StyHighNet: Semi-Supervised Learning Height Estimation from a Single Aerial Image via Unified Style Transferring

Area of Science:

• Computer Vision
• Remote Sensing
• Geospatial Analysis

Background:

• Estimating height from single aerial images is crucial for computer vision and remote sensing.
• Supervised methods face domain bias, limiting their application to new environments.
• Existing techniques struggle with adaptability to diverse urban landscapes.

Purpose of the Study:

• To propose StyHighNet, a semi-supervised framework for accurate single aerial image height estimation in new urban areas.
• To enable effective domain adaptation using unlabeled data by transferring multi-source images to a unified style.
• To reduce reliance on extensive labeled datasets for height estimation tasks.

Main Methods:

• Developed a novel semi-supervised framework (StyHighNet) integrating three sub-networks: style transferring, height regression, and style discrimination.
• Employed style transferring to create Unified Style Distribution Maps (USDMs) for consistent appearance distribution.
• Utilized unsupervised learning for style distribution function, enhancing flexibility and accuracy.

Main Results:

• StyHighNet demonstrated superior performance in both supervised and semi-supervised learning modes on Vaihingen and Potsdam datasets.
• The framework effectively transfers multi-source images to a unified style, providing additional supervision signals from unlabeled data.
• Achieved accurate height map prediction through the height regression sub-network utilizing USDMs.

Conclusions:

• StyHighNet offers a flexible and accurate solution for height estimation from single aerial images, particularly in new urban environments.
• The semi-supervised approach significantly reduces the need for labeled data, addressing domain adaptation challenges.
• The framework's adaptability extends to purely supervised learning by disabling the style discrimination sub-network.