Optimal ranges to evaluate sub-pixel classifcations for landscape metrics

Landscape metrics rely on classifications of remote sensing data, and errors inherent to the classification scheme will be propagated into any spatial pattern results. This issue is compounded for metrics derived from sub-pixel unmixing techniques, since a universal method for assessing the certainty of these soft classifications has not yet been accepted. This study investigates the role of sub-pixel classification accuracy on landscape metrics through a combination of mathematical, ecological, and remote sensing methods by evaluating the fragmentation of saltcedar, a weedy invasive plant species, in the Rio Grande basin. First, ecological curve fitting methods are adopted to model landscape metric response across sub-pixel land cover proportions, and the proportions affecting the greatest landscape structure changes are extracted. Second, the classification accuracy of the tessellated linear spectral unmixing technique (TLSU) is assessed at narrow fractional abundances to determine whether accuracy varies with land cover proportion. Lastly, the land cover proportions significantly influencing landscape structure are compared to the ranges of highest accuracy to examine how errors in the sub-pixel classification technique are propagated into metrics. Results show that curve fitting is an appropriate technique for modeling metric responses to sub-pixel land cover proportion, however optimal ranges differ depending on the particular metric. Classification accuracy varies across sub-pixel proportions, and pixels with lower fractional abundance exhibit higher mapping accuracies. Since the most accurate classification ranges are not always coincident with the optimal ranges for metric measurement, agreement should be tested before applying metrics to a research problem.