An Object Linked Intelligent Classification Method for Hyperspectral Images

Hyperspectral images have recently become one of the finest basis for highly accurate identification of objects. Such images, however, are very large in size and carry huge information. Processing and handling of such information is also quite resource savvy and require complex algorithms as well. In this paper, an intelligent classification method has been proposed. Using two independent, simple feed-forward artificial neural networks (ANN), a 426 band hyperspectral image has been first processed for dimensionality reduction to 15 principal components (PCs) using standardized principal component analysis (SPCA), containing 98.86% of the original information. In the second stage, these 15 PCs have been utilized to train another ANN and three different objects viz. road, soil and vegetation have been accurately identified using supervised learning. Only a portion of hyperspectral image data was used as training set and three unique signature patterns were created in the form of pattern library for road, soil and vegetation. Once trained, three independent hyperspectral images taken from San Joachim field site located in California (NEON Domain 17) were fed into the well trained two-stage ANN. The three aforesaid objects were correctly identified. The proposed method is fully scalable and a pattern library can be created to identify more classes of objects. Once trained, the proposed method does not require any more statistical process and all subsequent images can be processed on-board due to the method's implementation using ANNs.