Classification of Hyperspectral Image using Convolutional Neural Network to Detect Coastal Water Features

Abstract

In this study, classification method of hyperspectral images using convolutional neural network (CNN) is introduced. In general, two-dimensional image data are used as the inputs of CNN classification model. Therefore, 2-D input data for each of the image pixels under specific area were prepared by using the reflectance values in all spectral bands. Thus, two-dimensional grey scale image of 1296 pixels was generated for each pixel under the region of interest (ROI) area obtained from PIKA-II hyperspectral camera. A CNN network was constructed to distinguish human, sea, sand, ship and rock in the image. The CNN model training based classification for each target was performed which resulted in highest classification accuracy for the sand (86.4%).ch of the image pixels under specific area were prepared by using the reflectance values in all spectral bands. Thus, two-dimensional grey scale image of 1296 pixels was generated for each pixel under the region of interest (ROI) area obtained from PIKA-II hyperspectral camera. A CNN network was constructed to distinguish human, sea, sand, ship and rock in the image. The CNN model training based classification for each target was performed which resulted in highest classification accuracy for the sand (86.4%).