Prediction error aggregation behaviour for remote sensing augmented forest inventory approaches

In this study we investigated the behaviour of aggregate prediction errors in a forest inventory augmented with multispectral Airborne Laser Scanning and airborne imagery. We compared an Area-Based Approach (ABA), Edge-tree corrected ABA (EABA) and Individual Tree Detection (ITD). The study used 109 large 30 × 30 m sample plots, which were divided into four 15 × 15 m subplots. Four different levels of aggregation were examined: all four subplots (quartet), two diagonal subplots (diagonal), two edge-adjacent subplots (adjacent) and subplots without aggregation. We noted that the errors at aggregated levels depend on the selected predictor variables, and therefore, this effect was studied by repeating the variable selection 200 times. At the subplot level, EABA provided the lowest mean of root mean square error ($\overline{\mathrm{RMSE}}$) values of 200 repetitions for total stem volume (EABA 21.1 percent, ABA 23.5 percent, ITD 26.2 percent). EABA also fared the best for diagonal and adjacent aggregation ($\overline{\mathrm{RMSE}}$: 17.6 percent, 17.4 percent), followed by ABA ($\overline{\mathrm{RMSE}}$: 19.3 percent, 18.2 percent) and ITD ($\overline{\mathrm{RMSE}}$: 21.8, 21.9 percent). Adjacent subplot errors of ABA were less correlated than errors of diagonal subplots, which resulted also in clearly lower RMSEs for adjacent subplots. This appears to result from edge tree effects, where omission and commission errors cancel for trees leaning from one subplot into the other. The best aggregate performance was achieved at the quartet level, as expected from fundamental properties of variance. ABA and EABA had similar RMSEs at the quartet level ($\overline{\mathrm{RMSE}}$ 15.5 and 15.3 percent), with poorer ITD performance ($\overline{\mathrm{RMSE}}$ 19.4 percent).