Executive Summary : | With rapid changes in climate the need of remote sensing for earth observations is ever increasing. Hyperspectral imaging is one of such methods which plays a significant role in monitoring the land use and land cover. In recent years, hyperspectral (HS) imaging has attracted much attention to land used and land cover classification (LULC), forest applications and target detection etc. The methodologies developed in those works are often described as image classification algorithms. These methods are supervised and they require class labels of the training samples. This approach is expensive and time-consuming. Moreover, the learned classification model might not be transferable to new hyperspectral images. Due to the limited spatial resolution of a hyperspectral image, each hyperspectral pixel might cover several pure materials on the ground and classification of each pixels become more challenging. Furthermore, the classification maps are not accurate enough to analyse. To tackle this challenge, hyperspectral unmixing techniques are being developed. These methods can estimate the composition of each hyperspectral pixel by only utilizing the spectral signature of pure materials (endmembers). In remote sensing application, it is generally assumed that the spectra of the pure materials are mixed linearly and several linear unmixing techniques have been developed to find the contribution of each individual material. Mixed pixels are considered linear combinations of endmember signatures weighted by the proportional abundance of the constituents. When the endmembers of the hyperspectral image are available, the fractional abundances can be estimated by minimising the least squared errors between the actual reflectance spectra and the ones, reconstructed by the linear model. To have a physical interpretation of the estimated fractional abundances, one must assume that no endmember can have a negative abundance. This constraint is often described as the abundance non-negativity constraint (ANC). The second constraint is the abundance sum-to-one constraint (ASC), i.e., the observed reflectance spectrum is completely composed of endmember contributions. As a result, many endmember extraction algorithms have been proposed to maximize the volume enclosing simplex in the hyperspectral dataset. When endmembers are not available in the hyperspectral image (no pure pixel-scenario), virtual endmembers can be estimated by seeking the minimum volume linear simplex, which encloses the data points. In summary, the primary goal of this project is to extract/estimate endmembers and their fractional abundances in each pixel by only utilizing the observed hyperspectral image. The endmembers estimated by the methodology developed in this work will be compared with the ground truth endmembers acquired by utilizing low-cost and small-size hyperspectral cameras. |