How Computers Automatically Spot Disaster Damage Using Different Types of Images

A method for comparing two different types of images—like satellite photos and drone footage—to automatically identify areas damaged by natural disasters.

What it covers

This technology compares 'before' and 'after' images of a disaster zone, even when those images come from different sources like optical cameras and multispectral sensors. It first identifies specific objects in both images—such as buildings or vegetation—based on their color or light reflectance. It then filters out 'exceptional objects' that aren't relevant to the disaster, like moving cars or shadows, to ensure they don't trigger false alarms. Finally, it uses mathematical techniques like Change Vector Analysis and PCA K-means to highlight exactly where the landscape has changed, pinpointing the damage.

What it doesn't cover

• —Does not cover analysis of homogeneous image data (e.g., comparing two identical camera feeds).
• —Does not cover manual human-in-the-loop damage assessment without the automated filtering of exceptional objects.
• —Does not cover real-time streaming video analysis for non-disaster events.

The clever bit

The system doesn't just overlay images; it intelligently filters out 'exceptional objects' based on the disaster type before comparing them, preventing common false positives caused by temporary items like parked cars or changing light conditions.

Why it matters

In disaster relief, time is critical. Manual inspection of satellite or drone footage is slow and prone to error. By automating the comparison of heterogeneous data—meaning data from different sensors or platforms—this system helps emergency responders identify hard-hit areas faster, allowing for more efficient deployment of resources.

Real-world examples

• 1.Automated flood damage assessment using satellite imagery compared to drone photos.
• 2.Post-earthquake structural damage detection in urban environments.
• 3.Wildfire progression monitoring using multispectral sensor data.