Sparse Feature Learning of Hyperspectral Imagery via Multiobjective-Based Extreme Learning Machine.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral images (HSI) offer rich spectral and spatial data.
• Extreme Learning Machine (ELM) is common for HSI analysis but struggles with sparse feature learning due to random hidden layers.
• Existing methods often face challenges in achieving optimal sparse feature representation.

Purpose of the Study:

• To propose a novel unsupervised sparse feature learning approach for HSI analysis.
• To enhance feature extraction in HSI by addressing limitations of classical ELM.
• To improve the separability and sparsity of learned features.

Main Methods:

• Developed Evolutionary Multiobjective-based ELM (EMO-ELM) for unsupervised sparse feature learning.
• Formulated ELM Autoencoder (ELM-AE) construction as a multiobjective optimization problem (sparsity and reconstruction error).
• Utilized Evolutionary Multiobjective Optimization (EMO) and a curvature-based method to select optimal trade-off solutions.

Main Results:

• EMO-ELM demonstrated improved sparsity and reconstruction error balance.
• The proposed method showed reduced susceptibility to local minima and fewer trainable parameters compared to gradient-based approaches.
• Experimental results on real HSIs confirmed superior feature separability and sparsity.

Conclusions:

• EMO-ELM offers an effective approach for unsupervised sparse feature learning in HSI analysis.
• The method provides a robust alternative to traditional ELM and gradient-based autoencoders.
• EMO-ELM advances HSI feature extraction, leading to better classification and analysis outcomes.