Within the forestry and environmental science community, it is widely known that the forest understory plays an important ecological role in forest ecosystems. In this study we focused on the three main components of the forest understory in the Sierra Nevada Mountain Range of California: small trees, shrubs, and coarse woody debris. These three components play different key ecological roles in the forest ecosystem; acting as a source of food and habitat for many wildlife species, essential for the water balance of forest ecosystems, and providing fuels for fires. Because of their importance, forestland managers are in need of a viable method for mapping understory presence and absence across the landscape. This study demonstrates a technique for mapping understory presence and absence at the Yosemite Forest Dynamics Plot in California using a random forest model trained using 24 Aerial LiDAR-derived forest structure metrics. Using this technique, investigated 1) overall understory detection accuracy, 2) model accuracies for differentiating and classifying fine fuels and coarse woody debris, and 3) the minimum LiDAR pulse density required for accurate detection of understory through pulse decimation.
Results/Conclusions
In investigation 1) we found the model predicted understory presence and absence correctly with 71.57% accuracy, and a commission error of 20.54%. From these results we concluded that our techniques were viable for classification of understory. In investigation 2) we achieved prediction accuracies of just 33% for fine fuel presence and 22% for coarse woody debris. With this, we concluded that these techniques cannot be used to differentiate and classify fine fuels and coarse woody debris. Finally, in investigation 3) we found that a minimum of just 10 pulses per square meter is required for detecting understory presence with 71.14% accuracy. In investigation 1 and 3, ~20% of the error associated with the predictions was error of commission. We believe that these errors of commission are not simply model errors, but rather can be attributed to model detecting unmapped understory components, such as low lying tree branches, large rocks, and <1 cm DBH trees. Thus, we accepted the ~71% error rate and concluded that with more thorough mapping of all understory, the results would likely improve.