SAR-based terrain classification using weakly supervised hierarchical Markov aspect models.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Accurate terrain classification of remote sensing imagery is crucial for various applications.
• Existing models often struggle with local ambiguities and multiscale feature integration.
• Hierarchical approaches can improve spatial consistency in image labeling.

Purpose of the Study:

• To develop a computationally efficient graphical model for dense labeling of remote sensing images.
• To improve local label consistency and context-aware classification using hierarchical structures.
• To enable efficient training using reduced data labeling efforts.

Main Methods:

• Introduced the hierarchical Markov aspect model (HMAM), combining quadtree representations and aspect models.
• Utilized multiscale visual features and hierarchical smoothing for local label consistency.
• Incorporated probabilistic latent semantic analysis aspect models for broader image context.
• Employed a forwards-backwards inference pass for efficient local posterior computation.
• Used variational expectation-maximization for model training with pixel or keyword-level labels.

Main Results:

• HMAM demonstrated high accuracy and efficiency in labeling TerraSAR-X synthetic aperture radar data.
• Achieved significantly better results compared to single-scale aspect models.
• Showcased the effectiveness of combining quadtree and aspect model benefits.
• Keyword-level training reduced data provision costs with minimal accuracy loss.

Conclusions:

• HMAM provides an accurate and efficient solution for dense terrain class labeling in remote sensing.
• The model's hierarchical structure and context-aware approach effectively resolve local ambiguities.
• Keyword-level training offers a practical and cost-effective alternative for model development.