Localizing general models with classification and regression trees

摘要

Typically, in forest inventory the volume of tally trees is predicted with a volume model estimated at national level. Such a global model is not unbiased regionally if there is spatial variation in the tree form due to one or more unknown predictors. This regional bias could be reduced or removed if the models were localized to each region or subarea. The localization is easiest if the area can be divided into homogeneous areas with respect to stem form. This study tested whether the localization results depend on the way the division is made and on the size of the subareas. The study area was divided spatially into homogeneous subareas with residuals of the global model or with the local spatial index, G(i)*, or both with classification and regression trees, the leaves of which formed the subareas. In addition, two other spatial divisions were created: an administrative forest centre and spatially equal-sized subarea divisions. The localized models were compared with the global model. The root mean squared errors (RMSEs) of localized models were smaller in median and in mean, but maximum values exceeded the overall global model RMSE. The localization reduced local RMSEs on average by 1-6%. The differences between the spatial divisions were small, although the aggregate standard errors and RMSEs were slightly smaller in regression trees. Only 50 +/- 8% of the subareas were spatially homogeneous in regression tree divisions, which suggests that either the division criteria or the division method were inadequate.