The PlantCLEF 2025 competition promotes scientific research in biodiversity monitoring. Participants are invited to analyze high-resolution images, known as quadrats, to identify the plant species present. This task poses several challenges. For example, plants may occupy only a small proportion of the total image area, or they may have a morphology that makes them dificult to identify due to seasonal variations. Every detail counts, and the entire processing chain must be optimized, from loading the high-definition image to compiling the list of predicted species. Preprocessing, which transforms the source image into a tensor for the machine learning model, is the focus of this article's study and leads to the best classification performance in the PlantCLEF 2025 competition. As part of this participation, three approaches to analyzing the quadrat images were explored. The ifrst method, used as a reference, simply reduces the images to the expected model resolution, allowing the research to focus on optimizing preprocessing. The research results were then used to tile the images before analyzing them with a single-shot high-resolution inference. The Rust computer language was primarily used. The associated source code is available on a public repository.1.
Environmental issues and biodiversity conservation are becoming increasingly important, particularly
for monitoring the decline of native species and the emergence of invasive ones [
The Pl@ntNet team, which has been active for fifteen years in research related to the visual
identification of plant species, has developed a free application [
New automation capabilities for image analysis have been made possible by artificial intelligence,
which is gradually being integrated into various aspects of everyday life, such as medicine and
autonomous vehicles. In 2012, the ImageNet competition, which involved classifying images into 1,000
categories, was won by the AlexNet model [
In 2017, convolutional neural network technology was surpassed by Transformers [
Certain use cases of artificial intelligence require specific reliability and robustness criteria, regarding
the quality of predictions and proper software functioning. Research is underway to develop explainable
AI to increase confidence in the technology [
Participation in PlantCLEF 2025 aims to advance scientific research in image analysis and demonstrate
the potential of the Rust language, through the Candle library [
Implementing image analysis methods in Rust required reproducing the oficial code of the reference
method (Python script "Oficial Starter Notebook | Inference on Test Data") and processing the image
directly at the expected 518x518 resolution. This process highlighted the essential role of preprocessing,
which was then investigated further. Next, two approaches to high-resolution image processing were
explored, each leading to diferent technical challenges: the tiling method, used in the previous edition
of the competition [
A second approach, VaMIS (Variable Model Input Size) [
The goal of this competition is to improve the performance of plant classification, i.e., to identify the species visible within each quadrat based primarily on visual data rather than image metadata. Therefore, these metadata have been utilized to a limited extent.
The annual PlantCLEF competition aims to promote scientific research in plant classification. Since
last year [
A significant part of the research was devoted to data preprocessing. Various avenues of research have made it possible to optimize technical choices, such as the interpolation used during resizing or the choice of a subsampling scheme for certain color channels and the JPEG compression quality, performed after resizing, which can surprisingly be seen as a regularization factor.
During the five-week participation in the PlantCLEF 2025 competition, two approaches were explored for analyzing high-definition images with a VisionTransformer implemented in Rust: the classic tiling method and high-resolution inference methods known as VaMIS, for VAriable Model Input Size. VaMIS uses either global or window attention, or both, within the same model, referred to as hybrid attention. Since VaMIS-adapted models are highly memory-intensive in terms of GPU, it was necessary to reimplement an attention module calculation in CUDA, which is more VRAM-eficient.
A brief collaboration with the company Quandela should also be mentioned. The goal was to conduct
preliminary tests to evaluate and enhance the performance of multilabel classification using quantum
photonics and demonstrate the potential of this technology. The Python libraries Perceval, for designing
quantum circuits, and MerLin [
The PlantCLEF 2025 competition promotes innovation in ecology, focusing on biodiversity monitoring and plant species evolutionary dynamics. Based on the standardized quadrat protocol sampling method, the challenge motivates participants to develop automated approaches for analyzing high-resolution images of plant quadrats. The main objective is to identify various plant species among over 7,800.
The PlantCLEF 2025 monolabel training dataset remains the same as in the previous edition. It consists of observations of individual plants in southwestern Europe and covers 7,806 species. The dataset includes approximately 1.4 million images, which have been supplemented with additional images from the GBIF platform to include less represented species. The images are pre-organized into subfolders by species and divided into training, validation, and test sets to facilitate model training.
