Estimating Error in an Analysis of Forest Fragmentation Change Using North American Landscape Characterization (NALC) Data

Abstract

We describe an approach for estimating measurement error in an analysis of forest fragmentation dynamics. We classified North American Landscape Characterization (NALC) images in four path-row locations in the Upper Midwest to characterize changing patterns of forest cover. To estimate error, we calculated the differences in values of forest fragmentation metrics for overlapping scene pairs from the same time frame (or epoch). The overlapping image areas were subdivided into landscape partitions. We tested the effects of amount of forest cover, landscape phenology, atmospheric variability (e.g., haze and clouds), and alternative processing approaches on the consistency of metric values calculated for the same place and approximate time but from different images. Two of the metrics tested (average patch size and number of patches) were more sensitive to image characteristics and contained more measurement error in a change detection analysis than the others (percent forest cover and edge density). Increasing the landscape partition size moderately reduced the amount of error in landscape change analysis, but at the cost of reduced spatial resolution. Processes used to generalize the forest map, such as small-polygon sieving and majority filtering, were not found to consistently decrease measurement error in metric values and in some cases increased error. Predictive models of error in a forest fragmentation change analysis were developed and significantly explained up to 50% of the variation in error. We demonstrate how, in a change analysis, predicted error can be used to identify locations that exhibit change substantially greater than the error in value estimation.