Vegetation classification by two new iterative reallocation optimization algorithms

This paper presents two new non-hierarchical iterative reallocation optimization algorithms for vegetation classification. The OPTimal PARTitioning algorithm (OPTPART) optimizes the ratio of within-cluster similarity to among-cluster similarity; the OPTimal SILhouette algorithm (OPTSIL) optimizes the difference between the similarity of each sample to the cluster to which it is assigned and its similarity to the most similar cluster. The algorithms were tested on three vegetation datasets (Mt. Field Massif, Tazmania, Australia; Podyj/Thayatal National Park, Austria/Czech Republic; and Shoshone National Forest, Wyoming, USA) using three dissimilarity/distance matrices (Bray-Curtis, chord distance, and Hellinger distance) and compared to five other commonly used or recently introduced vegetation classification algorithms (flexible-β, TWINSPAN, PAM, ISOPAM, and K-means) using eight goodness-of-clustering evaluators. Five of the eight evaluators were species-based and operate on the distribution of individual taxa among clusters; three were community-based and operate on the compositional similarity of clusters. OPTPART was initialized from random partitions and from the results of a flexible-β classification as the initial partition; OPTSIL was initialized from partitions resulting from OPTPART, flexible-β, and K-means classifications. Algorithms were ranked from best to worst on each clustering evaluator for each dissimilarity/distance matrix for each dataset, and summarized by median ranks. OPTPART, SIL/OPT (OPTSIL from an OPTPART initial partition), and SIL/FLEX (OPTSIL from a flexible-β initial partition) ranked 1–3 respectively for results pooled across all three datasets and dissimilarity/distance matrices. OPTPART, SIL/OPT, and SIL/FLEX consistently ranked 1–3 across the individual datasets, although the order varied slightly by dataset.