A benchmark GaoFen-7 dataset for building extraction from satellite images

Area of Science:

• Remote Sensing
• Geospatial Data Science
• Computer Vision

Background:

• Accurate building extraction is vital for urban planning and analysis.
• Existing building datasets lack comprehensive coverage of diverse Chinese urban and rural landscapes.
• High-resolution satellite imagery is essential for detailed geospatial feature extraction.

Purpose of the Study:

• To develop a high-quality, high-resolution building dataset for China.
• To address the scarcity of building data covering both urban and rural Chinese environments.
• To facilitate advancements in building extraction algorithms and intelligent interpretation.

Main Methods:

• Creation of a new building dataset using GaoFen-7 (GF-7) satellite imagery.
• Dataset comprises 5,175 image tiles (512x512 pixels) covering 573.17 km².
• Inclusion of 170,015 buildings, with a focus on diverse urban (84.8%) and rural (15.2%) areas.

Main Results:

• The GF-7 Building dataset demonstrates high quality and diversity.
• Convolutional neural networks achieved over 93% overall accuracy using the dataset.
• The dataset proved effective for building extraction in both urban and rural settings.

Conclusions:

• The GF-7 Building dataset effectively supplements existing resources.
• It supports the development of advanced building extraction techniques.
• The dataset will aid in intelligent building interpretation across China.