Monitoring tree harvest in selection forestry with UAV data and machine learning

Sammanfattning

Forests are vital ecosystems that support biodiversity, regulate climate, and provide economic and social value. Their resilience is strongly influenced by harvesting methods, with selective cutting and rotation forestry having noticeably different impacts on carbon dynamics. Understanding these effects requires accurate data on the number and location of felled trees. Although modern harvesters can record this information, the functionality is often underused, leaving gaps in monitoring. Unmanned Aerial Vehicle (UAV) imagery offers a cost-effective way to collect high-resolution data, which, when combined with artificial intelligence (AI), enables automated detection of felled trees. Advances in machine learning and deep learning make it possible to process UAV imagery efficiently, reducing the need for extensive manual work. This thesis investigates which AI models are most suitable for identifying individual felled trees in UAV imagery. Three approaches were evaluated: a Random Forest, a Mask R-CNN with a ResNet-50 backbone, and a Mask R-CNN with a Swin Transformer backbone. The study addresses three main research questions: (1) how UAV remote sensing data can be prepared and processed for training AI models. (2) What accuracy can be achieved and whether the models are reliable enough for detecting individual felled trees. (3) Which of the tested models performs best. By answering these questions, the thesis explores the potential of AI to support sustainable forest management by improving the monitoring of harvesting practices. The results showed that the Swin Transformer struggled, but consistently outperformed the other models, achieving the highest average precision (0.34), recall (0.479) on the validation data, and performance when tested on a random sample of images, while Mask R-CNN struggled but showed potential. In contrast, Random Forest suffered from overpredicting and, since it classifies only pixels rather than entire objects, it is primarily suitable for mapping general trends. Although constrained by dataset size and image quality, the study demonstrates that deep learning models, particularly the Swin Transformer, show potential for detecting felled trees from UAV imagery, offering potential for improved monitoring in sustainable forest management.