Comparing semantically-blind and semantically-aware landscape similarity measures with application to query-by-content and regionalization

Local landscapes can be assessed for a degree of mutual similarity between their patterns using a methodology pioneered in the domain of Content-Based Image Retrieval (CBIR). They are represented by multivariate histograms of selected pattern features and their similarity is calculated as the similarity between those histograms. Potential similarity functions include semantically-blind measures - designed to compare only the patterns without taking into account intrinsic similarities between the features, and semantically-aware measures - that also take into account intrinsic (semantic) alikeness between landscape features. In this paper we compare the performance of these two types of landscape similarity measures. Two measures are evaluated, the semantically-blind Jensen-Shanon (JS) and the semantically-aware Earth Mover's Distance (EMD). The performance of these two measures is evaluated empirically by running tests pertaining to their potential practical applications: a query-by-content and regionalization into landscape types. Variants of the JS and EMD corresponding to 1D and 2D histograms are applied to the two test sites taken from the NLCD 2006. Each test site is divided into 400 3 × 3 km areal units serving as local landscapes. For each test site extensive query-by-content and regionalization calculations are run using different variants of JS and EMD. The results are compared using statistics and visual assessment. The EMD has been found to perform better than the JS, but the difference in performance is not pronounced, especially if a 1D histogram representation of landscapes is used. For the 2D histogram representation of landscapes the EMD consistently yields better results than the JS but suffers from high computational cost. The EMD also enables the comparison of landscapes having different legends.