Semi-Supervised Contrastive Learning for Remote Sensing: Identifying Ancient Urbanization in the South-Central Andes.

Area of Science:

• Digital Archaeology
• Machine Learning Applications in Cultural Heritage
• Remote Sensing for Archaeological Prospection

Background:

• Traditional archaeological surveys face limitations in scale and sampling efficiency.
• Manual analysis of satellite and aerial imagery offers broader views but is labor-intensive and subjective.
• Existing self-supervised learning methods struggle with highly imbalanced datasets common in archaeological feature detection.

Purpose of the Study:

• To develop a scalable framework for detecting archaeological features from satellite imagery.
• To address the 'long-tail problem' of rare archaeological features in large, imbalanced datasets.
• To improve the efficiency and accuracy of archaeological site identification using machine learning.

Main Methods:

• Proposed a novel semi-supervised learning framework integrating labeled and unlabeled satellite imagery.
• Utilized the inherent data imbalance to generate pseudo-negative pairs for contrastive learning.
• Applied the method to a dataset of 95,358 unlabeled and 5,830 labeled images for ancient building detection.

Main Results:

• The semi-supervised contrastive learning model achieved a balanced accuracy of 79.0% in detecting ancient buildings.
• Demonstrated a 3.8% improvement over state-of-the-art approaches on the imbalanced satellite image dataset.
• Successfully leveraged a small percentage of labeled data (<7%) within the semi-supervised setting.

Conclusions:

• The proposed semi-supervised framework effectively overcomes the challenges of detecting rare archaeological features in large-scale satellite imagery.
• This approach offers a more scalable and accurate alternative to traditional survey methods and purely supervised machine learning.
• The method provides a promising direction for advancing archaeological prospection through advanced machine learning techniques.