Self-supervised in-domain representation learning for remote sensing image scene classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Transfer learning from ImageNet shows limitations in remote sensing due to domain differences.
• Annotating remote sensing data is challenging, while unlabeled data is abundant.
• Self-supervised learning (SSL) offers improved representation learning over supervised methods.

Purpose of the Study:

• To pre-train in-domain representations for remote sensing imagery using contrastive SSL.
• To transfer learned features to related remote sensing datasets for scene classification.
• To identify key factors influencing effective in-domain feature extraction.

Main Methods:

• Utilized the SimSiam algorithm for self-supervised pre-training on remote sensing datasets.
• Transferred pre-trained weights to various downstream scene classification tasks.
• Conducted linear evaluation with limited samples per class.
• Analyzed feature pre-training using datasets with diverse attributes.

Main Results:

• Achieved state-of-the-art results on five land cover classification datasets.
• Demonstrated that higher-resolution pre-training datasets lead to more discriminative and general representations.
• Identified influential factors for selecting optimal datasets for in-domain feature learning.

Conclusions:

• Contrastive self-supervised learning, specifically SimSiam, is effective for remote sensing representation learning.
• In-domain pre-training significantly enhances performance on downstream tasks, especially with limited labeled data.
• Dataset resolution is a critical factor for learning robust and transferable remote sensing features.