Computer Vision-Based Damage Assessment Using Optical Remote Sensing Imagery

Developing a framework to localize and classify damage to buildings and roads in natural disaster events, by using computer vision algorithms on optical remote sensing imagery.

Requirements

  • M.Sc. in Machine Learning, Data Science, Computer Science, Mathematics, Telecommunications, or similar
  • Knowledge of Python
  • Software development skills
  • Basic concepts of image processing
  • Basic concepts of data science, concerning data analysis, processing and machine learning

Description

The increasing frequency and severity of natural disasters necessitate rapid and accurate damage assessment to aid humanitarian assistance and disaster recovery efforts. Traditional methods require in-person evaluations within 24-48 hours, posing significant risks to human life. Collecting this data is often dangerous, as it requires people on the ground to directly assess damage during or immediately after a disaster. With the increased availability of satellite imagery, this task has the potential to not only be done remotely but also automatically by applying computer vision algorithms.

This thesis proposes the application of neural network algorithms to automate the damage assessment process, thereby enhancing efficiency and safety. It aims to develop a framework for building and road damage assessment using state-of-the-art computer vision neural networks, including ResNet and Visual Transformers, which will be applied to optical remote sensing imagery, Synthetic Aperture Radar (SAR) imagery, and aerial imagery. The objective is to apply the models to pre- and post-disaster events to localize infrastructures such as roads and buildings, and to provide a damage classification, either as a score or damage class, providing valuable insights for disaster response agencies.

The methodology involves reviewing existing techniques, collecting relevant datasets, developing and training neural network models, evaluating their performance, and applying them to out of distribution dataset.

Contacts