Waterloo building dataset: a city-scale vector building dataset for mapping building footprints using aerial orthoimagery

Automated building footprint extraction is an important area of research in remote sensing with numerous civil and environmental applications. In recent years, deep learning methods have far surpassed classical algorithms when trained on appropriate datasets. In this paper, we present the Waterloo building dataset for building footprint extraction from very high spatial resolution aerial orthoimagery. Our dataset covers the Kitchener–Waterloo area in Ontario, Canada, contains 117 000 manually labelled buildings, and extends over an area of 205.8 km2. At a spatial resolution of 12 cm, it is the highest-resolution publicly available building footprint extraction dataset in North America. We provide extensive benchmarks for commonly used deep learning architectures trained on our dataset, which can be used as a baseline for future models. We also identified a key challenge in aerial orthoimagery building footprint extraction, which we hope can be addressed in future research. © Canadian Science Publishing. All rights reserved.