SAGN: Semantic-Aware Graph Network for Remote Sensing Scene Classification.

Area of Science:

• Remote Sensing (RS)
• Computer Vision
• Artificial Intelligence

Background:

• High-resolution remote sensing (HRRS) image scene classification is crucial but challenging due to diverse content, scale, and volume.
• Current deep convolution neural networks (DCNNs) often treat this as a single-label problem, ignoring hidden, diverse semantics and leading to inaccuracies.

Purpose of the Study:

• To propose a novel Semantic-Aware Graph Network (SAGN) to overcome the limitations of single-label classification in HRRS imagery.
• To effectively mine and utilize diverse, multi-scale semantic information within HRRS images for improved scene classification.

Main Methods:

• Developed SAGN comprising a Dense Feature Pyramid Network (DFPN) for multi-scale extraction, an Adaptive Semantic Analysis Module (ASAM) for mining semantics, a dynamic graph feature update module for relation exploitation, and a Scene Decision Module (SDM).
• Focused on leveraging diverse semantics rather than converting to multi-label problems.

Main Results:

• Extensive experiments on three popular HRRS scene datasets demonstrated the effectiveness of the proposed SAGN.
• The method successfully utilizes diverse semantics for accurate HRRS scene classification.

Conclusions:

• The proposed SAGN effectively addresses the challenge of diverse semantics in HRRS image scene classification.
• SAGN offers a promising approach for accurate and robust scene classification in high-resolution remote sensing applications.