Observed or Probabilistic? Evaluating the Effects of Ancillary Data as Predictor Variables in Remote Sensing Analyses of Vegetation

Abstract:

Standard remote sensing analyses of vegetation focus on spectral predictors of targeted map variables, but on the occasions these predictors fail to produce a required degree of accuracy, ancillary datasets such as topographic or climate variables are often added to the analysis. This approach, to "boost" remote sensing product accuracy through the inclusion of ancillary datasets is relatively widespread in the literature, but there have been few investigations into potential issues that can arise when choosing to use these variables. In this analysis, we investigate the degree to which the use of ancillary data when producing vegetation cover and land cover classification maps results in something more akin to a probabilistic map of vegetation as opposed to observed vegetation conditions. To accomplish this, we investigated the extent to which sites with high potential for a given vegetation condition, as determined by ancillary data, resulted in map values that were biased from actual observed condition. Furthermore, we investigated how the use of ancillary-data boosted vegetation maps in ecosystem analyses can set up cyclical logic problems. We discuss suggestions for best practices when developing, reporting, and using spectral-only vs. ancillary-data boosted vegetation maps.