Scholarlink

<p>Remote sensing images differ fundamentally from natural images due to their complex spatial structures, large intra-class variability, and subtle inter-class distinctions, posing significant challenges for remote sensing scene classification (RSSC). While existing convolutional neural networks (CNNs) have achieved notable progress, they often fail to capture the unique characteristics of remote sensing imagery and typically suffer from excessive parameter redundancy, resulting in high computational costs. These limitations are particularly critical given the scarcity and annotation difficulty of large-scale remote sensing datasets.To address these challenges, we propose AEBANet, a lightweight yet discriminative network specifically designed for RSSC. AEBANet first employs a depth-wise separable convolution (DS-Conv) backbone to ensure efficient feature extraction with reduced computational overhead. We then introduce the Adaptive Enhanced Branch Attention (AEBA) module, a novel lightweight design that jointly enhances channel-wise and spatial feature interactions while suppressing redundant background responses. Furthermore, we develop the Multi-Level Feature Fusion (MLFF) module to effectively integrate shallow fine-grained details with deep global semantics, thereby strengthening contextual dependencies and improving feature robustness. Extensive experiments on three benchmark datasets (NWPU, AID, and UCM) show that AEBANet achieves overall accuracies of 93.12%, 96.76%, and 99.52%, respectively. Ablation studies further confirm the necessity and effectiveness of each proposed component. In summary, AEBANet achieves competitive performance with substantially fewer parameters and lower computational cost, highlighting its potential as a practical solution for large-scale RSSC tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A lightweight enhanced branching attention model for remote sensing scene image classification

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Huiyue Wang,
  • Xianbin Wen,
  • Haixia Xu,
  • LiMing Yuan,
  • Xinyu Wang

摘要

Remote sensing images differ fundamentally from natural images due to their complex spatial structures, large intra-class variability, and subtle inter-class distinctions, posing significant challenges for remote sensing scene classification (RSSC). While existing convolutional neural networks (CNNs) have achieved notable progress, they often fail to capture the unique characteristics of remote sensing imagery and typically suffer from excessive parameter redundancy, resulting in high computational costs. These limitations are particularly critical given the scarcity and annotation difficulty of large-scale remote sensing datasets.To address these challenges, we propose AEBANet, a lightweight yet discriminative network specifically designed for RSSC. AEBANet first employs a depth-wise separable convolution (DS-Conv) backbone to ensure efficient feature extraction with reduced computational overhead. We then introduce the Adaptive Enhanced Branch Attention (AEBA) module, a novel lightweight design that jointly enhances channel-wise and spatial feature interactions while suppressing redundant background responses. Furthermore, we develop the Multi-Level Feature Fusion (MLFF) module to effectively integrate shallow fine-grained details with deep global semantics, thereby strengthening contextual dependencies and improving feature robustness. Extensive experiments on three benchmark datasets (NWPU, AID, and UCM) show that AEBANet achieves overall accuracies of 93.12%, 96.76%, and 99.52%, respectively. Ablation studies further confirm the necessity and effectiveness of each proposed component. In summary, AEBANet achieves competitive performance with substantially fewer parameters and lower computational cost, highlighting its potential as a practical solution for large-scale RSSC tasks.