A new band selection framework for hyperspectral remote sensing image classification

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Dimensionality Reduction (DR) is crucial for hyperspectral image (HSI) analysis to improve classifier accuracy and manage data redundancy.
• Conventional clustering methods for band selection (BS) struggle with limited data points relative to feature space dimensionality.

Purpose of the Study:

• To propose a novel DR framework for HSI that integrates band selection with spatial features.
• To address limitations of conventional BS methods by employing dual partitioning and ranking for effective information mining.

Main Methods:

• A DR framework combining band selection (BS) via dual partitioning and ranking with a hemispherical reflectance-based spatial filter.
• Classification using a Convolutional Neural Network (CNN) with three-dimensional convolutions on the reduced band subset.

Main Results:

• Achieved high classification accuracies: 99.92% on Indian Pines, 99.94% on Salinas, and 97.23% on KSC.
• Demonstrated effective band selection with low Mean Spectral Divergence values (42.4, 63.75, 41.2) across datasets.
• Outperformed state-of-the-art techniques in quantitative and qualitative evaluations.

Conclusions:

• The proposed DR framework significantly enhances HSI classification accuracy.
• The dual partitioning and ranking approach for band selection is effective in mining crucial information.
• The method shows potential for diverse HSI applications requiring high classification performance.