Estimating Tree Height Distribution Using Low-Density ALS Data With and Without Training Data

Abstract—This study applies an approach based on stochastic geometry for retrieval of forest characteristics from airborne laser scanning (ALS) in two situations: 1) without ground-measured training data and 2) with training data. The applied model treats the ALS echo heights as an outcome of a random process, express-ing the observed heights of canopy envelope as a function of stand density, the parameters of the tree height distribution, and the shape of the individual tree crown. The model was applied to a eucalyptus plantation dataset with known spacing, where the main interest was to estimate the plot-specific tree height distribution. Estimation without training data resulted in RMSEs of 2.9 and 0.9 m for mean and dominant heights, respectively. Estimation using training data resulted in RMSE’s of 1.4 and 0.8 m, respec-tively. In both cases, the estimates of dominant height were more accurate than with the reference method, but the estimates of mean height were less accurate (area-based approach; RMSEs 1.1 and 0.9 m, respectively). The model-based method was robust to substantial decrease in echo density from 1.4 echoes/m2 to 0.14 echoes/m2. Index Terms—Airborne laser scanning (ALS), eucalyptus, forest inventory, recovery, height, stochastic geometry. I