs were found even at = 100. This lack of a clear pattern suggests that the optimal might be more sensitive to interactions with other parameters, rather than being independently influential.
Regarding logit type, there was no conclusive advantage between using raw logits versus softmaxed probabilities. The top 10 runs included a near-even split between the two approaches (5 raw, 5 softmax), and this ratio remained balanced in the top 20 (11 softmax vs. 9 raw), indicating that neither consistently outperformed the other.
All results, including full parameter sweeps, are provided in the Appendix for further inspection.
One of the most surprising outcomes of our study was how often the baseline model outperformed pretext task driven models. Despite applying a range of pretext tasks and transformations, as well as other competitors, it was consistently the base DinoV2 model with variance-based post-processing that yielded the highest scores. This outcome could, however, be due to the foundation itself was simply too limited to benefit meaningfully from added complexity, either due to architecture or the subset of the dataset used for training. Our suspicion leans toward the latter. Without a more robust or discriminative model backbone, the benefits of post-hoc refinements seem capped, regardless of how efective the aggregation or filtering strategy may be.
This realization brings us to an important point: post-hoc methods may have their strongest impact when paired with an already competitive model. In our case the organiser-provided model, despite being a provided baseline, performed significantly better once we applied Bayesian model averaging and variance driven confidence filtering. Even so, simple aggregation as in the case with dinov2_simple_aggregation already posted a competitive score with 0.33856. This suggests a compelling argument for future work: using strong, pretrained models as fixed feature extractors and then relying entirely on lightweight inference-stage adaptations. This approach is particularly attractive in resource-constrained or low-data settings, where full fine-tuning is infeasible.
A particularly illustrative failure case was the plant/non-plant binary classifier model. Despite being trained on a dedicated, hand-labeled dataset of over 3,000 examples, the model failed to surpass even the baseline in final leaderboard scores. One possible reason is that the classifier became too certain; overconfident in rejecting ambiguous or noisy predictions. While this may reduce false positives, it also risks discarding correct, albeit uncertain, labels. Another reason we suspect is that the regression model assigns scores of ’0’, which results in the patch’s predictions being nullified completely where as a lesser penalty might still include the predictions at a lower weight.
In contrast, our post-hoc confidence methods (e.g., variance-based z-score filtering) allowed for just enough uncertainty to admit potentially correct outliers. This demonstrates a kind of useful noise tolerance, where not everything that is low-confidence should be discarded outright. Also evidenced by higher -values where predictions, while intuitively would have been discarded due to perceived irrelevance, may still contribute indirectly by amplifying classes that are certainly present.
Finally, the broader implication of these findings is a shift in emphasis. Rather than focusing solely on making the model better through training, it may be equally (if not more) productive to improve what we do after the model makes its predictions. Given that state-of-the-art vision transformers are increasingly capable out of the box, enhancing inference strategies might yield significant gains with far less computational cost. Especially in large-class multi-label problems like this one, confidence-aware methods provide a valuable mechanism for navigating uncertainty without retraining from scratch.
This work explored the potential of post-hoc methods to enhance prediction performance in the PlantCLEF 2025 challenge. Without altering or fine-tuning any model architecture, we investigated whether strategic filtering, ranking, and aggregation could push standard models beyond their baseline performance. The answer, surprisingly, was yes, and by a notable margin.
Although regrettably due to constraints we were not able to fully test the extent of fine tuning models or exploring SSL tasks more in depth, we opted to leave them in due to our motivations of using crafted pretext tasks for ecological applications, and provide a starting line for future work exploring these methods.
Using the organiser-provided DINOv2 ViT backbone, we implemented a lightweight post-processing pipeline based on Bayesian model averaging, z-score filtering, and class ranking by variance-derived confidence. This method alone was suficient to outperform more complex pretext task-based models, including those that incorporated domain-specific augmentations or external classifiers. Most notably, our best-performing unoficial run would have placed 2nd overall on the leaderboard, beating out ifne-tuned models despite relying solely on inference-stage modifications.
