Poligonal yol zincirlerinden yararlanarak uydu görüntüleri ile 3B yol modelleme

Determination of roads from satellite images has gained more research interest after recent advances in data-heavy machine learning methods, which are also accelerated by the increasing amount of available data. A major challenge of learning-based approaches is obtaining labeled data to train systems. In this study, we propose a method to quickly label roads via satellite images of any desired location. Our method leverages 2D road information from OpenStreetMap, an online community-provided resource of geolocation information. In this environment, paths are roughly defined as line segments without their exact shapes and sizes. In systems where roads are given only as line segments, as in OpenStreetMap, we propose a system to obtain real road widths and make 3D modeling accordingly. Using this rough information, we propose a simple interactive user interface where users can easily label road boundaries on presented satellite images. Using our approach, it is possible to quickly label regions with different road characteristics. After labeling, it can be fed into machine learning algorithms and then road determination tests can be performed. Such an approach allows training separate machine learning systems for different regions of the world; this would be advantageous over training a single system to identify all types of paths. With the results obtained, it is possible to model the roads in 3D again to improve them. With our labeling tool, after cities were labeled and included in machine learning, tests were conducted with the results obtained. The results of these tests, which were made with satellite images of different cities, were obtained. When cities were tested with their own data sets, more successful results were obtained in finding roads. Three cities were considered and all were co-educated first. Afterwards, cities dual trainings were held. The third martyr was fine-tuned and transfer learning was performed on these training results. In addition, the training sets were cut in half and fine-tuned. While the success rate of transfer learning is low in Rome from the selected cities, the success rate is high in the cities of Etlik and Berkeley, and this success increases as the iteration increases. In mixed tests, its success is low compared to the tests made with the cities themselves. transfer learning is also not good according to the results trained and tested with them. The success rate of fine-tuning has increased with dual training labels.