Subpixel Target Detection based on a Deep Neural Network

The objective of this paper is to develop an algorithm mathematically for solving the stochastic optimization problem of a deep neural network numerically using a gradient descent method with a variable step-size that is determined by the algorithm automatically. Neural networks are conventionally solved using gradient descent methods with a fixed step-size, which is selected by a user manually. Artificial neural networks have difficulty in hyperspectral imaging for target detection of rare targets. The difficulty is caused by the availability of only one target spectrum for training the target class in neural networks. The proposed algorithm is applied to subpixel target detection in hyperspectral imaging using a real hyperspectral image. Additional target pixels are generated using synthetic signature modeling. The decision boundary for the deep neural network is nonlinear for the transformed pixels obtained through whitening transformation. Experimental results show that the proposed algorithm converges and can perform well in subpixel target detection.