Classification of Building Damage Using a Novel Convolutional Neural Network Based on Post-Disaster Aerial Images.

Area of Science:

• Earthquake Engineering
• Computer Vision
• Remote Sensing

Background:

• Accurate building damage assessment is crucial for disaster response and loss estimation.
• Existing methods often lack the granularity needed, typically classifying buildings as only intact or damaged.
• There is a need for more sophisticated methods to classify varying degrees of building damage.

Purpose of the Study:

• To introduce a novel convolutional neural network, the Earthquake Building Damage Classification Net (EBDC-Net).
• To enable detailed assessment of building damage using post-disaster aerial imagery.
• To improve upon the limitations of binary (intact/damaged) classification systems.

Main Methods:

• Developed EBDC-Net, a convolutional neural network with a feature extraction encoder and a damage classification module.
• The encoder extracts semantic information for distinguishing damage levels.
• The classification module integrates global and contextual features for enhanced accuracy.

Main Results:

• EBDC-Net achieved high classification accuracy on a public dataset and a large-scale post-earthquake UAV dataset.
• Overall accuracy reached 94.44% for two categories, 85.53% for three, and 77.49% for four damage categories.
• The model demonstrated effectiveness in classifying buildings across different damage levels.

Conclusions:

• The proposed EBDC-Net offers an accurate and detailed approach to classifying earthquake-induced building damage.
• This method significantly enhances the ability to assess damage beyond simple intact/damaged distinctions.
• The findings support the use of EBDC-Net for improved disaster emergency response and loss assessment.