BioAnalyst: A Foundation Model for Biodiversity
Athanasios Trantas, Martino Mensio, Stylianos Stasinos, Sebastian Gribincea, Taimur Khan, Damian Podareanu, Aliene van der Veen
TL;DR
BioAnalyst is a multimodal foundation model designed for biodiversity analysis and conservation planning, capable of handling diverse ecological data for regional to national-scale applications in Europe.
Contribution
It is the first multimodal foundation model tailored specifically for biodiversity analysis, integrating species data with remote sensing and environmental variables.
Findings
Provides a strong baseline for biotic and abiotic tasks.
Acts as a macroecological simulator with yearly and monthly resolutions.
Open-sourced for ecological research advancement.
Abstract
Multimodal Foundation Models (FMs) offer a path to learn general-purpose representations from heterogeneous ecological data, easily transferable to downstream tasks. However, practical biodiversity modelling remains fragmented; separate pipelines and models are built for each dataset and objective, which limits reuse across regions and taxa. In response, we present BioAnalyst, to our knowledge the first multimodal Foundation Model tailored to biodiversity analysis and conservation planning in Europe at spatial resolution targeting regional to national-scale applications. BioAnalyst employs a transformer-based architecture, pre-trained on extensive multimodal datasets that align species occurrence records with remote sensing indicators, climate and environmental variables. Post pre-training, the model is adapted via lightweight roll-out fine-tuning to a range of downstream…
Peer Reviews
Decision·Submitted to ICLR 2026
- integrates diverse data types (remote sensing, climate, soils, species, etc.). - code is available and opensourced - comparison with microsoft’s aurora gives an initial external reference point. - generally well-written with detailed implementation details.
- Only two downstream tasks are shown (species forecasting; climate probing), with limited baselines (Aurora only). - Can include task-specific baselines. - why a swin transformer? please justify the backbone choice. - How are the encoders-decoders trained? Together with the whole model or are they pretrained beforehand? - There is no uncertainty quantifcation discussed in the work. It would be good to have some experiments around that. - Has the model been trained to pr
- Timely and societally valuable problem. Bringing foundation-model methodology to biodiversity and conservation is meaningful and impactful for environmental science. - Ambitious multimodal integration. The model aligns diverse ecological variables within a single latent representation, a nontrivial data-engineering achievement. - Transparent and reproducible. Code, model weights, and detailed appendices are provided. - Clear writing and motivation. The introduction and related-work sections
- Engineering integration rather than scientific novelty. The architecture is an adaptation of existing components (Perceiver IO, Swin Transformer) with minor modifications and a temporal-difference loss. While this is solid engineering, it lacks new methodological ideas or theoretical insights expected at ICLR. - Over-reliance on outperforming Aurora. The only strong empirical claim is that BioAnalyst “outperforms” Aurora on a few metrics. However, Aurora is primarily a climate model, not a bi
- The motivation is clear: there is currently no foundation model that has been pretrained on such a variety of variables relevant to biodiversity-related tasks.
1. I find the evaluation to be far from enough to drive the message of the paper. With only two tasks, one of them not directly related to biodiversity, it is hard to convince the readers that the proposed model actually serves as a foundation model. In addition, only one competing method is shown: Aurora, which is intended as a climate foundation model and is thus not a great fit for the species distribution task (I couldn’t find a comparison for the climate recovery task, where Aurora would pr
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.