On the choice of spatial and categorical scale in remote sensing land cover classification

Our interest in this paper is on the choice of spatial and categorical scale, and their interaction, in creating classifications of land cover from remotely sensed measurements. We note that in discussing categorical scale, the concept of spatial scale naturally arises, and in discussing spatial scale, the issue of aggregation of measurements must be considered. Therefore, and working towards an ultimate goal of producing multiscale, multigranular characterizations of land cover, we address here successively and in a cumulative fashion the topics of (1) aggregation of measurements across multiple scales, (2) adaptive choice of spatial scale, and (3) adaptive choice of categorical scale jointly with spatial scale. We show that the use of statistical finite mixture models with groups of original pixel-scale measurements, at successive spatial scales, offers improved pixel-wise classification accuracy as compared to the commonly used technique of label aggregation. We then show how a statistical model selection strategy may be used with the finite mixture models to provide a data-adaptive choice of spatial scale, varying by location (i.e., multiscale), from which classifications at least as accurate as those of any single spatial scale may be achieved. Finally, we extend this paradigm to allow for jointly adaptive selection of spatial and categorical scale. Our emphasis throughout is on the empirical quantification of the role of the various elements above, and a comparison of their performance with standard methods, using various artificial landscapes. The methods proposed in this paper should be useful for a variety of scale-related land cover classification tasks.