🌪️ Post-Hurricane Damage Detection with Deep Learning

A machine learning-powered solution for rapid post-disaster assessment using aerial imagery. This project leverages deep learning models—ResNet50, CNN, and EfficientNet—to classify structural damage caused by hurricanes. Designed for real-time predictions via a Streamlit web interface.

---

👩‍💻 Team Members

• Srinivas Saiteja Tenneti
• Namratha Prakash
• Lakshmi Sreya Rapolu

---

📌 Project Overview

Hurricanes in the U.S. cause an average of $21.5 billion in damage per event, with over 10 billion-dollar storms annually between 2015 and 2020. Accurate and rapid post-hurricane damage assessment is essential for emergency response, insurance processing, and recovery planning.

This project builds an AI system that:

• Detects structural damage from aerial images post-hurricane
• Utilizes transfer learning with ResNet50 and EfficientNet models
• Deploys a Streamlit web app for interactive image uploads and predictions

---

🛰️ Dataset

• Source: University of Washington Disaster Data Science Lab
• Location: Houston, TX (Post Hurricane Harvey)
• Images: 14,000 (7,000 damaged, 7,000 undamaged)
• Splits:
  • Train: 8,000
  • Validation: 2,000
  • Test: 2,000 (also tested on unbalanced and balanced subsets)

---

🛠️ Key Techniques Used

• Data Normalization & Augmentation (RandomHorizontalFlip)
• PCA for feature reduction
• Custom & pre-trained models
• Evaluation metrics: Accuracy, Confusion Matrix, F1-score
• Streamlit-based real-time interface for multi-image upload

---

📊 Exploratory Data Analysis (EDA)

To build a robust hurricane damage classification model, we performed in-depth exploratory data analysis (EDA) to uncover key visual and statistical signals differentiating damaged from undamaged structures.

---

🔍 Challenges in Damage Detection

• No Single Definition of "Damage": Can include debris, discoloration, roof collapse, or minor structural shifts.
• False Visual Triggers: Materials scattered for other reasons can look like damage.
• Intra-Class Variability: Buildings in the same class vary greatly in size, shape, and appearance.
• AI Ambiguity: Damage is often subtle or context-dependent, making detection by machines inherently challenging.

---

🖼️ Visual Inspection of Image Data

• Format: RGB, 128×128 pixels
• Content: Aerial view of rooftops and structures post-hurricane

Observations:

• 🌊 Flood patterns with unique texture and tone
• 🧱 Scattered debris and damaged rooftops
• 👀 Human-eye struggle: Subtle patterns not always easily visible

---

📈 Class Trends: Pixel-Level Averages

We computed mean grayscale intensity across images in each class:

Damage	No Damage
Brighter cores with dark surroundings	More uniform brightness across the image
Suggests collapsed or open roof areas	Indicates intact, cleaner structural surfaces

---

🌡️ Standard Deviation: Visual Variability

Pixel-wise standard deviation helps visualize variability:

Damage	No Damage
Lower variation across the image	Higher variation near core structure
Uniformity due to debris/flooding	Variation from visible rooftops and shadows

---

🧬 PCA Analysis: Dimensional Reduction

Principal Component Analysis (PCA) was used to extract key visual patterns:

Class	Components to explain 70% variance
Damage	19
No Damage	56

Figure 1: With damage: 19 principle components

Figure 2: No damage: 56 principle components

🧠 Insight: Damaged images have more visual consistency, making them easier for models to learn from.

---

🧠 Raw Pixel Intensity Distribution

We compared the mean intensity of the first 1,000 pixels across classes:

• Damage: Lower, noisier intensity — possibly due to shadows and debris
• No Damage: Higher and smoother intensity — cleaner rooftops

---

🌍 Geospatial Distribution of Images

• Aerial tiles are spread across Houston, Beaumont, and Victoria (Texas)
• Damage and no-damage classes cluster by location

⚠️ Risk: Models might pick up on geographic cues instead of real damage features

---

✅ Key Takeaways

• Strong statistical signals in pixel-level data
• Class imbalance handled
• Geographic clustering introduces potential bias
• PCA & intensity trends support model learning

---

🧠 Model Highlights

🏗️ Custom CNN (from scratch)

• Input: 128x128 RGB images
• Architecture: 3 Conv Layers + 4 FC Layers
• Accuracy:
  • ✅ Train: 99.48%
  • ✅ Validation: 96.25%

---

🦾 ResNet50 (Transfer Learning)

• Input: 224x224, ImageNet normalized
• Accuracy:
  • ✅ Validation: 99.50%
  • ✅ Test Set: 99.61%
• 🧠 Best model for generalization and deployment

---

🌱 EfficientNet Models

EfficientNet-B0:

• Accuracy: 99.30%
• Lightweight and fast but slightly underperformed vs. ResNet50

EfficientNet-V2-S:

• Frozen: 91.7% accuracy — very fast but limited learning
• Fine-tuned (Last 2 Blocks): 97.95% — efficient and effective

---

💻 Streamlit Web App

Interactive interface for uploading and classifying images.

Features:

• Multi-image upload with grid view
• Class predictions (damage / no damage)
• Confidence scores with visual indicators
• Session-wise prediction history
• Optional visualization of transformed model input
• Lightweight and runs locally or on any Streamlit-compatible server

---

📈 Sample Result Snapshot

• ✅ ResNet50 Confusion Matrix
  • True Positives: 7,980
  • False Negatives: 20
  • True Negatives: 985
  • False Positives: 15
  • Accuracy: 99.61%

---

🚀 Future Work

• Multi-class damage levels (minor/moderate/severe)
• Integrate Grad-CAM for visual attention maps
• Expand to detect other disaster types: fire, floods, earthquakes
• Incorporate geospatial overlays using GIS libraries

---

🔗 References

• Dataset on IEEE Dataport
• University of Washington Disaster Data Science Lab

---

“In the aftermath of a hurricane, every second counts. With AI-driven tools, response teams can act faster and smarter.” – Group 8

📬 Contact

For questions, contact any team member via this repository's issue tracker.