Learning representations of hyperspectral data for pixel‐wise image classification

Abstract

This study thoroughly examines the application and efficacy of Silhouette Distance Loss (SDL) as a novel approach for representation learning, specifically tailored to the classification of hyperspectral data. Hyperspectral imaging (HSI) classification presents unique challenges, primarily due to its inherently high dimensionality, complex spectral characteristics, and the general scarcity of labelled training examples. Addressing these challenges is essential with regards to improving classification accuracy. Thus, the core goal of this Final Year Project (FYP) is to systematically evaluate the effectiveness of SDL for pixel‐wise classification tasks on hyperspectral data. Said evaluation makes use of a variety of publicly available corpora, in order to ensure robustness and reproducability of the respective findings. Moreover, this research endeavours to benchmark the proposed method’s performance against State‐of‐the‐Art (SOTA) models which are currently prevalent in hyperspectral data classification tasks. The comparison makes use of average classification accuracy as the key metric. This is valid for such a case since the utilised training and validation sets are both balanced. By juxtaposing the results of the proposed approach against established models, this FYP aims to elucidate the strengths and limitations of the SDL function, and Supervised Contrastive Learning (SCL) frameworks in general. Ultimately, the study contributes significantly towards advancing current methodologies employed in hyperspectral image classification, thereby laying foundational groundwork that future research is able to build upon. This may potentially result in the development of more accurate, computationally efficient, and scalable classification algorithms, which are more viable for real‐world application.