Empirical models for predicting gamma diversity in macrobenthic communities of Korean tidal flats

초록

A biodiversity prediction model may play an important role in restoring tidal flats and designing artificial habitats. A model with higher accuracy requires abundant data on gamma diversity (a total number of species in a vertical transect of a tidal flat) and obtaining them is one of the prerequisites in developing the model. We developed multiple linear regression and artificial neural network (ANN) models that predict macrofaunal gamma diversity from fewer number of samples using averages of alpha diversity (averages of species number in sampling stations), beta diversity (averages of dissimilarity among stations) and tidal flat width (m) and compared their performances. Data based on 8 tidal flats ranging from 140 to 5000 m of widths and 25 to 116 species of gamma diversity were used in the diagnostic procedure. Data from Banweol, Buan in 1999 and Boreum Is., Incheon in 2000 were used in validity test procedure of a regression model. Feed-forward neural network composed of three layers (i.e., input, hidden and output layers, respectively) was applied. Network training was done by back-propagation algorithm and transfer function between layers was hyperbolic tangent function. To avoid overfitting and lack of generalizability problem, data were subdivided into multiple dataset (training set, 60% of the data; cross-validation set, 15%; testing set, 25%). In diagnosing regression model, two dissimilarity indices (percent dissimilarity, beta1; chord distance, beta2) were chosen for estimating beta diversity. Two regression models that differ in beta diversity were estimated and compared. They showed no significant differences in explanatory power (both of them were >96%) and p-value (<0.001), but model beta2 seemed to be more valid from a knowledge-based point of view (higher influence in tidal flat width and beta diversity). In validity test, correlation between observed and predicted values of model beta2 was 0.952. ANN model employed chord distance as a beta dive