Deep Learning Models for Spatial Spectral Feature Extraction and Classification in Hyperspectral Images

Abstract

Hyperspectral image (HSI) analysis is prominent to identify and distinguish spec- newlinetrally distinct substances on the earth s surface. However, HSI classification is a challenging task because of the large dimensionality of the data, inadequate datasets, huge data, and limited training samples. Several Deep Learning (DL) based architectures are being explored to resolve the aforementioned challenges and provide significant potential in the processing and classification of HSI data. In recent studies, CNN has been used for HSI data classification. The majority of these algorithms are based on 2D-CNN to extract the spatial and spectral features independently for HSI classification. In contrast, the HSI classification accuracy depends heavily on joint spatial and spectral information. The purpose of the current work reported in this thesis is to develop some new Deep Learning models for improving the accuracy of hyperspectral image classification. Unsupervised and semi-supervised approaches have rarely been explored. newlineThe focus of the approaches in this thesis is to study; the effect of unsupervised fea- newlineture learning both in spatial and spectral domains influence the classification accuracy, the effect of optimization of the parameters in existing DL models, the impact of spatial resolution in extracting joint spatial-spectral features, and the effectiveness of the DL models for feature learning with a limited training set. The proposed work is aiming to enhance the accuracy of pixel classification of HSI by extracting both spatial-spectral newlinefeatures from the HSI data cube efficiently. Four different deep learning models have newlinebeen proposed and published along with rigorous experimental analysis. newlineIn the first contribution, a hybrid CNN spectral partitioning model is proposed for newlinespatial-spectral feature learning and classification of HSI. Initially, PCA is employed to decrease the dimensionality redundancy. In this, 2D-CNN is used as first step for spatial feature extraction and then 3D-CNN is employed for spatia