Plant production system monitoring via multiple signal classification and sparse regression

In this paper, we propose a dictionary pruning technique for signal unmixing, which is a recent unmixing concept aimed at extracting the signal contribution of an individual class of endmembers to each observed pixel in a hyperspectral scene. We focus on a specific application in the area of plant production system monitoring. The pruning technique allows to infer the physical state of vegetation in a region of interest on the ground, by selecting appropriate spectra from a very large spectral library designed to model a wide variety of such possible states. On the one hand, Multiple Endmember Spectral Mixture Analysis (MESMA), a combinatorial technique, was intensively used to solve the signal unmixing problem. Sparse unmixing, a fast approach dealing with spectral libraries, was recently proposed as a reliable technique for spectral unmixing. The performances of both techniques are influenced by the large number of spectra collected in the libraries. The aim of this paper is two-fold: 1) to present a subspace-based method for library pruning, such that only spectra likely to be present in the image are used in unmixing, which is expected to improve MESMA results, and 2) to introduce sparse unmixing as a reliable signal unmixing technique, given the natural ability of these methods to deal with applications in which a low number of atoms is used to accurately describe every observed vector. The results obtained in a simulated environment show the potential of the proposed scheme to boost the signal unmixing performance both in terms of accuracy and running time.