A comparative analysis of kNN and decision tree methods for the Irish National Forest Inventory

Abstract

Two non-parametric estimation techniques were tested in two study areas in Ireland. For each area, plot level estimates of standing volume per hectare and basal area per hectare were computed from the National Forest Inventory field data and combined with SPOT 4 XS satellite imagery and a digital elevation model to form a set of observations. These observations were then used to predict variables across the satellite image using k-Nearest Neighbour (kNN) estimation and a Random Forest algorithm. Comparisons between the two techniques were assessed based on the estimation errors primarily using the Root Mean Square Error (RMSE) and relative mean deviation (bias). In both study areas it was found that the RMSE was lower for kNN than for RF. Overall, the RMSEs and mean deviations were lower in Study Area 1 when compared to Study Area 2, largely due to a difference in the number of available NFI reference plots.

Acknowledgements

This research is funded by the Irish Research Council for Science, Engineering and Technology (IRCSET) under the Embark Initiative. SPOT satellite images used within this research were provided by SPOT Image Distribution/MEDIAS France through the OASIS (Optimizing Access to SPOT Infrastructure for Science) Programme. The authors acknowledge the support of the Irish Forest Service for the provision of National Forest Inventory data. The authors would like to thank Dr. Ron McRoberts and two anonymous reviewers for extremely valuable comments and suggestions on the manuscript.