Dimensionality Reduction of Remotely Sensed Hyperspectral Image forClassification using PCA with Autoencoder Technique

摘要

Hyperspectral Imagery (HSI) is widely used in the application domains such as agriculture,environment, forestry and geology for the identification and observations which demands the efficientclassification accuracy. The supervised classification is a challenging task due to limited number of availabletraining samples compared to large number of spectral bands. This phenomena reduces the classificationaccuracy. To overcome this problem, the dimensionality reduction preprocessing step is adopted. This processreduces the number of spectral bands which leads to decrease in computational complexity and enhancementin classification accuracy. In this study, AEPCA (Auto Encoder and Principle Component Analysis) methodis proposed for dimensionality reduction of HSI. The performance of AEPCA is evaluated against AE(Autoencoder) and PCA (Principle Component Analysis) method. The dimensionally reduced components areclassified using CNN (Convolutional Neural Network) based classifier. The proposed model of dimensionalityreduction demonstrates superior classification accuracy due to effective combination of characteristics of AEand PCA. The noisy or corrupted pixels are recovered by AE Model and high dimensional image is representedby efficient fewer number of principle components by PCA is the potential advantage of AEPCA Model.