CONTRIBUTIONS OF OPEN SOURCE TOOLS FOR ROAD DATA PRODUCTION IN DEEP LEARNING APPLICATIONS

Due to the complex structure, design and type of material of the roads, it is difficult to perform with the models trained with the data of other countries in deep learning studies based on road extraction. Especially in Turkey, deep learning based existing data sets for the modelling are not suitable due to complex road structures. Thus, local data sets that represent the features of the region must be used to be able to achieve promising results. In this study, a labelled road data set for Istanbul was created by using high-resolution orthorectified aerial images which cover a wide variety of urban and rural areas. In this context, used imageries were obtained from the Istanbul Great Municipality. In the study, Mapbox and QGIS were employed as open source tools. Apart from the road data were eliminated via Mapbox and obtained global road data was overlaid orthorectified images in QGIS. After this process overlaid images were exported as tile files and labelled data was produced combining these tile files. The generated labelled data set was tested using U-NET Deep learning-based convolutional neural network architecture. As a result, 96 % training accuracy and 89.6 % validation accuracy have been achieved.