| Abstrakt: |
Forests provide several ecological, sociocultural, and economic advantages to humanity. Rapid loss of forest lands is a result of uncontrolled, unauthorized, and commercial use. Although several methods have been proposed to protect and control forests, none of them was sustainable. One promising approach to monitor forest lands is using satellite images and computer vision techniques. Therefore, researchers proposed tree detection methods for this purpose. In this study, we propose merging the result of these methods via weighted boxes fusion which is a post-processing technique. We aim to eliminate the disadvantages while maintaining the advantages of existing methods this way. Hence, we picked the swin transformer, RCNN, faster RCNN, YOLO, and DETR for tree detection from satellite images. Then, we fuse their results in a combinational way. While doing so, we always keep the swin transformer since it has the highest performance score. We tabulate the fusion results on a diverse dataset. We report the strengths and weaknesses of the proposed fusion method. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink