Adaptive Discriminative Regions Learning Network for Remote Sensing Scene Classification.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing scene classification (RSSC) is crucial for intelligent interpretation.
• Deep convolutional neural networks (DCNNs) have advanced RSSC but face challenges like complex land cover and scale variation.
• Existing methods struggle with redundant and noisy areas in scene classification.

Purpose of the Study:

• To propose an adaptive discriminative regions learning network (ADRL-Net) for robust RSSC.
• To enhance RSSC performance by effectively locating and utilizing discriminative regions.
• To mitigate the impact of complex scene composition, object scale variation, and irrelevant information.

Main Methods:

• Developed ADRL-Net with three modules: discriminative region generator, region discriminator, and region scorer.
• Employed a novel self-supervision mechanism to guide region identification.
• Iteratively refined informative regions through generator-discriminator feedback.

Main Results:

• ADRL-Net consistently outperformed state-of-the-art RSSC methods across four benchmark datasets.
• The network demonstrated robustness to complex scene composition and object scale variations.
• Effective focus on informative regions reduced interference from redundant areas.

Conclusions:

• ADRL-Net offers a significant improvement for remote sensing scene classification.
• The proposed self-supervision mechanism and adaptive region learning are effective.
• ADRL-Net shows strong potential for practical applications in remote sensing interpretation.