A Deep Learning approach to Hyperspectral Image Classification using an improved Hybrid 3D-2D Convolutional Neural Network

Abstract

In recent years, the task of Hyperspectral Image (HSI) classification has appeared in various fields, including Remote Sensing. Meanwhile, the evolution of Deep Learning, and the prevalence of the Convolutional Neural Network (CNN) has revolutionized the way unstructured, especially visual, data are processed. 2D CNN have proved highly efficient in exploiting the spatial information of images, but in HSI classification, data contain both spectral and spatial features. To make use of these characteristics, many variations of a 3D CNN have been proposed, but a 3D Convolution comes at a high computational cost. A fusion of 3D and 2D convolutions decreases processing time by distributing spectral-spatial feature extraction across a lighter, less complex model. An enhanced Hybrid network architecture is proposed alongside a data preprocessing plan, with the aim of achieving a significant improvement in classification results. Four benchmark datasets (Indian Pines, Pavia University, Salinas and Data Fusion 2013 Contest) are used to compare the model to other hand-crafted or deep learning architectures. It is demonstrated that the proposed network outperforms state-of-the-art approaches in terms of classification accuracy, while avoiding some commonly used, computationally expensive design choices.