Regionalization of multi-categorical landscapes using machine vision methods

Abstract

We introduce a novel method for regionalization of multi-categorical landscape or land cover pattern based on the principle of machine vision rather than clustering of landscape metrics. Maps of land use/land cover (LULC), such as, for example, the NLCD 2006, show spatially varying pattern of LULC categories. Using an LULC map as an input our method discovers and maps different landscape types (LTs), each dominated by a characteristic LULC pattern that reflects varied influence of natural and anthropogenic factors. At the core of the method are the concepts of landscape signature and landscape dissimilarity adapted from the field of machine vision. A two-dimensional histogram of LULC categories and clump sizes is used as a landscape signature and the Jensen-Shannon measure is used as landscape dissimilarity. We have also adapted the machine vision technique of object-bases image analysis, consisting of segmentation and clustering, to find the set of LTs regions. Such technique maximizes spatial contiguity of the regions while also providing necessary level of generalization. The method is applied to the study area located in the northern part of the U.S. state of Georgia using the NLCD 2006 data. Local landscape is defined over a square, 3 km × 3 km areal units; there are 8475 such units in the study area. Several different computational protocols for regionalization of these units are evaluated with the best protocol resulting in a discovery and delineation of fifteen LTs. The LTs are summarized from descriptive and quantitative perspective, their landscapes signatures are shown, and an example of each landscape is given. Delineated LTs agree well with perceptual patterns seen in the NLCD map. Non-urban LTs are arranged roughly in stripes running from SW to NE perpendicular to the regional gradient of elevation. Full resolution maps, the data, and our regionalization computer code are available at http://sil.uc.edu/downloads.html.