TL;DR
This paper introduces Tighnari v2, a novel framework for predicting plant species distributions that tackles the challenges of noisy and imbalanced datasets. By employing a "mixture of experts" approach and a new weakly supervised learning strategy, it achieves state-of-the-art performance, particularly in scenarios with significant geographic distribution shifts between training and testing data.
What Problem Does This Paper Address?
Predicting where plant species are likely to occur is crucial for biodiversity conservation and understanding ecological changes. However, this task is hampered by the nature of available data. Presence-Absence (PA) data, which clearly labels both present and absent species, is accurate but scarce and expensive to collect. In contrast, Presence-Only (PO) data, often gathered from citizen scientists, is abundant and covers vast areas, but suffers from significant "label noise" – meaning many "absent" records might actually be present species that weren't noticed or recorded. Furthermore, models trained on data from one geographic region often perform poorly when applied to another due to "distribution shifts." This paper aims to build a robust model that can effectively utilize both PA and PO data, mitigate label noise, and generalize well across different geographic regions.
Methods
The core of Tighnari v2 is a multimodal fusion framework that integrates various data sources: satellite imagery, tabular data (like environmental factors), and time-series data.
- Weakly Supervised Pseudo-Labeling: To leverage the abundance of PO data, the authors developed a novel strategy. They aggregate PO sample labels within the geographic coverage of satellite image patches. This effectively turns groups of PO observations into more reliable "pseudo-labels" that are geographically aligned with the satellite imagery, reducing label sparsity and minimizing negative label noise without introducing positive label noise.
- Mixture of Experts (MoE): To address distribution shifts, Tighnari v2 employs an MoE architecture. Test samples are partitioned based on their geographic proximity to PA training samples. Different "expert" models, trained on distinct datasets (e.g., in-distribution vs. out-of-distribution), are then used for inference on these partitions. This allows the model to specialize for different geographic scenarios.
- Enhanced Modality Fusion: The paper introduces a stackable serial tri-modal cross-attention mechanism. This design allows modalities to attend to each other sequentially, enabling the model to prioritize dominant modalities and capture higher-order dependencies, leading to better integration of diverse data types.
- Backbone Networks: The model utilizes Swin Transformer Base for satellite imagery, TabM network for tabular data, and Temporal Swin-Transformer for time-series data, building upon previous work (Tighnari).
- Asymmetric Loss (ASL): To handle the imbalance between positive and negative classes and the inherent noise in PO data, ASL is used instead of standard Binary Cross-Entropy. ASL offers asymmetric focusing parameters and negative sample clipping to improve robustness and learning from noisy data.
Datasets
The primary dataset used is the GeoLifeCLEF 2025 dataset, which includes:
- PA (Presence-Absence) training samples.
- PO (Presence-Only) training samples.
- PA test samples.
The paper also references the GeoLifeCLEF 2024 dataset for comparison and validation.
Key Insights and Ideas
- PO data is noisy but valuable: The paper clearly demonstrates that raw PO data contains significant label noise, but its sheer volume and spatial coverage make it indispensable for comprehensive species distribution modeling.
- Geographic distribution shift is a major challenge: The EDA reveals substantial differences between PA training and test sample locations, highlighting the critical need for models that can generalize beyond their training distribution.
- Mixture of Experts is key for OOD generalization: The MoE approach, by partitioning test data and using specialized models, proves highly effective in handling out-of-distribution samples, significantly boosting performance on the 2025 GeoLifeCLEF challenge.
- Serial cross-attention improves fusion: The novel tri-modal cross-attention mechanism allows for a more nuanced integration of different data modalities, outperforming simpler concatenation or parallel attention methods.
Key Results
- Tighnari v2 achieved third place in the GeoLifeCLEF 2025 challenge, which featured significant out-of-distribution characteristics in its test set.
- The model surpassed the second-place score on the GeoLifeCLEF 2024 leaderboard, where training and test distributions were more aligned.
- Ablation studies showed that the MoE model significantly improved performance on OOD datasets, particularly on the 2025 GeoLifeCLEF test set, compared to models trained only on PA data or a combined PA+PO dataset. For instance, the MoE model achieved a public score of 0.24493 on the 2025 GeoLifeCLEF, compared to 0.21860 for PA+PO and 0.20604 for PA Only.
Limitations and Future Work
- Pseudo-label limitations: The current pseudo-labeling strategy relies on prior rules, meaning the number of pseudo-labeled species might be less than the actual number. The authors acknowledge the lack of a semi-supervised process for iterative refinement.
- Class-dependent label noise: PO data's label noise is class-dependent, and the current work doesn't explicitly model or mitigate this.
- Overly generalized feature extraction: The multi-label nature of the task with a vast number of species leads to overly generalized feature extraction, negatively impacting performance.
- Long-tailed distribution: The long-tailed distribution of species (many appearing rarely) makes the model prone to overfitting on common species.
Future work directions include: * Allowing the model to predict an appropriate K (number of species to consider) for each sample based on tabular features, rather than using a fixed K. * Developing methods to anchor positive labels as immutable while allowing negative labels to transition to positives with a certain probability, potentially using a transition matrix to down-weight noisy negative labels. * Exploiting inter-label correlations by constructing a graph structure and potentially using Graph Neural Networks (GNNs) to refine predictions.
Summary updated