The PlantCLEF 2025 test dataset is a compilation of quadrat image datasets in various floristic contexts. It contains a total of 2,105 high-resolution images. The shooting protocols vary, including diferent angles and weather conditions. These images, produced by experts, allow to evaluate the ability of models to accurately identify plant species under various conditions, thereby testing their robustness.
A complementary dataset is also available [
The images are provided in JPEG format, which can introduce compression artifacts that modify the represented data by reducing its entropy, according to the principle of lossy compression. Two main parameters control compression: JPEG compression quality and the YCbCr subsampling scheme. The latter is the underlying color space used to decompose the luminance of the chroma channels. These compression parameters are discussed in a dedicated section below.
Tables 1 and 2 summarize the JPEG technical characteristics of the training dataset, which contains
images with a maximum size of 800px. Similarly, tables 3 and 4 report the JPEG compression parameters
of the quadrat test set.
3.1.4. Models
To facilitate access to the competition, two VisionTransformer [
The first model uses public DINOv2 weights pre-trained by SSL for the image encoder. Only the classification head was trained in a supervised manner. The second model was trained entirely in a supervised manner using the first model as initial checkpoint. Both models were trained on a server with A100 GPUs using the Timm library with Torch. The first model was trained for approximately 17 hours over 92 epochs, with a batch size of 1,280 images per GPU, and a learning rate of 0.01. The second model was trained for about 36 hours over 92 epochs, with a batch size of 144 images per GPU and a learning rate of 0.00008.
For the PlantCLEF 2025 competition, only the second model was used. Since all of its parameters were fine-tuned on the plant training dataset, it can be expected to perform better than the first model. 3.1.5. F1 metric The metric chosen to rank participants’ submissions is the macro-averaged F1 score per sample, which strikes a good balance between statistical recall and precision. The 2,105 images are grouped into transects, which represent samples from specific areas within selected sites. To mitigate biases related to oversampled areas, the score is first calculated for each transect in the test set and then averaged across transects to obtain the final score.
Several machine learning challenges must be solved to enable efective classification. The training data consists of images of individual plants or parts of plants with a single label, while the test data consists of images of vegetation quadrats with multiple labels. Therefore, the monolabel model must be adapted to perform multi-label classification. Additionally, the test images have fairly high resolutions, frequently ranging from eight to ten million pixels, compared to the model’s input size of approximately 250,000 pixels. The next section details diferent approaches to address this discrepancy.
Quadrat image analysis involves processing a substantial amount of information, including images containing nearly 10 million pixels and a tiling process that can multiply the number of inferences by 100 compared to the initial dataset. Additionally, the analysis predicts between one and 15 species out of 7,806 possibilities, generating a large number of potential combinations. In accordance with the oficial public Python script, the decision was made to limit the maximum number of predicted species to 15 in order to avoid substantially reducing statistical accuracy.
A clear pipeline for processing this data must be established. This pipeline can be described in four successive steps: 1. Preprocessing: conversion of a 3000x3000x3 quadrat image stored on hard disk in JPEG format to 518x518x3 tensor(s) in RAM (or VRAM).
First, the high-definition image file is loaded into RAM. Then, the JPEG format is decoded. Next, the image is resized and cropped. JPEG decoding can take into account the final position of the pixels. In other words, it is not necessary to decode all 10 million pixels of the initial quadrat image because this process is relatively intensive. One or more tiles (i.e., square areas within the original image) are extracted and resized to the model’s input resolution of 518x518 pixels or a higher resolution for the high-resolution inference approach. 2. The image encoder of the DINOv2 plant model: transformation of the 518x518x3 image tensor(s) into a vector(s) of size 768.
In a standard deep learning approach, the reduced image is provided as input to the model in the form of a tensor. The VisionTransformer image encoder calculates "deep features," which are lfoating-point vectors that summarize all the plant information contained in the original image. Then, it performs linear classification. It was decided that these deep features (more precisely, only the CLS token, which is used for classification) would be stored on a hard disk. The classification head is usually calculated immediately after the image encoder. However, it is preferable to divide the model inference into two steps. This is because calculating deep features is more resource-intensive than calculating the classification head. The latter can be recalculated quickly after loading the deep features from the hard disk without significant delay. Additionally, storing predictions by tile and species would require much more space than storing the respective deep features. Finally, testing other classification heads or fine-tuning only the head may be desirable, if necessary. 3. Linear classification and prediction aggregation: transformation of the 768-size vector(s) into a 7806-size score vector.
