Predicting the spatial and temporal distribution of plant species is a key challenge for biodiversity monitoring and conservation planning. In this report, we present the solution to the GeoLifeCLEF 2025 challenge, which requires predicting the presence of plant species using satellite images and time series, climate time series and other rasterized environmental data. Our approach utilizes multimodal network with encoders for images, climate and satellite time series. During inference, we apply a fixed probability threshold to produce multi-label predictions. Without any pseudo-labeling or ensembling, our model achieves a macro-F1 score of 0.218 on the public leaderboard and 0.192 on the private leaderboard, placing us 6th place. We analyze the impact of each modality and discuss ways for further improvement.
Species distribution modeling [
The GeoLifeCLEF 2025 competition presents two main complexities: extreme data
heterogeneity and challenging labeling constraints. It is similar to previous editions [
In this paper we introduce an adaptive prediction method that combines a tuned probability threshold with a top-k fallback mechanism. This approach provides more contextually relevant predictions than a fixed-k method while preventing overly sparse outputs. We demonstrate that a single, non-ensembled model can achieve a top-tier ranking. Our experiments show that strategically focusing on the richest data modalities and curating the training set is more efective than including all available data.
The GeoLifeCLEF 2025 dataset contains presence-absence (PA) and presence-only (PO) observations. PO data includes about 5 million observations and reports only the presence and not absence of certain plant species in specific areas. However, PA data combines around 90K surveys with about 5K unique species of the European flora and reports the presence and absence of plant species. The total number of surveys in the test set was approximately 16K.
In PA train data, species distribution exhibits extreme imbalance: 50% of taxa have 16 occurrences, while only 20% exceed 110 observations. Geographically, the data skew towards western Europe, a map of the locations can be seen in Figure 1. More detailed descriptions can be found on the competitions’ homepage1.
Each survey pairs GPS coordinates with some environmental data: • Satellite imagery: 128m×128m Sentinel-2 RGB/NIR patches (10m resolution) + Landsat time-series (6-band quarterly composites); • Climatic data: 19 bioclimatic rasters (1km resolution); • Soil variables: 19 SoilGrids properties (e.g., pH, organic carbon); • Human footprint: 16 time-dynamic pressure indices (1993/2009); • Land cover: (500m resolution) and elevation (30m resolution);
The competition utilizes a macro-averaged 1-score to evaluate species presence predictions. This metric provides equal weight to all species, mitigating bias from class imbalance by independently estimating each taxon’s performance. For each test plot with true species set , the model predicts ranked presence probabilities ,1, ,2, ..., , . After thresholding these probabilities to obtain binary predictions, the metric computes:
Macro 1 = 1 ∑︁ 1, =1
, where 1, =
1 ∑︁
, =1 , + ( , + ,)/2 Here ,, , and , are the true positive, the false positive and the false negative of the i-th input sample, respectively, while represents all 5K species. This formulation emphasizes balanced performance across rare and common taxa.
This section describes our proposed multimodal network and the methods that were tried during this competition.
Our solution for the GeoLifeCLEF 2025 challenge was based on the baselines provided by
the organizers in current and previous competitions. However, through experimentation, we
determined that our approach from the previous year [
Our final model, illustrated in Figure 2, is a multimodal neural network that exclusively
integrates three rich data sources: Sentinel-2 satellite imagery, bioclimatic data cubes, and
Landsat data cubes. The core of our architecture consists of three specialized encoders, one
for each data modality, based on a modified Swin Transformer [
The Landsat data encoder utilizes a Swin_t model trained from scratch. To prepare the data, we first apply an initial LayerNorm to stabilize the input distribution. The primary modification involves adapting the model’s first convolutional layer (the patch embedding layer) to accept 6 input channels instead of the default 3. This allowed to accommodate the rich multi-spectral and temporal information. To simplify the model structure and focus on feature extraction, we replace its final classification head with an Identity layer, which outputs a 768-dimensional feature vector.
Similarly, the encoder for the 19 bioclimatic variables (structured as a 4-channel input) uses a Swin_t architecture also trained from scratch. It is preceded by a LayerNorm layer, accommodating the diversity of Bioclim data, and its initial patch embedding layer is modified to handle 4 input channels. Like the Landsat encoder, its classification head is removed to output high-level features.
