Publication:
Spatial low-rank tensor factorization for hyperspectral image unmixing

Thumbnail Image
Authors
Navas-Auger, William J.
Embargoed Until
Advisor
Manian, Vidya
College
College of Engineering
Department
Department of Computer Science and Engineering
Degree Level
Ph.D.
Publisher
Date
2021-05-14
Abstract
This work presents a method for hyperspectral image unmixing based on nonnegative tensor factorization. While traditional approaches may process spectral information without regard for spatial structures in the data set, tensor factorization preserves the spectral-spatial relationship which we intend to exploit. We used a rank-(L, L, 1) decomposition which approximates the original tensor as a sum of R components. Each component is a tensor resulting from the multiplication of a low-rank spatial representation and a spectral vector. Our approach uses the spatial factors, to identify high abundance areas where pure pixels (endmembers) may lie. Finally, abundance maps are generated by applying Fully Constrained Least Squares for each inferred endmember. Results of this method are compared against other approaches based on non-negative matrix and tensor factorization. We observed a significant reduction of spectral angle distance for extracted endmembers and equal or better root mean square error for abundance maps as compared with existing benchmarks.

Este trabajo presenta un método para desmezclar imágenes hiperespectrales basado en factorización de tensores no negativos. Mientras que los métodos tradicionales procesan información espectral sin tomar en consideración las estructuras espaciales en el conjunto de datos, la factorización tensorial conserva la relación espectral-espacial que pretendemos explorar y utilizar. En nuestro trabajo utilizamos una descomposición de rango-(L, L, 1) que se aproxima al tensor original como una suma de R componentes. Cada componente es un tensor resultante de la multiplicación de una representación espacial de bajo rango y un vector espectral. Nuestro enfoque utiliza los factores espaciales para identificar áreas de alta abundancia donde pueden encontrarse pıxeles puros. Finalmente, se generan mapas de abundancia usando ”Fully Constrained Least Squares” para cada pıxel puro inferido. Los resultados de este método se comparan con otros enfoques basados en matrices no negativas y factorizaciones de tensores. Observamos una reducción de la distancia del ángulo espectral para los pıxeles puros extraıdos y un error igual o menor error (RMSE) para los mapas de abundancia en comparación con el punto de referencia existente.
Keywords
Remote Sensing,
Non-negative Tensor Factorization,
Hyperspectral Unmixing,
Low-Rank,
Signal Separation
Usage Rights
Except where otherwise noted, this item’s license is described as Attribution 3.0 United States
Cite
Navas-Auger, W. J. (2021). Spatial low-rank tensor factorization for hyperspectral image unmixing [Dissertation]. Retrieved from https://hdl.handle.net/20.500.11801/2772