Derivation of stand wise information from forest enterprise wide inventories for forest management planning.

Abstract

Current forest management planning in Lower Saxony derives its data from a two phase sample and a subsequent yield table guided ocular assessment of each stand. The spatial scale of forest management activities is at the stand level, thus stipulating reliable data both for planning and assessment purposes at this level. However, owing to economical constraints, the sample design is optimized at the enterprise level rendering sound statistical inference at the stand level impossible. Furthermore, the introduction of sivilcutural programs aiming at transforming typical species rotation forest management into mixed structured continuous forestry system stands entails an increasing imprecision of yield table based information derivation. Therefore, there is an increasing need for techniques that allow inference of reliable stand wise information and which should be able to reduce the total costs of the current data collection. In this article a prediction technique called "Most Similar Neighbor" (MSN) has been compared to the current forest management planning approach. The comparison is based on 44 stands of the woodland Selling, situated in southern Lower Saxony, comprising beech, spruce, and larch in varying mixture and age structure. The stands were Surveyed each with 20 systematically aligned plots using a Count factor of 1. Each stand were than predicted by the current forest management planning approach and by most similar neighbors. Most similar neighbors relies on canonical correlation analysis for a definition of similarity. Canonical correlation analysis is a generalization of regression. Given two sets of variables it seeks pairwise linear combinations of the variables of each group that are each maximally correlated as well as orthogonal. One set of design variables was measured oil 1927 terrestrial phase 2 sample plots whereas the other set of auxilliary variables was derived from a digitized version of the color infrared imagery regularly employed during phase 1. The auxiliary variables were computed for the phase 2 plots as well as for plots aligned on a 50 m times 50 m grid inside the investigated target stands. Each target plot of the investigated stands was then compared to all phase 2 plots and all measured variables of the most similar one were assigned to the target plot. MSN techniques cannot extrapolate outside the range of given design variables. In general it suffers from underprediction at the upper range and overprediction at the lower range. However, even with the simple derivation of auxiliary variables from the digital imagery there is only little systematic deviation in the prediction of spruce. The root mean squared error of the volume prediction showed that MSN already yields results for spruce comparable to the forest management planning. While yielding acceptable results for beech systematic underprediction of stands with large volume occurs. Prediction of larch is still unsatisfying. The study showed the potential of the MSN approach but additional research is needed to further the understanding of the impact of species prevalence, and evaluate the stability and optimization of the canonical correlation analysis.