Abstract
Existing deep learning-based hyperspectral image (HSI) classification works still suffer from the limitation of the fixed-sized receptive field, leading to difficulties in distinctive spectral-spatial features for ground objects with various sizes and arbitrary shapes. Meanwhile, plenty of previous works ignore asymmetric spectral-spatial dimensions in HSI. To address the above issues, we propose a multi-stage search architecture in order to overcome asymmetric spectral-spatial dimensions and capture significant features. First, the asymmetric pooling on the spectral-spatial dimension maximally retains the essential features of HSI. Then, the 3D convolution with a selectable range of receptive fields overcomes the constraints of fixed-sized convolution kernels. Finally, we extend these two searchable operations to different layers of each stage to build the final architecture. Extensive experiments are conducted on two challenging HSI benchmarks including Indian Pines and Houston University, and results demonstrate the effectiveness of the proposed method with superior performance compared with the related works.
Introduction
Hyperspectral images (HSI) capture hundreds of continuous spectral bands for every spatial pixel, creating a rich "spectral-spatial" cube. However, traditional deep learning models face two significant hurdles in this domain. First, fixed receptive fields in standard convolutions struggle to adapt to ground objects of varying scales and shapes. Second, the asymmetric dimensions of HSI—where spatial and spectral resolutions represent different physical properties—are often treated symmetrically, leading to suboptimal feature extraction. While Neural Architecture Search (NAS) can automate design, general-purpose NAS methods often overlook these HSI-specific characteristics.
Methodology: A2S-NAS
The proposed Asymmetric Spectral-Spatial NAS (A2S-NAS) framework utilizes a differentiable search strategy to discover architectures optimized for HSI. The core of the method is the A2SConv block, which features a nested search space designed to handle the data's inherent asymmetry.
Nested Search Space
The A2SConv block incorporates two levels of flexibility:
- Outer Search Space (Pooling): The model can choose between no pooling, spectral pooling, or spatial pooling for each layer. This allows for asymmetric downsampling, preserving critical information based on the data's redundancy.
- Inner Search Space (Convolution): The model searches for optimal 3D convolution parameters, selecting between different kernel sizes ($3\times3\times3$ or $5\times5\times5$) and dilation rates to dynamically adjust the receptive field.
To ensure search stability and efficiency, the framework employs Beta-Decay regularization, which helps the gradient-based search converge to compact, high-performing networks.
Experiments and Results
A2S-NAS was evaluated on two challenging HSI benchmarks: Indian Pines (IP), an agricultural dataset, and Houston University (HU), an urban dataset.
Performance and Visualization
The model achieved state-of-the-art performance, outperforming established hand-crafted and NAS-based models like SSRN and 3D-ANAS. On Indian Pines, it reached an Overall Accuracy (OA) of 97.17%, with significant gains in difficult classes like "Soybean-mintill." On the Houston University dataset, it achieved an OA of 92.20%, notably improving the identification of urban "Roads."
Key Insights
A key finding was the model's preference for asymmetry; searched architectures favored asymmetric operations in 50% to 83.3% of layers. This automated flexibility allowed the network to adapt its receptive field to the specific scales of ground objects, proving that "one-size-fits-all" symmetric processing is suboptimal for HSI.
Conclusion
A2S-NAS demonstrates that tailoring neural architectures to the asymmetric nature of hyperspectral data significantly improves classification accuracy. By automating the selection of pooling operations and receptive fields, the framework provides a robust solution for diverse remote sensing applications. Future research will focus on improving the computational efficiency of these 3D architectures and exploring cross-dataset generalization.
Summary updated