Abstract
Hyperspectral images (HSIs) provide rich spectral and spatial information, yet their utility in change detection is often hampered by complex noise and the inherent trade-off between local detail and global context. To address these challenges, this paper introduces HBFormer, a novel Transformer-based architecture. HBFormer integrates a Hybrid Attention Module (HAM) and a Bidirectional-Gated Feed-Forward Network (BGFN). The HAM utilizes parallel CNN and gMLP branches to extract local and global features simultaneously, effectively suppressing noise. Meanwhile, the BGFN enhances feature fusion across spatial and channel dimensions. Experimental results on three benchmark datasets demonstrate that HBFormer significantly outperforms mainstream models like BiT and CBANet, achieving superior accuracy and robustness in complex environments.
Introduction
Change detection in hyperspectral remote sensing involves identifying transitions in land cover between two timestamps. While HSIs offer hundreds of narrow spectral bands, they are susceptible to Gaussian and striping noise. Traditional Convolutional Neural Networks (CNNs) excel at capturing local spatial details but struggle with global context. Conversely, standard Transformers capture long-range dependencies but are sensitive to noise and may overlook fine-grained details. HBFormer is designed to bridge this gap by processing local and global information in parallel, preventing noise from polluting the global feature extraction process.
The HBFormer Architecture
The HBFormer model is structured into two main stages, each containing specialized blocks designed to balance feature extraction. The core innovations are the Hybrid Attention Module (HAM) and the Bidirectional-Gated Feed-Forward Network (BGFN).
Hybrid Attention Module (HAM)
Unlike standard self-attention, HAM employs a parallel structure:
- CNN Branch: Utilizes a bottleneck structure with Squeeze-and-Excitation (SE) blocks to focus on local spatial details and suppress high-frequency noise.
- gMLP Branch: Employs a gated Multi-Layer Perceptron to capture long-range dependencies and global context. By running these branches in parallel, the model ensures that local details do not compromise global relationships, and vice versa.
Bidirectional-Gated Feed-Forward Network (BGFN)
Replacing the standard FFN, the BGFN introduces two independent gating paths for spatial and channel dimensions. This allows the network to selectively emphasize informative spectral bands and spatial regions while filtering out irrelevant sensor noise.
Experimental Evaluation
The model was evaluated on three EO-1 Hyperion datasets: River, Farmland, and Hermiston. HBFormer consistently achieved the highest Percentage Correct Classification (PCC) and Kappa Coefficients (KC) across all tests.
Key Results
- Farmland Dataset: HBFormer reached a 98.40% PCC, improving the Kappa coefficient by 2.02% over the BiT model. It proved particularly effective at identifying small, dense change areas like narrow farm paths.
- Hermiston Dataset: The model achieved 96.48% PCC, significantly reducing false positives compared to CBANet by effectively managing the complex noise inherent in irrigated agricultural data.
- River Dataset: Achieved a PCC of 97.86%, the highest among all tested mainstream methods.
Conclusion
HBFormer demonstrates that parallel integration of local and global features is a highly effective strategy for hyperspectral change detection. By utilizing the HAM and BGFN modules, the architecture suppresses noise while maintaining high sensitivity to actual land-cover changes. Future work will explore the migration of this architecture to multi-source data fusion and the integration of high-level semantic labels to further enhance detection intelligence.
Summary updated