A Deep Learning Approach for Detection and Segmentation of Airplanes in Ultrahigh-Spatial-Resolution UAV Dataset

Advancements in unmanned aerial vehicle (UAV) technology have enabled the acquisition of images of a geographical area with higher spatial resolutions as compared to images acquired by satellites. Detection and segmentation of objects in such ultrahigh-spatial-resolution (UHSR) images possess the potential to effectively facilitate several applications of remote sensing such as airport surveillance, urban studies, road traffic monitoring crop monitoring, etc. Investigating these images for target extraction tasks turns out to be quite challenging, in the terms of the involved computation complexities, owing to their high spatial resolutions and information content. Due to the development of several deep learning algorithms and advanced computing tools, there exists a possibility of harnessing this information for computer vision tasks. Manual surveillance of airports or similar areas and manual annotation of images are cost-intensive and prone to human-induced errors. Therefore, there exists a substantial requirement of automating the task of keeping track of the airplanes parked on the premises of airports for civil and military services. With this paper, we propose a framework for detecting and segmenting such airplanes in UHSR images with supervised machine learning algorithms. To detect the target i.e., airplanes, MobileNets-deep neural network is trained, whereas to segment the target, U-Net-convolutional neural network is trained with our dataset. Further, the performance analysis of the trained deep neural networks is presented. The UHSR image dataset utilized in this research work is an airport dataset provided by SenseFly. Data is acquired by eBee classic drones, flying at a height of 393.7 ft., which provide 2D-RGB images with a ground resolution of 3.14 cm/px.