These results afirm a broader insight: post-hoc methods are an underappreciated tool in large-scale, large-set classification. While it is common to attribute performance gains to deeper networks or larger datasets, we show that thoughtful post-processing of predictions, particularly when incorporating uncertainty measures, can yield comparable if not superior improvements. Crucially, this comes with minimal computational burden and no retraining costs.
Looking ahead, future work may explore this further in two directions. First, by pairing post-hoc strategies with stronger or more targeted pretrained backbones, we may unlock even more performance. Second, by expanding the space of inference-time adaptations—perhaps incorporating learned priors, spatial or temporal context, species co-existence or adaptive thresholds, we might approach state-ofthe-art results with surprisingly simple setups.
During the preparation of this paper, the author(s) used ChatGPT in order to paraphrase, reword and to check grammar. After using this GenAI service, the author(s) reviewed the contents of the generated content and take(s) full responsibility for the publication’s content.
The resources of this project is supported by NEUON AI SDN. BHD., Malaysia.
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 exp_SOFTMAX_k1000_z2.0.csv exp_RAW_k5000_z2.0.csv exp_SOFTMAX_k100_z2.0.csv exp_SOFTMAX_k500_z2.0.csv exp_RAW_k1000_z1.9.csv exp_RAW_k1000_z1.8.csv exp_RAW_k500_z1.8.csv exp_RAW_k1000_z1.4.csv exp_RAW_k500_z1.4.csv exp_RAW_k100_z1.6.csv exp_RAW_k100_z1.7.csv exp_RAW_k100_z1.8.csv exp_RAW_k5000_z1.3.csv exp_RAW_k500_z1.9.csv exp_SOFTMAX_k5000_z1.3.csv exp_RAW_k100_z1.4.csv exp_SOFTMAX_k1000_z1.3.csv exp_SOFTMAX_k500_z1.3.csv exp_SOFTMAX_k100_z1.3.csv exp_SOFTMAX_k5000_z1.0.csv exp_RAW_k1000_z1.3.csv exp_RAW_k5000_z1.0.csv exp_RAW_k100_z2.0.csv exp_SOFTMAX_k1000_z1.0.csv exp_RAW_k500_z1.3.csv exp_SOFTMAX_k500_z1.0.csv exp_RAW_k5000_z1.2.csv exp_SOFTMAX_k500_z1.2.csv exp_SOFTMAX_k1000_z1.2.csv exp_SOFTMAX_k5000_z1.2.csv exp_RAW_k100_z1.3.csv exp_RAW_k500_z1.2.csv exp_RAW_k100_z1.9.csv exp_SOFTMAX_k100_z1.2.csv exp_RAW_k1000_z1.1.csv exp_SOFTMAX_k100_z1.0.csv exp_RAW_k100_z1.2.csv exp_RAW_k5000_z1.1.csv exp_SOFTMAX_k5000_z1.1.csv exp_RAW_k500_z1.1.csv exp_SOFTMAX_k1000_z1.1.csv exp_RAW_k1000_z1.0.csv exp_SOFTMAX_k500_z1.1.csv exp_SOFTMAX_k100_z1.1.csv exp_RAW_k1000_z1.2.csv exp_RAW_k500_z1.0.csv exp_RAW_k100_z1.1.csv exp_RAW_k100_z1.0.csv SOFTMAX RAW SOFTMAX SOFTMAX RAW RAW RAW RAW RAW RAW RAW RAW RAW RAW SOFTMAX RAW SOFTMAX SOFTMAX SOFTMAX SOFTMAX RAW RAW RAW SOFTMAX RAW SOFTMAX RAW SOFTMAX SOFTMAX SOFTMAX RAW RAW RAW SOFTMAX RAW SOFTMAX RAW RAW SOFTMAX RAW SOFTMAX RAW SOFTMAX SOFTMAX RAW RAW RAW RAW