Hyperspectral unmixing (HU) is a crucial step in the hyperspectral image (HSI) analysis. It aims at decomposing the observed spectrum at each pixel into a collection of constituent endmembers, weighted by their abundances. In a close spatial neighborhood, due to spatial autocorrelation, the abundances of an endmember tend to be similar to each other. Therefore, the spectra at the neighboring pixels are also very much alike. Recently, many studies have focused on graph-regularized nonnegative matrix factorization (NMF) approaches to address the blind HU problem. This is mainly because graph structures allow spatial-spectral relations to be effectively embedded in the NMF framework. This article takes one step further by defining each abundance map as a signal on a suitable graph, which enables the HU problem to be analyzed for the first time from a graph signal processing perspective. This article introduces a novel Laplacian regularizer based on the ₁-norm, where graph spectral analysis is utilized to show that the regularizer has natural piecewise smooth (PWS) signal promotion and noise rejection capabilities. A graph-based blind HU algorithm is developed by incorporating this regularizer and an l1/2-abundance sparsity constraint into the NMF problem. Since the featured regularizer exploits the PWS property of abundance maps, the proposed method is effective in HU. The experiments conducted on synthetic and real HSIs demonstrate that it has superior performance compared with several popular HU algorithms.
4 December 2023 in Computers in Biology and Medicine
Deep learning MRI reconstruction methods are often based on Convolutional neural network (CNN) models; however, they are limited in capturing global correlations among image features due to the intrinsic locality of the convolution operation. Conversely, the recent vision transformer models (ViT) are capable of capturing global correlations by applying self-attention operations on image patches. Nevertheless, the existing transformer models for MRI reconstruction rarely leverage the physics of MRI. In this paper, we propose a novel physics-based transformer model titled, the Multi-branch Cascaded Swin Transformers (McSTRA) for robust MRI reconstruction. McSTRA combines several interconnected MRI physics-related concepts with the Swin transformers: it exploits global MRI features via the shifted window self-attention mechanism; it extracts MRI features belonging to different spectral components via a multi-branch setup; it iterates between intermediate de-aliasing and data consistency via a cascaded network with intermediate loss computations; furthermore, we propose a point spread function-guided positional embedding generation mechanism for the Swin transformers which exploit the spread of the aliasing artifacts for effective reconstruction. With the combination of all these components, McSTRA outperforms the state-of-the-art methods while demonstrating robustness in adversarial conditions such as higher accelerations, noisy data, different undersampling protocols, out-of-distribution data, and abnormalities in anatomy.
30 September 2019 in Journal of the National Science Foundation of Sri Lanka
Continuous mapping of vegetation zones is vital to identify ecological changes as early as possible. In this paper, a novel, semi-supervised method is discussed that can be used to map boundary between two vegetation zones using satellite hyperspectral data. For this task, a hyperspectral image strip of Sri Lanka obtained by the Earth Observing-1 (EO-1) satellite's Hyperion sensor is used. In the proposed method, a Maximum Likelihood Classification is used to identify the pure vegetation pixels. Then, based on the degree of correlation among pixels containing vegetation, at different spatial coordinates, a possible region of existence of a boundary between the two major vegetation zones is obtained. Thereafter, a systematic procedure consisting of Fisher's Discriminant Analysis (FDA) and Spectral Clustering is used to separate the vegetation pixels in the narrowed down region into two vegetation zones. The results are validated through data obtained from the Survey Department of Sri Lanka.
Hyperspectral Imaging (HSI) utilises the reflectance information of a large number of contiguous spectral bands to solve various problems. However, the relative proximity of spectral signatures among classes can be exploited to generate an adaptive hierarchical structure for HSI classification. This enables a level by level optimisation for clustering at each stage of the hierarchy. The Umbrella Clustering algorithm, introduced in this work, utilises this premise to significantly improve performance compared to non-hierarchical algorithms which attempt to optimise clustering globally. The key feature of the proposed methodology is that, unlike existing hierarchical algorithms which rely on fixed or supervised structures, the proposed method exploits a mechanism in spectral clustering to generate a self-organised hierarchy. The algorithm gradually zooms into the feature space to identify levels of clustering at each stage of the hierarchy. The results further demonstrate that the generated structure tallies with human perception. In addition, an improvement to Linear Discriminant Analysis (LDA) is also introduced to further improve performance. This modification maximises the pairwise class separation in the feature space. The entire algorithm includes this modified LDA step which requires a certain amount of class information in terms of features, at the training phase. The classification algorithm which incorporates all novel concepts was tested on the HSI data set of Pavia University as well the database of Common Sri Lankan Spices and Adulterants in order to assess the versatility of the algorithm.
Hyperspectral image analysis has been an efficacious method utilized in remote sensing applications as it has the ability to discover richer information due to the presence of multiple spectral bands representing data, as opposed to the three bands in tri-color images. However, due to the dismal spatial resolution, a single pixel may comprise of a mixture of spectra belonging to multiple surface materials. Hyperspectral unmixing strives to extract the constituent spectra and their mixing percentages in each pixel. Independent Component Analysis (ICA) and Non-negative Matrix Factorization (NMF) are two prominent yet distinct methods used for hyperspectral unmixing. This paper proposes a novel method for hyperspectral unmixing, in which the fundamental notions of ICA are utilized to improve the accuracy of the standard NMF algorithm. This method was validated on standard and synthetic datasets and was observed to yield superior performance over the stand-alone implementations of ICA and NMF algorithms.