Hyper Spectral Image Analysis Using Wavelet Based Deep Learning Techniques

Abstract

In Hyper Spectral Imaging (HSI), sensors gather detailed spectral data across newlinemany narrow spectral bands, leading to high dimensionality. This dimensionality newlineproblem has a notable impact on HSI classification. Our study investigates the newlineapplication of a 2D-Convolutional Neural Network (CNN) for analyzing HSI data. newlineRecent works also show the use of 2D-CNN for HSI classification. Nevertheless, newlineachieving high-performance HSI classification relies on both spatial and spectral newlineinformation. Standard 2D-CNNs might struggle to effectively merge spectral newlineand spatial features for HSI classification. In this study, we introduce innovative newlinespectral-spatial techniques to improve the architecture of 2D-CNNs for accurate newlinecategorization and analysis of hyper spectral data. newlineInitially, this study explores combining a lifting-scheme-based DiscreteWavelet newlineTransform (DWT) with a 2D-CNN, known as the Discrete Wavelet 2D-CNN newlinemodel. In HSI processing, spectral-spatial feature extraction (FE) plays a vital newlinerole. By integrating DWT with 2D-CNN, emphasis on spectral-spatial FE newlinemay be enhanced. This study s uniqueness is in evaluating the accuracy of three newlineDWTs: Haar, Daubechies 4-tap orthogonal filter (D4), and Cohen-Daubechies- newlineFeauveau 9/7-tap bi-orthogonal filter (CDF-9/7-wavelet), for spectral FE. The newlinespectral features obtained from wavelet decomposition are combined with a 2DCNN, newlineretaining spatial information, to form a spectral-spatial feature vector for newlineclassification. Evaluation of the Discrete Wavelet 2D-CNN model s accuracy with newlinedifferent DWTs revealed that the D4 wavelet-based model achieved superior performance newlinecompared to other configurations. newlineIn a second step, an improved Discrete Wavelet 2D-CNN Model with Wavelet newlineiv newlineAndhra University, Visakhapatnam newlineNeural Networks (WNN) is presented as a 2-Stage Hybrid Model to enhance the newlinediscriminative power of classification models on HSI. The model uses a Discrete newlineWavelet 2D-CNN to extract spectral-spatial feature vector and a WNN classifier newlineto classify HSI using extracted features. Then