By Fonon Nunghe

Computing science student Navjot Kaur’s research paper is published in the Computer-Aided Civil and Infrastructure Engineering; a peer-reviewed journal that serves as a conduit for innovation in computer technology and civil and infrastructure engineering.

About the paper: Large-scale building damage assessment using a novel hierarchical transformer architecture on satellite images

This study accomplishes two main feats. Firstly, it presents a hierarchical UNet-based architecture which uses the transformer-based difference to perform well on both the damage classification and builds segmentation tasks, achieving state of art on large datasets like xBD and LEVIR-CD.

And it introduces a new data set (Ida-BD) for evaluating the model’s adaptation performance and provides a baseline for the damage assessment tasks and transfer learning on newly damaged regions.

From this, they were able to classify building damage based on satellite imaging in the aftermath of natural disasters. The first of the aforementioned feats works by relying on the difference of transformer-encoded features from pre-disaster and post-disaster images, and hierarchically builds the output from multiresolution features. This work can play a vital role in the carrying out of emergency services as the information sourced here in real-time can inform rapid response teams on their strategic approach towards administering disaster relief services. The model can be run and images can be produced to understand the highest areas of impact in a natural disaster, but Navjot says that the final step would be having a way to pass on those results to the right authorities to be followed up on the ground level.

A challenge this study takes on is in using models based on images from past disasters to assess the damage of new ones by testing the models on new data sets to evaluate the adaptation performance of the models. “Another purpose of this project was to present methodology for new disasters such that the learnings can be quickly transferred to that data set. Currently, what happens is that we have this huge data set, one of the most notable being xBD data set. That data has multiple sources of disaster images, but whenever a new disaster happens, we face the difficulty of how to transfer that knowledge to the new images—how to train the model on new images that are usually not labelled,” says Navjot.

Click below to learn more about the paper:

Large-scale building damage assessment using a novel hierarchical transformer architecture on satellite images