A Hyperspectral Image Classification Approach Based on Feature Fusion and Multi-Layered Gradient Boosting Decision...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep Neural Networks (DNNs) are prevalent for hyperspectral image classification but suffer from overfitting with insufficient data.
• DNNs require substantial data, extensive training time, and robust hardware, limiting their practical application.
• Existing methods often struggle with balancing spectral and spatial feature extraction for accurate classification.

Purpose of the Study:

• To propose a novel feature fusion and multi-layered gradient boosting decision tree (FF-DT) model for hyperspectral image classification.
• To address the limitations of DNNs, including overfitting and high computational demands.
• To improve classification accuracy and efficiency in hyperspectral imaging.

Main Methods:

• Feature fusion incorporating extended morphology profiles (EMPs), and linear/nonlinear multi-scale spatial characteristics.
• Development of a multi-layered gradient boosting decision tree (FF-DT) for classification.
• Experimental validation on the Pavia University, Indiana Pines, and Salinas hyperspectral datasets.

Main Results:

• The FF-DT model demonstrated superior classification accuracy compared to existing methods.
• FF-DT achieved improved training conditions and reduced time consumption.
• Effective fusion of spectral and spatial features contributed to enhanced classification performance.

Conclusions:

• The proposed FF-DT model provides an effective and efficient alternative for hyperspectral image classification.
• FF-DT mitigates overfitting issues common with DNNs, particularly in data-scarce scenarios.
• This approach offers a promising direction for advancing hyperspectral data analysis.