Autoencoder-Based Hyperspectral Unmixing with Simultaneous Number-of-Endmembers Estimation

Area of Science:

• Remote Sensing
• Data Science
• Signal Processing

Background:

• Hyperspectral unmixing is crucial for extracting information from hyperspectral data, impacting scientific, environmental, and industrial fields.
• Current challenges include accurately identifying endmember numbers, extracting endmembers, and estimating abundance fractions.
• Existing methods often struggle to effectively utilize both spatial and spectral information.

Purpose of the Study:

• To develop an advanced hyperspectral unmixing technique.
• To address the limitations of existing methods in endmember identification and abundance estimation.
• To leverage both spatial and spectral information for improved unmixing accuracy.

Main Methods:

• A convolutional neural network (CNN)-based autoencoder was employed to process hyperspectral images.
• A self-learning module with a fuzzy clustering algorithm was integrated to determine the number of endmembers.
• A novel approach was proposed to estimate endmember abundances using both autoencoder and clustering outputs.

Main Results:

• The proposed method demonstrated superior performance compared to existing techniques.
• Significant improvements were achieved, with a 47% enhancement in Spectral Angle Distance (SAD).
• A 42% reduction in root-mean-square error (RMSE) was observed, indicating higher accuracy.

Conclusions:

• The developed hyperspectral unmixing method effectively utilizes spatial and spectral information.
• The integration of CNN autoencoder and fuzzy clustering offers a robust solution for endmember and abundance estimation.
• This research provides a significant advancement in hyperspectral data analysis, with broad applicability.