Building footprint extraction from off-nadir Earth observation images

Project Description

Building footprint extraction from aerial imagery is essential for creating and continuously updating building inventories, supporting applications such as urban planning. Traditional methods often extract building roofs from aerial images under the assumption that they align with footprints, an assumption that fails in off-nadir imagery. This project introduces a novel multilabel learning approach that jointly predicts footprint, roof, and shape features of oblique buildings using a Vision Transformer (ViT) for accurate building footprint extraction from off-nadir aerial images. The study is a part of an ongoing research work on “Earth Observation and AI for urban building footprint extraction”. The outcome O1 of the study is available in the link below. See more on outcome O2 that was previously published out of this project.

Collaboration

Earth Observation and AI Research Group led by A/Prof. Dr. Jagannath Aryal of the Department of Infrastructure Engineering along with the research team from the German Remote Sensing Data Center (DFD), German Aerospace Center (DLR), Weßling, Germany have developed a deep learning-based solution for this project. The research collaboration was in part funded by the Bonn-Melbourne Research Excellence Fund 2023.

Qutput 1

Output 2