Silvilaser 2019 - Poster Presentations »
Method Comparison for the Assessment of Vertical Structure from ALS Data in Natural Forests
In the light of climate change and the related risks to forests, today more and more foresters tend to re-structure their forests “back” to a more natural state, which gives the forests higher resilience against wind-throw, drought and insect infestations. In addition, national park managers have legal obligation to prove, that their forests remain in their supposed natural state by repeated monitoring over time. In order to assess the success of forest management measures that intend to work in this direction objective and effective tools should be at hand. Forest status assessment is currently done by point-wise field assessments, which are time-consuming and thus costly and error-prone and subjective, all at the same time. ALS can provide the needed information area-wide and in an objective manner. This paper provides basic investigations on methods to monitor the vertical structure as one important indicator of the “naturalness” of the forest for a larger area. We have investigated a mixed forest in the Kalkalpen National Park in Austria by means of airborne LiDAR data. We evaluated three different sets of spatial reference units to calculate vertical structure: simple raster, automatically generated image segments and overlapping circles. Further, we compare, if the height distribution of the LAS returns or the height distribution of detected tree tops perform better in the assessment of the vertical structure. Finally, different classification methods including thresholding procedures, expert-based classification, regression analysis and random forest classifiers were compared for the purpose of classifying vertical stand structure. The individual results were evaluated against independently gathered reference field data. The results showed that simple, non-overlapping raster is an insufficient basic unit. Non-overlapping, automatically generated segments and overlapping circles showed similar results depending on the classifier. As for the use of LAS returns versus tree tops, the results are ambiguous: while forests with lower density and thus a low crown base height , methods based on LAS returns tend to confuse low branches with small trees. In contrast, when using detected tree tops , understorey trees are often omitted. Therefore, future work should be dedicated to the combination of both approaches. The overall best result was achieved using the segments and the LAS returns with an overall accuracy of 89% after plausibility check. However, this evaluation was done in an area with very few one-layered stands (= stands with low vertical structure). When expanding the reference data set to include more one-layer stands, the same method still achieved 73% overall accuracy , third to the same classification method, but using tree tops yielding 77%. In our analysis, the expert-based classification method performed best, followed by thresholding procedures. The regression-based method and the random forest classifier achieved comparably lower accuracies. One reason could be the amount of training data, which was limited due to limited field work available. More tests with extended reference data should be performed to unleash the potential, which might be inherent in machine learning and other AI classification approaches. Our results form the basis for an area-wide assessment of vertical structure as one main indicator in the monitoring of the naturalness of forest ecosystems both in managed as well as in protected forests.