🐦 Bird Trajectory Prediction with RNN/LSTM

🎯 Project Goal

This project aims to predict the future trajectories of migrating birds based on their historical GPS tracking data. By modeling bird movements, we can gain insights into migration patterns, habitat use, and potential conservation needs. Such predictive tools could also support ecological studies and wildlife protection by anticipating where birds are likely to travel.

🔬 Project Overview

The project implements a sequence modeling approach using Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTMs). Participants will start with an RNN model and then extend it to LSTM to improve accuracy. This hands-on project is designed as a learning exercise in time-series prediction and deep learning with PyTorch.

The notebook provided includes both a complete reference implementation and a challenge version with stripped/modified sections for participants to solve.

📊 Dataset

The dataset contains GPS tracking data for three migrating birds. Each record consists of features such as:

• Latitude
• Longitude
• Time information (timestamps)
• Derived features (e.g., speed, direction -- to be engineered by participants)

The task is to predict the next location(s) in the sequence based on past movement data.

🏗️ Project Structure

The main components of the project are:

1. Data exploration & visualization -- plot bird trajectories with Cartopy
2. Feature processing -- create and preprocess features for model input
3. Data preparation -- train/test split and sequence generation
4. Modeling -- RNN baseline, followed by LSTM for improvement
5. Training & validation -- loss curves, early stopping, hyperparameter tuning
6. Evaluation -- accuracy and trajectory prediction performance

🛠️ Setup and Dependencies

The project requires the following main libraries:

• PyTorch
• NumPy
• Pandas
• Matplotlib
• Cartopy
• Scikit-learn

Additional dependencies can be found in the notebook import statements.

🚂 Training

• The baseline RNN model is trained using sequence data from bird trajectories.
• Training includes optimization with gradient descent and early stopping.
• Participants will extend the training by experimenting with hyperparameters (learning rate, hidden units, sequence length, etc.).

📊 Evaluation

The models are evaluated based on:

• Trajectory prediction accuracy (threshold: ≥ 70% as a benchmark)
• Loss and convergence behavior across epochs
• Visual inspection of predicted vs. actual trajectories

👁️ Visualization

The project includes visualizations such as:

• Bird migration paths (using Cartopy for geographic context)
• Training and validation loss curves
• Comparison of predicted vs. actual trajectories

🧠 Challenges for Participants

This project contains several open issues for participants to solve, ranging from easy to hard:

• Easy: Data visualization, feature preprocessing, train/test split
• Medium: Early stopping, plotting loss curves, hyperparameter tuning, RNN initialization
• Hard: Replace RNN with LSTM to achieve improved prediction accuracy [Improvement over baseline <-20%]

By completing these challenges, participants will strengthen their skills in time-series modeling, RNN/LSTM networks, and PyTorch.

📍 Notes

This notebook is designed to be run on Google Colab for easy GPU access. Use the GitHub--Colab integration (File -> Save a copy in GitHub) to save your progress. If you have a powerful local GPU (e.g., Tesla T4 or better), you may also run the project locally.

Good luck, and happy modeling! 🚀🐦