Deep Learning for Feature Extraction in Remote Sensing: A Case-Study of Aerial Scene Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Scene classification from aerial images is crucial for remote sensing applications.
• Deep learning, particularly convolutional neural networks (CNNs), has significantly advanced image classification accuracy.
• Existing methods require effective feature extraction and classification strategies for aerial imagery.

Purpose of the Study:

• To develop and evaluate a novel two-stream deep architecture for aerial image scene classification.
• To leverage pre-trained CNNs for robust feature extraction from aerial images.
• To assess the performance of the proposed method against state-of-the-art techniques on benchmark datasets.

Main Methods:

• Utilized pre-trained CNNs for deep feature extraction from various network layers.
• Applied feature concatenation and dimensionality reduction to combine extracted features.
• Employed Support Vector Machines (SVM) for the final scene classification task.
• Experimented with diverse CNN architectures to optimize performance.

Main Results:

• The two-stream deep architecture demonstrated strong performance in aerial image scene classification.
• Feature concatenation and SVM classification proved effective for integrating deep features.
• The proposed method achieved competitive classification accuracies on the UC Merced and WHU-RS datasets.

Conclusions:

• The developed two-stream deep architecture offers a competitive and effective approach for aerial image scene classification.
• The integration of pre-trained CNN features with SVM provides a robust solution for remote sensing image analysis.
• This method shows promise for various remote sensing applications requiring accurate scene understanding.