Linear vs. Nonlinear Extreme Learning Machine for Spectral-Spatial Classification of Hyperspectral Images.

Area of Science:

• Machine Learning
• Remote Sensing
• Computer Vision

Background:

• Extreme Learning Machine (ELM) offers good performance but struggles with hyperspectral image (HSI) classification due to its lack of spatial information utilization.
• Spatial context is crucial for accurate HSI classification, yet traditional ELM methods do not leverage this information.

Purpose of the Study:

• To propose a novel framework for spectral-spatial classification of HSI by combining ELM with loopy belief propagation (LBP).
• To improve the recognition rate of HSI classification by incorporating spatial information into the ELM framework.
• To evaluate the effectiveness of linear ELM (LELM) against nonlinear ELMs for HSI spectral-spatial classification.

Main Methods:

• A new framework combining ELM with loopy belief propagation (LBP) for spectral-spatial classification of HSI is proposed.
• Linear ELM (LELM) was identified as a superior choice over nonlinear ELMs for this specific task through extensive analysis.
• Marginal probability distribution, utilizing all HSI information, is learned using LBP.

Main Results:

• The proposed spectral-spatial classification method maintains the fast processing speed characteristic of ELM.
• A significant improvement in classification accuracy is achieved compared to methods that do not utilize spatial information.
• Experiments on benchmark HSI datasets (Indian Pines, Pavia University) confirm the method's strong performance.

Conclusions:

• The integration of ELM with LBP provides an effective solution for spectral-spatial HSI classification.
• Linear ELM demonstrates better suitability than nonlinear variants for incorporating spatial context in HSI analysis.
• The proposed framework offers a computationally efficient and highly accurate approach for HSI classification tasks.