Hyperspectral agricultural mapping using support vector machine-based endmember extraction (SVM-BEE)

Area of Science:

• Remote Sensing
• Machine Learning
• Hyperspectral Imaging

Background:

• Extracting endmembers from vegetated areas in remotely-sensed images is challenging.
• Previous studies established Support Vector Machine-Based Endmember Extraction (SVM-BEE) as noise-tolerant and capable of semi-automatic endmember estimation.

Purpose of the Study:

• To compare the efficacy of SVM-BEE against N-FINDR and SMACC algorithms for endmember extraction.
• To evaluate SVM-BEE's performance on a real, agricultural hyperspectral scene.

Main Methods:

• Application of the Support Vector Machine-Based Endmember Extraction (SVM-BEE) algorithm.
• Comparison with N-FINDR and SMACC algorithms.
• Analysis of endmember extraction accuracy using Spectral Angle Mapper (SAM) classification and linear spectral unmixing.

Main Results:

• SVM-BEE successfully extracted vegetation and other endmembers from all image classes, unlike N-FINDR and SMACC.
• SVM-BEE demonstrated consistency in endmember estimation across multiple trials.
• SAM classifications using SVM-BEE endmembers were significantly more accurate than those using N-FINDR and SMACC endmembers.

Conclusions:

• SVM-BEE is a robust and accurate algorithm for semi-autonomous endmember extraction from hyperspectral images, particularly in complex agricultural environments.
• SVM-BEE offers superior performance compared to N-FINDR and SMACC for vegetation and diverse endmember identification.
• The accuracy of SVM-BEE facilitates improved classification and spectral unmixing results.