Work

Modélisation probabiliste et apprentissage profond pour le démélange hyperspectal

Year: 2025
Type: dissertation
Abstract: Hyperspectral unmixing is a crucial tool for analyzing and extracting features from hyperspectral remote sensing data. It aims to estimate, for each pixel, the spectral signatures of pure materials, known as endmembers, along with their associated fractional abundances. Over the past decade, it has been a highly active research area, with more than a hundred of publications annually since 2011. This reflects the growing interest in developing advanced methods for accurately decomposing mixed spectral signatures. In line with the broader trends in image processing and computer vision, where deep learning and neural networks have become mainstream, many recent unmixing approaches have embraced these techniques. This document presents two contributions to the hyperspectral unmixing problem from a deep learning perspective. In our first contribution, we place ourselves under the linear mixing model, and we address the challenge of limited labeled training data for machine and deep learning-based unmixing algorithms. Although supervised learning is among the most effective training frameworks for computer vision tasks, its application in hyperspectral unmixing has been limited due to the scarcity of labeled hyperspectral datasets. To overcome this, we propose a pipeline for synthetically generating labeled unmixing data directly from the hyperspectral image to be unmixed, enabling the self-supervised training of deep neural networks. We demonstrate the effectiveness of our data generation approach by training an unrolled unmixing method, LPALM, which achieves superior unmixing results compared to other state-of-the-art conventional and deep learning-based methods that rely on unsupervised training. In our second contribution, we explore the use of variational autoencoders (VAEs) to address the unmixing problem under the assumption of spectral variability. VAEs, through their latent representations, leverage a key property of spectral variability: the variability of endmember signatures is confined to a low-dimensional manifold within the high-dimensional spectral space. This characteristic suggests that the spectral signatures of most materials are influenced by a limited set of physico-chemical parameters. By capturing this low-dimensional structure, VAEs offer a powerful framework for modeling and addressing spectral variability in hyperspectral unmixing. Building on previous work that employed Beta distributions to model endmember variability, we introduce a novel VAE model with a Beta-distributed latent representations called Beta-latent VAE. %, and first assess its ability to capture the endmember distribution on a series of materials extracted from a hyperspectral image.We motivate the use of the Beta-latent VAE framework for the endmember variability estimation task through a series of experiments, and then build on this framework to propose three variations of VAE-based unmixing networks: two models operating at the pixel level and one spatial-spectral model. For the two last models, we also leverage on Dirichlet-distributed VAE framework to estimate our abundances via Dirichlet distributions, leading us to fully probabilistic VAE-based unmixing networks. We showcase the performance of our unmixing networks in comparison to other state-of-the-art variability-accounting unmixing methods on synthetic and real data and show very interesting results.
Author: Hadjeres, Rassim
Language: English
Open Access status: green
OpenAlex ID: W7112533862
Cites:
Cited by:
Related to: 10