TRANSFERABILITAS MODEL DISTRIBUSI SPASIAL PALEM LOKAL SULAWESI

Abstract

The aim of this study was to test transferability of the predictive models developed in North Sulawesi site using internal data dan external data (Buton island located on the south east coast of the mainland of Sulawesi). Logisitic regression method was used to build the predictive model. Three native palms of Sulawesi namely Pigafetta elata, Areca vestiaria, and Caryota sp were selected in this study. The relationship between environmental variables (e.g. derived from a Geographic Information System (GIS) and Remote Sensing (RS)) and each native palm species occurrence was quantified by the model. Several environmental parameters reflecting land cover, topographic features (elevation, slope and aspect), soil and socio-economic (accessibility) served as predictor variables. Model performance was measured by sensitivity (Sn), specificity (Sp), overall accuracy (S), Kappa statistic and Area Under the Curve (AUC). Predictive models of spatial distribution of the three native palm species were successfully developed using logistic regression method. Using the internal dataset, all models showed greater than 90% overall accuracy at predicting presence or absence which indicates that all models gave good performance. However, using Kappa index, the performance of the logistic regression model was poor. AUC values derived from the ROC procedure showed that the performance of all models was excellent. Nevertheless, when tested on Buton Island, overall prediction accuracy was low. Kappa statistic (0.06 to 0.24) also showed that none of the models performed adequately. It is obvious that predictive ability of the models evaluated with external data from a different island was low. This implies that transferring the models to other regions is not advisable. Unmodelled ecological-relevant processes and algorithmic error are suggested to be the reasons behind the poor transferability of the predictive models.