One of the critical tasks in monitoring of road infrastructure is the identification of road cracks. Recent efforts have been made in utilising Unmanned Aerial Vehicles (UAVs) to automate this task without interfering with the road network traffic and infrastructure. However, high-quality annotated datasets that allow the development of reliable deep learning models for this purpose, are scarce. Synthetic data generation offers a promising alternative to mitigate this issue by reducing annotation costs and enhancing the dataset’s variability. This paper represents a comparative study of two state-of-the-art deep learning models - UNet with EfficientNet and UNet with MobileNet- for road crack segmentation trained with three loss functions (Dice Loss, Focal Loss, and Weighted Binary Cross-Entropy Loss (WBCEL)) by using synthetic datasets generated with three different ways. Performance evaluation indicates that the best results were achieved using UNet with a MobileNet encoder, trained with WBCEL and synthetic images, yielding an mIoU of 63.52% tested on real crack imagery. A more granular analysis underlines how both synthetic data realism and the choice of the loss function impact segmentation accuracy. Our preliminary study concludes that well-designed synthetic data and appropriate loss functions have the potential to allow better generalization of the model to real-world scenarios.