Each deep feature vector (CLS token) is input into a linear classifier to obtain a vector of logits, and
then the SoftMax function is applied to get probabilities by species. Predictions are aggregated
from the tile level to the high-definition image level using the chosen method: maximum pooling
by species, average pooling, a given quantile [
To limit confusion and improve statistical precision, it is also possible to retain only one or two species per tile, as is done in the oficial code provided for the tiling method. 4. Submission calculation: transformation of the 7806-size score vector into a list of species.
Only the 15 most probable species at the high-definition image scale are retained (e.g., Top15), and only those with a score above the detection threshold are selected. The final list comprises the species chosen by the model as predictions for the given quadrat image. Other methods of calculating submissions are possible.
As part of the PlantCLEF 2025 competition, the research primarily focused on the initial stages of the pipeline: image preprocessing and selecting an image encoder process to analyze high-definition images. Two approaches were implemented for the latter: the classic tiling technique and the VaMIS approach, which analyzes each image with a single high-resolution inference. Due to insuficient computing power, these approaches were only tested in inference mode, i.e., without fine-tuning the deep learning model. Finally, prediction aggregation and submission calculation were briefly explored without specific analysis.
To implement these models and calculate inferences and submissions, Python and Rust code were produced.
The development of the Rust pipeline started with implementing the reference approach, which solves the problem of analyzing high-definition images by simply reducing them to the resolution expected by the VisionTransformer model (518 x 518 pixels). The initial goal was to reproduce the results of the oficial script provided for the "Oficial Starter Notebook | Inference on Test Data" competition. However, multiple tests and the launch of the code revealed a bottleneck caused by image preprocessing. Loading and decoding JPEG images with nearly 10 million pixels and reducing them to 518x518 pixels requires as much computational power as, if not more than, the VisionTransformer inference that follows image loading.
Therefore, the initial step involved performing external preprocessing of the images in the PlantCLEF 2025 test dataset to reduce their size upstream. This allowed to use the reduced images as input and test the Rust solution. The list of plant species predicted by quadrat varied significantly depending on the chosen preprocessing. In particular, preprocessing involving JPEG compression after resizing the high-definition image showed significant variations in performance.
Therefore, research on preprocessing was conducted along several lines: the interpolation used for resizing, and in the case of post-resizing JPEG compression, the YCbCr subsampling scheme and compression quality factor used by JPEG.
Tests were conducted using the reference approach, which reduces images to the expected dimensions of 518x518 pixels at the model input. This technical choice allows for multiple tests at reasonable computational cost. It also allows for better evaluation and comparison of the diferent preprocessing methods. Listing the plant species present in a quadrat image with a reduced resolution of 518x518 pixels is dificult; every pixel counts for identifying plants. This difers from a tiling process, in which tiles are extracted at the original resolution of the image by cropping it. In the latter case, the importance of technical preprocessing choices is reduced. The detection threshold was set to a one percent probability for all the experiments, which is the same value as the oficial 518x518 reference method.
Finally, the impact of preprocessing, including post-resizing JPEG compression, was measured on the two approaches implemented for PlantCLEF 2025: tiling and the high-resolution VaMIS approach.
Interpolation necessarily plays a role in the final performance of the classifier. Reducing the image size requires estimating the color of pixels in a grid that may not overlap the original geometry. This process makes assumptions when choosing a surface model. This introduces an additional source of noise that may afect the downstream deep learning model, thereby penalizing classification quality. The following interpolations were tested: nearest neighbor, bilinear, bicubic, and Lanczos. Many other interpolation methods exist and could have been evaluated as well. However, the decision was made to limit the evaluation to the most commonly used methods.
