Evaluation of a semi-automated approach for the co-registration of forest inventory plots and airborne laser scanning data

[Departement_IRSTEA]Territoires [TR1_IRSTEA]SEDYVIN

International audience

English. Continuous maps of forest parameters can be derived from Airborne Laser Scanning (ALS) data with the so-called area-based method. A prediction model is calibrated between local ALS statistics and forest parameters measured on field sampling plots. Unfortunately, inaccurate GPS positioning often leads to a bad matching of forest measures with ALS data, which results in poor calibration datasets. This article has two main objectives: first to present and test a novel semi-automated co-registration approach, and second to evaluate how the number of positioned trees per plot affects the efficiency of co-registration. The co-registration is based on the computation of correlation coefficients between the ALS-derived canopy height model and the rasterized forest plot map for different offsets applied to the GPS position. The plot map raster is computed by affecting to each pixel the height or diameter of the tree located in this pixel. Others pixels are set to zero. The final offset to be applied to plot coordinates is the one which results in the highest correlation coefficient. The validation dataset is composed of 139 17m-radius plots located in a 1000-ha forest of the Jura mountain (France). The algorithm was used to search the best offset within a 40m square centered on the GPS position for each plot. The resulting co-registration was evaluated visually. 127 plots (91.4%) were successfully co-registered by the algorithm. Six (resp. two) additional plots were correctly co-registered when the searched window was increased to 80 (resp. 200) meters. Three of the four remaining plots could be manually co-registered. For two of them the algorithm also found the correct position when trees felled between the inventory and the ALS flight were removed from the inventory. For the 138 identified plots, mean distance between the GPS and the validated position was 9.0 ± 8.7 m. Bootstrap cross-validations showed that the area-based prediction model for basal area calibrated with the validated positions was significantly better than when calibrated with the GPS positions. Root mean square errors were respectively 5.84 ± 0.04 and 6.94 ± 0.06 m²/ha. The influence of the number of georeferenced trees on the efficiency of automatic co-registration was assessed by including in the field plot inventory only the N largest or nearest trees, with N varying from 3 to 15. The search window was set to 40m around the validated position and a plot was considered successfully matched when the distance between the new and the validated position was less than two meters. When the three largest trees were used, 89% of the 138 plots were correctly co-registered. This proportion raised to 100% when six trees were used, but decreased to 99% when more than 14 trees were included. Using the N nearest trees yielded a lower proportion of correct co-registration with only 86% for 15 trees. These results show that the proposed approach is efficient and robust, as a high proportion of plots are properly co-registered with only diameter information of a small number of big trees. This semi-automatic method could thus contribute to better calibrated ALS prediction models while saving inventory time on the field. Future developments include the automatic assessment of co-registration liability, based on the distribution of correlation coefficients within the search windows, and the evaluation of the effect of plot radius on the accuracy of co-registration.