Auto-Encoder Based for High Spectral Dimensional Data Classification and Visualization

Abstract

The classification and visualization for surface objects receives a great deal of attention for high spectral dimensional data processing. A lot of methods were proposed and applied in this problem over the past decade. Whereas most of them still exist some challenge issues, include pre-treatment fussily, features extraction simplify, larger data processing difficultly and classification inaccurately. To solve these problems, we propose a novel method bases on deep learning which combines maximum noise fraction (MNF) and multilayers auto-encoders. Thereinto, MNF is used to reduce the inherent spectral dimensionality of high spectral dimensional data; auto-encoders and softmax logistic regression function are applied for extracting high-level features and visualizing types of objects. On the experiments, we compared with traditional linear SVM method by using a 200-bands AVIRIS dataset over Indian Pines and a 103-bands ROSIS dataset over Italy Pavia. The results indicate that the algorithm in paper statistically significantly improved the classification and visualization accuracies and efficiencies.