Satellite to Street : Disaster Impact Estimator

This paper presents "Satellite to Street," a deep-learning framework utilizing a modified dual-input U-Net with a class-aware weighted loss function to generate accurate, pixel-level damage maps from pre- and post-disaster satellite imagery, thereby enabling faster and more objective assessment of disaster impact severity compared to manual interpretation and conventional models.

The Gist

Imagine a massive natural disaster, like an earthquake or a hurricane, has just hit a city. Emergency teams are scrambling to figure out: Where is the damage worst? Which buildings are safe? Where do we send the rescue trucks first?

Right now, the way we answer these questions is a bit like trying to find a needle in a haystack by looking at the haystack with a magnifying glass, one square inch at a time. Experts sit in front of computers, staring at satellite photos taken before and after the storm, manually drawing lines around damaged buildings. It's slow, it's tiring, and by the time they finish, precious time has been lost.

This paper introduces a new "smart assistant" called Satellite to Street: Disaster Impact Estimator. Think of it as a super-powered, tireless detective that can look at the whole city in seconds and tell you exactly what happened.

Here is how it works, broken down into simple concepts:

1. The "Before and After" Photo Album

Imagine you have two photos of your neighborhood: one taken last week (before the disaster) and one taken today (after the disaster).

• The Old Way: A human looks at both photos and tries to spot the differences.
• The New Way: This AI system is like a detective who has a superpower. It doesn't just look at the photos; it stacks them on top of each other. It looks at the "Before" photo and the "After" photo simultaneously, pixel by pixel. This helps it spot even the tiniest cracks or missing roofs that a human eye might miss.

2. The "Smart Painter" (The AI Model)

The core of this system is a piece of software called a U-Net. You can think of this as a very talented digital painter.

• The Job: Its job is to take the satellite images and paint a "damage map" over them.
• The Colors: Instead of just painting everything red (damaged) or green (safe), this painter is very detailed. It uses a specific color palette to tell a story:
  • Gray: Nothing happened here.
  • Yellow: A little bit of damage (maybe a broken window or a small crack).
  • Orange: Major damage (the roof is gone, or the walls are leaning).
  • Red: Total destruction (the building is a pile of rubble).

3. Solving the "Crowded Room" Problem

One of the biggest problems with teaching computers to do this is that most buildings are not damaged. If you show a computer 100 pictures of buildings and only 5 are broken, the computer gets lazy and just says, "Everything is fine!" because it's easier to guess "fine" than to find the 5 broken ones.

The authors fixed this by giving the computer a special teacher's note (a weighted loss function). It's like a teacher telling the student: "Hey, don't ignore the 5 broken buildings just because there are 95 safe ones. Those 5 are the most important! Focus on them!" This forces the AI to pay extra attention to the disasters, not just the calm areas.

4. Why This Matters (The "Street" Part)

The title says "Satellite to Street." This is the magic part.

• Satellite: The AI looks from high above, seeing the whole city.
• Street: Because the AI is so good at spotting details, it can tell you exactly which street or block is in trouble.

Instead of a general report saying "The city is damaged," this system gives a graded report. It can say, "Block A has minor damage, Block B is destroyed, and Block C is safe." This helps rescue teams prioritize. They don't waste time checking safe blocks, and they don't ignore the ones that are completely gone.

5. The Results

The team tested this system on real data from past disasters.

• The Score: It was much better than older methods at finding the "orange" (major) and "red" (destroyed) zones.
• The Speed: It does in seconds what takes humans hours or days.
• The Accuracy: It can tell the difference between a building that just lost a roof and one that has collapsed entirely.

The Bottom Line

This project isn't trying to replace the brave human rescue workers or the experts. Instead, it's like giving them a super-powered flashlight. It shines a light on the most critical areas instantly, so the humans can focus their energy on saving lives rather than searching for where the damage is.

By turning complex satellite data into a simple, color-coded map, this system helps governments and charities make faster, smarter decisions when every second counts.

Technical Summary

1. Problem Statement

Natural disasters (earthquakes, floods, hurricanes, wildfires) cause widespread structural damage, necessitating rapid assessment to prioritize emergency response and resource allocation. Current assessment methods rely heavily on manual interpretation of satellite imagery, which is:

• Time-consuming and labor-intensive: Slowing down critical rescue operations.
• Subjective: Prone to human error and inconsistency.
• Scalability issues: Difficult to apply over large areas during major events.

While existing deep learning models (semantic segmentation and change detection) have automated parts of this process, they suffer from specific limitations:

• Class Imbalance: Undamaged buildings vastly outnumber damaged ones, causing models to ignore rare damage classes.
• Lack of Granularity: Most models only classify "damaged" vs. "undamaged," failing to distinguish severity levels (e.g., minor vs. destroyed).
• Subtle Feature Detection: They often struggle to capture fine-grained structural variations or subtle changes between pre- and post-disaster scenes.

