Combining canopy height and tree species map information for large-scale timber volume estimations under strong heterogeneity of auxiliary data and variable sample plot sizes

Abstract

A timber volume regression model applicable to the state and communal forest area of the federal German state of RhinelandPalatinate is identified using a combination of airborne laser scanning (ALS)-derived metrics and information from a satellitebased tree species classification map available on the federal state level. As is common in many forest inventory datasets, strong heterogeneity in the ALS data due to different acquisition dates and misclassifications in the tree species classification map had noticeable effects on the regression model's performance. This article specifically addresses techniques that improve the performance of ordinary least square regression models under such restricting conditions. We introduce a calibration technique to neutralize the effect of misclassifications in the tree species variable that originally caused a residual inflation of 0.05 in adjusted R 2 . Incorporating the calibrated tree species information improved the model accuracy by up to 0.07 in adjusted R 2 and suggests the use of such information in forthcoming inventories. We also found that including ALS quality information as categorical variables within the regression model considerably mitigates issues with time lags between the ALS and terrestrial data acquisition and ALS quality variations (increase of 0.09 in adjusted R 2 ). The model achieved an adjusted R 2 of 0.48 and a cross-validated root-mean-square error (RMSE cv ) of 46.7% under incorporation of the tree species and ALS quality information and was thus improved by 0.12 in adjusted R 2 (5% in RMSE cv ) compared to the simple model only containing ALS height metrics (adjusted R 2 = 0.36 , RMSE cv = 51.7%).

The R Package forestinventory: Design-Based Global and Small Area Estimations for Multiphase Forest Inventories

“…One is that a perfect tessellation by the supports strongly depends on the distance between the sample locations of s 1 and the support size. Since in practice the support size should ideally be chosen to achieve a best possible explanatory power of the regression model (Hill, Buddenbaum, and Mandallaz 2018a) a perfect tessellation might often not be feasible. In the infinite population frame, the supports are allowed to overlap if this seems necessary to acquire a sufficiently large sample n 1 to get a negligibly close approximation ofZ.…”

A Double-Sampling Extension of the German National Forest Inventory for Design-Based Small Area Estimation on Forest District Levels

“…The CHM was calculated as the difference between the digital terrain model and the digital surface model that were derived by a Delauney interpolation of the ground and first ALS pulses respectively. A more detailed description of the procedure can be found in Hill et al [41]. The CHM provided the most valuable information to be used in the OLS regression model for predicting the timber volume on the plot and cluster level.…”

“…Table 3 gives the classification accuracies [46] of the treespecies variable after calibration. More details on the processing of the explanatory variables and identification of optimal parameter settings for their calculation are described in Hill et al [41]. …”

“…Here, only a summary of the extensive analysis that was performed is provided but the reader can refer to Hill et al [41] if more details are desired.…”

“…As opposed to the ALS data, the availability of a country-wide tree species classification map has yet been unique among all German federal states. Whereas the study of Hill et al [41] already showed that the tree species information was able to improve the model fit, it has yet not been used to its full potential. One reason for this was the impossibility of modeling individual tree species within each ALS acquisition year, which would add further explanatory power.…”