Nearest neighbor interpolation is unique because it simply involves selecting the pixel in the source
grid that is closest to the target pixel. In contrast, bilinear, bicubic, and Lanczos interpolations are
convolution products, i.e., weighted averages of the pixels in the neighborhood of the target pixel
projected onto the source grid. Named after Hungarian mathematician and physicist Cornelius Lanczos,
the latter interpolation is calculated using the cardinal sine function and is known to reduce visual
artifacts such as blurring and aliasing. Figure 1 illustrates the filters associated with these convolution
products. [
The classification performance of preprocessing was evaluated with each of the four interpolations on grayscale test data images to avoid dependence on a particular color space and to focus on the geometric calculations involved in reducing high-definition quadrat images to 518x518 pixels. 4:2:2 4:2:0 4:1:1
YCbCr subsampling scheme Various preprocessing methods were considered, including recompressing the resized images to JPEG format. The initial aim was to reduce disk space and subsequently optimize classification performance. This required selecting the YCbCr subsampling scheme and compression quality used by JPEG.
The YCbCr color space, obtained via a linear transformation from RGB, distinguishes the Y luminance channel from the Cb and Cr chroma channels. To reduce data size while accounting for human visual perception, JPEG allows the chroma channels to be subsampled relative to the luminance channel according to standardized schemes: • 4:4:4 : No subsampling. The chroma and luminance components have the same resolution. • 4:2:2 : chroma is subsampled horizontally by a factor of 2. • 4:2:0 : chroma is subsampled horizontally and vertically by a factor of 2.
• 4:1:1 : chroma is subsampled horizontally by a factor of 4.
These schemes are summarized in Figure 2.
A fifth, less common scheme appears in the test data, in which chroma is subsampled vertically by a factor of two. It is denoted 4:2:2v in this article.
JPEG compression quality In JPEG compression, the quality factor is a number between 0 and 100
that determines how high frequencies are filtered in the discrete cosine transform (DCT, [
The prediction performance of the single inference reference method was assessed by reducing high-definition quadrat images to a resolution of 518x518 pixels. The tested preprocessing methods involve post-resizing JPEG compression, including a comparison of five subsampling schemes and JPEG compression qualities ranging from 75 to 100 percent.
Principle This paper explored the tiling method used in previous editions [
Each tile is fed into the VisionTransformer image encoder, which returns a vector of deep features stored on the hard drive. During the competition, diferent scales were calculated and aggregated (from scale one to scale ten). The preprocessing implements JPEG compression after resizing the high-definition image.
Prediction calculation and aggregation The model’s linear classification head is applied to the
deep features of each tile to obtain the logits. The probabilities for each species are obtained by applying
the SoftMax function. Probabilities at the image scale are obtained by applying maximum or average
pooling per species. Initially used for convolutional neural networks [
VaMIS (Variable Model Input Size) methods analyze high-resolution quadrat images using a single inference and adapt the machine learning model. This reduces computational costs compared to the tiling method, which performs multiple inferences at diferent scales.
In the VisionTransformer [
Window attention This method, known as Window Shifted Attention (WSA) [
Windows smaller than 518x518 can be chosen, provided the resized image size is a multiple of the window size. For example, a window VaMIS with a size of 1680x1680 and 5x5 windows of size 336x336 or 8x8 windows of size 210x210 pixels is possible.
Hybrid attention One limitation of window attention is that local tokens from diferent windows never interact with each other. Hybrid attention, which alternates between global and window attention, allows information to difuse better across the whole image by enabling all tokens to interact within given Transformer blocks.
The Segment Anything image encoder [
The three VaMIS methods described above were tested at a single resolution of 1554x1554 (equivalent to scale three of the tiling). Due to a lack of computing resources, the deep learning models were not ifne-tuned for these methods and were only used for inference. The preprocessing implements JPEG compression after resizing the high-definition image.
between all pairs of tokens in the image, resulting in an attention matrix. The attention matrix has 2 elements, where is the number of tokens. VisionTransformers divide an input image into patches of a ifxed size (e.g., 14 x 14 pixels for DINOv2) and then convert the patches into tokens. The number of tokens increases linearly with the image’s surface area and therefore quadratically with its dimensions (width and length) while maintaining its aspect ratio. Thus, doubling the dimensions of an image multiplies the size of the attention matrix by 16.
VaMIS methods rely on increasing the model input size, but the limits of the graphics card are quickly
reached using the explicit method to calculate the attention module because the attention matrix must
be stored in GPU memory. FlashAttention [
The above methods provide a vector of presence scores for each species at the image level. The list of species predicted by the model is obtained by thresholding; that is, by listing the species whose score (e.g. probability) exceeds a chosen detection threshold.
