Bias in land cover change estimates due to misregistration

Abstract

Listen

Translate article

Land cover change may be overestimated due to positional error in multi-temporal images. To assess the potential magnitude of this bias, we introduced random positional error to identical classified images and then subtracted them. False land cover change ranged from less than 5% for a 5-class AVHRR classification, to more than 33% for a 20-class Landsat TM classification. The potential for false change was higher with more classes. However, false change could not be reliably estimated simply by number of classes, since false change varied significantly by simulation trial when class size remained constant. Registration model root mean squared (rms) error may underestimate the actual image co-registration asccuracy. In simulations with 5 to 50 ground control locations, the mean model rms error was always less than the actual population rms error. The model rms error was especially unreliable when small sample sizes were used to develop second order rectification models. We introduce a bootstrap resampling method to estimate false land cover change due to positional error. Although the bootstrap estimates were unbiased, the precision of the estimates may be too low to be of practical value in some land cover change applications.