Introduction
This research presents a novel computer vision framework designed to predict the encounter rates of 684 bird species across the continental United States. By integrating high-resolution satellite imagery with large-scale citizen science data from eBird, the study frames species distribution as a multi-output regression task. The approach demonstrates that deep learning can effectively map biodiversity at scale by identifying habitat proxies within landscape imagery, significantly outperforming traditional models that rely exclusively on climatic variables.
Problem Statement
Accurate mapping of species distributions is vital for conservation efforts, particularly under the pressure of climate change. Traditional field monitoring is often prohibitively expensive and labor-intensive. While citizen science platforms like eBird offer vast amounts of data, they suffer from spatial biases toward popular birding locations. This work addresses these gaps by using remote sensing to "fill in the blanks," leveraging the landscape as a proxy for the bird communities likely to inhabit specific ecosystems.
Methodology
The researchers treat bird distribution as a supervised multi-output regression problem. The core architecture utilizes a ResNet-18 backbone, modified to process 4-channel input (RGB and Near-Infrared).
To enhance predictive accuracy, the model integrates 19 bioclimatic variables and 8 pedologic (soil) variables as additional raster layers. Geographical priors are incorporated to ensure ecological consistency through three techniques:
- Location Encoders (LE): A neural network that encodes latitude and longitude.
- Hard-masking (RM): Utilizing eBird range maps to zero out geographically impossible predictions.
- Soft-masking (SM): Applying a probabilistic correction based on regional prevalence.
The model is trained using a weighted cross-entropy loss function, which prioritizes "hotspots" with higher checklist counts to ensure the use of reliable ground-truth data.
Dataset and Experimental Setup
The study utilizes a curated dataset focused on the continental US during June (breeding season) to minimize migratory noise.
- eBird: Data from 8,439 hotspots and 684 species (2000–2021).
- Sentinel-2: 10m resolution satellite imagery (RGB and NIR).
- Environmental Data: WorldClim 1.4 bioclimatic data and SoilGrids pedologic data.
- Land Cover: 10m-resolution maps from ESRI.
Unlike previous benchmarks that associate a single species with a location, this dataset requires the simultaneous prediction of the joint distribution of hundreds of species.
Key Results
The optimal model configuration combined RGBNIR imagery, environmental data, and range map masking. This model achieved an adaptive top-k accuracy of 73.38%, meaning that for a site with N observed species, the model's top N predictions matched the ground truth approximately 73% of the time.
This performance significantly exceeded the "Environmental Baseline" of 68.86%, confirming that computer vision provides substantial added value by capturing land-use and habitat fragmentation details that climate data misses. Qualitatively, the model excelled at identifying common and range-restricted species, though it faced challenges with extremely rare species.
Discussion and Limitations
While the results are promising, several limitations remain. The current model is a temporal "snapshot" of June and does not account for seasonal migrations. Additionally, cloud cover in satellite imagery occasionally obscured critical habitat features. Future work aims to scale this approach to data-sparse regions, such as East Africa, and refine the soft-masking mathematical approach, which currently degrades performance by assuming independence between satellite features and location.
Summary updated