Demonstrating the transferability of forest inventory attribute models derived using airborne laser scanning data

Airborne laser scanning (ALS) is a reliable source of accurate information for forest stand inventory attributes including height, cover, basal area, and volume. The commonly applied area-based approach (ABA) allows the derivation of wall-to-wall geospatial coverages representing each of the modeled attributes at a grid-cell level, with spatial resolutions typically between 20 and 30 m. The ABA predictive models are developed using stratified inventory data from field plots, the requirement for which can increase the overall cost of the ALS-based inventory. Parsimonious use of ground plots is a key means to control variable costs in the operational implementation of the ABA. In this paper, we demonstrate how the prediction accuracy of Lorey's height (HL, m), quadratic mean diameter (QMD, cm), and gross volume (V, m3) vary when existing ABA models are transferred to different areas or are applied to point cloud data with different characteristics than those on which the original model was developed. Specifically, we consider three scenarios of model transferability: (i) same point cloud characteristics, different areas; (ii) different point cloud characteristics, same areas; and (iii) different point cloud characteristics, different areas. We generated area-based models using three modeling approaches: linear regression (OLS), random forests (RF), and k-nearest neighbour (kNN) imputation. Results indicated that the prediction accuracy of area-based models varied by attribute and by modeling approach. We found that when the models were transferred their prediction accuracy decreased, with an average increase in relative bias up to 22.04%, and increase in relative RMSE up to 29.31%. Prediction accuracies for HL were higher than those of QMD or V when models were transferred, and had the lowest average increase in relative bias and relative RMSE of <5% in the majority of cases. Likewise OLS models for HL had greater prediction accuracies when models were transferred compared to RF and kNN models, especially when the point cloud characteristics were similar. Conversely, we found that for QMD and V, RF models were found to be the most transferable in cases when models were applied to different areas with similar and different point cloud characteristics. While there is potential for cost savings by transferring models and reducing data acquisition costs, our results show the degree of transferability depends more on the attribute being modelled or the modeling approach applied, and less on the characteristics of the point cloud data.