Dimension reduction using PCA and LDA and 3D Convolutional Neural Network for Hyperspectral Image classification

Abstract

A Hyperspectral image is a collection of several hundreds even thousands of reflectance bands of different wavelengths that is captured by a satellite sensor. HSI has provided significant opportunities for material identification and classification because of its ability to contain rich information. But the processing of hyperspectral image is a challenging task because of its high dimensionality and data redundancy.That is why dimension reduction is necessary otherwise high dimensional data suffers from Hughes Phenomenon. Dimension can be reduced by using feature selection and feature extraction approaches.In here, supervised Linear Discriminant Analysis(LDA) & unsupervised Principal Component Analysis (PCA) have been used as preprocessing step .Then 3D CNN has been applied to classify Hyperspectral Image.

Dataset

(Indian Pines)
Image size: 145×145×200. 200 spectral bands. Wavelength range: 0.4−2.5𝜇𝑚. 16 classes.

University of Pavia
Image size: 610×340×103. 103 spectral bands. 9 classes.

WorkFlow

The steps of the proposed methods are:

1.Perform PCA to reduce dimension from the input dataset.
2.Perform 3D CNN to classify.
3.Perform LDA to reduce dimension from the input dataset.
4.Perform 3D CNN to classify.

Kaggle Code Link

For Indian Pine dataset:
PCA+3DCNN
LDA+3DCNN

For Pavia University dataset:
PCA+3DCNN
LDA+3DCNN

Classification Accuracy

For Indian Pine dataset:
classification_report of PCA+3dCNN of Indian Pine
classification_report of LDA+3dCNN of Indian Pine

For Pavia University dataset:
classification_report of PCA+3dCNN of Pavia University
classification_report of LDA+3dCNN of Pavia University

[Supplementary Material](HybridSN Exploring 3-D–2-D CNN FeatureHierarchy for HSI Classification.pdf)