CAW: A Remote-Sensing Scene Classification Network Aided by Local Window Attention.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Remote-sensing images present challenges in scene classification due to scale variations.
• Traditional convolutional neural networks struggle with complex spatial and textural information in these images, leading to suboptimal classification accuracy.

Purpose of the Study:

• To enhance remote-sensing image scene classification accuracy.
• To address the limitations of traditional methods in extracting complex spatial and texture information.

Main Methods:

• Introduction of a vision transformer network structure with strong global modeling capabilities.
• Implementation of a parallel network combining local-window self-attention and equivalent large convolution kernels for spatial-channel modeling.
• Utilizing the RSSCN7 and WHU-RS19 datasets for experimental validation.

Main Results:

• The proposed vision transformer network demonstrated improved accuracy in remote-sensing image scene classification.
• Ablation experiments, confusion matrix analysis, and heat map comparisons validated the network's effectiveness.
• The model showed enhanced performance in extracting both local and global features.

Conclusions:

• The vision transformer network offers a superior approach for remote-sensing image scene classification compared to traditional methods.
• The proposed spatial-channel modeling effectively improves feature extraction capabilities.
• The study confirms the network's potential for practical applications in remote sensing.