Unsupervised Hyperspectral Anomaly Detection with Convolutional Neural Networks

Abstract

In this study, we propose a convolutional neural network-based method for unsupervised hyperspectral anomaly detection. We obtain within- and between-cluster pixel pairs from hyperspectral images using the cluster maps obtained by an automatic clustering algorithm and constitute training data set by taking the differences of the pixel pairs. We design a multilayer convolutional neural network and train it using difference vectors. In prediction step, we apply a dual local window to hyperspectral image. For each pixel, we calculate the difference vectors between the center pixel and the surrounding pixels. By feeding the trained network with the difference vectors, we obtain prediction scores. In the last step, a weighted RX detector was obtained using prediction scores and used for anomaly detection. It has been observed that the experimental results conducted on four different real hyperspectral data show better results than the current CNN-based anomaly detector.