AUTOMATED BUILDING OBJECT SEGMENTATION FROM UAV ORTHOGRAPHIC IMAGERY USING DEEP LEARNING AND GIS TECHNIQUES

Abstract

This paper presents an automated model for segmenting building objects in orthographic images captured by unmanned aerial vehicles (UAVs) using deep learning and GIS techniques. The proposed method consists of two phases: Phase 1 employs a two-stage deep learning method (named YOLO-SAM) that uses YOLOv11 for coarse building segmentation, followed by the SAM2 (Segment Anything Model) algorithm to refine and enhance the results. Phase 2 further processes this shapefile using a series of geometry-based post-processing techniques, including the Douglas-Peucker algorithm for boundary simplification and curve smoothing to optimize the overall shape. The approach achieves high segmentation accuracy (~mAP of 0.75) and provides a scalable, practical solution for mapping in Vietnam. Unlike previous methods, this approach stands out by combining YOLOv11 and SAM2 for accurate, detail-rich segmentation, and by integrating this with GIS-based post-processing in a unified, efficient pipeline.