Using deep learning algorithms for built-up area extraction from high-resolution GÖKTÜRK-1 satellite imagery

Building extraction is a method used in high-resolution remote sensing image processing for urban planning and demographic analysis. This method plays a significant role in identifying urban structures and assessing damage in disaster situations. In this study, building extraction was performed using the U-net architecture with G & ouml;kt & uuml;rk-1 satellite imagery. This dataset was chosen because it has not been previously used in the literature for building extraction using deep learning methods. The impact of parameters such as the number of epochs, batch size, and learning rate on the results was evaluated. Additionally, the outcomes of the Adam (Adaptive Moment Estimation), SGD (Stochastic Gradient Descent), RMSprop (Root Mean Square Propagation), and Nadam (Nesterov-accelerated Adaptive Moment Estimation) optimizers were assessed. A training model was created using the Inria dataset, and the results were evaluated on G & ouml;kt & uuml;rk-1 satellite imagery. The findings indicated that increasing the number of epochs led to higher accuracy. The deep neural network model was applied to an urban area selected in Ankara using G & ouml;kt & uuml;rk-1 satellite imagery. Models for building segmentation and classification were both trained and tested. The results revealed that the U-net model achieved an overall accuracy of over 83.85% for building segmentation with the Adam optimizer. The intersection over union (IoU) ranged between 96 and 99%. The highest Dice similarity value (88.81%) was obtained with the RMSprop optimizer at 100 epochs.