Methods for estimating multivariate stand yields and errors using k-NN and aerial laser scanning

The objective of this study was to investigate techniques for integrating aerial LiDAR data into the forest yield information systems of a forest management company. Nearest neighbour (k-NN) estimation was identified as a useful approach and a 4 000 ha case study was undertaken to provide insight into the performance of the technique in a commercial environment. A field dataset was provided by 213 ground plots installed across the study area. Small-footprint discrete aerial LiDAR data were acquired concurrently and processed to derive descriptive metrics across the study area. A simulated annealing algorithm was produced and used to select important variables that were then used during model development. Sampling error for the k-NN predictions was estimated using an approach that accounted for spatial correlation in the reference dataset. An independent validation dataset was acquired from the forest manager's conventional high-intensity stand assessments within the study area projected to LiDAR acquisition date. Model validation showed excellent correspondence to the independent dataset with the responses total recoverable volume (relative mean deviation 4.3 per cent, RMSE 44.96 m3 ha−1), mean top height (relative mean deviation 1.3 per cent, RMSE 1.34 m), basal area (relative mean deviation 4.5 per cent, RMSE 2.06 m2 ha−1) and stand density (relative mean deviation −1.8 per cent, RMSE 47.39 sph). Log-product volumes were estimated using LiDAR metrics and k-NN estimation for all stands with encouraging accuracy and precision. Sampling error was estimated for all stands in the study area, including those with no reference plots. Sampling error estimates were small enough for stand estimates to be useful for the forest manager in all stands. For stands in the validation dataset the sampling errors for stand volume were smaller for the k-NN estimates (median confidence interval (CI) 29.13 m3 ha−1) than for the conventional inventory estimates (median CI 37.9 m3 ha−1). Imputation model error was examined and found to be insignificant.