Detection of aircraft and determination of aircraft types using deep learning methods in open source satellite images

The concept of deep learning has become very popular with the development of artificial intelligence in recent years. Especially with the development of computer hardware and cloud technology, deep learning has become accessible and applicable for everyone. The most important factor in the spread of object detection is that high resolution images can be easily obtained with the developments in satellite, camera, lens, etc. technologies. In this context, analysis studies particularly on satellite imagery have become widespread in both civil and military fields. Automatic object detection, object recognition and classification in the image is an important requirement. Detection of roads, buildings, plantations, airports, airplanes, military areas, hospitals, schools, forests and many other objects can be made in satellite images. Aircraft detection draws attention among these objects. Detecting the aircraft and determining the type of aircraft is important at the military strategic level. Based on this, in the thesis study, three different aircraft types were determined as military transport, combat and passenger type. In order for the object detection to work successfully, over a hundred civil and military airports were examined and a total of 529 original satellite images were downloaded. The data set containing a total of 2939 original aircraft images, including 917 transport planes, 1056 warplanes and 966 passenger planes was created. CenterNet Hourglass, YOLOR and YOLOv7 algorithms were used for detection of aircraft, and the success of the algorithms was evaluated with F1 score analysis. As a result, YOLOv7 achieved an F1 analysis score of 99.3%, YOLOR 94.2% and CenterNet Hourglass 79.6%.