Aircraft detection in satellite imagery using deep learning-based object detectors

Over the recent years, object detection in satellite imagery has become a crucial task in remote sensing applications. Specifically, the detection of aircraft is critical for military scenarios. Numerous computer vision techniques have been applied for this problem, however, robust and efficient detection of aircraft in satellite imagery poses several challenges, such as, variance of color, size, aspect ratio and orientation of aircraft and complex backgrounds. In this paper, we provide an in-depth review of deep learning methods for object detection applied to aircraft detection including both conventional and state-of-the-art techniques. We also provide a comprehensive experimental and comparative analysis of deep learning-based object detectors including Region-based CNN (RCNN), Fast RCNN, Faster RCNN and You Only Look Once (YOLO) encapsulating the most cited feature extraction networks including Alexnet, VGG-16, Resnet-18, Resnet-50, Resnet-101 and Inception-v3. The detailed quantitative comparison allows for selection of the most suitable set of object detector, backbone feature extraction network and hyperparameters for aircraft detection in satellite imagery with trade-off options between precision and detection time. The evaluation is based on mean average precision (mAP), precision versus recall, time complexity and computational complexity. Finally, we draw our conclusions and identify promising future work.