Deep convolution neural network approach for analysis of vegetation hyperspectral imagery

Abstract

Image information is collected using hyperspectral imaging technology in a number of narrow continuous bands of the electromagnetic wave spectrum. Hyperspectral images (HSI) hold information about the spectral reflectance of objectives and spatial resolution. They can be used to classify abundant spectral and spatial details for target assessment. The primary research question of this thesis is hyperspectral remote sensing image classification. Deep methods have become highly advanced thanks to advancements in high-performance computation and data acquisition. The immense majority of existing deep learning methods use hyperspectral images to understand feature maps in simple convolutional or fully connected layers. One of the most effective approaches for HSI classification is deep Convolutional Neural Networks (CNNs). All original pixels and obtained features are considered equal in this learning approach. Technologically advanced deep neural networks consistently outperform and establish new archives in all image processing issues, owing to their superior ability to extract feature representations via many nonlinear transformations. Moreover, three main impediments stand in the way of Convolutional Neural Networks (CNNs) being used explicitly for HSI classification. To newline