THE EFFECT OF DISTANCE CORRECTION FACTOR IN CASE-BASED PREDICTIONS OF VEGETATION CLASSES IN KARULA, ESTONIA

摘要

The aim of this study was to investigate the applicability of the distance correction parameter (DCP) integrated to the case-based prediction system CONSTUD to reduce the effect of spatial autocorrelation of training data in machine learning process. To achieve this, calculated similarity between observations is decreased by the so-called distance correction value (DCV - the quotient of DCP and distance between two observations). 50 machine learning iterations were carried through in the case of different DCP-s from 0 to 15 000 m using random samples generated from 450 training observations from southern Estonia (Karula National Park and its vicinity). Independent validation samples were used to estimate the effects of the use of each DCP. Machine learning results showed that the Cohen's kappa index of agreement decreased in accordance with the increase of DCP-s. The correspondences of field observations and predicted values followed the same trend. The explanation would be that with the increase of DCP-s successively more observations were rejected as useful ones. Conversely, no considerable decrease in correspondences of the predictions was recognized when DCP was increased. In our case, probably the most useful exemplars were chosen and the less useful ones were left beyond. As a result, scattered and probably spatially and thematically highly representative sample of observations remained. The border might be drawn at DCP from which the number of the in-between distances started to decrease considerably, but the correspondence in validation sample estimations as well as in training sample estimations remained relatively stable.