# Leveraging Deep Neural Network Compression Techniques for Real-Time
Hyperspectral Image Processing in Edge AI


Authors: Dheeraj Kumar, Leila Mozaffari

## Introduction

### Objective

To improve hyperspectral image analysis by integrating SSCNet \[1\] with
the FasterAI \[4\] compression technique, demonstrating efficiency and
performance on the HySpecNet-11k dataset. Hyperspectral imaging provides
rich spectral information across numerous bands, supporting applications
like remote sensing, agriculture, and medical imaging. However, the high
volume and computational demands of hyperspectral data necessitate
innovative compression and processing techniques.

### Challenges

- Managing large-scale hyperspectral datasets.
- Balancing reconstruction quality and compression efficiency.

### Contributions

- Integration of SSCNet with FasterAI pruning for compressing HSI
  compression model. to reduce model size while preserving high-quality
  image reconstruction.
- Validated on HySpecNet-11k, a large-scale hyperspectral benchmark
  dataset.
- Achieved significant reduction in model size and computational load
  with minor performance trade-offs.

## Methodology

### Dataset: HySpecNet-11k

HySpecNet-11k \[2\] is a large-scale hyperspectral dataset containing
11,483 image patches (128×128 pixels with 224 spectral bands) derived
from e Environmental Mapping and Analysis Program (EnMAP) satellite
data. It is designed for benchmarking learning-based compression and
analysis methods.

- Dataset Splits: Training (70%), Validation (20%), Test (10%).
- Preprocessing: Removed water vapor-affected bands, applied
  normalization, and used both patchwise and tilewise splits.

### Model: Spectral Signals Compressor Network (SSCNet)

SSCNet \[3\] uses 2D convolutions to compress spatial dimensions while
preserving spectral integrity.

- Encoder: Three 2D convolutional layers with parametric ReLU activation
  and max-pooling.
- Decoder: Uses transposed convolutions for reconstruction.
- Compression Ratio (CR): Defined by latent channels in bottleneck
  layer.

### FasterAI Pruning Compression Technique

- Remove redundant weights or neurons.
- Fine-tune to recover performance.
- Outcome: Smaller, faster model with minimal accuracy loss.

## Experimental Results

- Metric: Bits-per-pixel per channel (bpppc) vs. Peak Signal-to-Noise
  Ratio (PSNR).
- Pruned SSCNet achieved a PSNR of 42.98 dB at 2.53 bpppc.
- Reduced model size by 45% and computational complexity by 50%.

![](./images/Results.png)

### Comparative Analysis

- Outperformed traditional and learning-based methods in compression
  efficiency and speed.
- Visuals demonstrate minimal loss of fidelity in reconstructed
  hyperspectral images.

## Conclusion

- Effective reduction in memory footprint and computational demands for
  real-time edge AI deployment.
- Enables practical deployment of hyperspectral models in
  resource-constrained environments.
- Supports scalable analysis for large datasets like HySpecNet-11k.

## Future Work

- Test FasterAI compression on additional hyperspectral models.
- Explore dynamic pruning strategies.
- Apply other model compression techniques.

## References

- \[1\] M. H. P. Fuchs and B. Demir, “Hyspecnet-11k: A large-scale
  hyperspectral dataset for benchmarking learning-based hyperspectral
  image compression methods,” in IGARSS 2023-2023 IEEE International
  Geoscience and Remote Sensing Symposium, IEEE, 2023, pp. 1779–1782.
- \[2\] M. H. P. Fuchs and B. Demir, “HySpecNet-11k: A large-scale
  hyperspectral benchmark dataset.” Dryad, p. 63608947808 bytes,
  Jun. 26, 2023. doi: 10.5061/DRYAD.FTTDZ08ZH.
- \[3\] R. La Grassa, C. Re, G. Cremonese, and I. Gallo, “Hyperspectral
  data compression using fully convolutional autoencoder,” Remote
  Sensing, vol. 14, no. 10, p. 2472, 2022.
- \[4\] “FasterAI,” fasterai. Available:
  https://nathanhubens.github.io/fasterai/
- \[5\] Fuchs, M. H. P., & Demir, B. (2023). HySpecNet-11k: A
  Large-Scale Hyperspectral Dataset for Benchmarking Learning-Based
  Hyperspectral Image Compression Methods. arXiv preprint
  arXiv:2306.00385v2.