A New Feature Extraction Based on Local Energy for Hyperspectral Image

In hyperspectral classification, as the number of training samples to classify are limited, the accuracy of classifier decreases. One of the reasons for this phenomenon is the variability of spin-off extraction spatial features. This means that when the scene is rotated a bit, these features also change. It should be noted that these features are a local feature and ruin this situation, because there may be a class in two parts of the scene that is rotated relative to another. For this purpose, a new method for extracting spatial features has been proposed in this paper that is unchangeable to rotation. In this study, local energy has been extracted by local Fourier transform and structural information has been extracted by morphological attribute profiles (APs) to complete the extraction features. Energy information and spectral information in a scenario are stacked. Energy information, structure information and spectral information are stacked in another scenario. Then they are classified by support vector machine (SVM) classifier. The results express that the first scenario is beneficial for images without structural data, and the second scenario is more useful for urban images, which includes a lot of structural information. The proposed method are applied on three famous data sets (Pavia University, Salinas and Indiana Pines). The results demonstrate that the proposed method is superior to the other competition methods.