A Comparative Evaluation of Image Processing, Machine Learning, and Deep Learning Methods for Building Extraction from UAV Aerial Images

Automatic building extraction from aerial imagery is a fundamental problem in remote sensing and image processing, with broad applications in urban management, spatial planning, and three-dimensional city modeling. This study presents a comparative evaluation of three categories of methods for building extraction from UAV aerial images with ۲۰ cm ground sampling distance over an urban area in Iran: (۱) classical image processing methods including thresholding and region-based segmentation (Segment Mean Shift); (۲) machine learning methods including Support Vector Machine (SVM), Random Forest, and Maximum Likelihood; and (۳) deep learning methods represented by UNet. Training and reference data were prepared through manual labeling, and Precision, Recall, and F۱-score were used for quantitative evaluation at the pixel level. The results show that UNet achieved the best performance with an F۱-score of ۰.۹۱۸, while among the non-deep methods SVM performed best with an F۱-score of ۰.۶۳۳. Among the classical image processing methods, region-based segmentation (Segment Mean Shift) proved least effective overall with an F۱-score of ۰.۳۸۳, whereas a well-tuned global thresholding reached a markedly higher F۱-score of ۰.۵۲۴ thanks to a more balanced precision–recall trade-off. The study demonstrates that deep learning methods, owing to their ability to learn complex multi-scale features of Iranian urban fabric, offer a substantial advantage over both classical image processing and traditional machine learning approaches.

کلیدواژه ها:

Building segmentation ، Deep learning ، Machine learning ، UAV aerial imagery ، UNet.