Enhancing feature learning of hyperspectral imaging using shallow autoencoder by adding parallel paths encoding | JoVE Visualize

Area of Science:

Remote Sensing
Computer Vision
Machine Learning

Background:

Conventional image formats lack the spectral detail of Hyperspectral Imaging (HSI).
HSI captures hundreds of spectral features, but this richness introduces data redundancy and computational complexity for deep learning.
Effective feature extraction is crucial for HSI analysis, requiring methods that handle both linear and non-linear spectral characteristics.

Purpose of the Study:

To develop and evaluate a novel deep learning model for enhanced feature extraction in Hyperspectral Imaging.
To compare the performance of the proposed Dual-Path Autoencoder (D-Path-AE) against traditional linear methods and a Naïve Autoencoder (Naïve AE).
To assess the impact of dimensionality reduction techniques on the classification accuracy of HSI datasets.

Main Methods:

Proposed the Dual-Path Autoencoder (D-Path-AE) model, featuring concurrent encoding pathways for enhanced non-linear feature acquisition.
Implemented a down-sampling strategy within D-Path-AE to mitigate bias towards majority classes in unbalanced datasets.
Compared D-Path-AE with linear methods (PCA, ICA) and Naïve AE for dimensionality reduction, followed by classification using Decision Tree, SVM, and KNN algorithms.

Main Results:

The D-Path-AE model demonstrated superior performance in dimensionality reduction and feature extraction compared to linear methods and Naïve AE.
Classification accuracy using D-Path-AE reached up to 98.31% Overall Accuracy on the Pavia Center dataset with the KNN classifier.
The proposed model effectively captures complex, non-linear spectral features essential for accurate HSI classification.

Conclusions:

The D-Path-AE model offers a significant advancement in processing Hyperspectral Imaging data, overcoming limitations of existing methods.
The concurrent encoding pathways and down-sampling strategy contribute to robust and accurate feature extraction, even with imbalanced datasets.
This approach enhances the classification capabilities for HSI, paving the way for more effective applications in various fields.