A Deep Neural Network for Subpixel Target Detection Using Dimensionality Reduction in Hyperspectral Imaging

Using a deep neural network to detect a subpixel target in hyperspectral imaging is a difficult problem in remote sensing. The problem is caused by having only one full target pixel for training the target class and a weak target signal in subpixel target pixels for validation and having to solve a problem in high dimensional space resulting from a large number of spectral bands in hyperspectral imaging and a large number of weights in a deep neural network. The most recent algorithm is a deep neural network with two hidden layers, synthetic subpixel targets for training the target class, and variable stepsize for the gradient descent method and has been shown to be able to detect the subpixel target, but it is a high-dimensional problem resulting from setting the dimensions at the hidden layers to be the same as the number of spectral bands in order not to lose the weak target signal. In this paper an algorithm is developed by adding a hidden layer to the current neural network. Instead of using rectified linear unit, the dimensions at two of the hidden layers are set to be substantially lower than the spectral bands to reduce high computational cost and the potential problem with overfitting. Experimental results from two hyperspectral images, one with simulated subpixel target pixels for training and validating the algorithm and the other with simulated subpixel target pixels for training and actual subpixel target pixels for validation, have shown that the proposed algorithm can perform better than benchmark algorithms without losing the weak subpixel target signal.