🌿 Deep Learning for Rare Species Classification

Master’s in Data Science 2024/2025 – Group Project

This repository contains the source code, data processing pipeline, and model training workflow for our deep learning project on rare species image classification. The goal is to predict the biological family of a given organism based on its image.

The project is based on the BioCLIP dataset, introduced in Stevens et al., 2024, which provides rich metadata and high-resolution imagery sourced from the Encyclopedia of Life (EOL).

---

📁 Project Structure

📦 project-root/
├── data/
│   ├── raw/              # Original images and metadata.csv
│   ├── processed/        # Train/Val/Test splits
├── notebooks/            # Jupyter notebooks for EDA, training, evaluation
├── models/               # Saved model checkpoints
├── src/                  # Core modules (dataloaders, training loops, utils)
├── outputs/              # Evaluation results, confusion matrices, etc.
├── report/               # Scientific report (PDF or LaTeX)
├── requirements.txt      # Python dependencies
└── README.md             # This file

---

📊 Objective

Train a model to classify the family of an organism from its image using deep learning. Your model must generalize well to unseen data, so the test set should remain untouched during training and validation.

---

🔍 Dataset

• Source: Encyclopedia of Life (EOL)
• Metadata: Includes image paths, kingdom, phylum, and family labels
• Challenge: Highly imbalanced classes, inter-class visual similarity, fine-grained classification

---

🧠 Approach

We experimented with various techniques, including:

• Image preprocessing and augmentations
• Baseline CNNs and transfer learning with pre-trained models (e.g., ResNet, EfficientNet, ViT)
• Custom training and evaluation pipelines
• Hyperparameter tuning with Keras Tuner

For more details, refer to the report folder.

---

📈 Evaluation

Our evaluation strategy includes:

• Accuracy and top-k accuracy
• Confusion matrices
• Error analysis (e.g., misclassified species by visual similarity)
• Visual inspection of predictions

---

🛠️ Setup & Dependencies

Requirements

pip install -r requirements.txt

Python Libraries Used

• tensorflow / keras
• pandas
• numpy
• matplotlib / seaborn
• scikit-learn
• opencv-python

---

📄 Deliverables

• ✅ Source code and data processing pipeline
• ✅ Trained model and evaluation scripts
• ✅ 5-page scientific report (in English)
• ✅ Metadata and group info files (as per Moodle submission guidelines)

---

👥 Team

Replace the placeholder with your actual names and student IDs.

• Benedikt Ruggaber – 20240500
• Daan Van Holten – 2024002
• Jose Cavaco – 20240513
• Philippe Duntranoit – 2024004
• Joshua Wehr – 20240501

---

🧪 Future Work

Given more time, we would explore:

• Fine-tuning foundation models like BioCLIP or CLIP directly
• Multi-label classification for hierarchical taxonomy (e.g., kingdom → phylum → family)
• Integration of textual metadata as multimodal input
• Class balancing strategies (e.g., focal loss, SMOTE for embeddings)

---

⚠️ Academic Integrity

The collaboration between groups and the use of AI-generated code is strictly forbidden as per course policy.
Violation of this rule will result in immediate disqualification from the curricular unit in both epochs.