DEVISE AND DEVELOPMENT OF DEEP LEARNING MODELS FOR STRATIFICATION OF HYPERSPECTRAL IMAGES

Abstract

The classification of hyperspectral images (HSI) into categories that correlate to various land cover sorts such as water bodies, agriculture and urban areas, has gained significant attention in research due to its wide range of applications in fields such as remote sensing, computer vision, and more. This has led to the development of various deep learning models that include supervised and semi-supervised, for HSI classification. Among the aforemen tioned classes of models, the supervised networks have evolved to achieve almost perfect classification accuracy. Nevertheless, the process of obtaining labelled samples continues to pose a challenge in HSI classification, as the labelling remains a manual, time-consuming, and labour-intensive task, which necessitates the expertise of individuals to identify and label each pixel in the image. Furthermore, most of the existing supervised models are computationally slower due to the massive computations involved. To address these prob lems of the existing works, in this thesis, we propose five new deep learning-based models for HSI classification. In our first work, we propose a novel lightweight network, Xcep-Dense, a hybrid clas sification model that combines the core benefits of the extreme version of inception and dense networks. The Xception network employs depth-wise separable convolutions, and the 3D slicing phenomenon, which requires fewer parameters, is computationally efficient and provides excellent classification accuracy. The proposed network is configured with dense network and optimized to alleviate overfitting. Xcep-Dense’s performance is vali dated using two benchmark hyperspectral datasets, Indian Pines and Salinas. In our second work, we propose a Siamese network based deep learning model which implements one shot classification model and can work with limited samples and/or im balanced samples. The proposed Siamese network has a handcrafted feature generation network that extracts discriminative features from the image. Experimental findings demon iii strate that the proposed network is capable of improving the classification performance with an overall accuracy, with a small scale training data. To tackle the labelled limited samples problem, the third work introduces a novel semi supervised network constructed with an autoencoder, siamese action, and attention layers that achieves excellent classification accuracy. The proposed convolutional autoencoder is trained using the mass amount of unlabelled data to learn the refinement representa tion referred to as 3D-CAE. The added siamese network improves the feature separability between different categories and attention layers improve classification by focusing on dis criminative information and neglecting the unimportant bands. The efficacy of the proposed model’s performance was assessed by training and testing on both same-domain as well as cross-domain data and found to achieve 91.3 and 93.6 for Indian Pines and Salinas, respec tively. In our fourth work, we integrate autoencoders and Generative Adversarial Networks for enhancing feature representations and mitigating the constraints imposed by limited la beled data. Leveraging the power of semi-supervised learning paradigms, this innovative approach offers substantial progress in feature extraction, data augmentation, and classifi cation accuracy. It extends beyond traditional hyperspectral image classification boundaries by addressing zero-shot learning and integration of text embeddings to enrich feature rep resentations. The outcome is a precise classification framework that accommodates the intricacies of both same-domain and cross-domain datasets, ultimately pushing the bound aries of hyperspectral image classification. In our final work, we present an innovative semi-supervised framework that harmo niously combines unsupervised feature learning with the employment of graph-based con volutional neural networks (GCNs). Our approach harnesses the latent knowledge hidden within vast pools of unlabelled HSI data using autoencoders, which extract meaningful features. These features are then incorporated into a GCN-based architecture, leveraging spatial relationships among neighboring pixels. The fusion of autoencoder-based learning and graph-based techniques enables our model to achieve excellent classification accuracy, even in scenarios with minimal labelled samples.