For the tiling method, the threshold was slightly optimized by testing diferent levels and observing the public F1 score of the respective submissions. However, this method is costly in terms of the number of submissions required.
The detection thresholds for the VaMIS methods were chosen roughly, considering the average number of species detected. One run was performed for each VaMIS method to allow for initial validation of the Rust implementation.
Semi-automatic methods were tested occasionally to adjust the detection thresholds per transect because some transects are more dificult to analyze and require diferent thresholds. These thresholds were determined using a semi-automatic procedure with dichotomy to maintain an average number of species similar to a reference classification submission. A multiplicative parameter adjusts this number globally to optimize performance. These methods were primarily applied to tiling techniques.
Finally, manual optimizations were briefly explored through a detailed analysis of the species prediction probabilities generated by the model and subsequent submissions. For example, if a species frequently appeared within a transect, it was occasionally generalized across all images of that transect. Conversely, species appearing only once (or only a few times) within a transect were deemed outliers and subsequently excluded from the predictions. Such considerations were applied across various scales: the entire dataset, individual transects, and images from the same location within a transect. Given that these methods incorporate metadata rather than relying solely on visual data, performance metrics with and without these optimizations are provided to illustrate the quality and generalizability of the approaches outlined in this technical note.
Most of the test data (89 percent) is used to calculate the private F1 score, making it more robust than the public F1 score from a statistical perspective. To perform a relevant analysis of the results, this section mainly lists the private F1 statistics according to diferent methods and parameter sets. Figure 5 illustrates the classification performance of diferent interpolations. Nearest neighbor interpolation is the least efective, while Lanczos outperforms the others. It is notable that F1 score of 0.07906 is achieved with grayscale images (Lanczos interpolation), compared to a score of 0.12077 with the reference approach of the oficial script, which uses the three RGB channels.
Figure 6 illustrates the classification performance for diferent YCbCr subsampling schemes. Two "modes" or two maxima, are observed for diferent sets of parameters, summarized in Table 6. We ifnd that a JPEG quality of 100 percent does not improve multi-label classification performance for the considered model. Performance improves as soon as this quality is slightly reduced. In particular, all schemes perform remarkably well (private F1 score greater than 0.22) with a JPEG quality of 75 percent. 75 80 85 Quality 90 95 100
The 4:1:1 subsampling scheme, which reduces chroma by a factor of four relative to luminance, performs well for all JPEG qualities. In contrast, the 4:4:4 scheme, which does not subsample, requires significantly reduced JPEG compression quality (75 percent) to perform well.
The tiling schemes are summarized in Table 7. The tiling scheme that provides the best performance is the one with 91 tiles and a private F1 of 0.35013. Schemes with more tiles do not seem to improve performance. Note that the above results do not use metadata and only evaluate visual detection performance.
Based on the results of the 91-tiles scheme, the detection thresholds were subsequently optimized by studying the average number of species per transect. Manual optimization involves quickly analyzing the predictions and species lists visually at diferent levels of image grouping, such as the entire dataset, a transect, images from the same location within a transect, or an individual quadrat image. For instance, frequent species within a group of test images were generalized and outliers were removed. The performance results are summarized in Table 8.
Optimizing the threshold per transect improves the private F1 score by approximately +0.007. Manual optimization improves the private F1 score by an additional +0.0075 on top of this initial gain.
High-resolution inference methods without fine-tuning achieve the performance shown in Table 9.
Among these methods, the hybrid attention method based on the SegmentAnything image encoder
attention scheme [
All of the performance results shown below were obtained using a purely visual model. The detection threshold was the same for all of the images in the test dataset, and no metadata was used.
Two types of preprocessing were compared: Standard (without JPEG compression) and with JPEG compression after resizing. The Standard approach uses the Rust Image crate, which can be compared to Python/Pillow, while the JPEG compression approach uses the MagickRust crate, which uses Wand/ImageMagick and JPEG compression. Due to preprocessing, the performance delta indicated in Table 10 can be observed for the diferent tiling schemes.
For the VaMIS models, the performance deltas are listed in Table 11.
