Augmenting Hyperspectral Image Unmixing Models Using Spatial Correlation, Spectral Variability, And Sparsity

Abstract

Hyperspectral imaging sensors sample sunlight reflected from different targets on Earth's surface by utilising a series of contiguous narrow spectral channels. The higher spectral resolution of hyperspectral images (HSIs) comes at the cost of low spatial resolution; therefore, most pixels may consist of multiple targets. Spectral unmixing algorithms are essential in addressing the issue of low spatial resolution of HSIs by incorporating spatial correlation, spectral variability, and sparsity constraints. Moreover, unmixing methods can be used to measure the fractional abundance of pure materials (called endmembers) in a mixed pixel and are also helpful in enhancing the spatial resolution of HSIs.

In the first part of the thesis, sparse unmixing methods were improved by incorporating high adjacency effects and endmember spectral variability. Traditional total-variation-based sparse unmixing methods avoid high adjacency effects among the neighbouring pixels, which leads to over-smoothing and causes errors in the abundance estimation. A four-directional total-variation spatial regularisation approach is proposed to address these issues, which yields robust results when applied to low signal-to-noise-ratio images. Furthermore, spectral unmixing algorithms analyse the HSI by treating endmembers as independent entities in many remote sensing applications such as agriculture or mineral study. Therefore, traditional methods fail to estimate the fractional abundance of endmembers accurately. An endmember variability-based spectral-spatial weighted sparse regression unmixing method is proposed and demonstrated using a real airborne AVIRIS-NG HSI over the agriculture field, where fractional covers of red and black soil were estimated over sparsely vegetated areas. The experimental finding shows promising results as compared to other methods.

In the second part, the generalised bilinear mixing (GBM) model-based nonlinear unmixing methods were improved. Real HSIs are usually contaminated with complex mixed noises such as Gaussian noise, dead pixels, stripes, impulse noise, etc. The intensity of mixed noise may also vary band-to-band in HSIs, which reduces the accuracy of traditional GBM-based unmixing methods. A computationally efficient bandwise-GBM model is proposed to deal with these issues. The proposed technique reduces computation time while being comparable (and often better) to traditional GBM-based unmixing methods. Furthermore, traditional GBM-based unmixing approaches also reduce unmixing performance by ignoring spatial correlation among the neighbouring pixels. A super-pixel-guided weighted low-rank representation for the robust GBM model is proposed to overcome the above issues. This model employs an entropy rate superpixel segmentation approach to extract homogenous patches in the HSI that underlie the low-rank property. A weighted nuclear norm minimisation approach is introduced for each homogenous patch to estimate the low-rank property, which allocates smaller weights to larger singular values and higher weights to smaller ones. The proposed method significantly improves the fractional abundance estimation by incorporating spatial correlation and sparse noise constraints in the unmixing model.

Finally, spectral unmixing methods are utilised to improve the spatial resolution of HSI by employing high spatial resolution multispectral images (MSIs). Traditional unmixing-based fusion methods avoid noise effects in the modelling, which reduces the accuracy of fusion products. A robust coupled non-negative matrix factorisation is developed for HSI and MSI fusion, incorporating sparse noise effects in the unmixing models of HSI and MSI. Both unmixing problems are coupled by using the sensors' relative spectral response and point spread function.

The above study indicates that the proposed methods achieve robust performance by comprising spatial correlation, spectral variability, and sparsity constraints in the unmixing process.