Loading Events

« All Events

  • This event has passed.

Ph.D. Thesis {Colloquium}: CDS: 25th August 2022 : “Augmenting hyperspectral image unmixing models using spatial correlation, spectral variability, and sparsity”.

25 Aug @ 3:00 PM -- 4:00 PM

DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES

Ph.D. Thesis Colloquium 

_________________________________________________________________________________________________

Speaker                 : Mr. Touseef Ahmad

S.R. Number         : 06-18-01-10-12-13-1-10627

Title                       :”Augmenting hyperspectral image unmixing models using spatial correlation, spectral variability, and sparsity”

Research Supervisor:  Prof. Soumyendu Raha

Date & Time         : August 25, 2022 (Thursday), 03:00 PM

Venue                     : The Thesis Colloquium will be held on MICROSOFT TEAMS. Please click on the following link to join the colloquium:

_____________________________________________________________________________________________________

 

ABSTRACT

Modern   hyperspectral sensors can sample the sunlight reflected from a target on the ground with hundreds of contiguous narrow spectral channels. However, the increased spectral resolution comes at the price of a lower spatial resolution of hyperspectral imagery (HSI). In this context,  spectral unmixing methods play a major role in dealing with the problem of reduced spatial resolution in hyperspectral sensors. Sub-pixel information derived using the unmixing method can be utilised for estimating the abundance of pure materials present in the mixed pixel and for enhancing the spatial resolution of hyperspectral images. Keeping this in mind, this thesis investigates the possibilities for dealing with the problem of HSI’s reduced spatial resolution by taking spatial   correlation, spectral variability, and sparsity constraints into     account while unmixing the HSI. At first, efforts were made to improve   the spectral unmixing accuracy for both low and high SNR HSIs using a four-directional spatial regularisation    approach,  which utilises spatial correlation for sparse unmixing (SUnSAL-4DTV). The proposed method can deal with the problem of traditional total- variation  (TV) methods, which avoid high adjacency effects among the neighbouring pixels while unmixing. This leads to over-smoothing, which causes errors in the abundance estimation. The proposed method produces robust results when applied to low SNR HSIs. In the next part, spectral variability induced errors were minimised by considering the spectral variability within the endmember class. The majority of spectral unmixing algorithms analys  the  HSI by

treating   endmembers as independent   entities. Specifically, endmembers in agricultural fields are spatially and temporally varied, with strong spatial correlation among neighbouring pixels. Traditional methods fail to estimate the fractiona  abundance of red and black soil over sparsely vegetative areas. In this situation, a new method called spectral-spatial weighted unmixing with spectral variability (SSWU-SV) is suggested. It uses   both spectral and spatial weighting factors to make the solution even    more sparse. Real HSIs are contaminated by mixed complex noises such as Gaussian noise, impulse noise, dead pixels, lines, or stripes, which    affect the unmixing results. The intensity of mixed noise may also vary bandwise in HSI, which reduces the accuracy of traditional generalised    bilinear mixing (GBM)-based unmixing methods. To address this, a fast and robust variant of the nonlinear unmixing approach using band-wise GBM (NU-RBGBM) is proposed in this study, which reduces computation time while being comparable  (and often better) than traditional methods in terms of accuracy.   Traditional GBM-based unmixing approaches reduce unmixing performance by ignoring HSI spatial correlatio  and mixed noise. To address mixed noise and spatial correlationamong neighbouring targets, a super-pixel-guided weighted low-rank     representation model was proposed for robust GBM (RGBM-SG- WLRR).   The proposed model uses the nuclear norm to introduce low-rank property    and allocates smaller weights to larger singular values and higher weights to smaller singular values and accommodates mixed complex nois  and spatial correlation while performing unmixing. Finally, linear spectral unmixing was employed to improve the spatial resolution   of the HSI.  A robust   coupled non-negative matrix factorization (RCNMF) was developed for fusing HSI with high spatial resolution multispectral images (MSI). In the proposed method, unmixing problems for HSI and MSI were coupled using the relative spectral response and point spread function of the sensors. The introduced method   is robust as compared to the existing state-of-the-art methods when applied to noisy HSIs and MSIs fusion. Experimental results reveale  that  the   proposed methods   were   able   to   achieve Robust performance by incorporating spatial correlation, spectral variability, and sparsity constraints.

Details

Date:
25 Aug
Time:
3:00 PM -- 4:00 PM

Venue

Online