Journal
ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING
Volume 98, Issue -, Pages 145-156Publisher
ELSEVIER
DOI: 10.1016/j.isprsjprs.2014.09.016
Keywords
Hyperspectral; Absorption feature; Iron minerals; Vertical geology; Illumination conditions; Machine learning; Classification; Remote sensing
Categories
Funding
- Australian Centre for Field Robotics
- Rio Tinto Centre for Mine Automation
Ask authors/readers for more resources
In this study, we compare three commonly used methods for hyperspectral image classification, namely Support Vector Machines (SVMs), Gaussian Processes (GPs) and the Spectral Angle Mapper (SAM). We assess their performance in combination with different kernels (i.e. which use distance-based and angle-based metrics). The assessment is done in two experiments, under ideal conditions in the laboratory and, separately, in the field (an operational open pit mine) using natural light. For both experiments independent training and test sets are used. Results show that GPs generally outperform the SVMs, irrespective of the kernel used. Furthermore, angle-based methods, including the Spectral Angle Mapper, outperform GPs and SVMs when using distance-based (i.e. stationary) kernels in the field experiment. A new GP method using an angle-based (i.e. a non-stationary) kernel - the Observation Angle Dependent (OAD) covariance function - outperforms SAM and SVMs in both experiments using only a small number of training spectra. These findings show that distance-based kernels are more affected by changes in illumination between the training and test set than are angular-based methods/kernels. Taken together, this study shows that independent training data can be used for classification of hyperspectral data in the field such as in open pit mines, by using Bayesian machine-learning methods and non-stationary kernels such as GPs and the DAD kernel. This provides a necessary component for automated classifications, such as autonomous mining where many images have to be classified without user interaction. (C) 2014 The Authors. Published by Elsevier B.V. on behalf of International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS).
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available