Optimizing Building Footprint Extraction from UAV Imagery Using Pretrained Deep Learning Models: A Case Study in University of Phayao, Thailand

This study investigates an optimization of deep learning models, the ArcGIS pretrained model (Building Foorprint Extraction-USA), for detecting building footprints from UAV orthorectified images, emphasizing the influence of parameters including threshold, padding, and tile size. Experimental results show increasing the threshold value significantly reduces the false positives, with higher thresholds (0.7-0.9) achieving correctness levels of up to 90%, while the lower threshold values result in correctness around 40-60%. Padding values of 600–800 improve detection accuracy by minimizing false positives through the acquisition of additional contextual information surrounding building structures. Additionally, larger tile sizes improve detection efficiency, while smaller tiles provide finer resolution but can increase false positives. The findings reveal that balancing these parameters is critical for optimizing building footprint extraction, contributing valuable insights for geospatial analysis in urban planning and remote sensing applications.