EVALUATING YOLO CONFIGURATIONS FOR DETECTING DAMAGED SPRUCE TREES

Sammanfattning

This thesis investigates how different configurations of the YOLO26 computervision model affect the model’s ability to detect spruce bark beetle damaged trees from drone images over forest areas. This research serves as a step towards an automatic spruce bark beetle detection system that can be used to help diagnose forest areas. Eight different YOLO configurations were investigated with 𝑘-fold cross validation in which the best performing YOLO configuration was fully trained before being evaluated on a separate test set. Key findings include the importance of keeping a large image size for detecting damaged spruce trees from drone images, improved performance when only focusing on damaged spruce trees and treating the rest of the trees as background during training, significantly improved performance when using transfer learning, and further improvements when applying Random Brightness Contrast data augmentation. The best YOLO configuration evaluated on the separate test set achieved a precision of 72.24%, recall of 79.31%, mAP50 of 78.20%, and mAP50-95 of 42.85%. This promising result shows the potential for an automatic spruce bark beetle damage detection system in the future.