Ajoneuvojen tunnistus synteettisen apertuuritutkan drone-kuvista syväoppimista hyödyntäen
Sorsa, Jesse · Aaltodoc (Aalto University) · 2025
Automatic target detection from synthetic aperture radar (SAR) imagery has been increasingly important in applications such as urban planning, defense, and surveillance. Even though deep learning models have been widely utilized for object detection in SAR images, limited research exists on the use of drone-based SAR imagery, due to the lack of large-scale labeled drone SAR datasets. The aim of the thesis is to develop a deep learning model for detecting vehicles from drone-based SAR images. To achieve this goal, two deep learning architectures, the convolutional Faster R-CNN and transformer-based Deformable DETR, are trained and evaluated. The lack of drone-based training data is addressed by implementing several transfer learning techniques using publicly available data, including satellite-based SAR images as well as infrared aerial imagery. A small custom drone-based SAR dataset was collected and used to test the performance and generalization capability of the models. As a result, top performance was achieved with a transformer-based Deformable DETR model by utilizing a SAR-pretrained backbone as a transfer learning step. The results show that the best object detection model can achieve an AP of 0.69 on the drone-based SAR detection task, even when they are trained solely on satellite-based images. These findings demonstrate the feasibility of transferring knowledge from publicly available SAR datasets to custom drone-based SAR applications. Possible future research directions include collecting large-scale drone-based SAR datasets and fine-tuning the models on drone-based data to improve robustness.