Rooftop type classification plays a crucial role in numerous urban applications such as 3D city modelling, solar potential estimation, disaster management, and smart city development. Leveraging high-resolution UAV imagery and advances in machine learning and deep learning, this study conducts a comparative evaluation of traditional machine learning and deep learning algorithms for classifying rooftop types, gable, half-hip, and complex, using orthophotos from Rasht City, Iran. In order to achieve that, firstly a structured and annotated dataset of rooftops was constructed through manual digitization and image cropping, followed by extensive training data augmentation. In the second step, two handcrafted feature sets (HOG and HOG+LBP) were used to train five machine learning classifiers (SVM, Random Forest, KNN, XGBoost, and Logistic Regression). The best results among these was achieved by SVM (F1-score = 0.74), while KNN performed the weakest. Also, ensemble learning methods were utilized through the aggregation of diverse model predictions. Using ensemble learning techniques, particularly stacking, significantly enhanced classification accuracy, reaching an F1-score of 0.79. In the third step, deep learning models including a custom CNN and a MobileNetV2-based transfer learning approach were utilized. The latter achieved the highest overall classification accuracy (83.3%) and macro F1-score (83%), outperforming all traditional methods by learning spatial hierarchies directly from data. The results demonstrate that while classical methods benefit from interpretable feature design and faster training, deep learning showed better within-domain generalization on our Rasht dataset.