Deep Learning Model Comparison for Dense Building Segmentation in the City Using Aerial Imagery Data

City administrations in many developing counties, particularly in high-density cities, have expressed concern about automatic building identification as a foundation for urban growth planning and other reasons. Deep learning has gotten much attention in recent years as the most appealing technique to solve challenges in remote sensing. Semantic picture segmentation, which endeavors to classify each picture element into a collection of provided labels, is one use of deep learning. The purpose of this experiment is intended to assess the use of UNet and PSPNet models. Both models are utilized to recognize dense structures of city centres, which each pixel photo classifies as building or non-building label using deep learning. Based on DKI's research using aerial photographic pictures from the Pasar Minggu Subdistrict in the South Jakarta City Regency. The average accuracy for the Jakarta Provincial model in training was 85%.