Proposed semi-supervised approach. Few labelled flood-damaged buildings can be used to learn a semi-supervised model. Such models can be further adapted to new geography using domain adaptation techniques.
In the last decade, flood events have cost more than US$100 billion, with more than 100,000 people being killed, and 1 billion left without a home worldwide. A quick response can substantially reduce the damage caused by floods. However, the rapid provision of damage information for floods is challenging. In this project, we aim to develop an approach to rapidly assess building damage after flood events, by exploring three machine learning-based approaches:
a multi-sensor change detection method to provide near real-time damage assessment,
a large-scale building footprint damage assessment technique, benefiting from existing global urban mapping data, and
a semi-supervised and few-shot learning approach, to include the few labelled data collected in the early stage of the flood.
