Effects of class granularity and cofactors on the performance of unsupervised classification of wetlands using multi-spectral aerial photography

Abstract

This research focuses on improving image analysis using multi-spectral aerial photography and unsupervised classification with a high number of classes. A range of class numbers (15 to 240) were explored using an automated process for classification and validation using the kappa statistic as a measure of classification accuracy. The classification model was kept simple so that the different data treatments could be directly compared under an otherwise standardised method, acknowledging that a typical landcover analysis work-flow would normally build upon this data-preparation phase using a range of available post-classification refinement procedures. Although the accuracy of the unsupervised classification achieved in this simplified model was only moderate, this research has shown that unsupervised classification accuracy improves with increasing numbers of classes and with addition of two processing variants (addition of a normalised difference vegetation index and principal components transformation), and that similar performance to supervised classification can be achieved.

Keywords::

• remote-sensing
• classification
• landcover

Acknowledgements

We would like to thank the Waikato Regional Council spatial information team for use of GIS resources and logistic support of the field survey work, and Chris McKinnon and Daniel Tait for assistance in the field validating the vegetation classes, and Richard Barnett for providing boat and navigation for the river delta ground-truthing.