A method to obtain large quantities of reference data

Abstract

Project managers often struggle with the need of sufficient reference data to train and test for reliable classifications and budget concerns that restrain the amount of justifiable field data collection. For a forest study, we supplemented our 207 ground‐measured field sites with 4000 additional photo‐interpreted reference sites. We first used aerial photography to identify the extent of homogenous regions around field‐data sites and then picked additional reference points within these areas. This approach is based on the notion that similar‐appearing areas close to a measured vegetation plot will contain approximately the same mix and density of species as the known site. This resulted in clusters of additional data points around actual field locations. We avoided overestimating the classification accuracy due to spatial autocorrelation by using an entire cluster of reference points exclusively as training or test data.

Acknowledgements

We would like to acknowledge the efforts of Mark Rumble and the USDA Rocky Mountain Forest Service Research Station for providing vegetation data and support with the reference data collection, and Doug Baldwin and Krystal Price for collecting and processing field data. We would like to thank Dr Michael Weir and two anonymous referees for comments. This research was supported by NASA Food and Fiber Applications of Remote Sensing Programme, Grant #NAG13‐99021, South Dakota School of Mines and Technology, National Science Foundation/EPSCoR Grant #EPS‐0091948, EPS‐9720642 and the SD Space Grant Consortium.