Reduced 3-D Deep Learning Framework for Hyperspectral Image Classification

Abstract

In recent years, machine learning has achieved a breakthrough step due to re-branding of Convolutional Neural Networks (CNNs). Advancement in machine learning algorithms makes it easier to process big and information-rich images such as hyper-spectral images. Hyperspectral imaging (HSI) technology has also shown obvious increase in number of satellites and increased number of bands which lead to a huge amount of data generated every day. In this paper, we propose a reduced version for 3-dimensional convolutional neural network (3D-CNN) as a deep learning framework for hyperspectral image classification. The latest proposed CNNs models, especially 3D ones, have achieved near 100% of accuracy with benchmark hyperspectral data sets. Our proposed framework explores the effect of dimensions reduction on the performance with respect to total classification accuracy. In our experiments, two benchmarks HSIs are used to evaluate performance of reduced framework with different number of bands. The experimental results demonstrate that the reduced 3D-CNN framework has significantly reduced the time of training of CNN with more than 60% compared to the full bands training almost without affecting the accuracy of classification.