Abstract
Biodiversity is declining at an unprecedented rate, impacting ecosystem services necessary to ensure food, water, and human health and well-being. Understanding the distribution of species and their habitats is crucial for conservation policy planning. However, traditional methods in ecology for species distribution models (SDMs) generally focus either on narrow sets of species or narrow geographical areas and there remain significant knowledge gaps about the distribution of species. A major reason for this is the limited availability of data traditionally used, due to the prohibitive amount of effort and expertise required for traditional field monitoring. The wide availability of remote sensing data and the growing adoption of citizen science tools to collect species observations data at low cost offer an opportunity for improving biodiversity monitoring and enabling the modelling of complex ecosystems. We introduce a novel task for mapping bird species to their habitats by predicting species encounter rates from satellite images, and present SatBird, a satellite dataset of locations in the USA with labels derived from presence-absence observation data from the citizen science database eBird, considering summer (breeding) and winter seasons. We also provide a dataset in Kenya representing low-data regimes. We additionally provide environmental data and species range maps for each location. We benchmark a set of baselines on our dataset, including SOTA models for remote sensing tasks. SatBird opens up possibilities for scalably modelling properties of ecosystems worldwide.
Introduction
Biodiversity is currently in a state of crisis, yet our ability to monitor it effectively is hampered by a massive "knowledge gap." Traditional Species Distribution Models (SDMs) rely on labor-intensive field surveys that are difficult to scale geographically or across thousands of species simultaneously.
While remote sensing (satellite imagery) offers a "bird's-eye view" of habitats, and citizen science platforms like eBird provide millions of observations, these data streams are rarely integrated effectively for joint species modeling. SatBird seeks to bridge this gap by providing a high-resolution, multi-modal dataset that leverages "complete checklists"—reports where observers list every species they identified—to provide reliable presence-absence labels for hundreds of species at once.
Methods
The SatBird framework treats biodiversity monitoring as a supervised multi-output regression task. The goal is to predict a vector of "encounter rates"—the probability that a visitor will see a specific species at a given location.
The model architecture fuses high-resolution (10m) Sentinel-2 satellite imagery (RGB and Near-Infrared) with low-resolution environmental data, including 19 bioclimatic variables (temperature, precipitation) and 8 pedologic (soil) variables. To ensure ecological realism, the authors incorporate species range maps as binary masks during inference, preventing the model from predicting species in regions where they are known to be absent. To address spatial autocorrelation—where nearby training and test points look nearly identical—the researchers used DBSCAN clustering to ensure test locations are geographically distant from training locations.
Datasets
The SatBird dataset includes three primary components designed to capture diverse ecological contexts and data regimes:
- SatBird-USA-Summer: Approximately 122,000 hotspots and 670 species, focused on the breeding season.
- SatBird-USA-Winter: Approximately 53,000 hotspots and 670 species, focused on non-breeding distributions.
- SatBird-Kenya: Approximately 10,000 hotspots and 1,054 species, serving as a benchmark for "low-data regimes" where citizen science is less prevalent.
Labels are derived from over 80 million eBird records. By specifically filtering for "complete checklists," the dataset avoids the common pitfalls of "presence-only" data, allowing for much more robust statistical modeling of where birds are—and where they are not.
Key Results and Insights
Benchmarks using architectures like ResNets and Vision Transformers (e.g., Satlas, SatMAE) revealed several critical insights:
- Synergy of Pixels and Climate: Satellite imagery and bioclimatic data are highly complementary. Imagery captures local habitat structure (like forest fragmentation), while bioclimatic data captures the broader environmental requirements for survival. In the USA Summer dataset, a combined ResNet-18 achieved a Top-30 accuracy of 65.5%, significantly beating environmental-only (57.8%) and image-only (52.2%) baselines.
- Joint Modeling Benefits: Predicting 600+ species simultaneously allows the model to learn underlying correlations between species that share the same habitat, effectively regularizing the learning process.
- Transferability: The Kenya dataset demonstrates that models trained on data-rich regions like the USA can be fine-tuned to work in data-poor regions, a critical requirement for global conservation efforts.
Conclusion and Future Work
SatBird demonstrates that fusing remote sensing with citizen science data offers a scalable path for global biodiversity monitoring. While "complete checklists" provide high-quality ground truth, future work aims to incorporate observer expertise scores to further refine labels. Additionally, the authors plan to expand the framework to include multi-temporal imagery and other taxa, such as plants and insects, to create a truly holistic model of global ecosystems.
Summary updated