Latent Dirichlet Variational Autoencoder: a novel approach for hyperspectral image analysis and pixel unmixing exploring deep learning architectures

This thesis investigates deep learning-based methods for hyperspectral image analysis, focusing on pixel unmixing and classification tasks. Recognizing the challenges of high data dimensionality and limited labeled data availability, this research proposes innovative techniques to improve both the accuracy and efficiency of hyperspectral image interpretation. Initially, the impact of spectral band normalization and outlier removal on image segmentation scale selection is explored, leading to a robust method for Object- Based Image Analysis (OBIA). Subsequently, the research delves into the application of autoencoders for spectral dimensionality reduction, culminating in a comparative analysis demonstrating their efficacy in preserving crucial information for classification while achieving significant data compression. Building upon these findings, this thesis introduces the Latent Dirichlet Variational Autoencoder (LDVAE), a novel architecture specifically designed for hyperspectral pixel unmixing. The LDVAE model introduces an approach to hyperspectral pixel unmixing by incorporating a Dirichlet distribution within its latent space. This design enables LDVAE to effectively model abundance vectors, satisfying the inherent sum-to-one and non-negativity constraints, while simultaneously learning a low-dimensional representation of endmember spectra. The generative nature of LDVAE further allows for the synthesis of new hyperspectral pixels by reconstructing spectra from the learned Dirichlet distributions. Evaluations on benchmark datasets demonstrate that LDVAE achieves state-of-the-art performance in both endmember extraction and abundance estimation tasks. This thesis also introduces additional contributions to hyperspectral unmixing, addressing the challenges posed by limited labeled data and the potential for exploiting spatial information. Specifically, we extend the Latent Dirichlet Variational Autoencoder (LDVAE) framework in two key directions. First, recognizing the scarcity of labeled data and the inherent spatial coherence within hyperspectral imagery, we develop an iterative analysis-synthesis approach using the LDVAE (iLDVAE). This novel framework facilitates automatic endmember extraction and refines the unmixing process iteratively. Second, acknowledging the importance of spatial context, we propose SpACNN-LDVAE, which integrates the LDVAE with Convolutional Neural Networks (CNNs) and spatial attention mechanisms. This architecture effectively captures local spatial relationships between pixels, yielding a more informative latent representation for improved unmixing performance. The SpACNN-LDVAE enhances both endmember extraction and abundance estimation accuracy, particularly in scenes exhibiting complex spatial structures. These contributions provide robust and efficient tools for hyperspectral image analysis, offering potential benefits across various application domains, including agriculture, forestry, mineralogy, and environmental monitoring.

Name thesis:degree_name

Doctor of Philosophy (PhD)

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Ontario Institute of Technology

Year dc:date.issued

2025

Author dc:creator

• Mantripragada, Kiran

Advisor dc:contributor.advisor

• Qureshi, Faisal Z.

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10155/1972

OAI identifier oai:identifier

oai:ontariotechu.scholaris.ca:10155/1972