Multi-Objective Unsupervised Band Selection Method for Hyperspectral Images Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Hyperspectral imaging (HSI) generates vast datasets with high spectral dimensionality, making band selection critical for efficient analysis.
• Traditional band selection methods often use single objectives, potentially overlooking crucial information and leading to suboptimal object detection.
• Intelligent optimization algorithms are vital for addressing the combinatorial complexity of band selection in HSI.

Purpose of the Study:

• To propose a novel multi-objective band selection method for hyperspectral images (HSI).
• To enhance object detection accuracy by considering both band information content and correlation.
• To develop an unsupervised band selection model leveraging a multi-objective cuckoo search algorithm (MOCS).

Main Methods:

• Developed a multi-objective unsupervised band selection model (MOCS-BS) using a multi-objective cuckoo search algorithm (MOCS).
• Incorporated an adaptive strategy based on population crowding degree to optimize Lévy flight.
• Implemented an information-sharing strategy using grouping and crossover to balance global exploration and local exploitation.

Main Results:

• MOCS-BS demonstrated superior performance in hyperspectral image classification compared to state-of-the-art methods (NGNMF, MABC-BS).
• Classification experiments using Random Forest and KNN classifiers validated the effectiveness of the selected band subsets.
• The proposed method proved more effective and robust across four diverse HSI datasets.

Conclusions:

• The MOCS-BS method offers a significant advancement in hyperspectral band selection.
• Considering multiple objectives (information and correlation) leads to more accurate object detection.
• The developed adaptive and information-sharing strategies enhance the optimization process for HSI band selection.