2. Methodology

The proposed system is an end-to-end deep learning framework designed to generate pixel-level damage maps by analyzing paired pre- and post-disaster satellite images.

A. Data Preparation

• Dataset: Utilized the xView2 dataset, containing high-resolution (1024×1024) pre- and post-disaster satellite images with JSON annotations.
• Classes: The system targets five specific categories: No Damage, Minor Damage, Major Damage, Destroyed, and Unclassified.
• Preprocessing:
  • Images resized to 256×256 pixels using LANCZOS interpolation to preserve structural details while optimizing memory.
  • Normalization of pixel intensities to the range [0, 1].
  • Conversion of polygon annotations into pixel-level segmentation masks (single-channel images) where each integer value represents a damage class.

B. Model Architecture: Modified Dual-Input U-Net

The core of the system is a Dual-Input U-Net architecture, specifically designed to process two images simultaneously.

• Input Mechanism: The model accepts a 6-channel input (3 channels for pre-disaster RGB + 3 channels for post-disaster RGB), allowing the network to learn "change-aware" features directly.
• Encoder: Uses a sequence of DoubleConv blocks (two 3×3 convolutions followed by BatchNorm and ReLU) with feature sizes [64, 128, 256, 512].
  • Note: The paper mentions the use of an SE-ResNeXt50 encoder in the results section, which incorporates Squeeze-and-Excitation (SE) attention mechanisms and grouped convolutions to enhance feature representation.
• Decoder: Mirrors the encoder using ConvTranspose2d for upsampling.
• Skip Connections: Directly pass encoder feature maps to the decoder to preserve spatial details (edges, small buildings) essential for precise segmentation.
• Output: A 1×1 convolution maps features to the number of output classes (damage severity levels).

C. Training Strategy & Loss Function

• Class-Aware Weighted Loss: To address the severe imbalance between "undamaged" and "damaged" classes, a weighted loss function is employed. This forces the model to pay more attention to rare but critical classes (Major Damage, Destroyed).
• Data Augmentation: Includes horizontal flipping, brightness adjustments, and resizing to improve generalization.
• Metrics: Evaluation focuses on Pixel Accuracy, Intersection over Union (IoU), and Dice Score (specifically per-class to avoid masking poor performance on rare classes).

3. Key Contributions

1. Fine-Grained Severity Classification: Unlike binary models, this framework distinguishes between four levels of damage (No, Minor, Major, Destroyed), providing actionable intelligence for emergency responders.
2. Dual-Input Feature Fusion: By stacking pre- and post-disaster images as a 6-channel input, the model learns contextual differences more effectively than single-image or simple subtraction methods.
3. Addressing Class Imbalance: The implementation of a class-aware weighted loss function significantly improves the detection of rare, high-severity damage categories.
4. End-to-End Pipeline: The project integrates data preprocessing, model training, and web deployment, demonstrating a practical, scalable solution for real-world disaster response.

4. Results

The system was evaluated on the xView2 benchmark with a 70/30 train-test split.

• Quantitative Performance:
  • Overall IoU: 0.74
  • Overall Dice Score: 0.81
  • Class-wise Performance:
    • No Damage: IoU 0.82, Dice 0.89
    • Destroyed: IoU 0.75, Dice 0.83
    • Major Damage: IoU 0.58, Dice 0.72 (The most challenging class, yet the model achieved reasonable performance).
• Architecture Comparison: The SE-ResNeXt50 U-Net outperformed the standard ResNet-50 U-Net (0.74 mIoU vs. 0.69 mIoU), validating the benefit of attention mechanisms in capturing fine structural changes.
• Qualitative Analysis: The model successfully identified collapsed structures, roof displacements, and debris clusters that baseline models missed. The color-coded masks provided intuitive visualization of damage severity.
• Limitations: Performance degrades in cases of heavy occlusion (trees/shadows), uniform textures between pre/post images, or low-resolution inputs.

5. Significance and Impact

• Decision Support: The system provides decision-ready information by distinguishing severity levels, allowing emergency teams to prioritize resources (e.g., sending heavy machinery to "Destroyed" zones vs. medical teams to "Minor" zones).
• Speed and Objectivity: It offers a faster, consistent, and scalable alternative to manual inspection, reducing the time required for situational awareness.
• Complementary Role: The framework is designed to support, not replace, human experts, augmenting their capabilities with data-driven insights.
• Future Scalability: The paper outlines a path toward integrating drone imagery for street-level validation, multi-spectral/SAR data for all-weather capability, and cloud-based GIS integration for real-time operational use.

In conclusion, "Satellite to Street" represents a significant advancement in automated disaster assessment, moving beyond simple damage detection to a nuanced, severity-aware analysis that directly aids humanitarian response efforts.