Variable-Stepsize Multilayer Neural Network for Subpixel Target Detection in Hyperspectral Imaging

Abstract

Conventional algorithms for subpixel target detection of a rare target in hyperspectral imaging are derived from the generalized likelihood ratio test. Artificial neural networks are designed for classification, which needs large samples of pixels for training background and target classes, but have a problem with subpixel target detection, which typically has only one target pixel for training the target class. The current detection algorithm for subpixel target detection based on neural networks uses a single-layer neural network and stochastic gradient descent method with variable stepsize to solve the optimization problem. The objective of this paper is to improve the current algorithm by developing a detection algorithm for subpixel target detection based on a multilayer neural network and stochastic gradient descent method with variable stepsize. The decision boundary is linear for the single-layer neural network and nonlinear for the multi-layer neural network. Experimental results from two hyperspectral images, one with simulated subpixel target pixels for validating the algorithm and the other with actual subpixel target pixels for validation, have shown that the proposed algorithm can perform better than the conventional algorithms.

Keywords

• Hyperspectral imaging
• neural network
• stochastic gradient descent (SGD)
• target detection