Propagation of atmospheric condition parameter uncertainty in measurements of landscape heterogeneity

Remote sensing products are widely used in ecology and environmental science to understand how surfaces are composed and configured in a landscape and how they change through time. Land cover maps that describe the nature of habitats available to organisms have become critical tools in the study of anthropogenic impacts, such as fragmentation, on biodiversity and related ecological processes. Despite the increasing importance and widespread use of remote sensing derived landscape pattern indices measured using landscape composition and configuration metrics via land cover maps, these indices may conceal substantial and potentially problematic concerns about the uncertainty of reflectance data. This uncertainty derives from the presence of varying atmospheric conditions between the earth and the sensor and is often unreported. We asked how the magnitudes of measured landscape composition and configuration change and the dispersion in the changes caused by reflectance uncertainty compare. Using a Monte Carlo framework, we estimated multiple realizations of reflectance images of a simulated hyperspectral dataset depicting a peri-urban area. These estimates were influenced by uncertainty originating from atmospheric condition parameters (column water vapour, aerosol optical depth, and aerosol type). We generated multiple classified maps using Support Vector Machines and quantified standard landscape composition and configuration metrics. High dispersion in composition and configuration metrics caused by reflectance uncertainty strongly impacted our ability to detect and compare changes of the kind that natural scientists typically derive from such maps. We conclude that the propagation of uncertainty limits the extent to which current landscape patterns and Land Use and Land Cover change (LULC) analyses can be used for fine-scale decision-making for landscape management.