Deep Learning with Semantic Segmentation Approach for Building Rooftop Mapping in Urban Irregular Housing Complexes

Abstract

This research investigates the application of the Deep Learning (DL) U-Net architecture for building rooftop segmentation in densely populated urban areas with irregular housing patterns. The research explores the effectiveness of two loss functions-Binary Cross Entropy (BCE) and Dice Loss (DLs)-to optimize the segmentation accuracy. The present study utilized Small-Format Aerial Photography (SFAP) images processed into orthophotos with a final ground sampling distance of 5 cm. The study area, located in Bogor, Indonesia, features both regular and irregular housing patterns, making it an ideal testing ground for the segmentation model. The U-Net model, having been utilized EfficientNetB6 as the encoder and having been trained with augmented data, demonstrated stable performance across metrics, such as accuracy, precision, recall, and F1-score. The results show that the DLs function outperformed BCE, achieving an average Intersection over Union (IoU) score of 96.8% compared to the 87% score for BCE, indicating that DLs is more effective for this application. The study further enhances the segmentation results by converting the raster data into a vector format using the Ramer-Douglas-Peucker (RDP) algorithm, which simplifies and smooths the polygonal shapes of the segmented rooftops. The combination of the U-Net, DLs and RDP algorithm provides high accuracy results and high usability of the segmentation outputs in practical applications, such as urban planning and disaster management scenarios where accurate rooftop delineation is critical.