A Review on Deep Learning Approaches for Spectral Imaging

Deep learning algorithms have revolutionized the computer vision field in the last decade. They can reduce tedious feature engineering and have opened new possibilities of automated visual inspection. With deep learning techniques, the availability of large amounts of qualitative labeled data became more important than ever. The main share of computer vision research focuses on RGB images. With the advances in sensor technologies multi- and hyperspectral cameras have become more cost effective and accessible in recent years, allowing this imaging technology to be applied to new fields of application. This article gives an overview of approaches to apply deep learning techniques to multi- or hyperspectral data. Several state-of-the-art methods will be reviewed and problems and difficulties will be discussed. An overview of a selection of available datasets is presented. To give a broad and diverse insight, research from different fields of application are considered, namely the remote sensing domain, the agricultural domain and the food industry.