| Abstract [eng] |
Forest tree diseases can spread quickly and disturb forest ecosystems, which is why detecting and identifying tree damage at an early stage matters a great deal. Crown defoliation is among the indicators most often used to assess tree health. The conventional way of measuring it relies on field expeditions, during which specialists inspect each tree by eye and assign it a defoliation class. Such work is expensive, slow and subjective, which makes it impractical for covering large forest areas. If an automated system could reliably estimate the level of crown damage from images taken by unmanned aerial vehicles, the monitored area could be expanded and the whole process sped up considerably. This work studies how computer vision methods can be used to detect and classify forest tree crown damage automatically. The datasets analyzed in the study consist of forest images captured by UAVs, with tree crowns showing different degrees of defoliation. Related research, along with methods used in other fields, is reviewed to evaluate how well these approaches might transfer to the task of tree crown damage classification. The problem is approached using modern deep learning models, including convolutional neural networks, one-stage object detectors, and architectures built on transformers. The experimental part of the study aims to determine which computer vision architectures are best suited for the task by comparing their accuracy and processing time. A second question addressed here is whether splitting the task into two stages, where the objects are first detected and then classified, yields a more accurate estimate of tree damage than a single-stage model that does both at once. The findings contribute to the wider effort to automate forest health monitoring, improve early disease detection, and support more sustainable forestry management decisions. |
