Self-Supervised Learning with Trilateral Redundancy Reduction for Urban Functional Zone Identification Using...

Area of Science:

• Computer Vision
• Machine Learning
• Urban Analytics

Background:

• Supervised learning for urban scene classification requires extensive labeled data, which is often scarce.
• Existing methods struggle with the challenge of limited high-quality labels in urban datasets.

Purpose of the Study:

• To propose an innovative self-supervised learning framework, Trilateral Redundancy Reduction (Tri-ReD), to overcome label scarcity in urban scene classification.
• To introduce a novel data augmentation strategy, tri-branch mutually exclusive augmentation (Tri-MExA), to reduce augmentation-induced uncertainties.

Main Methods:

• Developed the Trilateral Redundancy Reduction (Tri-ReD) framework utilizing a novel 'trilateral loss' for self-supervised pre-training.
• Implemented tri-branch mutually exclusive augmentation (Tri-MExA) to enhance representation learning.
• Pre-trained models on 116,491 unlabeled street-view images and fine-tuned on labeled datasets (BIC_GSV, BEAUTY).

Main Results:

• The Tri-ReD framework achieved state-of-the-art (SOTA) performance in self-supervised pre-training for urban scene classification.
• Outperformed direct supervised learning approaches by an average of 19% in urban functional zone identification.
• Surpassed models pre-trained on ImageNet by approximately 11%.

Conclusions:

• The proposed Tri-ReD framework effectively learns essential urban scene representations without semantic labels, addressing the challenge of label scarcity.
• Tri-ReD is architecture-agnostic, demonstrating effectiveness with both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).
• This self-supervised approach offers a robust and adaptable solution for various downstream urban analysis tasks.