Potentials and limits of the k-nearest-neighbour method for regionalising sample-based data in forestry

Abstract

The k-nearest-neighbour (knn) method is known as a robust nonparametric method. It is used to estimate unknown values of data sets by means of similarity to reference data sets with known values. The spectral information of satellite remote sensing data can be used to provide the common characteristics in the knn estimation process. In forest sciences, the knn method is studied for its application potential. Some application examples are: (1) the estimation of parameters such as basal area, stem volume, number of trees per diameter class and tree species; (2) the estimation of forest debris and non-wood goods and services; (3) the production of wall-to-wall information for modelling, risk management and logistics. On the other hand, different limitations with respect to methodological characteristics as well as the selection of suitable parameters must be taken into consideration. The scope of this article concentrates on the discussion of the application potential and limits of the knn method in forestry with particular emphasis on management planning needs. The study is based on data taken from a forest inventory (FI) covering a test site near Rottenburg, in southwest Germany. Analysis results are compared with the traditional outcome of inventory data analysis and partly presented in thematic maps, which show identical spatial distribution patterns. For the map of six tree species, a map accuracy of 52.2% was found. The user’s accuracy for the prevailing tree species was between 52.6% for Picea abies and 69.4% for Quercus sp. A timber volume map for Quercus sp. clearly visualises the bias at the extreme ends of the volume distribution. The root mean square error (RMSE) for the total timber volume estimate was 30.9% for k = 5 and could be reduced to 22.6% for k = 20. For Quercus sp., however, the respective RMSE values were between 106.5 and 84.8%. Significant differences between FI and knn estimates were mainly found for rare classes with minor representation in the reference data.