Based on the results, Lanczos interpolation performed the best out of those tested. This is a common
ifnding in signal theory [
Interestingly, a JPEG quality of 100 percent with no subsampling (4:4:4 scheme) does not yield optimal performance in multi-label classification. The tests revealed two distinct "modes" in which the model performed well: One is the 4:1:1 sampling scheme with 94 percent JPEG compression; the other is the 4:2:2 scheme, which requires 85 percent compression.
These two modes are related to the characteristics of the data used to train the model, which include a quasi-unique 4:2:2 sampling scheme and two main levels of compression for the training images. 76 percent (for more than 3 images over 4) and 95 percent. These uniform JPEG parameters determine the entropy, or variance, and more generally the distribution of the RGB color channels of the input pixels during training. It is possible that the model has learned to process only images with these characteristics.
When a test image is resized from a high resolution, such as 3000x3000 pixels, to the model’s input size of 518x518 pixels—efectively dividing each side by a factor of approximately six—the resulting image exhibits a significantly higher equivalent JPEG quality compared to its original version. Furthermore, during tile extraction at a scale of six, which involves minimal resizing, two-thirds of the test images retain a compression quality of 98 percent (see Table 3). Overall, all inference methods tend to supply the model with images of higher quality than the one used during its training phase (see Table 1).
Figure 7 illustrates the impact of various compression methods on the YCbCr channels. The model was primarily trained using images with the 4:2:2 scheme (bottom row, middle image). The RGB images are almost indistinguishable to the human eye, following the principle of YCbCr subsampling. However, the model may be more sensitive to these variations than the human eye. Thus, changing the compression scheme between the training and test data constitutes an additional domain shift.
The VisionTransformer appears to treat additional data in the input image, beyond the entropy
expected by the model, as noise, which reduces its ability to focus on classification details. Figure 8
compares the 12 attention maps of the final block of the model when it is provided with an uncompressed
image and an image compressed with the optimal parameter set, "Scheme 4:1:1 — Quality 94." The first
map shows slightly noisier attention, while the second map shows more focused attention on certain
points. From the model’s point of view, JPEG compression can be seen as a regularizing factor for the
image, as the model expects precise entropy on the input pixels. It is noteworthy that the benefits of
JPEG compression in a diferent deep learning context have been previously discussed [
From the point of view of the final classification task, this lost data during compression after resizing
is certainly informative, particularly in reducing confusion between species thanks to the details of
neighboring pixels contained in the high frequencies of the JPEG discrete cosine transform (DCT, [
Figure 9 allows for a visual comparison of diferent submissions resulting from diferent preprocessing methods. Submissions with JPEG quality ranging from 91 to 95, as well as those with bilinear or bicubic interpolation, are similar. The ImageMagick submission is somewhat isolated, indicating specific preprocessing. The oficial submission is similar to preprocessing with the JPEG 4:2:2, 4:2:2v, and 4:4:4 schemes.
The tiling approach achieves the best results for high-resolution analysis. It gives a private F1 of 0.35013 with a scheme involving six scales. The high-resolution inference approach (VaMIS) achieves a private F1 of 0.22711 with the hybrid attention method, which alternates between global and window attention. These results improved with preprocessing and post-resizing JPEG compression, explaining a performance delta of +0.03354 and +0.05511, respectively.
Further research concerning the color space used during image resizing could be pursued to more exhaustively study preprocessing. Some color spaces use nonlinear transformations, such as LAB and LUV, and interpolation can produce diferent images and performance.
Preprocessing plays a significant role in the classification performance of computer vision models, especially when analyzing high-definition quadrat images. In these images, plants may occupy only a small area, making each pixel of the reduced image of great importance.
The model is conditioned by the training data. Classic training on the ImageNet dataset requires image normalization of the same name during the testing phase. It focuses on the first two moments of the input image tensor. More generally, the model is afected by the entropy of the input tensors. Therefore, special focus should be given to the format of the training images and their preprocessing.
For instance, having a monolabel training dataset with original, high-resolution images from the sensors before any resizing would allow to use various preprocessing techniques during training such as interpolation method or JPEG compression parameters. This would teach the model to consider the image details in the high frequencies of the JPEG DCT, improving visual classification performance and making the model more robust to preprocessing.
