Aircraft detection in very high-resolution satellite images using YOLO-based deep learning methods

Abstract

With the recent developments in remote sensing technology, satellite images with high spatial and temporal resolution have been becoming widely available. Very high resolution (VHR) satellite images are very appropriate data sources for geospatial object detection using deep learning algorithms. Airplane detection from satellite images is one of the significant application areas to support airspace inspection, airline traffic control, and defense applications. In this study, we compared various variants of YOLOv5 (You Only Look Once) models and the Scaled-YOLOv4 model for aircraft detection from satellite images. We implemented different hyperparameters, optimization algorithms, and data augmentation methods. Finally, based on the results of numerous experiments, we evaluated the advantages and disadvantages of both methods. Our analysis illustrated that the best mAP@0.50:0.95 value of 0.865 belongs to the YOLOv5x model with 16 batch sizes. Whereas, in terms of computational efficiency, the Scaled-YOLOv4 model has the shortest duration in the training.