Neurally augmented object identification scheme for hyperspectral images

Hyperspectral images consist of unique signature patterns for various physical objects. These unique signatures can be utilized to identify similar objects on a land used/ land cover. In this paper, a neutrally augmented methodology for efficient object identification using hyperspectral images has been proposed. In the proposed scheme, first, the training samples of the known objects (viz. road, soil, and vegetation) are extracted from the hyperspectral cube. An elementary two-stage feed-forward artificial neural network (ANN) was then employed for correct identification of road, soil, and vegetation, within the training image. In the first stage of ANN, principal component analysis (PCA) has been applied for dimensionality reduction (retaining about 99.5% information) whereas the second stage has a simpler 15 neuron feed forward (FF) back-propagation (BP) ANN for correct identification of different objects. Eventually, the proposed scheme has been verified with the image of San Joachim field site located in California (NEON Domain 17). This hyperspectral cube has a size of 477 ∗ 502 ∗ 426. All the corresponding pixels representing road, soil, and vegetation in the aforesaid image have been identified with 100% accuracy using the proposed scheme.