Abstract
Hyperspectral image (HSI) classification has seen significant advancements through deep learning, particularly with spectral-spatial approaches. However, these methods often face high computational costs due to large parameter scales in multi-channel networks or the inherent complexity of 3D filters. This paper introduces SSDC-DenseNet, an end-to-end spectral-spatial dual-channel dense network designed to be cost-effective in both training time and parameter scale while maintaining state-of-the-art accuracy. By replacing expensive 3D filters with a 2D dual-channel structure and utilizing dense connections, the proposed model achieves high-level feature exploration with a remarkably reduced computational footprint.
The Challenge of Hyperspectral Classification
Hyperspectral images (HSIs) capture hundreds of narrow spectral bands for every pixel, providing a rich "fingerprint" for material identification. However, this high dimensionality creates a "curse of dimensionality." To classify HSIs accurately, models must integrate spectral information (the light signature) and spatial information (neighboring pixel context). While 3D Convolutional Neural Networks (CNNs) are the standard for this task, they are computationally intensive and prone to overfitting on the relatively small labeled datasets available in the HSI field.
Proposed Architecture: SSDC-DenseNet
The SSDC-DenseNet architecture addresses these efficiency concerns through a cascaded three-block design that replaces heavy 3D math with a clever 2D dual-channel structure.
Block 1: Spectral Feature Extraction
The process begins with $1 \times 1$ convolutions that compress and extract initial spectral features. This step reduces the dimensionality of the input cube while preserving the spatial organization of the data.
Block 2: Multi-Scale Joint Spectral-Spatial Feature Learning
This is the core innovation of the model. Instead of using a single 3D filter, the signal is split into two parallel 2D paths:
- Spectral Channel ($1 \times 1$): Focuses on the spectral signatures of individual pixels.
- Spatial Channel ($3 \times 3$): Captures local spatial correlations between neighboring pixels.
Both channels employ Dense Connections, where each layer receives feature maps from all preceding layers. This promotes feature reuse and allows the network to be deeper and more expressive without a massive increase in parameters.
Block 3: Feature Fusion
The outputs from the dual channels are concatenated and passed through a final dense block. To further reduce the model's scale, Global Average Pooling (GAP) is used instead of traditional fully connected layers, significantly cutting down the parameter count and mitigating overfitting.
Experimental Evaluation
The model was tested on three benchmark datasets:
- Indian Pines (IN): 200 spectral bands, 16 vegetation classes.
- University of Pavia (UP): 103 spectral bands, 9 urban land-cover classes.
- Salinas: 224 spectral bands, 16 agricultural classes.
A $7 \times 7$ spatial patch size was used across all experiments, providing an optimal balance between local context and computational efficiency.
Key Results and Performance
SSDC-DenseNet demonstrated competitive performance against state-of-the-art models like the 3D-based FDSSC, while being significantly lighter:
- High Accuracy: The model achieved an Overall Accuracy (OA) of 99.47% on Indian Pines, 99.84% on University of Pavia, and 99.73% on Salinas.
- Parameter Efficiency: On the Salinas dataset, SSDC-DenseNet required only 132.9k parameters, a nearly 10x reduction compared to the 1254.1k parameters of the 3D-based FDSSC.
- Reduced Training Time: Training on the Indian Pines dataset took approximately 484 seconds, compared to over 1655 seconds for the 3D FDSSC.
Discussion and Limitations
While SSDC-DenseNet is highly efficient, its performance is sensitive to hyperparameters such as the spatial patch size and the number of layers within the dense blocks. For instance, datasets with lower spatial resolution, like Indian Pines, may require a slightly larger model scale to reach peak performance. Future research could explore automated hyperparameter tuning or evaluate the model's robustness against sensor noise in real-world aerial surveys.
Conclusion
The SSDC-DenseNet proves that a dual-channel 2D approach can effectively mimic the performance of 3D convolutions for hyperspectral classification. By separating spectral and spatial learning and utilizing dense connections, the model achieves high accuracy with significantly lower computational costs. This efficiency makes it a strong candidate for deployment in resource-constrained environments, such as on-board satellite processing or edge-computing devices for remote sensing.
Summary updated