Data reduction by randomization subsampling for the study of large hyperspectral datasets.

Area of Science:

• Data science
• Image analysis
• Machine learning

Background:

• Hyperspectral images (HSI) contain vast amounts of data, making analysis computationally expensive.
• Principal Component Analysis (PCA) on large HSI datasets requires significant RAM and processing power, especially for time-series analysis.
• Data reduction techniques are crucial for efficient HSI analysis.

Purpose of the Study:

• To explore data reduction methods for faster analysis of time-series hyperspectral images.
• To compare the effectiveness of randomized sub-sampling PCA (RSPCA) and randomized PCA (rPCA) in preserving analytical information.
• To provide a didactic comparison of the benefits and drawbacks of these two data reduction techniques.

Main Methods:

• Implementation and comparison of randomized sub-sampling PCA (RSPCA).
• Implementation and comparison of randomized PCA (rPCA) based on local-rank approximations.
• Evaluation of information retention at various compression levels.
• Performance timing analysis for both methods.

Main Results:

• Both RSPCA and rPCA significantly reduce computation time for HSI analysis.
• rPCA demonstrates a more robust preservation of analytical information compared to RSPCA at similar compression levels.
• The study quantifies the trade-off between compression degree, information retention, and analysis speed.

Conclusions:

• Data reduction methods like RSPCA and rPCA offer efficient solutions for analyzing large hyperspectral image time series.
• rPCA is recommended for applications requiring higher fidelity in data preservation.
• These methods facilitate simultaneous analysis of temporal effects and other factors in HSI data.