The Levenshtein metric, a new means for soil classification tested by data from a sand-podzol chronosequence and evaluated by discriminant function analysis

Abstract

Soil profiles are difficult to compare by statistical methods because sampling depths and intervals and the number of samples per profile may vary. This paper discusses a novel method for handling the problem. Chemical data from 52 soil profiles, which formed a chronosequence of podzols from Fraser Island, south-eastern Queensland, was classified using the Levenshtein metric. The soils varied from undifferentiated sands to deeply weathered podzols with many metres of bleached A, horizon. The data used were the proportions by weight in the soil samples of Na, K, Ca, Mg, Fe and Al extracted by mild acid digests. A polythetic divisive program was used to form seven groups from a matrix of six attributes and 573 samples. These groups were used to code profiles into strings of digits whose length equalled the number of samples in a profile. The Levenshtein metric then formed these strings into a dissimilarity matrix, which in turn was used to produce groups by an agglomerative hierarchical procedure. The groups produced by this procedure were checked using discriminant function analysis applied to depth function parameters derived from the original data set used in the pattern analysis procedures. These parameters were derived from a model which is consistent with ideas of soil genesis, hence their use should favour a classification related to soil development. The classificatory procedure could also be weighted for the depth to the B horizon, a procedure that was considered to be a direct weighting for the degree of soil development. Discriminant function analysis showed that the groups produced could be discriminated on the basis of depth function parameters regardless of any depth weighting. The groups were consistent with the known geochronology, but contained an important component due to other factors such as parent material and surface accumulation due to plant activity. Classification involving the Levenshtein metric proved to be a sound means of taking adequate account of the variation present in the data. Supplementary procedures provided by TAXON enable fresh insights into the nature of the individuals and their interrelationships to be obtained.