Automated generation of semi-labeled training samples for nonlinear neural network-based abundance estimation in hyperspectral data

Abstract

As the initial stage of a supervised classification, the quality of training has a significant effect on the entire classification process and its accuracy. In hyperspectral data analysis, a judicious selection of training samples can be tremendously difficult due to the presence of subpixel targets and mixed pixels, in particular, when no prior knowledge about the data is available. For instance, the Multi-Layer Perceptron (MLP) neural network can provide very accurate nonlinear estimations of fractional abundances, provided that the training set contains all possible mixture conditions. However, the requirement of large volumes of training data is a serious limitation in remote sensing because, even if classes concurring to a per-pixel cover class mixture are known, proportions of these classes are very difficult to be estimated a priori. This paper investigates, explores and further proposes solutions to resolve the issues above. Specifically, we develop a nonlinear neural network-based mixture model, coupled with unsupervised algorithms for automated generation of semi-labeled samples that can be effectively used for mixed pixel classification. These unsupervised algorithms, intended for situations where ancilliary information is difficult to be collected prior to data analysis, rely on the principle that patterns that lie close to the location of decision boundaries in feature space are more informative than patterns drawn from the class cores. Computer simulations and real experiments are conducted for performance analysis of nonlinear unmixing techniques based on training samples. KeywordsHyperspectral imaging, Nonlinear mixture analysis, Training samples, Semi-labeled samples, Multi-layer perceptron.