Areal Sample Units for Accuracy Evaluation of Single-date and Multi-temporal Image Classifications

Areal Sample Units for Accuracy Evaluation of Single-date and Multi-temporal Image Classifications

Kim Lowell1, Alex Held2 , Tony Milne3, Anthea Mitchell3, Ian Tapley3, Peter Caccetta4, Eric Lehmann4, Zheng-shu Zhou4

1. Cooperative Research Centre for Spatial Information, Dept. of Infrastructure Engineering, University of Melbourne, Carlton, VIC 3053 AUSTRALIA (klowell@crcsi.com.au)
2. CSIRO Canberra, ACT 2601 (alex.held@csiro.au)
3. Cooperative Research Centre for Spatial Information, School of Biological, Earth, and Environmental Sciences, The University of New South Wales, Kensington, NSW 2052 ( t.milne@unsw.edu.au, a.mitchell@unsw.edu.au, hgciant@bigpond.net.au)
4. CSIRO Mathematics, Informatics & Statistics, Floreat, WA 6014 (peter.caccetta@csiro.au, eric.lehmann@csiro.au, zheng-shu.zhou@csiro.au)
 

Abstract: Areal sample units are explored as an alternative to conventional point-based samples for map accuracy assessment. Three analyses are examined: confidence limits on estimates of the total amount of each landcover, regression of map data versus reference data, and a two-part confusion matrix approach. It is concluded that areal sample units provide some advantages compared to point-based samples, but are nonetheless subject to problems associated with providing robust accuracy assessment metrics for rare classes.

Keywords: landcover change, rare event sampling, regression, modified confusion matrix

AttachmentSize
LowellAccuracy2012.pdf395.71 KB