Reducing extrapolation bias of area-based k-nearest neighbor predictions by using individual tree crown approaches in areas with high density airborne laser scanning data.

K-nearest neighbor (kNN) approaches are popular statistical methods for predicting forest attributes in airborne laser scanning (ALS) based inventories. Their main upsides are the simplicity to predict multivariate response variables and their freeness of distributional assumptions on the conditional response. One of their largest draw-backs is that predictions outside the range of the reference data inherently result in an under- or overestimation. This property of kNN approaches is known as extrapolation bias and aggravates with an increasing number of neighbors (k) used for the prediction. This study presents one possibility to reduce extrapolation biases of predictions based on the area-based approach (ABA) by using individual tree crown (ITC) approaches within those specific areas of a low density ALS acquisition where the point density might be sufficiently high for using ITC methods. In the proposed strategy, additional (or artificial) reference plots augmented field measured plots. Artificial plots were created by applying ITC segmentation to a canopy height model derived from high density ALS data. The response variable biomass per hectare was predicted for every segment following a semi-ITC approach. The segment predictions were aggregated on the artificial plot level. The artificial plots were then treated in the same way as the original reference data to make predictions in areas with low density ALS data based on the ABA. It was hereby assumed that the predicted plot level response on the artificial plots is equivalent with the observed plot level response on the original reference data. The data consisted of 110 reference plots with a smaller data range than the 201 independent validation plots. Considerable extrapolation bias was visible if only the reference plots were used for the prediction. Almost no extrapolation bias was found if the prediction was based on reference plots augmented by artificial plots. The root mean squared error (RMSE) of the biomass predictions based on the reference plots was 39.1%. The RMSE reduced to 29.8% if the reference plots were augmented by artificial plots.