Abstract

In the present work, we perform spectral mixture analysis using Chi-square minimization (χ2 minimization) procedure and test the feasibility of applying an inverse technique, neural network (NN) approach, for the spectral unmixing. The training of NN is carried out using the Levenberg-Marquardt algorithm (LM) with the initial weights for training being chosen randomly. The experiments are performed in the laboratory by mixing young, matured and dead leaves of a sequoia tree in various proportions and reflectance spectra of these mixtures are recorded. The proportions are chosen to model a few near-real situations like different kinds of vegetation in a forest (by mixing young leaves and matured leaves) and trees damaged in a forest fire or affected by certain virus (by mixing matured and dead leaves) and a combination of all these (by mixing young, matured and dead leaves). The spectral mixture analysis employing χ2 minimization and the inverse procedure utilizing NN with two hidden layers yielded consistent results in accordance with the proportion of each kind of leaf.