Samonadzorovano in aktivno učenje za detekcijo objektov na satelitskih slikah

Object detection in satellite imagery remains challenging due to small object sizes, large scale variations, and the high cost of producing expert annotations. This thesis investigates how self-supervised learning (SSL) and active learning (AL) can be combined to address these challenges. We evaluate four SSL-pretrained transformer backbones on the DIOR dataset: DINOv2, DINOv3, SatMAE, and ScaleMAE. For active learning, we compare random sampling, several uncertainty-based strategies, and a diversity-based clustering method. Experiments were conducted across labeling budgets up to 20%, with results reported using the COCO evaluation protocol. Among the SSL models, ScaleMAE achieved the strongest performance, reaching mAP50 = 76.0 ± 0.7 %, substantially narrowing the gap to supervised pretraining. Furthermore, we show that uncertainty-based acquisition strategies begin to outperform random sampling once moderate annotation budgets are available. These findings demonstrate that combining SSL and AL can reduce annotation requirements while maintaining competitive detection performance in remote sensing.