BBD: a new hybrid method for geospatial building boundary detection from huge size satellite imagery

Abstract

Buildings that are constructed without the necessary permits and building inspections affect many areas, including safety, health, the environment, social order, and the economy. For this reason, it is essential to determine the number of buildings and their boundaries. Determining the boundaries of a building based solely on its location in the world is a challenging task. In the context of this research, a new approach, BBD, is proposed to detect architectural objects from large-scale satellite imagery, which is an application of remote sensing, together with the geolocations of buildings and their boundaries on the Earth. In the proposed BBD method, open-source GeoServer and TileCache software process huge volumes of satellite imagery that cannot be analyzed with classical data processing techniques using deep learning models. In the proposed BBD method, YOLOv5, DETR, and YOLO-NAS models were used for building detection. SAM was used for the segmentation process in the BBD technique. In addition, the performance of the RefineNet model was investigated, as it performs direct building segmentation, unlike the aforementioned methods. The YOLOV5, DETR and YOLO-NAS models in BBD for building detection obtained an f1 score of 0.744, 0.615, and 0.869 respectively on the images generated by the classic TileCache. However, the RefineNet model, which uses the data generated by the classic TileCache, achieved an f1 score of 0.826 in the building segmentation process. Since the images produced by the classic TileCache are divided into too many parts, the buildings cannot be found as a whole in the images. To overcome these problems, a fine-tuning based optimization was performed. Thanks to the proposed fine-tuning, the modified YOLOv5, DETR, YOLO-NAS, and RefineNet models achieved F1 scores of 0.883, 0.772, 0.975 and 0.932, respectively. In the proposed BBD approach, the modified YOLO-NAS approach was the approach that detected the highest number of objects with an F1 score of 0.975. The YOLO-NAS-SAM approach detected the boundaries of the buildings with high performance by obtaining an IoU value of 0.912.