A Two-Stream Deep Fusion Framework for High-Resolution Aerial Scene Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Aerial scene classification is crucial for interpreting high-resolution remote sensing images.
• Effective feature representation and classification methods are key to improving accuracy.
• Existing methods face challenges in capturing comprehensive scene information.

Purpose of the Study:

• To develop a novel two-stream deep architecture for enhanced aerial scene classification.
• To leverage both original image data and saliency-enhanced features.
• To improve the accuracy and robustness of aerial image analysis.

Main Methods:

• Utilized two pre-trained Convolutional Neural Networks (CNNs) for feature extraction.
• Employed saliency detection to process aerial images for a second feature stream.
• Implemented two feature fusion strategies to combine features from RGB and saliency streams.
• Applied the Extreme Learning Machine (ELM) classifier for final classification.

Main Results:

• The proposed two-stream architecture achieved significant classification accuracy improvements.
• Performance was validated on four diverse and challenging datasets (UC-Merced, WHU-RS, AID, NWPU-RESISC45).
• The method outperformed existing state-of-the-art approaches across all tested datasets.

Conclusions:

• The developed two-stream deep architecture effectively enhances aerial scene classification.
• Combining original image features with saliency-based features boosts performance.
• This approach offers a promising direction for high-resolution remote sensing image analysis.