Zhou, Chao; (2023) HSI Data Unmixing Using Machine Learning Techniques. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text UCL_PhD_Thesis_Final_copies.pdf - Accepted Version Download (11MB) | Preview

Abstract

Hyperspectral image (HSI) unmixing is a challenging research problem that tries to identify the constituent components, known as endmem- bers, and their corresponding proportions, known as abundances, in the scene by analysing images captured by hyperspectral cameras. Recently, many deep learning based unmixing approaches have been proposed with the surge of machine learning techniques, especially convolutional neural networks (CNN). However, most of these meth- ods rely on the general-purpose CNN structures and it is unclear how to design an efficient network for unmixing purposes. In this work, we first address the structural issue by proposing new unmixing net- works that leverage algorithm unrolling techniques to the Alternating Direction Method of Multipliers (ADMM) solver of a constrained sparse regression problem underlying a linear mixture model. However, like many other methods in the literature, there is no guarantee that the net- work could generate physically meaningful unmixing results. To solve this problem, we proposed a novel blind unmixing network using dou- ble DIP techniques (BUDDIP) which consists of two DIP sub-networks to estimate the endmember and abundance respectively, which are coined as EDIP and ADIP. The network is trained in an end-to-end manner by minimizing a novel composite loss function. Finally, we pro- pose a novel unmixing algorithm that can address both issues, simul- taneously. Specifically, we first propose a novel MatrixConv Unmixing (MCU) Model for endmember and abundance estimation, respectively, which can be solved via certain iterative solvers. We then unroll these solvers to build two unfolding based sub-networks, which are coined as UEDIP and UADIP, to generate the estimation of endmember and abundance, respectively. The overall network is then constructed by assembling these two sub-networks. To further improve the unmixing quality, we also add explicitly a regulariser for endmember and abun- dance estimation, respectively. Experimental results on both synthetic and real HSI data show that the proposed method achieves state-of- the-art performance compared to other unmixing approaches.

Download activity - last month

Export as CSVJSONXML

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item