An Analysis of Spatial-Spectral Dependence in Hyperspectral Autoencoders

Hyperspectral imaging is central for remote sensing, and much research has been done on analysis methods for land surveillance using space- and air-borne imaging systems. Proximal hyperspectral imaging is also widely used in plant and agriculture science. It allows the remote capturing of leaf reflectance information in order to determine and classify plant health and disease. With the high information density in hyperspectral images, it becomes increasingly important to apply sophisticated feature extraction in order to reduce image dimensionality while retaining useful information. Autoencoders are one of the primary methods for deep learning-based feature extraction in hyperspectral images. We investigate different setups of autoencoders to encode the spatial and spectral dimensions in different orders and ways. To our surprise, the best turns out to be a 3D CNN, where the spectral dimension is treated in the same way as the spatial dimensions.