Comparative analysis of training strategies for neural network-based spectral unmixing of laboratory-simulated forest hyperspectral scenes

In this work, we address the use of neural networks for nonlinear mixture modeling of hyperspectral data by focusing on different training strategies which can automatically generate mixed training samples without a priori information. The proposed approach is compared to the standard, fully constrained linear mixture model using a database of laboratory-simulated forest scenes acquired by the Compact Airborne Spectrographic System (CASI), in which the areal fractions of the main constituents were calculated by the SPRINT canopy model. Our experiments demonstrate that simple multilayer perceptron (MLP) neural networks, when trained using a few mixed training samples, can provide good mixture characterization in different types of vegetation environments.