Effect of Denoising on Vectorized Convolutional Neural Network for Hyperspectral Image Classification

Abstract

The remotely sensed high-dimensional hyperspectral imagery is a single capture of a scene at different spectral wavelengths. Since it contains an enormous amount of information, it has multiple areas of application in the field of remote sensing, forensic, biomedical, etc. Hyperspectral images are very prone to noise due to atmospheric effects and instrumental errors. In the past, the bands which were affected by noise were discarded before further processing such as classification. Therefore, along with the noise the relevant features present in the hyperspectral image are lost. To avoid this, researchers developed many denoising techniques. The goal of denoising technique is to remove the noise effectively while preserving the important features. Recently, the convolutional neural network (CNN) serves as a benchmark on vision-related task. Hence, hyperspectral images can be classified using CNN. The data is fed to the network as pixel vectors thus called Vectorized Convolutional Neural Network (VCNN). The objective of this work is to analyze the effect of denoising on VCNN. Here, VCNN functions as the classifier. For the purpose of comparison and to analyze the effect of denoising on VCNN, the network is trained with raw data (without denoising) and denoised data using techniques such as Total Variation (TV), Wavelet, and Least Square. The performance of the classifier is evaluated by analyzing its precision, recall, and F1-score. Also, comparison based on classwise accuracies and average accuracies for all the methods has been performed. From the comparative classification result, it is observed that Least Square denoising performs well on VCNN.