Spectral-Spatial Hyperspectral Image Classification Using Cascaded Convolutional Neural Networks

Abstract

In this paper, we present a framework for classifying hyperspectral images using convolutional neural networks. In order to compare the importance of spectral and spatial information as well as the effect of dimension reduction prior to classification, we propose different architectures utilizing 2D and 3D convolutions. We introduce a novel coarse-to-fine two-stage fusion architecture employing a cascade of consecutive neural networks to deal with the problem of separating classes which are very similar to each other. Using this approach, we can boost the classification accuracies significantly with little additional effort. The results of the proposed framework indicate that further research regarding cascaded architectures could help in distinguishing similar classes in hyperspectral images.