To leverage the high-resolution multispectral Sentinel-2 imagery (Red, Green, Blue, and
Near-Infrared bands), we employed a Swin_t model pre-trained on the ImageNet-1K dataset
[
The 768-dimensional feature vectors produced by each of the three Swin_t backbones are
ifrst passed through separate projection heads. Each projection head consists of a Linear layer,
BatchNorm1d, a GELU [
These projected features are then concatenated into a single 3000-dimension vector. This
fused vector is fed into a final classification head—a multi-layer perceptron (MLP) with GELU
[
To manage the highly imbalanced and long-tailed distribution of species occurrences in the training data, we first filtered the dataset. We focused on plant species with more than 5 recorded occurrences, which reduced the number of target classes from the original 5,016 to a more manageable 3,425. This occurrence threshold was empirically determined through experimentation to optimize the trade-of between class coverage and model stability.
The model was trained on the Presence-Absence (PA) data for 12 epochs. We employed the AdamW optimizer with a learning rate of 8e-5 and a Cosine Annealing Learning Rate Scheduler (CosineAnnealingLR) to promote stable convergence. Given the multi-label nature of the problem (multiple species can be present at one location), we used Binary Cross-Entropy (BCE) loss. Training was conducted with a batch size of 300.
For the final predictions on the test set, we developed a hybrid, threshold-based strategy that deviates from the baseline approach of simply predicting a fixed number of the most probable species. Our method is designed to be more adaptive to the local biodiversity of each observation point. After passing the test data through the model, we apply a Sigmoid function to the output logits to obtain a probability score between 0 and 1 for each of the 3,425 species. We then apply a probability threshold of 0.18. Any species with a score exceeding this threshold is classified as present. This threshold was carefully tuned on a validation set to balance precision and recall. To ensure a minimum number of predictions for each observation and avoid generating overly sparse results, we implemented a fallback mechanism. If the number of species predicted using the 0.18 threshold is fewer than 14, we discard those predictions and instead select the 14 species with the highest probability scores for that specific observation. This fallback value was chosen as it approximates the median number of species per plot in the training data, providing a data-driven baseline that prevents overly conservative predictions while being robust to outliers.
To tune hyperparameters and validate our architectural choices, we partitioned the oficial training set into a training subset (80%) and a validation subset (20%). The training subset was
used for model optimization, while the validation subset provided an unbiased estimate of performance for model selection. For comparing diferent model versions during this development phase, we benchmarked performance using the top-25 most probable species for each observation. This standardized metric allowed for a consistent and direct comparison between models, removing the potential bias introduced by our custom probability thresholding strategy (described in Section 4.2). The detailed settings of training are shown in Table 1.
The dataset exhibits a strong long-tailed distribution, where most species have far fewer presence records than absence records. We explored several mitigation techniques. While methods like applying a pos_weight to the Binary Cross-Entropy (BCE) loss were tested, they did not yield significant improvements. Our most successful strategy was a multi-faceted approach combining data curation and regularization. We constrained the training problem by focusing on species with an occurrence count greater than 5. This reduced the number of classes from 5,016 to 3,425, allowing the model to learn more robust features for species with a reasonable number of examples. Table 3 shows the impact of species filtering on model performance. Our second strategy involved the hybrid inference approach detailed in Section 4.2: we applied a tuned probability threshold and used a top-14 fallback for observations with sparse predictions. The performance of diferent thresholding strategies is presented in Table 4.
A key finding from our ablation studies was the superior performance of the Swin Transformer
[
A key strategy for improving overall score was diferent model regularization. To prevent
overfitting and improve generalization, we integrated standard image augmentations (such as
rotation and random brightness/contrast adjustments). We utilized Dropout [
In this paper, we presented the multimodal deep learning framework that secured 6th place in the GeoLifeCLEF 2025 competition. Our approach integrated Sentinel-2 imagery, Landsat timeseries, and bioclimatic data cubes using a network of specialized Swin Transformer encoders. We demonstrated that this architecture is particularly efective for processing complex, rasterized environmental data. The success of our method hinged on three key contributions: the tailored Swin Transformer backbones, a pragmatic data curation strategy that prioritized signal quality over data quantity, and a novel hybrid inference method combining a tuned probability threshold with a top-18 fallback.
To support reproducibility and encourage further research, the complete source code for our solution is publicly available on Kaggle3. Looking ahead, several avenues for enhancement exist. While our model achieves strong performance, a more exhaustive hyperparameter optimization could yield further gains. The most significant opportunity for improvement, however, lies in data enrichment. A key future direction would be to revisit the integration of the Presence-Only (PO) data, potentially using advanced techniques to correct for sampling bias, which could substantially improve the model’s generalization across rare species. Finally, incorporating additional environmental data layers and exploring more sophisticated data augmentation methods remain promising areas for future research.
During the preparation of this work, the author used DeepSeek to check grammar and spelling, as well as to paraphrase and reword the text. After using this tool/service, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content.