Comparing Deep Learning Performance for Aircraft Detection in Satellite Imagery

The recent public availability of satellite imagery has pushed forward the field of deep learning allowing the application of novel object detection methods. However, object detection in satellite imagery, such as aircraft detection, presents significant challenges including large image variability due to orientation, scale, and shape. This work makes a summarized comparison of known object detection algorithms including You-Only-Look-Once (YOLO) and RetinaNet-ResNet variants performing aircraft detection on the publicly available RarePlanes dataset. RarePlanes was assessed to be a highly effective training dataset with similar bounding box size distributions across classes. Results indicate that YOLO variants generally provided better performance than RetinaNet-ResNet variants in identical testing scenarios. YOLO variants were especially effective when using high numbers of classes and with classes containing large numbers of annotations.