SemiRS-COC: Semi-Supervised Classification for Complex Remote Sensing Scenes With Cross-Object Consistency.

Area of Science:

• Computer Science
• Remote Sensing
• Machine Learning

Background:

• Semi-supervised learning (SSL) leverages unlabeled data for model training.
• Satellite Earth observation images present unique challenges for SSL due to complex scenes.
• Existing SSL methods struggle with unreliable pseudo-labels caused by background interference and intra-class variations in remote sensing data.

Purpose of the Study:

• To propose a novel semi-supervised classification method, SemiRS-COC, for complex remote sensing scenes.
• To generate reliable object-level pseudo-labels by exploiting object similarities within and across images.
• To enhance the performance of semi-supervised classification in remote sensing applications.

Main Methods:

• Developed a Local Self-Learning Object Perception (LSLOP) mechanism to address background object interference.
• Introduced a Cross-Object Consistency Pseudo-Labeling (COCPL) strategy to handle intra-class differences.
• Utilized feature space similarities between foreground objects for robust pseudo-label generation.

Main Results:

• SemiRS-COC effectively generates reliable object-level pseudo-labels.
• The method demonstrates superior performance compared to state-of-the-art techniques on three benchmark remote sensing datasets.
• Improved classification accuracy in complex remote sensing scenes with limited labeled data.

Conclusions:

• SemiRS-COC offers a robust solution for semi-supervised classification in complex remote sensing scenarios.
• The proposed LSLOP and COCPL mechanisms significantly improve pseudo-label quality and model performance.
• This work advances the application of semi-supervised learning for analyzing large-scale satellite Earth observation data.