ISBDD Model for Classification of Hyperspectral Remote Sensing Imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Traditional Diverse Density (DD) algorithms struggle with low classification accuracy in hyperspectral imagery due to mixed pixels.
• Existing DD algorithms learn feature vectors that may not effectively represent ground cover types.

Purpose of the Study:

• To introduce a novel Instance Space-based Diverse Density (ISBDD) model for enhanced hyperspectral image classification.
• To address the limitations of feature vector representation in the standard DD algorithm.

Main Methods:

• The ISBDD model computes Diverse Density (DD) values for each pixel, rather than learning a single feature vector.
• Classification is performed directly based on the computed DD values for individual pixels.
• The model was evaluated using airborne hyperspectral data from AVIRIS and PHI sensors.

Main Results:

• The ISBDD model achieved high classification accuracy, reaching 97.65% on AVIRIS data and 89.02% on PHI data.
• Kappa coefficients of 0.97 (AVIRIS) and 0.88 (PHI) demonstrate significant classification performance.
• The ISBDD model effectively handles interference from mixed pixels in hyperspectral data.

Conclusions:

• The Instance Space-based Diverse Density (ISBDD) model offers a superior approach to hyperspectral image classification compared to traditional DD methods.
• The pixel-wise DD value computation strategy enhances the representation of ground cover types.
• The proposed ISBDD model shows strong potential for accurate analysis of hyperspectral remote sensing data.