Although array factorization methods have been vastly investigated as a source separation tool in disparate areas like chemometrics or remote sensing, studying multiple data set simultaneously, factorizing very large data set and identifying the extracted components solely from their spectra remain challenging tasks.
In this talk, I will present some recent related works on matrix and tensor factorization. First, I will show how a library of known spectra can be used inside a factorization algorithm to improve identification performances, based on a new formulation of the sparse coding problem. Second, I will introduce multiway data fusion as a general framework that can tackle both variability in the data and multimodality. Several exemples of data fusion models applied to chemometrics and spectral unmixing of time-dependent hyperspectral images will be discussed. This will be illustrated on hyperspectral images of the surface of the earth.