Global-Local Attention Network for Aerial Scene Classification

Yiyou Guo, Jinsheng Ji, Xiankai Lu, Hong Huo, Tao Fang, Deren LiIEEE Access 2019 · DOI

Introduction

Classifying aerial imagery presents unique challenges compared to standard ground-level photography. The primary obstacles are high intraclass variance—where scenes of the same type (e.g., residential areas) look vastly different due to geography or lighting—and low interclass variance, where different categories (e.g., commercial vs. residential) share many identical features like roads and trees.

Standard Convolutional Neural Networks (CNNs) often prioritize global context at the expense of the fine-grained local details that distinguish these subtle differences. This paper introduces GLANet, a lightweight, end-to-end architecture designed to bridge this gap by capturing both broad environmental context and specific local landmarks.

Methods: The GLANet Architecture

GLANet utilizes a pre-trained VGGNet-16 as its backbone. However, the authors significantly modified the architecture by removing the memory-intensive Fully Connected (FC) layers, which typically account for the bulk of a model's parameters. In their place, they introduced two specialized attention branches:

  1. Global Attention (GA) Branch: Utilizing a "Squeeze-and-Excitation" (SE) block, this branch evaluates feature channels to determine which are most critical for the overall scene context.
  2. Local Attention (LA) Branch: This branch employs spatial residual attention modules to generate a mask that highlights discriminative regions—such as a specific airplane or a baseball diamond—while suppressing irrelevant background noise.
The GLANet architecture featuring Global Attention (GA) and Local Attention (LA) branches.

Multi-Stage Supervision and Rank Loss

To ensure the branches learn complementary rather than redundant information, the model uses a multi-stage supervision strategy. Losses are calculated for the global branch, the local branch, and the final combined output. A key innovation is the Rank Loss function, which mandates that the combined "global-local" prediction must be more confident than either individual branch, forcing the network to integrate both types of features effectively.

Key Insights and Ideas

  • Efficiency: By replacing FC layers with attention mechanisms, GLANet reduces the model size from 528MB to approximately 56MB, making it nearly ten times smaller than VGGNet and highly suitable for edge deployment.
  • The "What" and "Where": The GA branch identifies what features are present globally, while the LA branch identifies where the most important details are located.
  • Decoupled Supervision: The Rank Loss ensures that the local branch actually focuses on local details rather than defaulting to global patterns.
Heatmap visualization of the Local Attention branch focusing on discriminative scene elements.

Experimental Results

GLANet was evaluated on three major datasets: AID (10,000 images), NWPU-RESISC45 (31,500 images), and PatternNet (38 classes). These datasets were chosen for their complexity, as older benchmarks have become saturated.

  • AID: Achieved 95.02% accuracy, outperforming standard VGGNet and more complex fusion models.
  • NWPU-RESISC45: Reached 93.45% accuracy, a significant improvement over off-the-shelf deep features.
  • PatternNet: Hit a near-perfect 99.70% accuracy.

Qualitative analysis via confusion matrices showed that GLANet is particularly effective at distinguishing between subclasses of residential areas (dense, medium, and sparse), which are traditionally difficult for CNNs to separate.

Conclusion and Future Work

GLANet demonstrates that attention-based feature fusion is both more accurate and more efficient than traditional deep architectures for aerial scene classification. Future research directions include expanding the model to handle multi-temporal data and non-optical imagery, such as Synthetic Aperture Radar (SAR) or hyperspectral data, to further enhance robustness in diverse environmental conditions.

Summary updated

BibTeX
@article{guo2019global,
  author = {Yiyou Guo and Jinsheng Ji and Xiankai Lu and Hong Huo and Tao Fang and Deren Li},
  title = {{Global-Local Attention Network for Aerial Scene Classification}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2918732},
}
Figure