fbpx

Innovation Hub

Innovating FEMA Disaster Response with Deep Learning Model for Building Damage Detection
Case Study Salesforce.com much more than CRM
Reading Time: < 1 minute

Challenge

FEMA is responsible for response and recovery efforts after natural disasters. This often involves analyzing aerial imagery of affected areas. The current manual review process can take several days to complete, leading to delayed responses and increased and high costs.

Solution

REI has developed a deep learning model capable of automatically detecting and assessing damaged buildings using satellite images. This model was trained on over 3,000 images of damage caused by Hurricane Michael, utilizing more than 300 GPU hours to iterate through multiple deep learning models and develop a deep neural network that achieves an impressive 80% accuracy

The U-Net architecture employed in this model is built on down-sampling and up-sampling convolutional neural networks, connected by skip connections. In order to optimize the model, several hyperparameters were fine-tuned, including:

    • Neural Network Architectures
    • Learning Rates
    • Iterations
    • Image Augmentation
    • Optimizers

Impact

This deep learning approach enables automated, real-time monitoring and damage assessments during natural disasters, significantly reducing the manual effort and time needed to initiate response and recovery efforts.

Capabilities Shown

  • Deep Learning
  • Computer Vision
  • Model Optimization
  • Satellite Imagery Processing & Analysis