Random N-finder (N-FINDR) endmember extraction algorithms for hyperspectral imagery

Area of Science:

• Remote Sensing
• Signal Processing
• Data Analysis

Background:

• N-finder algorithm (N-FINDR) is a standard for endmember extraction.
• N-FINDR faces challenges including endmember determination, computational complexity, and irreproducible results due to random initializations.

Purpose of the Study:

• To re-design the N-FINDR algorithm to overcome implementation issues.
• To develop a reproducible and automated endmember determination method.

Main Methods:

• Implemented N-FINDR as a random algorithm (RN-FINDR) where each run is a realization.
• RN-FINDR terminates when the intersection of consecutive runs stabilizes, automatically determining the number of endmembers (p).

Main Results:

• RN-FINDR ensures that true endmembers appear in all realizations, regardless of initial random endmembers.
• The algorithm automatically determines the optimal number of endmembers (p) through the intersection stability criterion.
• Validation performed using synthetic and real hyperspectral image data.

Conclusions:

• RN-FINDR provides a robust and reproducible solution for endmember extraction.
• The method addresses critical limitations of the original N-FINDR algorithm.
• RN-FINDR demonstrates significant utility in hyperspectral image analysis applications.