Sources of error in a video-derived reference data set for accuracy assessment of land-cover maps in the Amazon

Título	Sources of error in a video-derived reference data set for accuracy assessment of land-cover maps in the Amazon
Autores	Rebecca Powell Nicholas Matzke Carlos De Souza, Jr. Matthew Clark Izaya Numata Laura Hess Dar Roberts
Ano de publicação	2003
Download

Powell, R., Matzke, N., Souza Jr., C., Clark, M., Numata, I., Hess, L., & Roberts, D. 2003. Sources of error in a video-derived reference data set for accuracy assessment of land-cover maps in the Amazon. Anais do XI Simpósio Brasileiro de Sensoriamento Remoto (pp. 2877 – 2879). Belo Horizonte: Instituto Nacional de Pesquisas Espaciais.

Abstract

Using a reference data set of high-resolution geocoded images generated from mosaicked digital videography, we have examined the sources of errors introduced during the accuracy assessment of a regional land-cover map classified from Landsat TM imagery of central Rondônia. We compared the reference data produced by five trained interpreters to quantify the degree of subjectivity in labeling of reference data and to identify which classes are most likely to be confused by interpreters. Errors were attributed to sources including geolocation difference between reference data and classified map, change between dates, mixed pixels, and disagreement between interpreters. Interpreters disagreed on the classification of almost 30% of the total samples; mixed pixels accounted for a large proportion of the disagreement. Agreement between any two interpreters in labeling of 30×30 m reference samples averaged less than 50% for the second growth forest class. Approximately 10% of the error in agreement between map and reference data could be attributed to geolocation error. These findings suggest that 1) definitions of land-cover types such as “second growth” are more subjective and variable than is commonly assumed; 2) validation data sets that include only non-mixed, nonedge samples are likely to result in overly optimistic accuracy estimates.