Akademik Veri Yönetim Sistemi
20th International Symposium on Environmental Pollution and its Impact on Life in the Mediterranean Region, Athens, Yunanistan, 26 - 27 Ekim 2020, ss.200-201, (Özet Bildiri)
Deep learning is a specific part of machine learning and represents the number of layers that are used for modeling the data. Successful results in image classification and image segmentation studies have been obtained by using this technique. Although it has become the main approach of many new applications, it still contains problems that need to be overcome. One of them is semantic labeling which is the most important activity and widely used in object extraction from remotely sensed data by using deep learning techniques. The main problem of semantic labeling is the assignment of a class label to every pixel. Generally, much of the labeling is still time consuming, costly and slowly performed by human experts. Additionally, deep learning requires a lot of training data because of the big number of complex parameters needed to be tuned by a learning algorithm. For automatically labeling of objects from aerial or satellite imageries such as roads, should be developed. Moreover, in the current datasets, mostly contains systematically planned urban areas that are generally obtained roads have uniform shapes in deep learning applications. However, in Turkey, roads are not in a uniform shape, especially in rural areas. Thus, extracting these features is becoming more difficult. Using these data set asthe training data may bring several learning problems for other counties as well. To overcome this situation, generalized data are generated for deep learning applications. During the data set preparation, each country should be evaluated separately considering their land use/cover characteristics. In this study, the road labeling dataset obtained from an aerial image covers a wide range of urban and suburban regions was generated. During this process, the availability of recently orthorectified imageries and open access data were considered. Deep learning-based convolutional neural network architecture was tested on a generated dataset. Additionally, using the availability of generated dataset were evaluated in wide range areas with different conditions.