ALPAI Project: Semi-Supervised Learning for Post-Earthquake Image Classification

🔍 Objective

This project focuses on automating the classification of post-earthquake building images using semi-supervised learning (SSL). The aim is to reduce dependence on expensive and time-consuming manual labeling while maintaining high classification accuracy.

---

📂 Dataset

• Source: PEER-Hub ImageNet (phi-Net)
• Contains real-world earthquake reconnaissance images.
• Two classification tasks:
  • Task 1 – Scene Level
    • Class 0: Object-level (component like column or wall)
    • Class 1: Pixel-level (surface close-up)
    • Class 2: Structural-level (entire building/multiple buildings)
  • Task 2 – Damage State
    • Class 0: Damaged
    • Class 1: Undamaged

---

🧠 Methodology

• Model: Transfer learning using VGG16 (ImageNet pretrained)
• Baseline: Trained on labeled data only
• SSL Strategy: Pseudo-Labeling
  • Generate pseudo-labels for unlabeled images
  • Retrain model using both labeled and pseudo-labeled data
  • Use a weighted loss:
    [ \mathcal{L} = \mathcal{L}{\text{labeled}} + \alpha(t) \cdot \mathcal{L}{\text{unlabeled}} ]
  • Where α(t) increases over epochs to trust pseudo-labels more gradually

---

📎 License

This project is for educational and research purposes.