Localization of growth estimates using non-parametric imputation methods

The purpose of this study was to examine different non-parametric imputation methods to reduce regional biases in growth estimates. Growth estimates were obtained using non-parametric k-nearest neighbour imputation (k-NN) to predict future 5-year diameter increment over bark at breast height for Scots pine (Pinus sylvestris L.) and Norway spruce (Picea abies). The Mahalanobis distance function was chosen as the most suitable measure of similarity, and then it was modified using weights provided by linear regression analysis. The use of weights from linear regression facilitated the examination of the correlation structure of the variables and allowed for transformations of the independent variables. Localization of the non-parametric estimates was obtained through a variety of methods, in particular, by using spatial coordinates as independent variables, by restricting the selection of neighbours to a circular area around the target tree, and by restriction the selection of neighbours to a local database. The localized estimates using spatial measures were then compared with non-spatial imputation and also with estimates from a parametric growth model. Results were then compared by vegetation zones in Finland. The differences between the non-spatial k-NN estimates and the localized spatial estimates were negligible when summarized to the stand level, and localization did not reduce the regional biases relative to the non-spatial k-NN estimates. Regional biases in northern Finland and in south-western Finland were reduced substantially using the non-parametric estimates rather than the parametric growth models, however, and the mean biases in all of the regions were quite similar, while the mean biases of the growth estimates obtained with the parametric model varied notably between the regions.