Additional tests could be conducted for high-resolution image analysis. The tiling approach with scales beyond the native resolution of the quadrat images and/or with overlap should be explored further. Indeed, tiles that framed the plants more precisely facilitated identification. Similarly, singleshot high-resolution inference approaches (VaMIS) could benefit from fine-tuning the model’s input size. As previously indicated for tiling, smaller windows for methods involving window attention would refine the spatial framing of each plant within an image.
Additionally, contest participants most likely focused on tiling schemes and methods for aggregating predictions from these tilings. Some may have even tested and fine-tuned other deep learning models. However, preprocessing optimization is not a research avenue usually prioritized to improve the classification performance of a computer vision model. Therefore, we can expect the solutions proposed by other participants to be combined with the preprocessing optimization proposed here to achieve cumulative gains in classification performance.
Finally, since preprocessing is relatively independent of the final multi-label classification task for these high-definition images, the study presented in this paper may be applicable to other competitions and computer vision use cases beyond the scope of the PlantCLEF 2025 competition.
I would like to thank the Pl@ntNet team for organizing this competition, which enabled us to test new methods and avenues of research. Some of these methods were unexpected and daring, but they were all relevant and instructive.
Thanks are also extended to Jean Senellart and Grégoire Leboucher from Quandela [
Gratitude is expressed to Rosine Choupe for facilitating the preprocessing experiment in Pl@ntNet by assisting with field photography and sharing her equipment to diversify the sensors used.
Finally, thanks are due to Guillaume Gomez for reviewing the source code and for providing accessible,
comprehensive, and up-to-date Rust language documentation [
During the preparation of this work, the author used Mistral AI’s LeChat in order to: Grammar and spelling check. The author also used DeepL Translation and DeepL Write to facilitate the translation of this document. After using these tools, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content.
During the final week of the competition, a late collaboration was established with Quandela, a company specializing in quantum photonics and ofering machine learning solutions that leverage this technology.
The Python libraries Perceval and Merlin [
Although the preliminary results were not conclusive at this stage, the experiments conducted as part of the participation in PlantCLEF 2025, in collaboration with Quandela, are briefly outlined here to facilitate their continuation or to inspire future research.
The computations were performed according to the principles of quantum photonics: the quantum circuit, known as the BosonSampler, consists of two main types of components: phaseShifters and BeamSplitters. The floating-point input values are encoded into the superposed input state given to the circuit to perform quantum computations. The phase shifters and beam splitters will then modify the probability of observing a photon in one specific mode, creating a superposed output state diferent from the input one. Those output probabilities are then converted back into floating-point values. This is called strong simulation, all of this is calculated thanks to a classical computer, calling this method a "quantum inspired".
It is also possible to run the same algorithm on an actual quantum computer, running the same experiment multiple times (called shots) measuring the output state, and thus obtaining an estimation of the output probabilities. See Figure 10 for an example of quantum circuit.
A first approach aimed to analyze the 2,105 quadrat images and predict the species present among the 7,806. The classification method used resizes the test images to the model’s input size of 518x518 pixels. The quantum layer was inserted between the image encoder and the linear classifier, with input and output dimensions set to 768 floating-point values. The core idea behind this method is to create a new embedding based on the Vision Transformer one, using quantum to explore new areas of the latent space. This process involved semi-manual techniques, including manually testing diferent quantum circuits, as the Quantum Layer was not fully optimized for GPU execution at that time, leaving some room for improvement on this project. Indeed, optimizing such a high-dimensional quantum circuit using gradient descent proved challenging due to the large volume of data to be processed, especially in a limited amount of time. The linear classifier was fine-tuned using a limited set of training images (a few hundred thousand), selected by capping the number of train images of each of the 7,806 species.
A second approach was explored to reduce computational complexity and allow for multiple tests:
the study was restricted to the 15 quadrat test images from the LISAH-JAS transect and the 78 species
detected by the oficial public Python script submission (which processes images directly at a resolution
of 518x518 pixels). Two successive linear layers with dimensions 768 → 32 → 78 were trained (initialized
by SVD factorization of the weights of the initial model’s linear classifier) to limit the number of
parameters according to the bottleneck principle ([