Coral Patch Classification using Deep Learning

Overview

This project presents a deep learning pipeline for coral reef benthic classification using annotation-guided patch extraction and transfer learning.

Instead of classifying entire underwater reef images containing multiple coral categories, the dataset was transformed into single-label image patches using expert-provided annotation coordinates. This significantly improved label quality and enabled effective supervised learning.

A ResNet18 model was trained on a balanced dataset of 60,000 coral image patches and achieved 82.38% validation accuracy across six benthic classes.

---

Dataset

Original Dataset

• 418,310 coral annotations
• 2,455 annotated reef images
• Image Resolution: approximately 1116 × 906 pixels
• Multiple coral categories present within each image

Selected Classes

The six most representative benthic classes were selected:

1. Crustose Coralline Algae (CCA)
2. Macroalgae
3. Off
4. Porites
5. Sand
6. Turf

Balanced Patch Dataset

• Patch Size: 224 × 224
• Total Patches: 60,000
• Classes: 6
• Samples per Class: 10,000

---

Methodology

1. Annotation Processing

Coral annotation coordinates were extracted from the provided CSV annotations.

2. Patch Extraction

For every annotation:

• Annotation-centered crop
• Patch size: 224 × 224
• Label inherited from annotation class

3. Dataset Balancing

An equal number of samples were selected for all six classes to prevent class imbalance.

4. Data Augmentation

• Horizontal Flip
• Vertical Flip
• Rotation
• Color Jitter
• ImageNet Normalization

5. Model Training

Model:

• ResNet18 (Transfer Learning) Training Strategy:
• Phase 1: Train Classification Head
• Phase 2: Fine Tune Entire Network

---

Results

Overall Performance

Metric	Value
Validation Accuracy	82.38%
Macro F1 Score	81.79%
Classes	6
Dataset Size	60,000 Patches

Training Curves

Training vs Validation Loss

Training vs Validation Accuracy

Per-Class Accuracy

Class	Accuracy
Crustose Coralline Algae	54.90%
Macroalgae	87.75%
Off	93.65%
Porites	92.20%
Sand	93.70%
Turf	72.10%

---

Visualizations

Sample Coral Patches

Confusion Matrix

Per-Class Accuracy

Project Structure

Coral-Patch-Classification/
│
├── notebooks/
├── results/
├── images/
├── data_info/
└── README.md

---

Technologies Used

• Python
• PyTorch
• OpenCV
• NumPy
• Pandas
• Matplotlib
• Scikit-learn
• Jupyter Notebook

---

Future Work

• EfficientNet Comparison
• Coral Segmentation Models
• Attention-Based CNN Architectures
• Calibration Artifact Removal
• Larger Multi-Class Coral Classification