Ph.D. Tezi Görüntüleme

      The biggest difficulty in the classification of satellite images with deep learning is the lack of sufficiently labeled data. The usability of the crowdsourcing approach to solve this problem was investigated and an application was developed to demonstrate its applicability. The crowdsourcing method was implemented through the web platform, which was prepared to enable users to generate data in a dynamic structure. Users were asked to create tagged data for the desired classes using the help document provided. The control of these datasets was determined by giving scores to the polygons drawn by the users for each class. The generated dataset contains a total of 260262 pixels. The data obtained were overlapped with the original image to examine visually. Finally, 40 pieces of images were classified with the proposed spectral and spatial CNN model, which proved to be successful after a series of tests. The fact that the average general accuracy values are above 95% indicates that this problem can be solved by crowdsourcing. The use of crowdsourcing for post-classification accuracy assessment was performed through the module presented in the same interface. For this purpose, users were asked to enter class values for 1000 randomly generated reference points. The class value of each of the 1000 points is entered three times by three different users, and the class value of each point determined by a majority voting method. The results prove that reference points for high spatial resolution multispectral images can be generated by crowdsourcing.