Urban Green Space Segmentation in Planet Scope Satellite Images Based on the Improvement of U-Net Architecture

Urban Green Space is one of the important areas in urban areas which is characterized by the presence of vegetation and is a factor in determining temperature and air quality in urban areas. It plays an important role in balance, comfort, health, sustainability, and improving the quality of human life. Research on deep learning using satelliteimagery, especially in urban green space segmentation, is still very rare. Therefore, we conducted research with deep learning in urban green space segmentation using satellite imagery. We made improvements from previous researchersby using U-Net architecture to segment urban green space with planet scope satellite images in Sleman Regency and Yogyakarta City. The dataset that we have has a resolution of 3m and 4 bands, namely Red, Green, Blue and Nir. From the results of the training model, our method gets the smallest loss with a value of 0.012 for training and 0.0006 for validation. Our method obtains an IoU score of 90.68% which outperforms previous studies with our dataset.