This study presents a low-cost and data-efficient framework for alligator crack detection and geospatial visualization using smartphone imagery, synthetic data generation with a Denoising Diffusion Probabilistic Model (DDPM), and YOLOv11m. Pavement images and GPS coordinates were collected across urban roads in Tehran under varying lighting conditions. To address severe data scarcity, a DDPM was trained using only 60 real alligator crack images to generate additional synthetic samples. The generated images were evaluated using Fréchet Inception Distance with 64-dimensional features (FID-64), Kernel Inception Distance (KID), and nearest-neighbor similarity analysis. An ablation study compared the initial dataset with conventionally augmented and DDPM-augmented datasets, both involving a 100% expansion of the training subset, under identical training conditions across five random seeds. The DDPM-augmented model achieved mean precision, recall, mAP50, and mAP50–95 values of 0.907, 0.843, 0.865, and 0.664, respectively. Compared with conventional augmentation, the clearest improvement was observed in mAP50–95, which increased from 0.602 to 0.664, whereas additional gains in precision, recall, and mAP50 were modest. These findings suggest that DDPM-generated images can provide benefits beyond increasing training sample quantity, particularly under stricter localization criteria. Finally, the detection outputs were integrated with GPS coordinates for geospatial visualization, and an independent case study in Gorgan was used to demonstrate the framework in a different geographic location. The proposed approach provides a practical basis for low-cost pavement monitoring.
