The fight against the ”eight-toothed spruce bark beetle” (Ips Typographus) is a crucial task in forest management. Since the conditions for spruce bark beetles are becoming better due to climate change and monocultural farming, it is necessary to develop more efficient techniques to fight their spreading. There is some recent research in the area of bark beetle or dead wood detection, however, all of the projects have slightly different goals and experiments with different approaches. Using recent research from semantic image segmentation and image classification with methods from deep learning, a novel approach to detect and identify infestation of spruce trees by the eight-toothed spruce bark beetle was developed in this thesis. More concretely, this thesis covers the localization of sick spruce trees from a bird eye’s perspective and the identification of infested spruce trees directly from the trunk, with the help of RGB images taken by drone and smartphone cameras. The development of a mobile application allows users to perform inference in real-time and therefore leads to a more efficient and fast removal of infested trees. With only 35 annotated images available for the segmentation task and 80 annotated images for the image classification task, it was necessary to use techniques like transfer learning or data augmentation to artificially enlarge the dataset. The solutions presented in this project and the suggestions for future work should guide further research in this area and could be viewed as an advanced baseline for such work.
Early Detection of Spruce Bark Beetles using Semantic Segmentation and Image Classification
Advisor(s): Josef Scharinger
Johannes Kepler University, MS Thesis, June 2020.