Early detection of bark beetle attacks: Integrating Segment Anything Model (SAM) zero-shot segmentation and spectral indices for tree health assessment

英语. Forests offer vital ecosystem services but face threats from various stressors, including climate change and insect infestations. The European spruce bark beetle (Ips typographus L.) poses significant risks to Norway Spruce (Picea abies). Early detection of bark beetle infestations is crucial for damage control but challenging with traditional methods. This study aims to utilize modern remote sensing technologies, particularly high-resolution Unmanned Aerial Vehicle (UAV) imagery, combined with the Segment Anything Model (SAM) to segment individual spruce trees and detect early signs of bark beetle attacks. Field studies were conducted in Mulatorp, a nature reserve in southeast Sweden, capturing UAV images over several months. The SAM, a state-of-the-art deep learning model for image segmentation, was used to segment individual spruce trees from RGB UAV data. The study aimed to assess SAM's zero-shot capabilities, refine its segmentation parameters, and compare its outputs with manually validated data. Additionally, the study sought to develop a time-series of vegetation indices to detect early signs of bark beetle infestations. Results indicated that SAM's box prompts yielded better segmentation accuracy than point prompts, though the model often merged canopies and missed some trees. Despite the high spatial resolution of UAV imagery, SAM detected only 37% of all trees and 33% of Norway spruce trees, with an IoU of 0.55 for spruce trees. The Green Chromatic Coordinate (GCC) was identified as the most effective vegetation index for early detection, showing significant differences between healthy and attacked trees as early as June. The findings suggest that while SAM has potential for remote sensing applications, its current zero-shot capabilities are insufficient for precise tree segmentation without further refinement and training. The study highlights the importance of integrating advanced segmentation models with UAV imagery for effective forest health monitoring and early intervention in bark beetle infestations. Future research should focus on enhancing SAM’s segmentation accuracy and expanding field-validated datasets to improve early detection frameworks.

英语. Forests are crucial to the world, offering numerous benefits such as providing timber, regulating the climate, and serving as recreational spaces. However, these forests face threats from climate change and with insects like the European spruce bark beetle, which attacks Norway Spruce trees and causes significant damage. The year of 2018 showed the devastating effects of spruce bark beetle attacks during drought periods, making it vital to explore ways to detect these infestations early. This project explored a way to detect these harmful beetles early, using advanced technology such as high-resolution images taken by drones and an artificial intelligence tool called the Segment Anything Model (SAM) released by Meta. This combination aimed to identify individual spruce trees and detect early signs of bark beetle infestations. Conducting research in Mulatorp, a nature reserve in southeast Sweden, detailed images of the forest were captured over several months. These images were then analyzed using SAM to see if it could accurately identify the trees, and UAV images were used to assess health of attacked and healthy trees over time to spot early infestations. The images were used to calculate spectral indices, which are commolu used as indicators of tree health. The results showed that SAM needs improvement, it is at present unable to identify individual trees in a forest setting. It managed to correctly identify about a third of the trees. One of the most effective spectral index found was the Green Chromatic Coordinate (GCC), which could distinguish between healthy and infested trees as early as June. This early detection is crucial for preventing widespread damage. As a result, we were able to distinguish between healthy and attacked trees early. Despite its potential, the study highlighted several limitations of SAM. The model often grouped trees together and missed individual trees, highlighting the need for further refinement and training. Additionally, SAM struggled with complex forest images, which affected its accuracy. These limitations suggest that while SAM and drone imagery hold promise, more work is needed to enhance their effectiveness for forest health monitoring. The findings highlight the potential of combining drone imagery with advanced AI models to gain insight on forest health. By refining these technologies, forest health can be better monitored, infestations detected early, and actions taken to preserve these vital ecosystems. This project showcases the importance of using modern technology to tackle environmental challenges and ensure forests remain healthy.