The effects of spectral dimensionality reduction on hyperspectral pixel classification: A case study

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Hyperspectral imaging captures detailed spectral information, leading to high-dimensional data.
• Dimensionality reduction is crucial for efficient processing and analysis of hyperspectral pixels.
• Pixel classification is a key task in hyperspectral image analysis.

Purpose of the Study:

• To systematically investigate the impact of various dimensionality reduction techniques on hyperspectral pixel classification accuracy.
• To evaluate the trade-offs between compression rate, signal reconstruction, and classification performance.
• To identify optimal compression strategies for hyperspectral pixel classification.

Main Methods:

• Applied five dimensionality reduction methods: Principal Component Analysis (PCA), Kernel PCA (KPCA), Independent Component Analysis (ICA), Autoencoder (AE), and Denoising Autoencoder (DAE).
• Compressed 301-dimensional hyperspectral pixels using these methods at varying compression rates.
• Performed pixel classification on the compressed data and evaluated accuracy, compression rates, and reconstruction errors using three diverse hyperspectral datasets (urban, suburban, forest).

Main Results:

• PCA, KPCA, and ICA demonstrated strong signal reconstruction but lower classification scores at compression rates exceeding 90%.
• AE and DAE achieved better classification accuracy at a 95% compression rate.
• The performance of AE and DAE degraded as compression rates approached 97%.

Conclusions:

• Both the choice of dimensionality reduction method and the compression rate significantly influence hyperspectral pixel classification performance.
• AE and DAE show promise for high-compression hyperspectral pixel classification tasks.
• Careful consideration of compression strategy is essential for designing effective hyperspectral image analysis pipelines.