Comparing the accuracies of forest attributes predicted from airborne laser scanning and digital aerial photogrammetry in operational forest inventories

Airborne laser scanning (ALS) data provide a detailed representation of forest canopy structure and are highly suitable for forest inventory applications. Providing three-dimensional data at a lower cost, digital aerial photogrammetry (DAP) has emerged as an alternative to ALS. Previous studies have compared the utility of ALS and DAP data for predicting forest attributes, however none of those studies used data acquired as part of large-area operational inventories. We used ALS, DAP and field data obtained from 836 plots and five operational inventories conducted in southeastern Norway. We compared local and regional modelling approaches for predicting basal area, number of stems, volume, Lorey's mean height, and dominant height. First, we developed district-, forest type-, and data source-specific nonlinear prediction models for all forest attributes (local models). Second, we fitted forest type- and data-source specific nonlinear models with pooled data from all districts (regional models), in which we included dummy variables to account for district-specific effects. We compared the accuracies of ALS- and DAP-based predictions made with local and regional models. Finally, we assessed how a reduced number of calibration plots affected the accuracies of ALS- and DAP-based predictions made with regional models. In general, ALS-based predictions made with local models were most accurate. District-specific effects needed to be accounted for in the regional models. ALS models required substantially fewer calibration plots than DAP models. The accuracies obtained with regional ALS models fitted with 50% of the available data were better than the accuracies obtained with local DAP models fitted with all the available data. Thus, by fitting regional ALS models with data from multiple inventories, field efforts can be reduced substantially while still obtaining better prediction accuracies than by fitting local DAP models.