ORBITaL-Net: A labeled training library for large-scale building feature extraction.

Area of Science:

• Geospatial artificial intelligence
• Computer vision
• Remote sensing

Background:

• Machine learning models require diverse, large-scale datasets for robust performance.
• Existing building outline datasets often focus on urban areas, limiting generalizability.
• Geospatial data is crucial for urban planning, disaster response, and environmental monitoring.

Purpose of the Study:

• To introduce the Oak Ridge Building Image and TrAining Label Net (ORBITaL-Net) dataset.
• To provide a globally diverse dataset of building outlines and corresponding imagery for machine learning.
• To facilitate the development of scalable and generalizable geospatial machine learning techniques.

Main Methods:

• Generation of nearly 1.5 million building outlines from ~128,000 training tiles.
• Utilization of very high-resolution multispectral overhead imagery (2010-2020).
• Inclusion of samples from 72 countries across diverse geographic settings and land uses.

Main Results:

• Creation of the ORBITaL-Net dataset, featuring extensive building outline labels.
• The dataset covers diverse conditions, including urban/rural land use, varied terrain, and imagery quality.
• Open-source release of labeled building outlines paired with reference imagery.

Conclusions:

• ORBITaL-Net offers a unique, globally diverse resource for training geospatial AI models.
• The dataset's focus on variability enhances the development of generalizable computer vision techniques.
• This resource supports advancements in AI for geospatial applications beyond urban centers.