On using landscape metrics for landscape similarity search

• We compare different implementations of metrics-based landscape similarity search.
• Numerical experiments on land cover data compare different similarity measures.
• Robust similarity measure requires proper normalization and weighting of metrics.
• Similarity based on co-occurrence features outperforms metrics-based similarity.
• Landscape regionalization is sensitive to small differences in similarity measure.

Landscape similarity search involves finding landscapes from among a large collection that are similar to a query landscape. An example of such collection is a large land cover map subdivided into a grid of smaller local landscapes, a query is a local landscape of interest, and the task is to find other local landscapes within a map which are perceptually similar to the query. Landscape search and the related task of pattern-based regionalization, requires a measure of similarity – a function which quantifies the level of likeness between two landscapes. The standard approach is to use the Euclidean distance between vectors of landscape metrics derived from the two landscapes, but no in-depth analysis of this approach has been conducted. In this paper we investigate the performance of different implementations of the standard similarity measure. Five different implementations are tested against each other and against a control similarity measure based on histograms of class co-occurrence features and the Jensen–Shannon divergence. Testing consists of a series of numerical experiments combined with visual assessments on a set of 400 3 km-scale landscapes. Based on the cases where visual assessment provides definitive answer, we have determined that the standard similarity measure is sensitive to the way landscape metrics are normalized and, additionally, to whether weights aimed at controlling the relative contribution of landscape composition vs. configuration are used. The standard measure achieves the best performance when metrics are normalized using their extreme values extracted from all possible landscapes, not just the landscapes in the given collection, and when weights are assigned so the combined influence of composition metrics on the similarity value equals the combined influence of configuration metrics. We have also determined that the control similarity measure outperforms all implementations of the standard measure.