A comparison of nested analysis of variance and geostatistics for characterizing the spatial structure of patchy grassland: a study using simulated landscapes

Ecological processes act upon landscapes over a wide range of temporal and spatial extents. These processes interact to create complex and ever-changing landscape patterns. While remote sensing data (e.g. AVHRR, TM) has often been used to characterize such heterogeneity at the landscape scale, these sensors cannot resolve spatial pattern to the detail required for finer scale analyses, such as the linking of biogeochemical cycles to communitylevel processes (e.g. patch formation). In such cases, it is often necessary to sample heterogeneity in situ. However, because data collection and analysis methods are generally designed to meet certain statistical criteria and perform with certain statistical properties, the sampling techniques used to characterize spatio-temporal heterogeneity will strongly influence perceived landscape structure. If the assumptions that an investigator makes about the sampled landscape are untrue, then the resultant characterization of landscape structure is likely to be misleading. In this study, we use three examples to highlight this potential problem. More specifically, we compare geostatistics and nested random-effects ANOVA for characterizing landscape heterogeneity under conditions where the landscape homogeneity assumptions on which geostatistics are based are violated, and where sampling budget is limited. We simulate a variety of landscapes whose statistical structures are known, and analyze the spatial structures estimated by geostatistics and nested ANOVA. Our results indicate that for all simulated conditions, nested design provides a more stable and rigorous characterization of structure than that provided by geostatistics. Furthermore, the statistical properties of the nested ANOVA design allow testing for significant differences in patchiness between spatial resolutions.