Deep Belief Networks for Feature Fusion in Hyperspectral Image Classification

Abstract

Hyperspectral data classification is a great challenging method for remote sensing. In recent years, the researchers have had a great attention to the feature fusion of hyperspectral data. In this paper, based on distinctive advantage over machine learning, we suggest a novel technique to classification of hyperspectral images, which employs deep belief networks (DBNs) to fuse spectral and spatial features together. In the light of the above-mentioned descriptions, DBN be able to extract the hierarchical features from raw data, which are cost-effective for classification based on support vector machine (SVM). First, we verify the eligibility of DBN, and SVM-based classification and then, suggest a new framework, stacking the spectral and spatial features, fuses features by DBN, and classifies them by SVM to get most accuracy. First, we extract spatial features by applying principal component analysis (PCA) and extended morphology (EMP) and append at the end of spectral features, then fuse and classify achieved features by the suggested method. The experimental test results demonstrate the suggested method yields to most accuracies results.