Forest attribute model extrapolation to adjacent areas by means of image based canopy height models

Dominant tree height, stand density index and standing timber volume are essential information for the characterization of forest stands. As terrestrial measurements are very time consuming and field inventories are often missing, the use of height information automatically derived from high resolution remote sensing data, e.g. digital stereo images, has great potential as an alternative technique to gain knowledge about the forest conditions. The main goal of this study was to arrive at a better exploitation of aerial stereo images by the following tasks: (i) derivation of dense 2.5d point clouds and surface models from digital stereo images via semi-global matching and orthoimage computation, (ii) area-based estimation of dominant tree height, stand density index and standing timber volume at the plot level and (iii) extrapolation of these estimation models to an adjacent forest area. The study is conducted within the Steigerwald test site in the northwestern part of Bavaria, Germany. For two large forested areas (northern part 7742 ha and southern part 3469 ha), field measurements are available from a forest inventory conducted in 2010. In total, 1875 circular inventory plots (each with a size of 500 m²) are located in the northern part and 835 plots in the southern part. The plots are arranged on a regular grid of 200m x 200m in both areas. For each plot, we calculated the dominant tree height, the stand density index and the standing timber volume. Digital aerial images (UltracamXP, PAN & RGBI, 20 cm GSD, forward overlap 75%, side overlap 30%), acquired in May 2011, and a bare-earth digital terrain model (DTM, 1m resolution) from ALS-data are used. Up to now, the following analyses were conducted according to task (i): First, for both data sets, Semi-Global Matching was applied to obtain the absolute elevation of the canopy surface, i.e. 2.5d point clouds and a digital surface model (DSM) with 1m resolution. Second, canopy height models (CHM) were derived by subtracting the ALS-based terrain heights. Third, using the image-based DSM for rectification, we calculated orthoimages (RGBI) with 1m resolution. In order to meet task (ii), subsequently the following explanatory variables were computed separately for each inventory plot: height metrics, amongst others percentiles of the height distribution; variability metrics describing the variability of heights; canopy cover metrics representing the spatial occupancy based on the CHM; color and texture metrics based on the spectral information of the orthoimages. We applied the non-parametric Random Forests approach to generate models for the estimation of dominant tree height, stand density index and standing timber volume based on all plots in the northern part of the test site. Model evaluation was conducted via repeated 10-fold cross-validation and the model accuracies were examined by means of root mean squared error (RMSE) and bias. According to task (iii) of this study, we scrutinize the predictive performance of the models when extrapolating to adjacent forest areas. It is assumed that reasonable predictions to areas outside the calibration test site strongly depend on the similarity of the forest stands with respect to the considered attributes. Thus, the forest attributes derived from the field measurements were compared in terms of their distributions in the two test areas. The non-parametric Wilcoxon rank sum test was applied for the statistical comparison since none of the three attributes was normally distributed in both areas. The tests revealed, that the distributions from both test areas were not significantly different for dominant tree height (p = 0.4849), stand density index (p = 0.7184) and standing timber volume (p = 0.3099). Proven the similar forest conditions of the two areas, we will use the models built and calibrated in the northern part of the test site to predict the forest attributes for all inventory plots located in the southern part. Consequently, all plots in the southern part will be used as an independent test set. Again, the accuracy will be examined in terms of RMSE and bias to find out if an extrapolation of the area-based models to neighboring forests is feasible. We conclude that the extrapolation method might substantially support forest decision makers for upcoming planning and monitoring tasks, especially if no terrestrial inventory information is available for those forests.