How do accuracy and model agreement vary with versioning, scale, and landscape heterogeneity for satellite-derived vegetation maps in sagebrush steppe?

Maps of the distribution and abundance of dominant plants derived from satellite data are essential for ecological research and management, particularly in the vast semiarid shrub-steppe. Appropriate application of these maps requires an understanding of model accuracy and precision, and how it might vary across space, time, and different vegetation types. For a 113 k Ha burn area, we compared modeled maps of different vegetation cover types created from satellite data to ‘benchmark” models based on intensive field sampling (~1500-2000 plots resampled annually for 5 years) for three new satellite-derived models: USDA Rangeland Analysis Platform (RAP), the USGS Rangeland Condition Monitoring Assessment and Projection (RCMAP), and USGS fractional estimate of exotic annual grass cover (USGS-fractional-EAG). We assessed out-of-sample point accuracy and asked if and how accuracy changed each year due to vegetation shifts, new images, and model improvements (i.e. model versions). We also assessed how map agreement between satellite-based and field-based models changed with scale of application, topography, and time since fire.

Accuracy and map agreement varied considerably among the vegetation types and across time and space (r² ranging from 0 to 0.53), and some of the variability was predictable. All models tended to over or underestimate cover when field-measured cover was relatively low or high, respectively, i.e. a “false moderating effect”. Accuracy was greater and improved with newer versions of RAP (+0.05 to 0.29 r²) compared to RCMAP and USGS fractional model estimates, and in some cases was greater than field-based models. Variability in map agreement tended to decrease with larger areas sampled (particularly in areas >12 km), and this scale dependency was more evident in RAP and USGS-fractional-EAG models. Creating a “fair” basis for comparison of spatial models of low-statured semiarid vegetation derived from satellite compared to field data is not trivial because scaling the field data to the scale of large satellite pixels (or downscaling satellite-based models to field scale) requires modeling and associated model uncertainty. Accuracy can vary considerably and understanding the variation can help guide application of the models to the appropriate time, place, and variables.