MASTER’S THESIS: The Practical Applicability of a CNN for Automated Building Damage Assessment
WHAT IS IT ABOUT
Conducting a building damage assessment is common practice after disasters caused by natural hazards. While these assessments are now done manually, being able to do them automatically, by the means of a model, holds a lot of potential. This research aims to understand the ability of a specific model, based on a Convolutional Neural Network (CNN), to create this assessment in a realistic practical setting.
WHO IS THE AUTHOUR
Tinka Valentijn conducted this research for her MSc in Machine Learning at Aalto University. During her studies, she learnt the mathematical details of the workings of machine learning models while having the passion to leverage technology for societal issues. In the course of this research she explored this intersection by researching the use of a machine learning model in a humanitarian context.
WHY IS THIS WORK NEEDED
Complementing the current manual damage assessments with an automated building damage assessment holds a lot of potential. The assessments can be done faster and updated regularly with less human resources whilst also maintaining to a constant definition of damage. Despite the potential, none of the previous studies done on automated damage assessment have been applied in practice.
The main reason that these studies have not made it to production is that they do not take the practical context into account. For the model to be applicable, it should for example be known
- How the model performs on a wide range of disasters.
- How the models output can compliment the other factors to consider when assessing damage (design research conducted in parallel)
Furthermore, limited data availability should be taken into account. Thus, this research explores those factors with the goal to develop a model that does hold practical value.
HOW WE WORKED TOGETHER
This research was conducted in the context of the larger project we have within 510 on developing an automated damage assessment tool. This tool is being developed by a group of volunteers under supervision of staff members, and is fully open–source.
WHAT ARE THE MAIN FINDINGS
It was shown that the quality of the building damage assessment by the model heavily differs per disaster. While the assessment produced for the Joplin tornado was of high quality, that for the Nepal flooding showed disappointing value. What caused this difference remains an open question since this research showed that it was not significantly influenced by the type of damage (wind vs water), geographical location (North America vs Asia) or the differences in image parameters (such as the type of satellite).
The performance of the model in a realistic data setting was explored, where it was assumed that realistically satellite imagery before and after the disaster struck is available, but that no manually gathered damage labels are available for a part of the impacted area. This is unlike previous research, where it is often assumed that there is access to this manually labelled data. This study showed that without this labelled data, the model is still able to produce a high-quality assessment for a set of disasters, including the Joplin tornado. This result is very promising for actual use of the model since it shows that its contributions can be of high value in a realistic data setting.
- The results of this thesis are going to be published as a journal article.
- the findings of this research are being used to further develop 510’s model for automated damage assessment.
- The Model & Human centred design research will be amalgamated with the goal of bringing this model to production.
Written by Tinka Valentijn