A Detailed Analysis of Datasets Used in HSI in the Context of Mixture Models for Unmixing

This research article presents a detailed analysis of Hyperspectral (HS) imaging datasets, which are commonly used in algorithms for HS Unmixing. While numerous popular datasets are utilized for benchmarking these algorithms, there is a notable scarcity of research that conducts a thorough analysis of these datasets. To bridge this gap, we conducted an in-depth analysis using four real and two synthetic HS datasets: Samson, JasperRidge, Urban, Apex, Synthetic Matern, and Synthetic Spherical. Our study specifically investigates the effectiveness of Linear and Bi-Linear Mixing Models in reconstructing these HS images from given ground truths. Surprisingly, we discovered spatial variability in endmember scales across the aforementioned datasets, challenging the common assumption of spatial consistency in HSI images according to unmixing algorithms. This observation was substantiated through comparisons of reconstructed pixel power with ground truth pixel power. This was further establised by assessments of pure pixels within the dataset. Notably, such variability was absent in synthetic datasets. This research contributes to the advancement of understanding of HSI algorithms and underscores the importance of considering spatial variability in endmember scales for accurate data reconstruction and analysis in remote sensing applications.