Assessment of building extraction and geometric corrections from UAV image using Mask R-CNN deep neural network algorithm

High-resolution spatial data acquired by unmanned aerial vehicles (UAVs) are increasingly used as a method of extracting buildings in urban areas, supporting planning, disaster management, and navigation-based applications. This research applies the Mask R-CNN deep learning algorithm to automatically extract buildings within the Saruhanlı district in Manisa Province, Turkey, using UAV-derived orthophotos. A unique contribution of this research is the first systematic comparative evaluation of three post-segmentation geometric correction techniques (default, hybrid, adaptive) for building footprint refinement, addressing a critical gap in spatial accuracy improvement. The model correctly identified 113 out of 118 buildings in the test area, achieving a detection accuracy of approximately 95.8%. The average precision was 93.43%, and the F1 score reached 95.42% at an intersection-over-union (IoU) threshold of 0.5, quantifying the overlap between reference and predicted building footprints. Following automatic extraction, the three geometric correction techniques were employed to refine building geometries. To explore the spatial relationship between the actual building shapes and the automatically extracted and geometrically corrected shapes, regression analysis was performed. The resulting R² values were 0.873 for the default, 0.923 for the hybrid, and 0.942 for the adaptive technique, indicating that adaptive method outperformed the other approaches.