A Comparison of Machine Learning and Deep Learning in Hyperspectral Image Classification

Abstract

In recent years, hyperspectral remote sensing has become popular in various applications. This technology can capture hyperspectral images with a large terrestrial data. In this paper, the feasibility of applying various machine learning and deep learning techniques to perform classification on hyperspectral images are investigated and compared. Particularly, a total of three popular machine learning classifiers namely supports vector machine (SVM), K-nearest neighbors (KNN) and artificial neural networks (ANN) are used for hyperspectral imagine classification, followed by another two deep architectures in convolutional neural networks (CNN). Three benchmarking datasets of hyperspectral images are used to evaluate the classification performances of suggested machine learning and deep learning techniques, namely: Indian Pines (IP) dataset, Salinas dataset, and Pavia University (PU) dataset. Extensive simulation studies reveal the excellent performance of 3D CNN deep learning in solving larger datasets with better classification accuracy despite the longer training time is required. However, it is not really the case when the dataset is not large enough. This is because deep learning is data-hungry architecture. Furthermore, the 3D CNN deep learning models employed in this study have shown more advantageous as compared to other machine learning models for having simplified pre-processing stages such as feature extraction in solving the classification problems of hyperspectral images.