Skip to Main content Skip to Navigation

NMF Hyperspectral Unmixing Of The Sea Bottom: Influence Of The Adjacency Effects, Model and Method

Abstract : Sea bottom unmixing is a challenging task for the analysis of coastal zones. Actually the upward photons are attenuated and diffused by the water column layer, giving low signal to noise hyperspectral data. A classical approach is to perform inversion of the water column using semi-analytical parametric models and estimation process, and obtain the water column constituents (chlorophyll, suspended matter, dissolved organic matter and bathymetry), and the coefficients of pure materials reflectance spectra (endmembers), given in spectral libraries, for each pixel. We consider here the case of unknown endmembers, and we suppose that the water column components have been obtained by a classical inversion method or in-situ measurements. For each observed pixel the upward luminance is analysed and decomposed into three terms, respectively issued after interaction with the target bottom pixel, its neighbours, and the water column. We show that in some conditions the adjacent pixels effect is not negligible, due to the diffusion in the water column, and we develop in accordance a new mixing model for the sea bottom. We propose a non-negative matrix factorisation based unmixing method to solve the problem, and present results for hyperspectral data simulations.
Complete list of metadatas

https://hal-insu.archives-ouvertes.fr/insu-02398213
Contributor : Catherine Cardon <>
Submitted on : Saturday, December 7, 2019 - 8:10:55 AM
Last modification on : Wednesday, February 12, 2020 - 4:03:03 PM

Identifiers

Citation

M. Guillaume, L. Juste, X. Lenot, Y. Deville, B. Lafrance, et al.. NMF Hyperspectral Unmixing Of The Sea Bottom: Influence Of The Adjacency Effects, Model and Method. 2018 9th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), Sep 2018, Amsterdam, Netherlands. pp.1-5, ⟨10.1109/WHISPERS.2018.8747064⟩. ⟨insu-02398213⟩

Share

Metrics

Record views

54