Duplex-Hierarchy Representation Learning for Remote Sensing Image Classification.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing image classification (RSIC) is crucial for analyzing aerial imagery.
• Deep learning models have advanced RSIC but struggle with intra-class diversity and inter-class similarity.

Purpose of the Study:

• To address the challenges of diversity and similarity in RSIC.
• To propose a novel duplex-hierarchy representation learning (DHRL) method for more discriminative feature learning.

Main Methods:

• Utilized a pretrained ResNet for feature extraction from paired images.
• Mapped features into a common space to reduce intra-class scatter and increase inter-class separation.
• Employed discrimination loss in the label space and a confusion score for guided representation learning.

Main Results:

• The DHRL method demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant effectiveness on two challenging remote sensing image scene datasets.
• Successfully learned discriminative representations for improved RSIC.

Conclusions:

• The proposed DHRL method effectively overcomes limitations in current RSIC approaches.
• DHRL offers a promising direction for enhancing the accuracy and robustness of remote sensing image analysis.
• The method's ability to learn from duplex-hierarchy spaces proves highly effective.