Bird Species Distribution Modeling: Predicting Habitat Shifts with AI

Overview

This research project explores how climate change is reshaping bird migration patterns and habitats. By integrating satellite imagery and climate variables, we propose a hybrid AI pipeline to predict bird species presence across varying environments.

We leverage:

• Convolutional Neural Networks (CNNs) implemented from scratch and in PyTorch for spatial feature extraction from satellite images.
• Tabular classifiers (Random Forest, Gradient Boosting) using our own implementation and scikit-learn, combined with our own custom preprocessing and feature engineering pipeline.
• Real-world ecological data, including satellite data and bird occurrence labels from the eBird database.

Approach

We designed two parallel models:

1. CNN-based Vision Model (custom implementation compared with PyTorch): Uses satellite images to identify features like forestation, water bodies, and urbanization that influence bird migration patterns using ResNet models.

2. Tabular Feature Model (custom implementation compared with scikit-learn): Uses structured environmental variables (e.g., temperature, elevation, latitude/longitude) to predict presence/absence using Random Forests and Gradient Boosting classifiers.

These models were evaluated individually and comparatively to study their effectiveness in modeling species distribution.

Key Contributions

• A CNN trained on landscape imagery for species presence classification
• A tabular model using ecological features for probabilistic bird occurrence prediction
• A comparative evaluation showing strong predictive performance (~85% average accuracy)

Repository Structure

The research project features the following components:

• notebooks: Directory containing the image classification and tabular models
• images: Directory containing images of species distribution on a map as well as feature importance
• finalPaper.pdf: File containing the write up of our methods

Technologies Used

• Python 3
• PyTorch
• scikit-learn
• NumPy, Pandas, Matplotlib
• Sentinel, eBird

Reproducibility

In order to use the repository you can clone it then request an eBird API key as that’s our primary data collection method. Set the API key as EBIRD_API_TOKEN then you can proceed to run both the CNN and the tabular model.