Unsupervised Feature-Learning for Hyperspectral Data with Autoencoders

Abstract

This paper proposes novel autoencoders for unsupervised feature-learning from hyperspectral data. Hyperspectral data typically have many dimensions and a significant amount of variability such that many data points are required to represent the distribution of the data. This poses challenges for higher-level algorithms which use the hyperspectral data (e.g., those that map the environment). Feature-learning mitigates this by projecting the data into a lower-dimensional space where the important information is either preserved or enhanced. In many applications, the amount of labelled hyperspectral data that can be acquired is limited. Hence, there is a need for feature-learning algorithms to be unsupervised. This work proposes unsupervised techniques that incorporate spectral measures from the remote-sensing literature into the objective functions of autoencoder feature learners. The proposed techniques are evaluated on the separability of their feature spaces as well as on their application as features for a clustering task, where they are compared against other unsupervised feature-learning approaches on several different datasets. The results show that autoencoders using spectral measures outperform those using the standard squared-error objective function for unsupervised hyperspectral feature-learning.

Nondestructive Phenolic Compounds Measurement and Origin Discrimination of Peated Barley Malt Using Near-Infrared Hyperspectral Imagery and Machine Learning

“…Followed by the same procedure of data selection, in total 24000 background samples and 6750 barley samples are generated in the training and testing dataset with s=200. In addition, three conventional and four deep learning models are used for benchmarking, including PCA [27], Folded PCA (FPCA) [32], 1-D Singular Spectrum Analysis (1DSSA) [48], Deep convolutional neural network (CNN) [49], Stacked auto-encoder (SAE) [50], Deep recurrent neural network (RNN) [51], and Auto-CNN [52]. These stateof-the-art approaches are selected for two main reasons, i.e.…”

Brightness Invariant Deep Spectral Super-Resolution

“…Finally, the argument can be made that the RGB normalization layer might not be necessary when a scale invariant loss function is considered. Examples might be the spectral angular error, spectral information divergence [ 24 ] or spectral derivative based loss functions [ 25 ]. It should be noted that from a purely theoretical point of view there is no guarantee that training a network with a loss function that is invariant to changes in brightness will make the fully-trained network invariant to changes in brightness.…”

Characterizing snow surface properties using airborne hyperspectral imagery for autonomous winter mobility

“…The high number of spectral channels associated with hyperspectral imagery presents a challenge to the training and performance of machine algorithms due to what is commonly known as the "curse of dimensionality" (Windrim et al 2019). The high number of spectral bands drastically increases the number of data points, which in turn can dramatically increase the computational requirements to train a given machine learning algorithm.…”