Development of deep learning neural network classifier models for hyperspectral image classification

Abstract

The technology behind hyperspectral imaging is that it collects and processes the data across the electromagnetic spectrum. The images are captured from the high resolution hyperspectral sensors and the aim is to attain the spectrum pertaining to each pixel in the image of a captured scene. These hyperspectral images possess a rich content of spectral information and possess the capability to classify different features than the ordinary optical images. The applicability of HSI is wide and more prominent in agriculture, environmental, military, medical applications and in mining sector. These are not like the ordinary images and HSI images are rich with spectral information and this information reflects the physical structure and the basic chemical configuration of the said object. Considering the importance of these hyperspectral images and the need to classify these images, this thesis has contributed in developing deep learning based neural classifier models and the simulation process was done and the results were reported. The contributions carried out in this thesis are as given below. Feed Forward Neural Networks (FFNN) with multi-layer structure are used more than decades for image processing applications. The two deep learning models developed in this thesis based on FFNN includes the deep back propagation neural network classifier (DBPNN) and the deep radial basis function neural network classifier (DRBFNN) and the layer structure is designed with the convolutional layers for feature extraction. newline