Hyperspectral open set classification with unknown classes rejection towards deep networks

Abstract

ABSTRACT Recent research in deep learning has significantly improved the performance of hyperspectral classification. However, nearly all experimental evaluations based on deep networks have taken the form of ‘closed set’, where all testing classes are known in training time. Obviously, it is unreasonable for real-world-applications, where unknown classes in training time may appear during testing. The realistic scenarios require classifiers not only to classify the known classes, but to reject the unknown classes, which is referred as open set classification (OSC). Considering the increased applications in real world, the adaption of deep networks-based methods towards OSC can be of vital importance. Unfortunately, it has not been well addressed by existing deep network-based algorithms. To tackle it, we present a method based on deep networks for hyperspectral OSC, by introducing an OpenMax layer which can estimate the probability of an input being from unknown classes and classify known classes. Experiments conducted on five hyperspectral datasets show that the number of misclassifications of the unknowns made by traditional three-dimensional Convolutional Neural Network (CNN) can be reduced significantly using proposed method. This paper highlights the open set challenge in hyperspectral classification and verifies the effectiveness of OpenMax method for solving such problem.