SCAR-Net: A Selective Channel Attention with Residuals Network for High-Resolution Remote Sensing Scene Classification

Abstract

Accurate classification of high-resolution remote sensing images is a critical task in Earth observation, with wide-ranging applications in urban planning. However, challenges such as high intra-class variability, and complex spatial arrangements of small objects within large background regions often hinder traditional convolutional neural networks (CNNs). To address these limitations, we propose a novel Selective Channel Attention Residuals network (SCAR-Net) that integrates lightweight residual blocks with a channel attention mechanism based on squeeze-and-excitation (SE) units. This new architecture enables the network to selectively emphasize discriminative feature channels while maintaining robust multi-scale spatial feature extraction. By recalibrating channel-wise responses and leveraging residual learning, the proposed SCAR network effectively suppresses irrelevant background information and enhances class-specific features. The proposed SCAR network was evaluated on three widely used remote sensing datasets, and consistently outperformed several advanced CNN-based models. Furthermore, Grad-CAM++ visualizations confirm that SCAR accurately focuses on semantically meaningful regions in the images.