Blindly selecting method of training samples based hyper-spectral image's intrinsic character for object recognition

Based on the intrinsic assembling feature of the hyper-spectral images, we present a method to select the training samples for object recognition without any other previous knowledge. Firstly, we use the Parzen's window method to find the easily separable dimensions of the hyper-spectral images, then gain the smallest representative sample sets of all objects through intersecting the data of the same object of each easily separable dimensions, and obtain the object's number and the training data sources for the neural networks (NN) at the same time; secondly, train the neural network ensembles using the data selected from the representative sample sets to label the other data. Lastly, we analyzed the hyper-spectral images to detect red tide using this method, which proved this method could recognize the red tide effectively.