We introduced a deep-learning pipeline for automatic localization of caravanserais in satellite imagery. A database of 203 Iranian caravanserais was created and used to train YOLOv5 via transfer learning. To reduce false positives, misdetected locations were added to the database as a new class and the model was retrained. The final [email protected] on the test imagery was 91.43%.
Available on Civilica: civilica.com/doc/1955368