Nonlinear mixture models for analyzing laboratory simulated-forest hyperspectral data

Abstract

The interpretation of mixed pixels is a key factor in the analysis of hyperspectral imagery. A commonly used approach to mixed pixel classification has been linear spectral unmixing. However, the question of whether linear or nonlinear processes dominate spectral signatures of mixed pixels is still an unresolved matter. In this paper we describe new methodologies for inferring land cover fractions within hyperspectral scenes, using nonlinear mixture modeling techniques based on support vector machines and neural network-based techniques. A comparative analysis of these mixture estimation methods to the standard linear mixture model has been carried out using a database of laboratory simulated-forest scenes. For the simulations, canopies of both opaque and translucent trees were simulated using objects mounted on stems. Two tree densities (sparse and dense) and three background colors (dark, white and green) were considered. Hyperspectral images of these simulated scenes were acquired by the Compact Airborne Spectrographic Imager (CASI), and the areal fractions of the main constituents calculated by the SPRINT canopy model were used for comparison. Our quantitative and comparative analysis reveals that nonlinear approaches outperform linear mixture model-based approaches, particularly in the scenes with translucent trees. As a result, this investigation suggests that nonlinear mixture models are needed to account for the multiple scattering between tree crowns and background for the laboratory simulated-forest scenes used in this study.