HETMCL: High-Frequency Enhancement Transformer and Multi-Layer Context Learning Network for Remote Sensing Scene...

Area of Science:

• Computer Science
• Artificial Intelligence
• Remote Sensing

Background:

• Remote Sensing Scene Classification (RSSC) is crucial but challenging.
• Transformer models excel at global dependencies but struggle with high-frequency local details.
• Existing methods often fail to comprehensively utilize both high-frequency and low-frequency information.

Purpose of the Study:

• To propose a novel method, HETMCL, for improved RSSC.
• To effectively capture and integrate high-frequency and low-frequency features in remote sensing data.
• To enhance the performance of Transformer-based models in RSSC.

Main Methods:

• Utilizing Convolutional Neural Networks (CNNs) for low-level spatial structure extraction.
• Implementing an Adjacent Layer Feature Fusion Module (AFFM) to bridge semantic gaps between layers.
• Introducing a High-Frequency Information Enhancement Vision Transformer (HFIE) with a High-to-Low-Frequency Token Mixer (HLFTM) for high-frequency detail capture.
• Employing Multi-Layer Context Alignment Attention (MCAA) to integrate multi-layer features and contextual relationships.

Main Results:

• HETMCL achieved state-of-the-art Overall Accuracy (OA) on benchmark datasets.
• Achieved 99.76% OA on UCM, 97.32% on AID, and 95.02% on NWPU.
• Outperformed existing methods by up to 0.38% in OA.

Conclusions:

• The proposed HETMCL method effectively learns comprehensive features from both high-frequency and low-frequency information.
• HETMCL demonstrates superior performance in Remote Sensing Scene Classification compared to existing approaches.
• The integration of CNNs, HFIE, and MCAA offers a promising direction for advanced RSSC.